New Jersey played a leading role in inventing modernity, but it never built a mythology around it, while Silicon Valley can't stop telling its story. This reveals an interesting pattern: innovation hubs rise through specialized advantages, only to become traps that prevent adaptation. But New Jersey discovered something profound by accident. In moving beyond its achievements and simply existing without needing to prove anything, it found its authentic self.

Growing Up Without a Story

I grew up in New Jersey.

I never really thought much of it. I went to high school in Central Jersey and then college at Rutgers in New Brunswick, just a few minutes from where I grew up.

In many ways, New Jersey was an ideal place to grow up. It was culturally diverse, bustling, and living close to New York and Philadelphia gave anyone who wanted it world-class global exposure to arts and culture. Moreover, New Jersey was a place where successive new waves of immigrant communities came, worked hard, generation after generation, and achieved the American dream.

But New Jersey also had this gritty cultural element that was completely authentic and incredibly endearing. One of my favorites: Fat cat sandwiches at the grease trucks at Rutgers (sandwiches with cheese sticks, two hamburgers, fries, all loaded on a giant bun). At one time, a version of this innovative caloric monstrosity was named the best sandwich in America.

Fat cat sandwiches represent this live-in-the-present existence that typifies New Jersey for me.

Moreover, characterizations of New Jersey “life” with the success of The Sopranos and the controversial show Jersey Shore (whose cast was a bunch of New Yorkers!) ostensibly gave people a “window” into Jersey life that just reinforced the gritty aspects of its culture. And of course, who can ignore The Boss.

What I especially loved about New Jersey was that people didn’t sugarcoat anything. It was real. New Jersey is tough, fun, authentic, and unfiltered. And the state has leaned into this idea of itself. Even the official New Jersey Twitter account picks fights with Delaware and casually responds to tweets with “your mom” and “JERSEY wuz here.”

Midway through college, though, I was ready to get out. The whole scene wasn't for me: wearing all black clothes from the Banana Republic outlet and clubbing every Friday; then heading back on Monday to some back-office IT job at a finance company in New York. None of it seemed appealing.

I felt like opportunity wasn't going to meet me in New Jersey. I got into graduate school at Rutgers (but also in New York and Philly), but decided to take a chance and leave for Pittsburgh and Carnegie Mellon.

Chasing Lost Glory in Pittsburgh

Pittsburgh was radically different from Central New Jersey, and Carnegie Mellon embodied much of this cultural energy.

You could feel the hunger: greatness had been lost and needed to be reclaimed.

Pittsburgh was fighting hard to reinvent itself, determined to become the most inventive region in America. You could see the transformation happening: buildings in Oakland were being scrubbed clean with decades of smog finally coming off.

Pittsburgh also had a wealth of amenities from its past as one of the world’s great industrial hubs: museums, concert halls, mansions, the Steelers (they were in the Super Bowl three times when I was there; they won twice), and most of all, its people’s fighting spirit.

But what I loved was that Pittsburgh kept its authenticity, the same quality I appreciated about New Jersey.

Here is a promo video from Carnegie Mellon, “Today we work.” Telling the story of the creative working-class ethic driving the university and the city.

San Francisco Bay Area and the Myth Machine

Six years later, I moved to Stanford, California. In the San Francisco Bay Area, history was still in the process of being made. Silicon Valley was really only 50 years old by this time.

I arrived in 2010, just as the tech rebound from the financial crisis was getting underway. There was this constant buzz, but it wasn't about where the Revolutionary War was won—it was about where the Future would be won, where tomorrow was being invented.

California had this unshakeable self-confidence that it was the cradle of America's tomorrow.

Stanford was at the epicenter of this myth-making.

California has built a near-complete mythology around itself and tells these stories really well. It gets others to tell these stories too (I’ve done my share of this, too).

Do you have a dream?

Got a garage?

You, too, can create the future and become a billionaire.

The Research Triangle’s Clean Slate

Then we moved to the Research Triangle area of North Carolina.

The Research Triangle is also a place in the making, but it's different. Unlike Pittsburgh, it doesn't lean on its past. Unlike New Jersey, it isn't gritty. And unlike California, it's not buzzing with the kind of energy and ideas that come from everyone feeling like they're building a rocket ship (or have a lottery ticket in their pocket).

North Carolina wants to move beyond its past (it was once a slaveholding state). The Research Triangle, in particular, is a place where you go to start fresh with your new house, your corporate job at a respectable company, and your high-quality of life.

If you think these are just my stereotypes, here's a ChatGPT-generated image with a simple prompt: "Generate an image of what you believe the typical person from New Jersey, California, and the Research Triangle Park in North Carolina looks like."

These cultural types and stories told about them are so pervasive that this is now embedded in the billions of parameters of modern LLMs. Digital stereotypes.1

A few years ago, driving down Route 1 in New Jersey past the strip malls and the old Regal Cinema where I worked, then to Route 130 past my high school, I realized something.

I lived in New Jersey until I was 23, but New Jersey never really told me its story.

Jersey was just there.

Meanwhile, California wouldn't shut up about Silicon Valley, and Pittsburgh and North Carolina were busy trying to elbow themselves into the spotlight by “reinventing” themselves.

But New Jersey? It never talked about what it actually built. Nobody mentioned that before Silicon Valley was even a dream, New Jersey was inventing the modern world.

When Jersey Invented Tomorrow

Robert Gordon's The Rise and Fall of American Growth makes a very interesting observation: that the rate of American economic growth has been in decline for some time. In particular, there was a special century between 1870 and 1970 in which some of the most important technological breakthroughs in human history happened. After the 1970s, growth slowed.

During this truly transformative period, we gained the foundational technologies: electricity and the lightbulb, the internal combustion engine and automobile, powered flight, running water and indoor plumbing, air conditioning, communication systems for voice and data across vast distances, and the semiconductor and transistor enabling the digital age. Similarly, the chemical industry gave us plastics and synthetic materials (think everything you’re using today, including the clothes you are wearing), while antibiotics and modern vaccines revolutionized medicine.

Our physical landscape also changed: our cities were reshaped by skyscrapers, subways, and elevators.

This was the invention of modernity itself.

New Jersey played a starring role in this transformation.

The state was central to the global technological revolution from the earliest days of the Industrial Revolution.

Starting Up New Jersey’s Innovation Era: The Late 1800s

The gifts of New Jersey are perhaps too voluminous to mention in the context of a blog post. Several researchers have written about this in various forms, and there's always a list floating around once in a while in a local New Jersey newspaper about all the great inventions coming out of New Jersey.

Perhaps the most important of all of New Jersey's institutions for innovation was Thomas Edison and his industrial research laboratory in Menlo Park. This was Menlo Park, New Jersey, rather than California.

Edison's invention factory produced a series of breakthrough technologies during the late 1800s. From the phonograph in 1877, which allowed for the mass production of sound, to the carbon microphone for telephone transmitters in 1877, the most powerful of these inventions was perhaps the development of a practical incandescent light bulb in 1879. Edison tested thousands of filaments in a trial-and-error process (that he would repeat to solve other problems) to figure out how to create a light bulb that would be economically and practically viable. This one invention alone shifted humankind's relationship between night and day, transforming life as we know it.

Even foreign companies such as the Marconi Company, based in the United Kingdom, incorporated its US entity, the Marconi Wireless Telegraph Company of America, in 1899 in New Jersey. The first station in America was located in New Brunswick.

Creating Modernity in the Early 1900s

By the Early 1900s, it was more than just Edison who was making breakthroughs in his New Jersey laboratory. Brothers Robert Wood Johnson, James Johnson, and Edward Johnson founded Johnson & Johnson in New Brunswick, New Jersey in 1886. Johnson & Johnson, known for consumer healthcare products today, developed a set of breakthroughs that would transform surgery, including the first industrial sterilization process, the Band-Aid, and the first maternity kit. These breakthroughs revolutionized medicine, preventing the deaths of perhaps millions of individuals due to infections caused during surgery or even minor scrapes. Johnson & Johnson, in response to the 1918 flu pandemic, invented and then distributed a mask that would help prevent the spread of the flu.

It was more than just technology and medicine where New Jersey excelled. Joseph Campbell founded Campbell's Soup Company in Camden, New Jersey. Campbell's hired John T. Dorrance, who was a chemist with degrees from MIT and Germany, to develop a process for condensing soup, as the heaviest ingredient was water. What's interesting about this scientist-turned-businessman, Dorrance, is that he eventually ended up buying Campbell Soup and turned the business into one of the greatest American brands ever.

Post World War II and the Digital Revolution

If all of this sounds impressive, there was more to come.

Post World War II, New Jersey was to become one of the greatest fountains of innovation in human history. In particular, AT&T's Bell Laboratories in Murray Hill, New Jersey, was to become one of the epicenters of the digital age.

One of my favorite books for a general audience on innovation is “The Idea Factory.” Many have written about Bell Laboratories in one way or another as a fountain of innovation driven by large corporate labs staffed with brilliant basic scientists, from physicists to chemists to mathematicians, and engineers, who would solve both long-range and short-term problems for the monopolistic American Telephone and Telegraph.

Bell Labs was responsible for a dizzying number of inventions, from the transistor to the Telstar satellite communication system, to the Unix operating system, the first mobile phone, and major advancements in fiber optics. In addition to these breakthroughs in hardware, there were conceptual breakthroughs too that happened in New Jersey, including Claude Shannon's development of information theory, which he began at MIT but developed further at Bell Labs.

Post World War II, New Jersey was also the scene of major pharmaceutical breakthroughs, including the development of Valium in 1963, which skyrocketed Hoffmann-La Roche to become one of the major pharmaceutical giants.

New Jersey was a powerhouse beyond just corporate research. Two of New Jersey's major universities, Princeton and Rutgers, were hotbeds of innovation in their own way. The Institute for Advanced Study in Princeton, New Jersey, an independent research institute at one point headed by Robert Oppenheimer and staffed by people such as Albert Einstein, proved crucial to global technological progress. Princeton University housed John von Neumann, who pioneered computer architecture and game theory, Alan Turing during his foundational work in computer science, and John Nash with his Nobel Prize-winning work in mathematics and economics. At Rutgers there was Selman Waksman, who discovered streptomycin (a major treatment for TB) and won the Nobel Prize, along with key advances in ceramics, pharmaceuticals, and environmental science. The Institute for Advanced Study hosted Einstein, Oppenheimer, Kurt Gödel, and Freeman Dyson.

But now, when you think of innovation, you think of California rather than New Jersey. What happened?

Why wasn’t Silicon Valley in New Jersey?

My colleagues Ashish Arora, Sharon Belenzon, and their co-author Andrea Patacconi document that starting about the same time Gordon identifies as the end of the special century, from 1870 to 1970, you began to see a pretty drastic decline in the production of basic academic publications by corporations. I've discussed this before in a different post, but the decline of science in corporations was a pretty significant moment in the history of innovation.

But it also hints at what happened to New Jersey.

Despite the fact that some of the world's most advanced corporate research laboratories, Bell Labs, RCA, and Roche, were based in New Jersey, the ecosystem that generated breakthrough innovations appeared to falter.

Little by little, New Jersey's edge in the innovation game was being lost.

What California Did Right

At the same time, California was beginning to build an ecosystem in its own right. As always, the fundamental input to any innovation ecosystem is talented men and women, both individually brilliant and collectively spectacular.

Perhaps one of the founding stories of Silicon Valley is the movement of William Shockley from Bell Labs in New Jersey to Palo Alto, California, where he became a professor at Stanford University.

As I've written before, Stanford at this time was being led by Frederick Terman, who was slowly transforming the university from an intellectual backwater to one of the greatest universities in human history.

Shockley's intemperate behaviors, the self-confidence of the young scientists working for him, and a chance encounter with Arthur Rock led to the creation of Silicon Valley: startups, first Fairchild Semiconductor, then Intel, and the financing mechanism of venture capital. This replaced the need for the deep pockets of large corporations to fund long-term basic science aimed at maintaining monopoly power, with clear blessings from the federal government and the restrictions that came with it.

Several factors drove this shift from a corporate research model to a disaggregated startup ecosystem: a steady flow of motivated individuals graduating from universities in California and a consistent migration of people moving west from New York and New Jersey. Changes in capital markets and various other developments, such as Stanford's creation of the technology park on its campus to enable students to start companies using technologies developed in university labs, also contributed.

One of the most notable failures of the corporate research model for me was Xerox PARC. Xerox, a traditional East Coast company, established a research center in Palo Alto. PARC was an innovative and brilliant place, and several books have been written about it, including one of my favorites: “Dealers of Lightning.” PARC was where the graphical user interface, Ethernet, laser printing, and the mouse were developed. However, Xerox and its management, often called the “toner heads,” only focused on their existing markets and allowed Silicon Valley startups to take these ideas and develop them further. As a result, both Apple and Microsoft implemented many of Xerox’s innovations in their own products.

The new ecosystem, built with these various complementary components, was able to turn ambition, intelligence, and money into businesses and ideas that reached product markets faster than any slow-moving old corporation back in New Jersey.

Once these complementary institutions reached a critical mass, Silicon Valley was more than just a collection of startups; it was a complex set of interconnected institutions with significant feedback loops (see Saxenian’s Regional Advantage). A startup, for instance, generates returns for its founders and investors while also training the next generation of founders who now have built connections with each other and can source both money and human capital.

The Great Brain Drain

By the 1980s, the narrative around Silicon Valley had itself become a powerful competitive edge.

Every day, the Silicon Valley chronicles keep being written: about how young people with their brains, ambition, and the support of an ecosystem could make a mark on the world.

It began with Hewlett and Packard in a garage; then the Traitorous Eight broke away from Shockley to create Fairchild Semiconductor and later Intel, the startup of startups; then the next wave of startups came with Apple, Sun Microsystems, and then Google; and now the AI boom is here, based on the same fundamental story: Go West.

New Jersey as America’s Education Factory

What happened to New Jersey?

It could hardly compete with Silicon Valley; how could anyone?

Despite this transformation, New Jersey is not just another has-been backwater: it remains a vibrant, growing state. It remains #1 in terms of population density, #11 in total population, #3 in per capita income, and numerous other rankings.

But most importantly, New Jersey remains one of the best places in the entire country to get an education. The state consistently ranks at or near the top in many educational rankings. US News & World Report ranks it number one in the nation. Another website (datapadas.org), which gathers data from various sources, also ranks New Jersey number one, followed by Massachusetts and then Florida.

California, on the other hand, isn’t really doing that great in this regard.

What's striking about New Jersey is that statistics suggest a large percentage, somewhere between 30 and 50%, of New Jersey's high school students attend college out of state. There are excellent options for bright students from New Jersey, both in neighboring states, with New York hosting several Ivies, Connecticut having another, Pennsylvania offering some great universities, along with Maryland, and even further south here at Duke University. 

New Jersey is a net exporter of human capital. There's a brain drain going on.

In contrast, very few outsiders come to New Jersey for college each year. The main attraction is probably Princeton University, which is less of a local New Jersey college (even though it used to be called The College of New Jersey) than an international institution.

New Jersey's talent exodus seems like a sad coda to the disappearance of its innovation economy over the past half-century.

The Costs and Benefits of New Jersey

Having left the state, right after receiving my excellent and affordable public university education, I understand the economic calculus that causes people to leave.

People leave New Jersey for many reasons. I've highlighted some of the factors on the “benefits” side of the equation. New Jersey seems to lack the opportunity of other states when it comes to the frontier of the economy: technology, innovation, and even finance. Those things are happening in California, Boston, and New York.

But there's another side to this equation as well: costs. New Jersey is also a very expensive place to live. Property taxes are among the highest in the country. State taxes are also high. It's difficult for young people to put down roots.

In this way, North Carolina had significant advantages for quite some time (though it is also getting expensive here). It was a place where you could leave New Jersey, stay in the same time zone, enjoy better weather, and maintain an affordable middle-class lifestyle with a solid middle-class job.

California, though the situation was somewhat different, still offered significant benefits, even though the costs were also high (a typical house in Silicon Valley now costs well over $1 million). So if that cost-benefit balance works for you, it could be an excellent place to live. If you’re ambitious, it's worth considering. However, you need to understand that California has an up-or-out economy. There’s little room for the normal person with a regular job.

The Rise and Fall of Innovation Economies

But the rise and fall of regions is more than just a story of New Jersey. It appears to be a recurring pattern in economic geography that suggests something more fundamental about the rise and decline of innovation ecosystems worldwide.

We've seen these stories before. For example, the rise and subsequent fall of Manchester's textile industry. Manchester dominated global textile production for over a century. It had remarkable competitive advantages and specialized infrastructure, both economic and technical. It seemed like it would be the textile capital of the world forever.

But soon, competition arrived in the form of America. Cheap cotton from the American South shifted production away from Manchester to the United States, and then farther afield, eventually reaching Asia. Manchester's population peaked in the 1930s at 766,311 during the interwar period and dropped to nearly half that by 2001, with 392,819 residents. Today, the population of Manchester has grown again, but it is Manchester in name only. It is a different city with different people and different industries.

This appears to be a universal pattern—the cycle of birth, death, and reincarnation of major innovation hubs across different regions. We saw this in Detroit, where the population dropped from a peak of 1.8 million to 600,000 today, in Pittsburgh, and later in New Jersey. These cycles also trace back even further to the rise and fall of other important regional economies: Rome, Greece, and the Indus Valley civilization.

This pattern and its samsara-like rhythm are instructive and, in some ways, a predictable outcome of the forces of efficiency and the inertia that result from developing a well-oiled economic machine with complementary assets that are highly specific to a particular purpose. Specialized infrastructure, tacit knowledge, dense networks of suppliers and talent—all optimized for producing a specific type of value—drive these complementarities, which generate increasing returns, fostering a virtuous cycle and competitive advantage.

But this optimization can become a trap that may be hard to escape.

And once the game is set, an equilibrium emerges. We may start to see the gaming begin, which always happens when a place shifts from primarily creating value to extracting value.

Soon, new places emerge that have figured out how to generate more value using a completely new system of value creation. People, money, and other resources start going where the action is.

Then the cycle repeats.

“Just Being” New Jersey

As the cycle makes another rotation, maybe New Jersey has discovered something profound.

While Silicon Valley relentlessly pursues the next unicorn, Pittsburgh grinds to reclaim its storied history, and North Carolina tries to win the next corporate back office, New Jersey has achieved something special: the liberation that comes from having nothing to prove.

New Jersey gave the world the light bulb, the transistor, and satellites.

And now it seems completely content with the fact that you don’t know this.

I’ve started to rethink the fat cat sandwich from Rutgers' grease trucks. The fat cat is the embodiment of “just being”: excessive, authentic, and couldn't care less about what you think.

What I learned from living in New Jersey is perhaps that not every place needs to be an innovation hub chasing agglomeration and increasing returns. To many, New Jersey's lost glory and its outflow of human capital may be viewed as a tragedy.

However, it may actually be New Jersey's way of giving back to America, not through a flashy app that goes viral, but through the human capital (and the attitude that comes with it) that makes it possible.

The endless cycle of regional rise and fall, as seen in Manchester, Detroit, and possibly Silicon Valley next, hints that trying to stay permanently relevant is like trying to hold water. The more regions squeeze, the faster they lose their advantage. New Jersey seems to have learned this by chance: by loosening its grip, it discovered an unexpected balance.

In my mind, New Jersey exists in an eternal present, neither mourning too hard for the past nor reaching too far for the future, simply doing what needs to be done.

Maybe New Jersey's greatest gift wasn’t the transistor, the light bulb, or its many students who go elsewhere.

Its gift might be the understanding that a place doesn’t need to prove anything to matter; it can simply exist.

When we talk about “impact,” we often mean different things: sometimes, impact means transforming one individual’s life, and sometimes it means reaching many people with smaller changes. Like in the design of experiments, our efforts at impact tend to face a tradeoff between effect size and scale. It's natural to want both, but this can lead to an incoherent strategy. Consider instead choosing your focus: either intensive, personalized work that transforms a few lives, or scalable systems that benefit many. Either path can be meaningful as long as you know where you are headed and design accordingly.

The Problem of Impact

I was going to call this post “Middle Age and the Problem of Impact.” But I figured that might hit a bit too close to home.

Like many readers, I presume, I am at a point where many of my peers have accomplished more in their careers than they ever imagined.

So the obvious next question is:

What’s next?

Usually, the answer is something like:

Impact.

But what does that actually mean?

While “impact” captures something important we all care about, the term can be imprecise. Breaking it down into more specific components—namely, effect size and scale—and using the language and methodology of experimental design might help us think more clearly about how to achieve our goals.

Why should we think in terms of experiments?

When starting this reframing exercise, it might be helpful to consider how researchers, at least in the social sciences, think about the ideal ways we can both enact and measure change in the world.

In social science, we are often interested in understanding cause-and-effect relationships. Oftentimes, these relationships are analyses of things that go on in the world independent of us as researchers. For instance, we might ask: what is the impact of innovation on firm performance? Or what is the impact of being in the center of a social network on the ability to generate novel ideas?

A challenge with answering whether this cause-and-effect relationship is real (e.g., whether the "if x then y" relationship is causal) is that there could be many other factors at play simultaneously. So we might confuse correlation, which means two things happening at the same time, with causation, where one causes the other.

To get a better handle on causality, researchers often run randomized controlled experiments.

What is an experiment, and how do you design one?

Let’s say you have a theory that X causes Y.

For instance, you believe a new teaching method helps students learn better and want to test if this is true. A natural starting point is comparing student performance before and after implementing the new method. However, the problem with this approach is that other factors could explain any improvement you observe.

• Students might have grasped more foundational concepts by the time you implement the new method.

• The assessment method itself might have changed.

• Or students might have supplemented their learning with Khan Academy because the initial teaching was ineffective.

So a simple pre-post comparison cannot determine the causal effect.

A Randomized Controlled Trial helps deal with this problem of causal ambiguity. By randomly assigning some students to the new teaching method and others to the old one (e.g., flipping a coin to determine whether you get the intervention or not), you control for all these other confounding factors. The students and their experiences are identical except for the flip of the coin and this one change: the intervention, the new teaching method.

At the endpoint, you compare the two groups and their outcomes rather than comparing before and after. Did the new teaching method move the needle relative to the counterfactual? Both groups experience all the same external changes and are similar (or “balanced”) on both unobservable and observable traits, but only one receives the intervention.

The magic of randomization is that the two groups become essentially identical in every way that matters, except for the one thing you're testing.

This difference between the two groups in terms of the outcome is what tells you whether your theory, this innovation in teaching, actually translated into meaningful change, or whether you were fooling yourself.

Did your intervention have an effect?

Now imagine if you ran this experiment and measured outcomes. The table below (with simulated data) shows the results of a two-sample t-test comparing the control group (n=517, mean=49.81) to the treatment group (n=483, mean=54.99). The difference of 5.18 points between groups is statistically significant (p<0.01), providing strong evidence that the new teaching method caused improved learning outcomes.

The graph below visually demonstrates this treatment effect as a rightward shift in the entire distribution of scores.

That difference between the treatment group and the control group is the quantity of interest for us: your treatment effect.

This is YOUR effect on these students’ learning.

The treatment effect is the causal impact of your intervention measured as the difference in outcomes between treated and control groups, but more fundamentally, it represents your ability to create meaningful change in the world: in this case, the improvement in students' learning that exists because of something you designed and implemented.

The effect size of 5.18 points could have been larger or smaller depending on countless factors: the specific design choices you made, the context in which you implemented the intervention, and the particular students involved, making this final number a unique outcome of all these factors.

The tradeoff between effect size and scale

Recently, I discussed an impressive field experiment at the National Bureau of Economic Research. The study tracked Ugandan high school students for 9 years after early-life entrepreneurship training (impressive!). The researchers found large, persistent treatment effects well into the ninth year. For anyone familiar with entrepreneurial training literature, these results stand in stark contrast to most programs, which typically show much smaller or null effects.

The size of the eventual treatment effect represents a deliberate design choice by the researchers: given a fixed budget constraint, they chose to invest heavily in teacher training and curriculum localization rather than spreading resources thin to reach more students with more generic content. This allocation decision reveals their theory of change, that deep, contextualized preparation of instructors would generate larger treatment effects than the typical approach of maximizing reach with off-the-shelf materials. This is a particular bet on the impact they wanted to have.

To me, the design of the intervention illustrates the fundamental trade-off between effect size and scale. The researchers chose to invest heavily in instructor training and localization, making their program harder to scale but dramatically more effective.

Scaling becomes difficult because localized content may not translate across contexts, high-quality instructor training takes substantial time and resources, and adapting materials to each new setting requires deep local knowledge. While they couldn't easily roll this out to thousands of schools, the students they did reach experienced real, lasting change. Meanwhile, generic programs can reach far more people but with smaller impacts. Interestingly, the total effect (effect size times the number of students reached) could be similar for both approaches.

The figure below captures the trade-off that we often face in designing interventions: maximize effect size or maximize scale.

Of course, this framework presents a simplified model. In reality, some interventions do achieve both meaningful depth and broad scale, though these tend to be exceptions that often involve technological breakthroughs or fundamental innovations. The framework is most useful as a starting point for thinking about tradeoffs in our normal day-to-day actions, not as an absolute law.

Why does this trade-off exist?

At a glance, it seems like we should be able to take interventions designed for maximum effect size and simply replicate them broadly. But when you consider the logistical and technical constraints, as well as the statistical properties of experiments and when they actually produce effects, the trade-off between effect size and scale becomes less surprising.

Limited resources. The most obvious constraint is limited resources. Large-effect size interventions frequently rely on intensive inputs, both financial and human capital. You can see the difference in effects when you compare actively coaching and mentoring people with trained facilitators versus sending emails with high-level content on the same topics. This intensity of the first approach also enables better implementation, ensuring the intervention is delivered as designed. The counterpoint is, however, that with the money it takes to provide intensive support for a handful of people, you could reach thousands, even millions, with an email intervention.

Properly targeting your interventions. Beyond resources, targeting plays a critical role in determining effect size. A recent study in the American Economic Review found that targeting financing interventions to individuals with latent entrepreneurial ability nearly tripled the effect size, but acquiring such targeting information can be quite costly (Hussam, Rigol, and Roth 2022).

At scale, targeting precision is lost: eligibility criteria loosen, populations become more heterogeneous, and the average treatment effect gets diluted by less-engaged or marginally relevant participants.

Heterogeneous treatment effects. Relatedly, interventions affect people differently: what helps some may not help others, and can even cause harm. As interventions expand to include broader populations, heterogeneity in both baseline characteristics and contextual conditions increases. This leads to greater variance in outcomes and a flattening of average treatment effects.

Measuring the right outcome. Another crucial design choice is what to measure as your outcome. As your population expands, you need more generic measurements like wages, which not everyone might be optimizing or care about. But with a narrow, targeted population, your measurement can align closely with what participants actually value, creating a tighter coupling between treatment and meaningful outcomes. Similarly, more generic outcomes and the ability to collect data on them may introduce noise that makes your effect hard to detect, especially if it's small.

Feedback loops and experimentation. Finally, rarely is intervention design a one-time exercise; you're engaged in ongoing experimentation where each intervention generates learning that shapes your next iteration, making it crucial to think through how your measurement choices and observed outcomes will inform future design decisions.1

Designing a Path To Impact

Just like in field experiments, a strategy is often premised on first finding a position on the curve, and then deciding how to allocate resources to move incrementally closer to that position.

Indeed, this strategic positioning between effect size and scale parallels classic business trade-offs: differentiation versus cost leadership (Porter), specialization versus generalism (Hannan and Freeman), or, in statistical modeling terms, the bias-variance tradeoff.

Trying to maximize both effect size and scale simultaneously can create tensions that make it harder to excel at either. It’s like a coffee shop trying to be both a cozy neighborhood spot and a chain: it's plausible but challenging.

So you have to pick a point and move toward it.

If you've decided to go for high effect size (Path A), you will need to target precisely, tailor your approach, and be relentless in solving edge cases to figure out how to deliver impact for the few people you're trying to help. By the time you have impact, your “intervention” will be bespoke and narrow, and most of what you've done might be hard to generalize and even invisible to outsiders.

On the other hand, if you decide to go for high scale (Path B), you need to build a completely different kind of infrastructure, one that focuses on distribution, delegation, and generality, with the ability to operate without your involvement in making people's lives better.

Strategic Incoherence and the Bias Towards Scale

As such, the desire for “impact” often conceals a fundamental misunderstanding of the trade-off discussed above: wanting both a large effect size and a broad scale simultaneously is rarely possible.

This pursuit might sometimes reflect our natural desire for visible impact. Scale provides tangible evidence of our work, which is understandably appealing even when smaller-scale work might be more transformative.

So, our bias toward scale arises not because scale necessarily represents the most effective path to change (though it can be for some problems), but because it produces the most visible outcomes.

It satisfies our innate need for external validation.

However, failing to recognize that effect size can naturally decrease as interventions expand, and prioritizing what is visible over what is effective, can be self-defeating.

So is there a way out of this dilemma?

One approach is to start at one place, say, a highly specific solution for a narrow group, solve that really well, and then generalize.

This is a common strategy for startup founders, something Paul Graham has written about: start with things that don't scale but solve problems, then slowly expand while accepting that generalization will dilute effectiveness.

Alternatively, we can start broadly and then narrow down. This is like a foundation model for generative AI that helps lots of people a little bit, but no one incredibly well, then gradually, a company can add complexity through fine-tuning or RAG systems to increase accuracy for them at the expense of scale and generality.

Principles for Designing Interventions

Given the strategic positioning choice between effect size and scale, and understanding why we might become biased toward scaling, we probably should turn to a practical question: how do we design interventions that move the needle?

Design a Strong Intervention

One trick I now use in designing field experiments is to assume that most systems are in equilibrium, held together by economic, sociological, and psychological forces such as incentives, norms, and habits. These forces generate the stable patterns that keep the world turning (albeit dysfunctionally at times).

Any intervention aiming to change behavior must fight this equilibrium, and lighter-touch interventions like emails or nudges can work, but they often need to be especially well-designed or well-timed to overcome these existing forces.

To truly make a difference, we need to first understand the equilibrium and then design interventions with the right force to shift it.

Find the Invisible Zeros

One insight from Michael Kremer's O-Ring theory (named after the faulty O-Ring on the Challenger) is that complex processes often fail due to hidden bottlenecks that cause performance to drop to zero. These hidden bottlenecks destroy the value of everything else.

These zeros hide in different places depending on the problem you are trying to solve.

When designing interventions at a small scale, we should consider zeros that hide in implementation details: facilitators who don't show up, curricula assuming knowledge students lack, and follow-up systems that break down. For scaled interventions, zeros appear in structural assumptions: assuming that algorithms will prioritize your content or distribution channels that miss your population entirely.

Finding and fixing these zeros matters more than optimizing other parts of your intervention.

In sum, the multiplicative nature of impact can be unforgiving.

Find Responsive Subjects

As the Hussam et al. paper suggests, behavioral change follows different response functions for different populations. Some groups are more receptive to the kind of intervention you are designing because they sit near the tipping points where even small pushes create significant shifts (e.g., giving money to high-ability entrepreneurs unlocks a fundamental constraint, while giving it to low-ability ones does not), while others are locked in stable equilibria that negate even large interventions.

This creates an optimization problem: given any budget, you maximize impact by targeting those with the steepest response gradients, the people ready to change. The principle applies whether you're going for depth or breadth.

Measure What Matters

Have a clear theory of change before you intervene: what mechanism will drive the outcome you care about? And, is this the outcome you care about?

Without this clarity, you cannot distinguish between interventions that work through your proposed mechanism versus those that create noise or move irrelevant proxies.

This requires specifying a clear “model” of the system you are trying to change. If you believe mentorship drives career advancement through skill development, then measure skill acquisition, not just meeting attendance.

When you measure outcomes aligned with your causal theory, you can create feedback loops that reveal whether your mechanism operates as theorized, allowing you to iterate based on evidence.

Push when Change is Cheaper

Most systems alternate between periods when change is cheap and periods when change is expensive.

Even small interventions during the “change is cheap” moments can shift entire trajectories. At the same time, the same effort during periods of stability (e.g., fighting the equilibrium) gets washed out without having an effect.

For individuals, these moments often occur during transitions, periods of uncertainty, or when motivation varies. The key to effective intervention design is recognizing the right time to push.

Moving the Needle

“Impact” is the word we use because we currently don’t have a better one for the idea that we want to do things that matter. In the end, what we care about is some combination of effect size and scale.

This shift in framing leads to some interesting questions, ones we can better wrestle with using existing frameworks from the social and statistical sciences.

What kind of effect am I trying to have? Do I want a big effect on a small number of people, or a smaller effect on a large group?

Once we see that there is often an inherent trade-off, it becomes more challenging to avoid “forcing” a choice.

This is because the things we must do to maximize effect size are often the very things that make scale harder, and vice versa. These are different problems, with different solutions.

It helps to be deliberate and thoughtful about these choices. Understanding the problem deeply and designing accordingly tends to yield better results than hoping something will work out.

We're all trying to make a difference in our own ways. Thinking with an experimental frame is just one approach that might help us be more intentional about the kinds of change we want to create, and more at peace with the tradeoffs we choose.

TL;DR.

Knowledge exists in two forms: explicit (can be written down) and tacit (embedded in people and relationships). While explicit knowledge survives in books and databases, tacit knowledge lives only in our minds, hands and our connections to others. This includes the intuition, judgment, and craft wisdom that separates masters from students. Like family stories that exist in just one relative's memory, this knowledge can disappear when relationships break or experts retire without apprentices. The Tasmanians once knew how to make fire and craft bone tools, but isolation slowly erased these skills over generations. Today, as aging populations carry specialized knowledge, institutions dismantle, and AI replaces collaborative networks, we face a similar risk of forgetting.

Tracking Down Vanishing Knowledge

When I was a teenager, I was obsessed with genealogy. During family trips to India, I would track down distant relatives and ask them hundreds of questions about our ancestors. Names, stories, the origins of our family, what people did for a living, what they ate, and who I should talk to next. I was always on the lookout for little treasures: a photograph tucked away in piles of documents in an attic, a story that lived in only one person's mind, a historical document that provided a missing link to some distant branch of the family in a faraway place. I always carried a notebook with me (which I still have), where I would document all the little facts I learned, fearing that I too would forget these things one day.

As I’ve gotten older, and many of those same relatives have now passed on, what strikes me most is how fragile this knowledge was to begin with. These relatives were often the sole keepers of entire branches of our family history. When they died, this knowledge would vanish too.1

The Tasmanians

This fragility of knowledge turns out to be a much larger human phenomenon. My co-author and friend John-Paul Ferguson, when we were writing The Lives and Deaths of Jobs, shared what is now one of my favorite articles titled “Ten Thousand Years of Solitude” by Jared Diamond. The article is about collective forgetting among the Tasmanian people.

The Tasmanians had advanced technologies 10,000 years ago. They had bone tools for making needles and warm clothing, fishing techniques that provided 10 percent of their diet, and fire-making skills.

However, rising seas cut them off from the mainland, trapping just 5,000 people on the island.

Slowly, knowledge began to vanish.

By the time Europeans arrived, the Tasmanians had stopped eating fish. Their bone tools had disappeared around 3,500 years ago. They had lost the ability to make fire and instead carried firebrands, relighting them from neighbors when they went out.

Their knowledge had slipped away, one generation at a time, with no neighboring tribe to relearn from and no way to recover what was lost.

The Tasmanian case offers something that social scientists are always seeking: a controlled comparison. In this case, we know what technologies they started with and what they had lost by the end, giving us a much clearer picture of how knowledge vanishes over time.

Diamond's account is so compelling because it highlights several specific mechanisms behind knowledge decay: population thresholds below which skills can't be sustained, isolation that prevents relearning, and how useful practices can drift into taboos, ultimately leading to the disappearance of productive capabilities.

This prospect of societal forgetting seems impossible in the information age: everything, every click, is stored and shared worldwide in fractions of a second.

Yet, we face significant forces today that could make it happen: aging populations carrying specialized knowledge that dies with them, the deliberate dismantling of complex institutions that helped create the prosperity of the last 150 years, market forces and competitive pressures that make rare knowledge more valuable while also limiting its spread, and now, potentially, hierarchical corporate-owned AI systems replacing horizontal networks of specialists.

What is knowledge?

The concept of knowledge has been analyzed by philosophers for thousands of years. Plato's definition of “justified true beliefs” often serves as a helpful starting point2, and many philosophers have built upon and refined (and disputed!) this view over the centuries. (And if you’ve been around philosophers, you know that every word is scrutinized: What does justified mean? What counts as true? What is a belief?3)

From the perspective of an organizational researcher, the most significant modern advancement in our theory of knowledge comes from Michael Polanyi, a chemist-turned-philosopher who spent his later years grappling with a fundamental question: “why we can know more than we can tell?” I finally read his treatise, The Tacit Dimension, a few years ago and think about it nearly all the time.

Polanyi made a crucial distinction between two fundamentally different types of knowledge.

Explicit knowledge, which can be written down and transmitted through books, articles, or instruction. Think of the Pythagorean Theorem: Any 7th grader can memorize a² + b² = c² and apply it to solve problems.

Tacit knowledge is something entirely different in nature. It is knowledge that resides in our minds and hands, in both our practice and our intuition. It's what separates someone who can recite the Pythagorean Theorem from a carpenter who truly knows how to square a corner, or an architect who understands proportion and spatial harmony in ways that shape every design decision. And of course, a mathematician who has spent decades working with geometric proofs. In all these cases, the formula itself is just the beginning of what it means to truly know something.

This distinction between explicit knowledge and tacit knowledge is everywhere. In my field, there's a marked difference between a first-year student who has taken a few econometrics classes and a theory seminar and someone with decades of experience who has an intuitive feel for data, models, and their connection to the structure of a scientific theory and its predictions.

Where Tacit Knowledge Lives

Where these types of knowledge are located also matters.

Explicit and codifiable knowledge can be stored for long, uninterrupted periods, in libraries, databases, and, now, the cloud4.

Tacit knowledge is far more delicate.

It primarily exists in two places: within people (not just their minds, but also in their hands, feet, and muscles) and in relationships among people.

When scholars say that knowledge lives “in people,” what they mean is that it is embedded in their practice, muscle memory, and judgment.

The master sushi chef, Jiro Ono, possesses decades of tacit knowledge about rice temperature, fish quality, and timing that no cookbook can capture. Moreover, he probably can’t convey all the things he “knows.”

Similarly, the sarod masters in Amjad Ali Khan's lineage, both past and future, carry knowledge about instrument construction and playing techniques that have been refined through generations of practice. This produces a sound that takes decades to perfect in one person and centuries to reach its current state.

But perhaps more importantly, tacit knowledge lives in relationships, in the social fabric that connects people.

The idea that knowledge resides “in networks” is central to many organizational theories (and, obviously, in Sociology and Anthropology too).

March and Simon, in their 1958 masterpiece “Organizations,” were among the first to argue that organizations adapt through routines or programs that emerge from repeated interactions between people with bounded rationality.

Building on this foundation, Nelson and Winter's “An Evolutionary Theory of Economic Change” (1982) provided the most systematic treatment of organizational routines. Their key argument was that knowledge does not reside solely in individual minds, but in the practiced interactions between people engaged in interdependent work. (For a thorough review of the literature on knowledge and knowledge transfer, the work of Linda Argote is a must-read.)

For example, a surgical team's tacit knowledge emerges from their repeated collaboration, their unspoken communication, and their ability to anticipate each other's needs.

This relational dimension explains why knowledge loss in organizations and society can be so sudden, and so devastating.

When relationships dissolve, when teams disband, when master craftspeople retire without apprentices, when communities scatter, the tacit knowledge embedded in those connections can vanish with them. The knowledge held in human networks requires constant activation and use to survive.

Hierarchical vs. Network Knowledge Systems

We can consider knowledge systems at two ends of the spectrum: hierarchical and network-based. Recognizing this difference helps explain why some knowledge is more prone to forgetting.

Hierarchical systems follow a top-down, one-to-many model.

For instance, textbooks, online courses, and AI systems are quite hierarchical in nature. These systems excel at transmitting explicit information efficiently and at scale. A single textbook can teach millions of students; it’s a 1:m system.

An AI system can instantly provide answers to billions of queries. The efficiency and scale are remarkable, but these systems can only transmit what can be explicitly documented or programmed (it’s an interesting question as to what a non-documented but robust statistical relationship between concepts that emerges in these models counts as explicit or tacit?).

Network systems operate through peer-to-peer (1:1), many-to-many (m:m) relationships. For instance, research laboratories, craft workshops, or apprenticeship programs are network-type systems. In these systems, knowledge flows through interaction, observation, co-creation, and practice. A master craftsperson does not just tell an apprentice how to shape wood; they work side by side, sharing countless little adjustments, corrections, and insights that accumulate into knowledge. These networks are slow and costly to maintain, requiring sustained relationships.5

The critical difference lies in what each system can transmit:

• Hierarchies excel at distributing codified knowledge but cannot convey the intuition, judgment, and embedded know-how that constitute tacit knowledge. You can learn music theory and chord progressions from a YouTube video, but you cannot learn the timing, feel, and musical instincts required to actually jam with other musicians this way.

• Networks, by contrast, are specifically designed for this kind of tacit knowledge transfer through repeated interaction and shared work.

This distinction in how knowledge is stored highlights a vulnerability: once a knowledge network collapses, its tacit core can vanish forever.

When the last master craftsman retires without an apprentice (or is laid off when a private equity firm takes over), or when research teams disperse due to funding cuts, or when communities of practice dissolve because no one wants to organize them, the tacit knowledge embedded in those relationships disappears.

No amount of documentation or AI can recover what was never codified in the first place.

The Economic Value of Tacit Knowledge

Tacit knowledge is interesting from both philosophical and sociological perspectives, but it is also the foundation of durable competitive advantage in many industries.

Consider ASML's extreme ultraviolet (EUV) lithography machines, which enable the production of the world's most advanced computer chips (e.g., those used in AI). ASML’s dominance in this space provides a textbook example of how tacit knowledge can create insurmountable competitive moats even in high-tech manufacturing.

The key ASML technology consists of machines (often costing up to $370 million) that manufacture transistors through “printing.” The precision is incredible: the machines can print transistors that are the length of five DNA strands (or approximately ten thousand times narrower than a human hair, according to the article I link to). Given how delicate and nuanced this production technology is, it is not surprising that it took nearly two decades for this technology to hit the market (with billions in investments). What was learned over these nearly two decades? Lots of tacit knowledge that came from solving specific problems, both big and little ones.

Furthermore, these problems are not solved solely by ASML’s scientists or engineers, but are solved collaboratively with thousands (yes, thousands) of suppliers, whose IP constitutes nearly 85% of all the IP (this linked article is superb!) that goes into the final product. Moreover, its buyers (Samsung, Intel, TSMC) also work with ASML to solve specific problems that they face. The knowledge is embedded in very bespoke networks inside ASML, among its suppliers, and with its buyers. In an interesting CNBC article, the CEO talks about the embeddedness of ASML’s knowledge, as per Granovetter:

“We’re unique to some of our customers, and some of our supplies are unique to us,” Wennink said. “And those almost symbiotic relationships, some people say, are worse than being married because you cannot divorce.”

That is embeddedness.

This knowledge is so subtle and tacit that ASML also embeds customer support engineers with their machines (the full WSJ article is worth a read because it highlights the surprising solutions that come from tacit knowledge—e.g., using two Home Depot buckets to fix a multi-hundred-million-dollar machine).6 These engineers are trained for over a year on all the subtleties of these incredibly complex machines. So if problems arise, they can be solved on the spot.

As one might imagine, this knowledge is remarkably valuable, and people have attempted to steal it or replicate it in other ways by investing billions in the development of similar technology. Thus far, to no avail.

There is, however, a flipside to this economically valuable trove of tacit knowledge that ASML's dominance rests on. If key holders of unique, tacit knowledge leave, the network begins to fray, or reckless cost-cutting removes important but invisible nodes, ASML may lose critical knowledge; more importantly, the world may regress in terms of its technological capabilities because the knowledge is embodied in people and relationships.

Examples of such technological forgetting are plentiful. For example, the loss of knowledge about Roman concrete as the Roman empire disintegrated is a notable example of a significant and superior technology that was lost until modern science was able to reverse-engineer it (described here in a more straightforward manner). Similarly, Roman knowledge on how to build domes was also lost until the Renaissance, when it was revived through the ingenuity of Brunelleschi.

Likewise, the gradual decline in the skill of Ukiyo-e woodblock printing in Japan was driven by multiple factors, including the emergence of competing technologies, public disinterest, and government policies that discouraged its practice (see the section on its decline on the Wikipedia page). Yet, the art form survived—though not in its original form—due to unexpected reasons, particularly interest from Western artists such as Whistler, van Gogh, and Cassatt. In recent times, David Bull, a Canadian printmaker who has been based in Japan for several decades, has also worked to preserve the tacit knowledge of this endangered art form (see, for instance, this three hour video where he makes a woodblock print from start to finish).

If you watch David Bull’s hands in this video you can understand Polanyi’s refrain: “we can know more than we can tell.”

The Economic Paradox of Tacit Knowledge

When I read about the ASML example, it suggests to me that we should not be too enamored by the competitive implications of tacit knowledge held by a single, dominant firm, nor too nostalgic about the unique wisdom of the experienced craftsperson.

This is because a fundamental economic tension is at play.

Tacit knowledge can be incredibly valuable both socially and economically. As a result, sharing that knowledge creates competition and reduces the holder's ability to capture value. That is, the subtle skills, judgment calls, and hard-won insights that make tacit knowledge valuable become less unique when it is shared with others. This economic dynamic creates powerful incentives for knowledge hoarding. And, in turn, this hoarding behavior is precisely what ensures the knowledge’s eventual disappearance.7

I’m not sure if this paradox has been supported with data, but if it is, it would be quite tragic.

However, today the forces that threaten tacit knowledge extend far beyond the endogenous economic incentives described above.

Our Tasmanian Dilemma: An Era of Mass Forgetting?

What we're witnessing today may be a perfect storm of endogenous (see above) and exogenous forces at play, which can cause societies to forget important, productive knowledge.

Here are just a few:

Dismantling of Core Institutions

Perhaps the most urgent issue is the large-scale dismantling of core public institutions. For example, agencies like the NIH are consolidating from 27 institutes to 8 with a 40% budget cut, ending over 800 research projects, and breaking apart 70-year-old interconnected research networks. These institutions are webs of tacit knowledge, and losing one node can be devastating; losing hundreds or even thousands of crucial nodes in the knowledge network may be nearly impossible to recover from. It is not just the NIH; dozens of institutions are being dismantled, and tacit knowledge — embedded in people’s minds, practices, and their relationships — may slowly be forgotten. This is perhaps the least of it, as the ripple effects of this institutional loss will be felt more widely.

The Aging Workforce

Another pressing issue is the problem of aging workforces in developing economies (my colleagues Yoko Shibuya, Ines Black, and Maria Zhu, and I) have an integrative review titled “The Aging Firm” (happy to share it if you are interested) and an early-stage empirical paper examining the firm-specific factors that lead to skewed (older) age distributions inside firms. A key tension a firm faces is getting work done now (and competing!) and passing on that knowledge to the next generations.

Japan, for instance, has seen a marked inversion of its population pyramid which threatens the loss of important production knowledge in the economy.

The United States hasn’t inverted yet, but its population pyramid is now more rectangular. We have already seen a decline in manufacturing throughout the US, and it’s not clear that policy or financial incentives can bring it back quickly, as the TSMC experience in Arizona demonstrates.

As the world ages, and the torch of tacit knowledge isn’t passed down, we may witness a monumental scale of forgetting.

Artificial Intelligence and the Hierarchical Knowledge Systems

Finally, the long-term effects of AI on knowledge worker productivity raise serious concerns that go beyond just short-term efficiency gains documented in some research studies.

Large language models excel at generating novel combinations from codified knowledge. However, these technologies raise important issues for organizations and the economy as a whole. These technologies risk replacing the natural growth of young talent and the networks of practice that have traditionally linked both explicit and tacit knowledge sharing. For example, you can ask ChatGPT to explain a mathematical theorem, but it cannot teach you how to think like a mathematician, build mathematical intuition, or identify which problems are worth pursuing. The widespread use of these systems may accelerate the erosion of human networks that store knowledge.

That is, when we can instantly access synthesized information, we may lose the slower but essential processes through which expertise is cultivated through study, observation, and human interaction.

Enrique Ide's compelling new working paper, “Automation, AI, and the Intergenerational Transmission of Knowledge,” explores this tension and should be essential reading for anyone grappling with how knowledge flows between generations in an age of artificial intelligence.

What Can Be Done to Prevent Forgetting?

While these are quite significant forces and most of these are things that individuals cannot solve on their own, there is probably a lot we can do to preserve valuable tacit knowledge in our own backyards.

At the organizational level, we can all start being more thoughtful about who holds what knowledge within our organizations and what networks store the relational knowledge that enables things to run smoothly.

For transmitting crucial, productive, tacit knowledge, we should think about how to create apprenticeship programs that go beyond traditional mentoring (or classroom education). Apprenticeship programs should be designed so experienced practitioners work alongside newcomers on real problems rather than sage-on-stage lecturing.

Similarly, companies can redesign incentives to reward knowledge sharing rather than hoarding (check out research by Alex Oettl on helpfulness here and here, as well as the brilliant work on awards and social contributions by Jana Gallus).

We should consider promotion criteria that value the development of others.

There is also much we can do to preserve the hard-won knowledge we have acquired. We too should focus on mentoring the next generation in our fields, and contribute on a 1:1 basis in our professional communities.

Most importantly, we should probably recognize that preserving knowledge is not just about our legacies, but about maintaining the collective tacit knowledge that enables societal progress.

When I reflect on my fascination with genealogy, one thing is clear: each of us may be the sole keeper of some valuable knowledge.

We have a responsibility to ensure it doesn't die with us.

“You wasted $381,500 on an education you coulda got for $0.00 and a few annoying Grammarly ads.”1

TL;DR

AI will make it easier than ever to generate high-quality educational content at little or no cost. If content is all you offer, it will be hard to compete. Universities have long survived by bundling content with credentials, community, and access to opportunity. That model still holds value, but each part of the bundle is facing new pressure. The challenge now is to rethink what makes the educational experience worth paying for and to invest in the parts that truly matter for learning and growth.

The AI Video Revolution

One of the most fascinating capabilities of artificial intelligence, which is advancing rapidly, is its ability to generate images and, more recently, videos. I've seen many impressive videos created by AI, and there's been a clear leap in the ability to produce videos from prompts over the past two years.

I recall being impressed a few years ago by a homemade video titled Harry Potter by Balenciaga. The person who made it clearly had creative skills, imagination, and, it appears, a decent set of technical skills to put this together.

However, as with all technological advances, this already impressive effort probably represented the most limited AI video we would ever see.

Today, models produce much higher-quality video and audio. More importantly, however, they are also far more accessible to the average person. We now have an entire TV channel featuring very creative (and quite impressive) AI videos generated with Google Veo.

I have no doubt that AI video generation technology will have a significant impact on many industries.

In this post, I will explore how AI-powered video generation could impact a sector I care deeply about: education.

Why Education Deserves Special Attention

Right now, I see a lot of discussion about AI video generation focusing on raw model power. There are entire genres of YouTube posts, LinkedIn demonstrations, and Substack posts that compare the subtle virtues of different video models (VEO3 vs. SORA vs....).

While I believe that raw model power is an essential factor in understanding how this technology will affect organizations and the economy (and society), I think that reasoning about the impact of AI starting from “disembodied capabilities” is much less valuable than thinking through how a technology will fit into the organizational and economic systems of specific industries. (In case you are interested in a deeper dive into how to think about organizational design in the context of AI, check out my working paper with Sampsa Samila and Alex Oettl [substack post here and SSRN Link here])

Without the specific details of sector-specific analyses, it is difficult to grasp, even in general terms, the impact this technology will have.

While many industries are likely to be impacted by AI's video generation capabilities, education—encompassing primary, secondary, and higher education—may face particularly severe consequences.

The Rise of Video as the New Lecture

By now, most students and teachers are accustomed to videos being a key part of educational content. Whether assigned by a high school math teacher to review concepts like L'Hôpital's Rule or used in MBA programs and college courses where professors record lessons to explain accounting, business strategy, or finance concepts, video has become a vital element of the learning experience.

Video serves as both a complement and a substitute for traditional in-person lectures or explanations, which were typically delivered live. Video-based educational content has already had a significant impact on education, and artificial videos are expected to have a similar effect.

Easily accessible educational video content has been available in various forms for over twenty years.

What caused it to spread rapidly was the ability to stream videos worldwide at a low cost and with ease. Streaming allowed anyone to distribute videos, resulting in a surge of content from all parts of the world and from anyone.

Over time, the ability to distribute content led to the endogenous emergence of innovations that further lowered the cost of creation. Specialized tools like Loom, as well as editing software such as Canva, allowed people to produce higher-quality videos. For those earning likes, clicks, and ad revenue, investing in better tools and production processes became worthwhile because the economic opportunity was present due to the shift in “Distribution” from 0 to 1.

Nevertheless, creating high-quality video content still required a lot of effort. You have to write a script, record it, edit it, and add visual features. While some tasks, such as adding subtitles, have become easier over time, they still require effort and skill.

The video revolution brought about by AI, combined with the ease of distribution through platforms like YouTube, in particular, means that content creation, especially video content creation, is now near costless (AI can produce content such as scripts and stories, as well as video).

Today, you don't even need a recording device, high-end equipment, or any editing skills to produce videos that once required someone with deep skill and expertise.

Making a B-School Explainer with VEO and HeyGen

I wanted to see if we could do something more “everyday” with AI video generation: create a believable video that explains an academic concept in a way that would be helpful to a student.

To do this, I signed up for a free trial of Google AI so I could test VEO2 and VEO3.

My goal was to have the AI create an explainer video for Porter's Five Forces, a basic business strategy concept I discussed in a different post.

VEO2

I actually started with VEO2, not realizing it didn't have sound. I'll paste it just for fun, because it's both funny and interesting (text is still a problem in these videos).

Here’s my prompt: “Generate a video of a professor (Indian) on a blackboard explaining Porter's five forces to students in an elite MBA program.”

VEO3

I then used this same prompt in VEO3.

Frankly, it isn’t that impressive.

I think what I'm asking this AI to do is quite complex and multidimensional. The overall video elements are impressive, and the voice is very good.

However, the clip is extremely short (I realize you are asking the model to do a lot and so have to string together small clips into a larger sequence, but this seems like a hassle to me).

There's also a significant misalignment between what the person is saying and what’s on the board. Additionally, what’s on the board is misspelled and incorrect.

This does not mean that AI-generated videos for education are doomed.

Educational content delivery tends to be a complex effort (you need the content, script, Blackboard writing, speaking, and all of these elements must align to give a high-quality experience to the person watching).

Even with its impressive capabilities, VEO3, as a monolithic general model, is challenged by the superadditivity of delivering educational content.

Moving to Modular Production

ChatGPT + HeyGen

So, I decided to try creating a video that either eliminated or modularized some of the components required to produce an educational video, moving from superadditive production to a more linear approach.

I found a cool technology called HeyGen, which I've subscribed to, and it appears to be much simpler than VEO. It essentially features a series of avatars where the AI only moves a few body parts (such as the mouth and hands), while the rest of the avatar remains still. All you need to do is provide a script, which you can create yourself or get from another source.

What I did here is simple. I used a prompt in ChatGPT 4o to generate a script explaining Porter's Five Forces. Then, I pasted it into HeyGen, and it created the video after I selected the persona and voice. Here’s the prompt that generates the script.

Write a 5-minute script explaining the business concept of Porter's Five Forces. The video should explain the concept to MBA students, provide examples from recent news, and also discuss the limitations of the concept. Please include only the part of the script that needs to be read and nothing else.

If anything, this is a significantly lower-tech version of the video than VEO3. However, splitting "script + video” into two dedicated steps, using ChatGPT for writing and HeyGen for rendering, allows each model to perform at its best.

The video is watchable and informative.

Interestingly, an integrated system like VEO3, which attempts to handle both tasks simultaneously, appears to struggle in matching the quality of a specialized disaggregated approach with substantially weaker visual rendering.

ChatGPT + HeyGen

Here’s a version where I adjusted a few parameters in ChatGPT and selected a different avatar in HeyGen, resulting in a pretty decent Hindi language video.

Write a 5-minute script explaining the business concept of Porter's Five Forces. The video should explain the concept to MBA students, provide examples from recent news, and also discuss the limitations of the concept. Please include only the part of the script that needs to be read nothing else. The script should be in Hindi and the examples need to be localized to the language's context.

Again, not bad. Very watchable.

Just-in-Time Learning?The New Frontier of Educational Content

My quick question to myself is whether I can use some of these videos, with some refinements, in a class.

I believe the answer is yes.

The ChatGPT + HeyGen videos are actually quite good and could be improved further through iteration on the prompt, the video avatar, and perhaps incorporating slides in the background (which I believe HeyGen allows).

All of this suggests that the marginal cost of producing a video has decreased even further. Since the cost of distribution is already very low, and now the cost of creation is also declining, prices will decline rapidly as competition inevitably intensifies.

Moreover, this process can be automated into code that generates an entire course programmatically using AI. Here’s a condensed pseudocode of what this might look like.

concepts = OpenAI_API.generate("List key n-gram concepts for [Intro Business Strategy]")
FOR concept IN concepts:
    script = OpenAI_API.generate("5-min script explaining [" + concept + "] to MBA students")
    FILE.save("scripts/" + concept + ".txt", script)
    video_url = HeyGen_API.create_video(script, avatar_settings)
    youtube_link = YouTube_API.upload(video_url, OpenAI_API.generate("Title for [" + concept + "]"), OpenAI_API.generate("Description for [" + concept + "]"))
    save_to_site(concept, youtube_link)
launch_website()

But a much more exciting possibility is the ability to produce just-in-time explainer videos. For example, you could say to the app, “explain Institutional Theory to me,” and receive a high-quality video within minutes (and as compute and models improve, in fractions of seconds). [business idea: a platform that hosts educational videos generated by students that can be shared with anyone.]

You can also imagine that as computing continues to improve and become more affordable, these videos will become interactive, allowing you to engage in a conversation with the avatar.

In a way, we already have this with the current chatbots, just with text and audio. For instance, on longer solo drives, I sometimes “converse” with ChatGPT to learn about a new topic (my daughters have renamed my number on their iPad “Nerd Alert”).

As I consider this, a key challenge will be optimizing the backend to prevent the repeated creation of redundant videos and to highlight the best version among many, particularly when the content is largely identical. This won’t be all that different from what happened with streaming and internet content distribution (e.g., with the founding of companies like Akamai, which made internet content distribution even more efficient).

The MOOC Moment Revisited

What are the implications for the education industry from the substantially decreasing cost of high-quality video content generation and distribution?

When I was an assistant professor at Stanford University, the school organized a faculty session where a group of faculty from the engineering school discussed massive open online courses, or whatever they were called at the time. There was a large-scale online course on machine learning at that time, offered at Stanford, that went gangbusters, and several faculty members were starting to take these courses and spin them out into separate companies (e.g., Coursera, Udacity, etc.).

At the time, I was focused on figuring out how to teach my class of 35 students, so I didn't fully understand the bigger implications of what was happening.

But these faculty members were working to make education accessible worldwide, nearly instant, and available at a fraction of the cost of attending a place like Stanford. It was an ambitious and inspiring mission, and MIT had already demonstrated demand for it through its OpenCourseWare project.

A few months later, we had a faculty meeting where our dean asked how institutions like ours should (or should not) respond to online education. I remember that meeting very clearly. At the time, the prevailing sentiment was that online education did not align with our mission. Still, the door was left open with a wait-and-see attitude.

Around the same time, a nonprofit called Khan Academy was also making waves in the primary and secondary education worlds. Its approach was somewhat different from the emerging massive open online courses, as its content was hosted on YouTube, rather than a walled-off platform. Khan Academy adopted a straightforward method: Salman Khan, the instructor, recorded lessons using a digital tablet and posted them online. Anyone worldwide with internet access could watch them for free.

The promise of educational technology, especially video-based content available both freely or at a low cost, anywhere, was inspiring.

Does More Educational Video = More Learning?

So, what’s the impact of this rapid growth in online educational content?

My initial thought is to examine the descriptives. This acts as a kind of gut check. (With the obvious caveat that correlation does not equal causation)

Have student educational outcomes improved over time due to the availability of free and easily accessible educational videos?

The descriptives don't seem promising.

Analyzing PISA scores reveals that, despite a substantial rise in online educational content at all levels, primary, secondary, and tertiary, performance on standardized tests remains weak. Scores are declining in nearly all OECD countries.

Furthermore, scores were already falling even before the COVID pandemic occurred.

EdTech might not have caused the decline, but it also doesn’t seem to have helped.

A substantial body of high-quality evidence also examines the impact of educational technology across various contexts. The results here are also pretty mixed.

Some of the strongest evidence of positive impact comes from targeted, personalized interventions. For example, Banerjee et al. (2007) and Muralidharan et al. (2019) show that adaptive, computer-based tutoring systems in India produced significant short-term improvements in student learning, particularly for those who struggled the most. Math scores increased by nearly half a standard deviation in some cases, which is pretty large for an effect size! However, these effects faded: Banerjee et al. found that a year after the intervention, the effect sizes had dropped by nearly 80%.

However, when examining broader, large-scale technology rollouts, such as providing laptops to every child or expanding broadband access in schools, the evidence becomes much less promising.

Studies from Peru (Cristia et al., 2017), Uruguay (Yanguas, 2020), and the U.S. (Goolsbee and Guryan, 2006) found no significant improvements in test scores, even though students clearly used the technology more. These programs often lacked meaningful instructional content or support for teachers, which seems to have limited their effectiveness. Another study by Belo et al. (2013) found that broadband access was actually detrimental to educational achievement, mainly because students spent time watching YouTube videos during school hours.

In some cases, technology may have actually worsened situations. A notable study in Romania by Malamud and Pop-Eleches (2011) found that giving low-income families vouchers to buy home computers led to lower grades for students, likely because the computers were used more for games than for homework.

So, my main takeaway from the research appears to be that EdTech is neither inherently good nor bad. Its value depends entirely on how it is used, how it is implemented, and the instructional model supporting it. There are both success stories and many, many cautionary tales.2

Free Content, Expensive Degrees

Beyond the impact of educational technology on student learning, which already has a vast and complex body of literature that many are still trying to reconcile, there is a less-studied area, I believe, that warrants exploration. This relates to the impact of the explosion of cheap educational content on the broader educational industry.

In attempting to understand the impact of zero-priced educational videos on the educational industry, I’m trying to reconcile several key facts.

The first point is that a vast amount of educational content is now available online, much of it free of charge.

Simply visit YouTube and search for any subject you’re interested in. You can find a video on it.

The second fact is that MOOCs, especially those that require payment or subscription, never really took off as some people had expected. Many of these platforms have undergone multiple CEO changes, and despite numerous attempts, they’ve struggled to establish a sustainable business model. Some, like Byjus, turned out to be frauds.

Take Coursera, for example. When it went public, it was valued significantly higher than it is now. Since then, it has lost approximately 80 percent of its value.

The same trend applies to lower-cost competitors like Udemy, which has also decreased by about 75 percent of its value since going public.

All you need is standard economic logic to predict these outcomes.

When there's a very close competitor, such as free online videos, it's challenging to persuade people to pay for your product.

Coursera, in particular, was selling a bundle that included the video content and a certificate. Assuming the value was the maximum of the two components, and given that the content was free on YouTube, the Coursera credential wasn’t worth much, which is why the stock dropped. 

Why has In-Person Education Survived?

In-Person Enrollment has Remained Flat or is Increasing

The interesting puzzle is this: despite all these challenges, if you examine enrollment figures, especially in degrees that people are not required to pursue and must pay for, such as college or graduate school, the numbers have remained fairly steady. In some cases, they have even gone up.

For instance, based on data from the National Center for Educational Statistics, enrollment in four-year and two-year colleges has at worst remained flat.

In the case of post-baccalaureate degrees, there has been a steady upward trend over time, particularly among females.

However, to understand these trends, you need to examine them more closely to see what is really happening beneath the surface.

What is clear is that although the average sticker price of college has increased over time, the net price, which is the cost after discounts, financial aid, and other forms of support, has actually decreased. This is despite the fact that the cost of running a college has increased dramatically (e.g., services, scholarships, staff, etc.).

The Classroom is Not What it Used to Be

The third point is undeniable if you've recently taught in a classroom. The learning environment has “evolved.”

Students frequently fail to complete their assignments and often lack focus during lessons. Many are physically present but mentally distracted (some never show up, even once).

These issues are not isolated to large, non-selective schools; similar trends are observed across various institutions. [here, here, here, here, …]

The classroom experience appears to have undergone significant changes for many individuals, including both students and instructors. I’ve read at least a dozen op-eds from professors across the country, from various institutions, all saying the same thing: students do not participate, do not complete the readings, and do not engage in class.

I believe there are many reasons for this.

Here's one of mine: Why pay attention in class when most of the material is easily accessible outside of class for free? The class then becomes a hurdle, something to get past so you can earn your degree. The goal is to do just enough to pass. Grades are often inflated, so the strategy is to meet only the minimum requirements, get the credential, and move on to the job market (but why would anyone want to hire you?).3

At the end of each semester, I often receive multiple emails from students who, to be honest, seem completely unaware of their own disengagement. They request participation credit despite never attending class. A few years ago, I received an email from a student who emailed me to tell me how much they enjoyed listening to my “classes while [they] cook, clean, and handle other basic chores around the house."

(Research suggests that there may be a negative effect of lecture capture on attendance for some students.)

The Higher Ed Bundle Under Pressure

So, what exactly did free content do?

I think it's helpful to consider this from the perspective of a basic theory of substitutes.

Coursera, Udemy, and Udacity (which was eventually bought by Accenture) were mostly platforms for content. A credential from Udemy or Udacity doesn't hold much weight. The same is true for Coursera. Employers still prefer to hire candidates with degrees from traditional universities.

Free content available on platforms like YouTube directly competes with paid online content. When a close substitute is available for free, both the demand for and the price of the paid version typically decrease.

That is essentially what happened here.

The case for traditional education is much more subtle. Traditional universities weren’t selling content like the online vendors. They sold a package: You also receive a credential with labor market value, as well as support services, faculty, peer groups, and resources provided by the institution.

However, a significant portion of the content is now freely accessible. You can find similar lectures, explanations, and problem sets online with little effort.

This changes the value proposition for in-person educational institutions.

Many students no longer needed to attend classes primarily for the content (and recorded lectures made freely available by the professor don’t help either!).

If they are attending, it is for something else. In the simplest case, they are there only to get the credential. And even the value of that credential is becoming less certain, so students are increasingly offered scholarships and other incentives to enroll.

For many students, especially international ones, the value of a U.S. college degree is also less about the education itself and more about what it opens up: access to the U.S. job market.

Who Pays, and for What? Rethinking the Value Proposition of Education

I believe free online content must now be taken as a given, and it will become increasingly plentiful and notably improved thanks to AI.

In many instances, this content surpasses what most professors can produce independently. The videos I made earlier are significantly better than what I could have created myself.

Pandora’s box is now open.

We won’t be able to eliminate the free content students can access, and competing solely on content seems like a fool’s errand. Just look at what free content did to Coursera and Udemy. Brutal.

Moreover, lectures are an experience good, so why would someone pay hundreds of thousands of dollars upfront for content quality they can't easily assess beforehand, especially with so much free content available online to watch at home in your pajamas?

So what happens next?

To me, the main issue is that universities will need to figure out what they can bundle with their content to justify and maintain current prices. With the rise in scholarships and in-kind spending on services, actual “margins” in higher education are shrinking.

Whether we can sustain current price levels depends on the value of the other components in the bundle. Are they worth what students are paying?

What I am about to say does not directly concern artificial intelligence, but rather whether the surge of AI-generated content in education can be managed effectively, considering the various other components in the bundle and the potential threats to the value of these other elements from policy shocks or competitors.

• The first major factor is the big elephant in the room: the growing uncertainty about access to American higher education for foreign students. Public policy is shifting in ways that may make it harder for students to come or to feel confident they will be able to stay and work after graduation. If students cannot come, are afraid to come, or lose access to the U.S. labor market after arriving, then the value of the bundle they are buying decreases.

• Even for students already in the United States, the value of the credential itself may decline. Some states, such as Pennsylvania, have already begun to eliminate college degree requirements for many public sector jobs. And there are a host of new startups that are working with companies to provide skill-based credentialing. If these skill-based assessments or alternative credentials gain wider acceptance, there will be further downward pressure on the prices students are willing to pay for traditional degrees.

This naturally raises the question: what additional elements can be bundled with content that people are genuinely willing to pay for?

At an abstract level, there are two options on the table: (a) take the bundle as a given and reduce prices or (b) figure out how to increase willingness-to-pay by including higher-value items in the bundle.

I’m pretty sure that educational leaders do not want (a); so we’ll have to figure out (b).

This is already a very long post, so I will leave it for another day to think through what we can bundle with our content to survive the competition from AI.

Ultimately, I remain romantic about education. Knowledge brings economic, social, and personal power.

Watching a video, especially one generated by AI, is not “learning” or “education.”

The real challenge for us will be to inspire students to push themselves harder than they have ever done, to engage fully with ideas, and to aspire not just to get jobs, but to do great things.

How do we deliver on this promise?

This week I have a slightly different type of post. My collaborators (Professor of Innovation and Strategy at the Georgia Institute of Technology) and Sampsa Samila (Professor of Strategy at IESE, Barcelona) have a new working paper “From Model Design to Organizational Design: Complexity Redistribution and Trade-Offs in Generative AI.”

If you are interested in catching up on the absolutely latest research (from CS, Strategy, Economics, and spanning academic studies and industry reports) on Generative AI and what it means for organizations. Check out the full paper here. We learned a lot writing it, and hope that you find both our framework useful and the literature review a nice upgrade to your knowledge about AI in the field.

Here is the punchline

Generative AI appears to offer a free lunch: incredible power (Generality & Accuracy) through a simple chat box (Simplicity). But this user-facing simplicity is an illusion.

The paper introduces the Generality-Accuracy-Simplicity (GAS) framework to argue that the fundamental trade-off between these three elements hasn't disappeared: it has been relocated.

Where?

The complexity is shifted from the user to the organization, re-emerging as hidden infrastructure costs, new compliance burdens, the need for specialized talent, and a persistent "accuracy ceiling."

Our key takeaways:

• Complexity is Relocated, Not Eliminated: AI's ease of use for the individual creates immense, often invisible, complexity for the firm. This is why Enterprise AI doesn’t seem like the easy “win” that ostensibly should come from just purchasing site licenses to your favorite chatbot.

• The "Accuracy Ceiling" is a Core Strategic Constraint: Competitive advantage comes not just from adopting AI, but from designing workflows and cultivating human expertise to manage its inherent limitations.

• The New Competitive Moat is Mastering This Complexity: Sustainable advantage will come from designing new “intelligent” workflows, building complementary human expertise, and making deliberate choices about where to operate on the Generality-Accuracy frontier. The latter is particularly crucial. If your firm conducts high-accuracy, low generality tasks, you have considerable distance to traverse to make AI useful for you. This distance—the hidden complexity you’ll need to develop and manage—isn’t costless.

• Human Judgment is the New Bottleneck: As the cost of generating content plummets for everyone (not just you!), the value of critically evaluating, refining, and contextualizing it skyrockets.
  

This is a shift from viewing AI as a simple cost-reducer to seeing it as a catalyst for deep organizational redesign.

We hope this framework helps leaders, researchers, and builders navigate this new landscape.

You can read the full paper on SSRN: From Model Design to Organizational Design: Complexity Redistribution and Trade-Offs in Generative AI.

TL;DR.

Technology and incentives reshaped the music industry. Solo artists replaced bands because they were cheaper, able to produce faster, and easier to manage. This shift eliminated the creative friction that once defined much of music's richness. A similar dynamic may now be unfolding in knowledge work. AI reduces the cost of solo production, making teams less attractive unless they consistently outperform in quality. If expectations of quality decline over time, the economic case for collaboration may weaken, and we may not realize what we have lost.

Bands are Disappearing

I recently watched a YouTube video by Rick Beato, in which he made an interesting observation: there are almost no bands on the charts.

There are lots of solo artists. But very few bands.

What are bands? Bands are teams of people bringing together their skills, often as equals, to create music. I've been thinking a lot about his observations about bands over the last few weeks. Are we witnessing something similar in the world of knowledge work because of AI?

Bands Were Once Ubiquitous and So Were Fans

When I was a kid, all I listened to were bands, old and new: Led Zeppelin, The Doors, Radiohead, Soundgarden, Rage Against the Machine, and later, Massive Attack.

My first concert was Rage Against the Machine at the Roseland Ballroom in New York City in 1996 (I was 15). After school, I walked for two hours from my house in North Brunswick to Cheap Thrills, a record store in New Brunswick, New Jersey, to buy tickets.

A few months later, I went to see Soundgarden and saw a band member (maybe Chris Cornell?) storm off stage. Then, a few months later, Soundgarden broke up for good.

This friction, as explosive as it was, is what made bands, well, bands. Each person brought something distinctive. The music was an outcome of a superadditive function.

So, why have bands disappeared?

The Economics That Made Bands Disappear

According to Beato and others, the disappearance is caused by a mix of two forces: technology and incentives.

It’s useful to sketch out the band story and explore how far this analogy helps us understand the ways AI and incentives may weaken teamwork in other industries.

Success in Music is Power-Law Distributed

Music is meant to evoke a feeling. If it creates a desired feeling (happiness, hope, inspiration, even anger), people desire more of it and are willing to pay for it.

“Willingness to pay” (WTP) can be seen as the value generated by a piece of music.

Yet, even with very high WTP, the prices of records, CDs, and even streaming subscriptions remain relatively fixed. At a $15 price point, you get widely varying types of music with differing willingness to pay.

What drives profitability in this industry, then, is quantity.

The problem with music, as with many creative goods, is that willingness to pay (or success) is not normally distributed. It's very skewed. This is called a “power-law” distribution. Most artists get few listens. But a tiny fraction explodes.

Consider the number of streams on Spotify. It is a power law distribution. A few artists get most of the listens, and most get very little.

You Need Multiple Shots on Goal to Win

Given this distribution of success, a record company can only profit if it achieves a few mega-hits since, statistically, most artists tend to fizzle out. (This is not all that different from venture capital, book publishing, academic papers, tweets, apps, and consumer products)

This is further complicated because, ex-ante, you don’t know whether an artist is going to be a hit. That is, on average, it’s hard to pick winners. It depends on a lot of interacting factors. And thus is susceptible to a weakest link problem (e.g., x1 * x2 * x3… xn), and if any x is missing, you get a dud.1

How Do You Cut Costs in Music?

With power-law distributions, you need multiple attempts to succeed. In most cases, the results are “0.” Sometimes, when luck is on your side, you achieve the extreme outcome necessary to recoup your investments. The economics of streaming also reinforce the payoffs for massive hit singles.

A label’s incentive in this environment is to have as many shots on goal as possible. If they have a hard budget constraint, say K, they will try to squeeze in as many artists, with some cost c, as possible in that budget. This means that you need to have very low c per artist to buy enough lottery tickets to get a hit.

What should the label do? It should cut out expenses and figure out a way to spread costs across many different bands. To do this, they could:

• Hire a songwriter who composes for multiple artists rather than commissioning custom songs for each project.

• Use a drum machine instead of hiring a live drummer.

• Replace natural, organic instrument recordings with high-quality synthetic or virtual instruments.

• Smooth out the voice of a bad singer with a pretty face with Autotune

In fact, the solo artists themselves can do much of this in their bedroom with an Apple MacBook Air. They can sing someone else's song (or their own, if it is any good) and tell the computer to generate drum sounds, guitar, and bass. The computer will comply.2

It is not hard to see that the cost of a solo artist c(s) is substantially lower than the cost of a band c(b).

In addition to input costs, you also reduce the “interaction cost.” If the solo artist is in full control (they tell the computer music software precisely what to play) or the music is purely the vision of the producer, there won't be any conflict between team members, shortening production cycles.

The music then becomes the linear vision of the artist. No pushback.

And, indeed, this is what we've gotten: Hozier, Ed Sheeran, Sabrina Carpenter: economically viable, built for scale, and low creative friction.

Compare this to 1980, where nearly 50% of the top 10 are teams.

Losing Creative Friction

The economic system of music doesn’t favor bands. And it seems that the incentives were ALWAYS this way.

So what changed?

To produce music, you needed bands because you did not have viable sonic substitutes. Today we do, so the outcome has changed.

We get fewer bands.

Do Teams Create Different Music Than Solo Artists?

When teammates are replaced by technology, something else also disappears.

Most individuals have a sense of self and a desire to assert their vision on the world.

Individuals will resist the undue influence of others. They will push back. This friction can be both constructive and destructive.

In music, this creative tension has served as a wellspring of tremendous sound.

Two examples come to mind.

The Police are notorious for the friction between Stewart Copeland and Sting. In fact, there are several YouTube videos showing Sting and Copeland arguing with each other (here and here, though sometimes in jest). In a fantastic interview with Beato, Copeland notes that Sting's willingness to push for his vision developed after many of his songs became hits. It was a belief in himself that allowed him to advocate for his ideas within the band, The Police.

Copeland, considered one of the greatest drummers of all time, did not want to be merely an arm of Sting’s vision and, of course, pushed back. This friction led to sometimes dysfunctional interactions among team members, but also to incredibly sublime music.

Listen to Spirits in the Material World. Every instrument and every sound is intentional. You can tell that each sound is coming from someone who cares, someone who owns what they're producing, someone who wants their voice, their sound, to be heard. First, focus on the drums only. Then the bass. Each person is making the music theirs. Go back and listen to Hozier, Sheeran, or Carpenter, and you can hear the difference.

Listening to them is like squinting, but with your ears.

Another example of creative friction is Mezzanine by Massive Attack. It is perhaps one of the greatest albums ever recorded. A genre-defining work.

It also led to the “breakup” of Massive Attack due to creative differences between Robert Del Naja (in front) and Mushroom (Andrew Vowles, in the middle).

In an interview around the time of the release of this album, Del Naja talks about how creative friction is necessary for breakthroughs to happen. Mushroom, sitting in the middle, looks upset.

“There is always tension in music. Sometimes it’s a good thing, sometimes it’s a bad thing. When it is a strong creative process and we all feel strongly about what we are doing, we’re gonna come up against each other, you know. Quite heavily sometimes, you know. And there is going to be a lot of aggravation, but that’s just the way we’ve always been….It’s part of it, really, part of the process.”

After Mezzanine, Massive Attack was no longer a band. The “band” was a project mainly of Del Naja and his musical vision. Not that I'm the most sophisticated listener of music, but nothing matches their earlier albums as a band: Blue Lines, when their vision aligned, and Mezzanine, a powerful work of creative friction.

At their best, bands aren’t merely linear collections of skills; they are multiplicative production functions brimming with conflict and energy in ways no single person can replicate.

Will Artificial Intelligence Kill Teamwork?

I believe we can learn a great deal from the music industry about how incentives and technology, particularly AI, will impact teamwork in knowledge-intensive industries.

Teamwork makes up a large fraction of people’s time

“Team” work makes up a substantial proportion of how people spend their time at work. A recent survey found that knowledge workers spend 42% of their time collaborating with others (more than solo time, 35%). I imagine if the question were something like: “What percentage of your time contributes to team-level outputs?” the number would be larger.

The big question for me is: How will AI affect collaborative knowledge work?

Team Types and The Impact Of AI

I think this likely depends on what problems teams exist to solve.

In terms of production:

1. Type A teams are teams where most people can do most jobs. But teams exist for three purposes: (a) scale, since one person doesn’t have the time to do everything so they delegate; (b) flexibility, since there is variability in who is available to do work, when; (c) learning, because teams are designed to develop the next generation of talent.

2. Type B teams are teams of specialists where each person owns a distinct, non‑overlapping skill set. The team exists for three purposes: (a) depth, because  no individual can master every niche; (b) coordination, because clear ownership reduces handoff errors; (c) innovation,  specialists can bring in advances that a generalist doing the same thing can’t. (There is a version of Type B teams with similar skills, but different “perspectives.”)

3. Type C teams are teams where people enjoy working with each other, independent of functional reasons.

These are not pure types but rather points on a spectrum. Academic research in the social sciences aligns more closely with Type A (most researchers are taught to do everything, e.g., write, model, analyze data, etc.); physician teams in complex cases tend toward Type B (anesthesiologists, surgeons, etc.), and Type C teams are distributed throughout.

I believe AI currently poses the greatest threat to Type A teams. As models improve, AI may also threaten Type B teams (though I’m less sure when this will happen, given the current weaknesses and strengths of foundation models).

I hope Type C teams will remain unaffected by AI's influence. However, I’m not sure. Many people are using AI for companionship and support.

How AI is Affecting My Teamwork

When I think of my own research teams, they are mostly Type A teams (with some members having relative strengths but not deep specialization). While they have remained the same size so far, AI has already had profound effects:

• Low‑level tasks (data cleaning, literature searches, grammar edits) are primarily handled by AI (custom scripts; ChatGPT, Grammarly). This has freed up time to think at a higher level about other problems.

• My best PhD students have extremely strong AI‑usage skills: prompt crafting, troubleshooting AI output, coding, and verifying accuracy independently. They are also extremely smart in various other dimensions.

• I’ve found that my PhD students have become substantially more autonomous; they solve roadblocks (e.g., structuring data, thinking through empirics, finding relevant literature, etc.) using AI rather than escalating to me.

• Overall, I focus a lot more on “structuring” and strategic choices rather than nitty-gritty stuff, while each team member can execute autonomously with far fewer bottlenecks.

Because so much of the lower-level work is made easier, we are spending a lot more time on the big picture.

At some point, it is not hard to imagine turning this patchwork AI team into an integrated and autonomous research team. Perhaps this team may be more auditable than a human team.3

Will AI Replace the Band: Rethinking Collaboration in Knowledge Work

The big question is whether this shift on the supply side will lead to the same dynamics we saw in the music industry—e.g., toward solo creators substituting teammates with technology?

If we think about it from a demand‐side perspective, solo individuals > teams when at least a few things are true:

• Hits are unpredictable ex ante;

• The market’s payoff for a “hit” is convex;

• There are strong Matthew effects (success begets success).

So, when do teams survive?

I think it boils down to the following:

• Teams have a higher probability of getting a hit than solo artists, and this is known ex ante

• Consistency matters more than extremes

• The payoff from the increased likelihood of success overcomes the higher cost of the team.

In other words, if you know a team can consistently meet the quality threshold often enough (even at a higher cost), then its “hits-per-dollar” outperforms a solo+AI lottery ticket. However, if uncertainty is extremely high, payoffs are heavily skewed, and solo+AI costs are so low that you’d prefer to take more inexpensive shots, then teams become less viable.

We May Not Notice What We've Lost

On the demand side, if quality thresholds are fixed and set outside our control (e.g., exogenously), teams will survive only when their combined expertise raises the chance of meeting that bar enough to justify higher costs.

In that case, it is a simple economic calculation: if a team’s output clears the standard far more often than a solo creator with AI, then paying for a team makes sense.

If a solo creator can hit the same mark almost as frequently but at a fraction of the cost, then the team option loses out.

The Danger of Lowering Our Expectations

But everything changes if quality thresholds are shaped by what the audience comes to expect or by relative comparisons.

What if our expectations of “good” or even “amazing” are lowered through some social process (e.g., grade inflation)? In this case, gradually, we might settle for whatever solo + AI can deliver.

The bar moves downward to match cheaper, AI‑driven work, and teams lose their clear advantage.

Even if a team could produce something objectively better, no one would value (or even know how to assess) that higher level if expectations have already fallen. Once quality standards become endogenous and trend downward, there is no economic rationale left to pay the premium for collaboration.

In other words, even economically viable teams endure only so long as the market insists on a higher standard.

When we lower our standards, we may not like what we get.

The scarier thing is that we might not even notice.

Let me not end on a sad note. One antidote to this downward spiral is to develop our taste for the beautiful things that are still out there. These are the pieces of superadditive magic we find in music, sports, science, and at work, too.

These awe-inspiring things may be harder to come by, but they do still exist. We have to notice them, appreciate them, and realize that they may take a bit of conflict to make.

TL;DR

I tried vibe coding a Retrieval-Augmented Generation (RAG) system to write my Substack posts. It sort of worked. Which made me wonder: What happens when we can build functioning systems without understanding how? This post explores the tension between magic and comprehension and what it might imply for software, markets, and beyond.

Introduction

One of the most fascinating aspects of AI is its coding ability (the Claude 4 coding demo from a few days ago is worth watching!).

This capability and its application are colloquially known as “vibe coding.”

This week, I carried out an experiment in vibe coding to understand what the excitement was all about. 

What is vibe coding? It consists of creating software with AI, guided by your intuition and ideas, and then refining it iteratively with sequential prompts. 

It's magic.

My goal wasn’t to build software per se (though that is interesting too); rather, I wanted to understand what it might mean if everyone had this power at their disposal.

That is, what does the advent of vibe coding mean for how we build products, companies, and society as a whole?

Using AI to Write a Substack about AI

First, I needed something to vibe code.

It seemed that vibe coding and my sabbatical goal of writing Substack posts about AI were a perfect combination to experiment with.

What if I didn’t have to write my posts, but instead built a system to write them all for me?

I’m not talking about just telling ChatGPT to write a post about topic X, but actually building a retrieval-augmented generation system (RAG) that would take in current headlines from news, social media, academic articles, and whatever else, then write a complete article for me, like me.

So, taking my previous post on Patchwork AI seriously and building on the idea that AI will be integrated into workflows, I thought I would begin with a comprehensive schematic of what a generative AI system for writing Substack posts for superadditive.co would look like.

Ingestion and Inspiration

So, there would be multiple layers of code, a mix of ingestion from APIs that would bring in content from across the internet, including Hacker News, Reddit, arXiv, etc.

This content would feed into a topic engine. This engine would then evaluate each piece of content based on my personal lens, audience interest, and then score each possible topic for relevance.

Idea Generation and Evaluation

Ideas would be generated with a focus on diversity and quality.

I don’t know exactly how, but it would happen.

Writing

These ideas would then be scored using some rubric, and the full idea for the substack article would be outlined.

A writing process would ensue. My agent minions would auto-generate SEO and A/B tests and input some aspects of my strategy and artifacts to write a more personalized post.

Launch and Feedback

Then, I would publish on Substack, learn from user engagement, and update my strategy and tactics.

Let’s Start Vibe Coding

Okay, I was going a little wild here as I thought through what I could possibly do. I’m not going to spend all this time building this big system because I wouldn’t want to read those Substacks anyway (at least I don’t think I would want to read this).

However, it was helpful to get those ideas out there.

I’m going to build a much simpler system using a free API, specifically the arXiv API, and then use my OpenAI subscription to build the RAG system.

Prompt 1: Build Me a RAG System for Substack Posts

1. I would like to pull recent articles from the archive with an API for a given keyword that I choose. 2. I would like to save these into a csv file. 3. I would like to cluster the articles into 5 topics. 4. Using the text of the articles in the topic, I would like to have open AI generate ideas for 10 possible substack post ideas for each topic. The posts should be related to innovation, economics, industrial organization, organization theory, entrepreneurship. 5. The ideas should be output into a txt file with the relevant articles in the cluster that was just described at the bottom, with a link to the original articles. 6. Here is my openai API: "OPEN AI API GOES HERE"   


http://export.arxiv.org/api/query?search_query=ti:%22generative%20ai%22&sortBy=submittedDate&sortOrder=descending&max_results=100 The output is: 

<?xml version="1.0" encoding="UTF-8"?> <feed xmlns="http://www.w3.org/2005/Atom"> <link href="http://arxiv.org/api/query?search_query%3Dti%3A%22generative%20ai%22%26id_list%3D%26start%3D0%26max_results%3D30" rel="self" type="application/atom+xml"/> <title type="html">ArXiv Query: search_query=ti:"generative ai"&amp;id_list=&amp;start=0&amp;max_results=30</title> <id>http://arxiv.org/api/AVnAV8tSzmKVQVQnaeljoO7C0ec</id> <updated>2025-05-13T00:00:00-04:00</updated> <opensearch:totalR
…..
….
>

At first, it just generated the output for me.

I wanted the code (typo included for accurate representation of my process)!

i just want code generated to do these tasks, with each step explaine

Prompt 2: Make it Modular

ChatGPT produced lots of code. It was hard for me to process it with “vibes.”

So, I asked it to make the code a bit more modular so I could scan it for understanding.

i'd like to make the code more modular. In particular, I would like to separate out the prompt so that it can be better.

Actually, re-reading this prompt, I realized that I made no sense. But it did a pretty good job turning my wish into modular code…

Now I have my MVP code for my Superadditive.co substack post generator.

Okay, this is pretty cool. Will it work? Let me try it out in Google Colab and see if it works.

Post 1.0: Cringe!

Here’s the introduction for the Substack it produced. If you want to read the entire piece, you can scroll to the bottom.

Imagine yourself in a bustling city, surrounded by intricate networks of infrastructure, each system functioning autonomously yet interdependently. This sprawling urban landscape represents the complexity of our current sociotechnical systems, where artificial intelligence (AI) has become an integral part of our daily lives, just like the electric grid or public transportation. Yet, as we increasingly rely on these AI systems, we're confronted with a paradox: the very technology that promises to improve our lives also brings with it new challenges and risks. This article explores the rise of Generative Artificial Intelligence (GenAI), a powerful new technology that is revolutionizing industries but also raising concerns about accountability, security, and education.

Nevertheless, this was magic. Wingardium Leviosa.

The output is not good. I would not read it or encourage anyone to read it because it is not based on any real “idea.”

However, as a minimum viable product, it’s promising.

So far, I know my prompting stinks. But that is what makes it even more impressive.

It’s a vibe.

Prompt 3: Adding Personalization

Overall, it's very cool that ChatGPT could build this so fast. The content is pretty generic and uninspired, even though it was generated from recent research pulled from arXiv.1

I decided to vibe code my way to a slightly better setup.

One idea I had was that to make it more personalized, I should probably tell the system something about myself. Otherwise, of course, it's going to produce something generic. So, I asked it to incorporate my background into the post. I uploaded my CV and pasted the code that came after Prompt 2.

i'd like to add a function that injects my personality into the writing of the substack. Here is my CV.

It generated the code and told me to include this piece of code that would incorporate my personality into the writing.

def personalize_voice(base_prompt: str, personality_descriptor: str = None) -> str:
    """Injects a personal writing voice into the base Substack prompt."""
    
    default_descriptor = (
        "You are Sharique Hasan, a professor of strategy and sociology who blends deep academic rigor "
        "with accessible storytelling. Your writing style is reflective, often drawing from personal experience, "
        "field experiments, and empirical work to illustrate insights. You connect theory to practice with clarity, "
        "emphasizing real-world implications, especially in entrepreneurship, innovation, and social networks."
    )
    
    descriptor = personality_descriptor if personality_descriptor else default_descriptor
    
    personalized_prompt = base_prompt.replace(
        "You are a Substack writer specializing in", 
        f"{descriptor}\n\nYou are writing a Substack article specializing in"
    )
    
    return personalized_prompt

prompt = format_substack_prompt(cluster_id, summaries)

base_prompt = format_substack_prompt(cluster_id, summaries)
prompt = personalize_voice(base_prompt)

But I had no clue where to paste it in, and frankly, I was too lazy to understand what was going on. So, I asked it to do it for me with two words.

full code

Yay! It produced the full code.

I ran the code again, and it still produced pretty poor, and in this case, quite cringeworthy writing.

### Substack Post for Cluster 1

**Opening Paragraph**

As I stroll through San Francisco, a city I call home, I can't help but notice the intricate urban layout that encapsulates its history, culture, and aspirations. The serpentine roads, the Victorian houses, and the steel-and-glass skyscrapers all speak to the city's past and future. Each architectural detail, each junction, and each open space is a testament to meticulous urban planning. But with the dawn of the AI era, we are witnessing a paradigm shift in how cities are designed and planned. Could generative AI redefine urban planning, and what are the implications for society at large?

I thought maybe the problem was that I was using the old model, GPT-4, so I asked it to upgrade to the best model available. It didn’t. It used 4-turbo. I wanted 04-mini, so I did it myself.

can you upgrade the chatgpt model to the best model out there. 

# ========= GENERATE BLOG POST =========
def generate_post_ideas(summaries, cluster_id, openai_client):
    base_prompt = format_substack_prompt(cluster_id, summaries)
    prompt = personalize_voice(base_prompt)
    response = openai_client.chat.completions.create(
        model="o4-mini-2025-04-16",  # Updated to latest best model
        messages=[{"role": "user", "content": prompt}],
        temperature=1
    )
    return response.choices[0].message.content

The writing in this version was way better. Frankly, it scared me because it started sounding like me. However, the story is NOT true.

Last spring, I found myself in a small conference room at a community center in rural Virginia, listening to a group of farmers debate whether ultraviolet light could cure COVID-19. I’d been invited to speak about information flows in crisis moments, but the talk evolved into an impromptu fact-checking workshop: we sifted through Facebook posts, dissected forwarded audio clips, and traced dubious medical claims back to satirical news sites. One farmer confided he’d always trusted what a neighbor shared on WhatsApp. When I explained how generative AI can spin entirely new, but patently false, stories at the click of a button, his eyes widened. We’d crossed a threshold: misinformation had ceased to be an accident of rumor; it was now a technologically amplified phenomenon, blurring the line between honest mistake and deliberate manipulation.

Prompt 4: Making it Sound More Like Me

Frankly, though, the full post wasn’t something I would be interested in reading because it still lacked an element that I couldn’t really put my finger on.

Here is a real substack that I wrote (first one) followed by an autogenerated substack with a RAG system I vibe coded. Give me some ideas about how could modify the code below to get the AI generated post, to have the same kind of feel and interest to the human one. 

It suggested more stuff, including:

• 🗣️ Inject Real Voice into Output via Style Transfer + Memories

• 🤯 Include a “contrarian move” requirement in the prompt

• 📉 Add a ‘Meta-commentary’ Layer

• 📚 Make the system rate its sources for emotional or intellectual resonance

• 🧵 Encourage narrative scaffolding in the blog post generation

• 🪞 Don’t simulate your voice — simulate your process

✨ Final Thought

Your experiment is valuable not because it automates your Substack, but because it reveals what makes your writing irreplaceably human: recursive thought, lived stakes, intellectual humility, and a sense of friction between knowing and not knowing.

In other words, vibe coding can mimic your structure.
But not your struggle.
And your struggle is the content.

Would you like help rewriting your AI prompt based on the feedback above — perhaps as a new modular template?

What I Learned Through This Experiment

I’m sure these other modules could be added. I didn’t incorporate them because I learned something during this activity that I thought was worth thinking through a bit more than creating a fully functional system to replace me in writing Substack posts about AI.

First, let me take a moment to reflect on what I learned about the quality of the posts generated by AI through this process.

• Currently, the post the AI has written is not very readable.

• I’m pretty sure that with better prompts, more modules, and more data sources, the RAG system could create a very interesting post that I wouldn’t mind reading.

• But it wouldn’t be my writing.

• With all the caveats, there is real economic power in the vibe-coded RAG system “I built” in just a few minutes.

When Thought Becomes Software

I started this post with an idea. Could I replace myself with a stream of generative AI code to write Substacks for me while I enjoy piña coladas all summer long, collecting likes and subscribers without doing anything?

However, what startled me was how I managed to achieve all this without fully understanding what I was doing. The outcome turned out to be much more complex than I realized.

Of course, I could invest time in grasping this complexity, but adding features and solving real problems is far more exciting than fully understanding my actions.

The ability to introduce complexity that I didn’t understand with a few missives is awe-inspiring. But what are the consequences?

The Danger of Uncomprehended Complexity

Software is everywhere. It powers our cars, sets our toasters, orchestrates activities on the cloud, and tracks our movements through our watches. Everywhere. What are the consequences of rapidly proliferating software in the world? Software that we don't understand.

Generally, before the era of vibe coding, people created software. In many cases, it was written by engineers who (mostly) grasped, at a high level, what was happening when their code ran.2

However, vibe coding will dramatically change the ratio between understanding and complexity.

Historically, our limits on understanding acted as natural checks on the complexity we could create in human-designed systems. (Yes, interactive or emergent systems have superadditive complexity.)

Vibe coding will allow us to produce highly complex systems, but at a radically lower level of understanding.

In other words, software’s complexity is no longer constrained by our comprehension in the same way. Complexity is now super cheap.

Marching Toward Abstraction and Opacity

This isn't entirely new. Technological progress is often a march toward increasing levels of abstraction, allowing us to create complexity without fully understanding everything. Abstraction allows us to build ever more complex systems without fully needing to know or understand lower-level mechanisms: C++ compiles to assembly, which maps to logic gates, which operate through the movement of electrons…..

Without this march, we would not have the bounties of the modern world.

Vibe coding seems to be the next logical step. We are moving away from the need for humans (e.g., programmers) to be intermediaries in translating ideas into computer code. Vibe coding allows us to abstract away more than just low-level implementation; it also allows us to create abstraction layers for intent and architecture.

What happens when the systems we rely on become opaque even to their builders?

The Economic Effects of Coding Magically

Many products have become unreasonably complex over the years. They are impossible for consumers to repair. In many cases, they are too expensive and impossible even for experts to fix.

Indeed, the modern economy, in many ways, is based on the notion that such technologies cannot be repaired. Instead, when things don’t work, we replace them.

In most cases, replacing products is cheaper and easier than fixing them, not simply because they are unrepairable, but rather because the incentives of the economic system have made replacement more efficient than repair.3

Three Possible Effects of Magical Software

I wanted a framework to consider what all of this meant, and while I don't yet have a fully formed set of ideas, three things came to mind as I thought through the market forces that could shift due to this innovation.

Three things that come to mind:

• Lower Barriers to Entry. The first point is that we can build quite complex things very cheaply. This will lead to a significant reduction in entry costs. You no longer need to learn coding to build systems and connect complex AI pipelines. You saw that I was able to do it with 243 words, and a “vibe” about what I wanted. Obviously, I had a baseline level of knowledge that this could be done and how to run it once completed, but this isn’t rocket science. When tasks become easier, barriers to entry are lowered, leading to increased competition.

• Lower copying costs. Second, vibe coding is about first understanding the functionality you want and then being able to reverse engineer something that produces that functionality.
  
  If I were to consider this in the context of competition, assuming there was no IP protection—and for many software products, IP protection is quite flexible anyway—I could simply look at a website, take a few screenshots, and then figure out how to reverse engineer the functionality of the software using vibe coding. In fact, I could create a bot that scours the web for new companies, learn their functionality, and then codes a product with similar functionality. I’ve also tried uploading patents and asking ChatGPT to tell me how to create a technology that has the same functionality, without violating the patent.
  
  This means that copying costs decline dramatically for any company developing technology.

• Increased variance of outcomes. Finally, the third component concerns complexity. I produce code that I do not fully understand, and I could make it even more complex without comprehending it at all. As the number of components in a system increases, the potential complexity grows exponentially due to the possible interactions among those components. Given that we don’t fully grasp each of the components or how they interact, there’s considerable uncertainty regarding the system's behavior. Outcome variance should increase dramatically.

What Might Happen to Software Markets?

We are likely to see many more players entering the software market, and exclusivity will diminish as copying becomes easier and cheaper, particularly with global coding costs decreasing. Technical skills alone won’t provide a competitive edge. At the same time, outcomes will become more unpredictable, with significant breakthroughs on one end and catastrophic failures on the other, which will probably lead us toward certification systems akin to an FDA for software to restore some trust. Additionally, since creating new software will be cheaper than repairing old systems, the market will shift toward constant replacement. Repair simply won’t be worth it anymore, except in specific edge cases.

The Fragility of Rapid Complexity

The vibe coding experience was fun. I think I’m going to build more software with it.

However, my biggest takeaway is that I can create incredibly complex working artifacts without fully understanding them.

This is both amazing and scary.

As vibe coding and AI-generated startups proliferate, we must ask ourselves what the overall consequences of this transformative technology will be.

The fact that these systems appear to function on the surface but are epistemically opaque raises a host of issues.

What are the unintended risks of scaling systems whose complexity we can't comprehend, either individually or collectively?

And do we really want systems that produce “throwaway code” or systems that can’t be fixed when they break?

My oven broke over a year ago, and I haven't had it replaced. I have better things to do with my money. I asked someone to come and repair it, and they said, 'No, you’ve got to throw it away.' I’m worried we will end up with a lot of broken ovens.

I guess that's what magic is—outcomes without understanding. Magical worlds are fine, but somewhere inside of me, it makes me feel like the last several hundred years have been geared toward removing magical thinking in favor of reason and understanding.

Are we ready to go back?

Version 1.0 - Straight out of the vibe

### Substack Post for Cluster 0

**Opening Paragraph:**

Imagine yourself in a bustling city, surrounded by intricate networks of infrastructure, each system functioning autonomously yet interdependently. This sprawling urban landscape represents the complexity of our current sociotechnical systems, where artificial intelligence (AI) has become an integral part of our daily lives, just like the electric grid or public transportation. Yet, as we increasingly rely on these AI systems, we're confronted with a paradox: the very technology that promises to improve our lives also brings with it new challenges and risks. This article explores the rise of Generative Artificial Intelligence (GenAI), a powerful new technology that is revolutionizing industries but also raising concerns about accountability, security, and education.

**TL;DR:**

Generative AI, while a promising tool for innovation, presents unique challenges due to its complexity, transparency issues, and potential security risks. Successful GenAI implementation requires a balance between innovation and governance, necessitating new frameworks for enterprise architecture management (EAM) and a focus on equipping learners with skills to engage with GenAI effectively. Meanwhile, the evolution of AI security practices and the development of AI agents with integrated cross-domain abilities are critical to harness the full potential of GenAI and mitigate its risks.

**Main Body:**

Generative AI is a rapidly evolving technology that is reshaping industries, from software development to education. Its power lies in its ability to generate high-quality outputs, such as images or text, from simple inputs. For example, visual GenAI models can create stunning images from textual prompts, offering limitless possibilities for creativity and innovation. However, the complexity of these AI systems and their opaque decision-making processes raise concerns about accountability and transparency. In a world increasingly reliant on AI, understanding the underlying mechanisms of these systems is crucial to manage potential risks and ensure responsible AI use.

On the other hand, the integration of GenAI into enterprise settings poses unique challenges. Organizations face significant obstacles in scaling GenAI, including technological complexity, governance gaps, and resource misalignments. Traditional enterprise architecture management (EAM) frameworks often fall short in addressing these challenges due to their inability to accommodate the unique requirements of GenAI. For example, balancing innovation with compliance and managing low data governance maturity are specific issues that must be tackled when adopting GenAI.

The education sector is another area where GenAI is making significant inroads. Tools like ChatGPT are revolutionizing teaching and learning, demanding a new set of skills from students and educators alike. AI literacy, critical thinking, and ethical AI practices are just some of the competencies needed to engage effectively with GenAI. But the integration of GenAI into education isn’t just about equipping learners with the right skills; it also involves reshaping curriculum design and pedagogical strategies to foster inclusive and responsible GenAI adoption.

**Model or Mental Framework:**

The complexity of GenAI can be likened to a bustling cityscape, with its intricate web of systems and networks. Just like urban planners need to understand the city's infrastructure to manage it effectively, we need a deep understanding of GenAI to harness its potential responsibly. This involves viewing GenAI not just as a tool but as an "artificially created nature," a metaphor that underscores the need for a precautionary approach to GenAI risks.

A useful model for addressing these risks is the Advanced Threat Framework for Autonomous AI Agents (ATFAA), which organizes agent-specific risks. This model helps us understand how GenAI agents' autonomy, persistent memory access, complex reasoning, and tool integration create novel security risks. Coupled with the SHIELD framework, which proposes practical mitigation strategies, we can begin to address GenAI’s unique security challenges.

In the context of education, the Aspire to Potentials for Learners (A2PL) model offers a way to cultivate Self-Directed Growth, a competency that enables learners to drive their own developmental pathways within GenAI-supported environments. This model highlights the importance of learner aspirations, complex thinking, and summative self-assessment in fostering a sustainable learning ecosystem.

**Takeaways:**

Embracing GenAI calls for a delicate balance between innovation and governance. Organizations must tailor their EAM frameworks to GenAI-specific challenges, and educational institutions should equip learners with the necessary skills to engage effectively with GenAI. As GenAI agents become more common, we must adapt our security practices to address their unique risks and continue to explore the integration of cross-domain abilities to advance toward Artificial General Intelligence (AGI).

**Closing:**

In conclusion, as we navigate the bustling cityscape of GenAI, we're faced with the challenge of ensuring that this powerful new technology serves us well without compromising our security or integrity. As we continue to explore the potential of GenAI, what new frameworks and strategies will emerge to guide us? How will we ensure that everyone, from enterprise leaders to learners, is equipped to engage with GenAI effectively? These are questions that will shape our journey through the evolving landscape of Generative Artificial Intelligence.

---

### Related Articles

- Accountability of Generative AI: Exploring a Precautionary Approach for

"Artificially Created Nature" (http://arxiv.org/abs/2505.07178v1)

- Enterprise Architecture as a Dynamic Capability for Scalable and

Sustainable Generative AI adoption: Bridging Innovation and Governance in

Large Organisations (http://arxiv.org/abs/2505.06326v1)

- Safer Prompts: Reducing IP Risk in Visual Generative AI (http://arxiv.org/abs/2505.03338v1)

- NGENT: Next-Generation AI Agents Must Integrate Multi-Domain Abilities

to Achieve Artificial General Intelligence (http://arxiv.org/abs/2504.21433v1)

- Fostering Self-Directed Growth with Generative AI: Toward a New Learning

Analytics Framework (http://arxiv.org/abs/2504.20851v1)

- Securing Agentic AI: A Comprehensive Threat Model and Mitigation

Framework for Generative AI Agents (http://arxiv.org/abs/2504.19956v2)

- Generative AI in Education: Student Skills and Lecturer Roles (http://arxiv.org/abs/2504.19673v1)

- The Role of Generative AI in Strengthening Secure Software Coding

Practices: A Systematic Perspective (http://arxiv.org/abs/2504.19461v1)

Version 1.1: With Personalizaiton

### Substack Post for Cluster 1

**Opening paragraph**

As I stroll through San Francisco, a city I call home, I can't help but notice the intricate urban layout that encapsulates its history, culture, and aspirations. The serpentine roads, the Victorian houses, and the steel-and-glass skyscrapers all speak to the city's past and future. Each architectural detail, each junction, and each open space is a testament to meticulous urban planning. But with the dawn of the AI era, we are witnessing a paradigm shift in how cities are designed and planned. Could generative AI redefine urban planning, and what are the implications for society at large?

**TL;DR**

Generative AI models are opening new avenues in urban planning, synthetic data generation, and complex data structuring. These models offer the capacity to automate traditionally manual processes, enhance the quality of generated data, and facilitate a more efficient transfer of unstructured to structured data. However, this cutting-edge technology also brings forth challenges, including data quality assessment, potential misuse, and intellectual property infringement. This post explores these ideas in depth, providing a comprehensive overview of how generative AI is transforming multiple sectors while highlighting the need for robust frameworks to ensure ethical and responsible usage.

**Main body**

Urban planning, traditionally a manual and labor-intensive process, is witnessing a technological revolution with the advent of generative AI. For instance, a recent study used a state-of-the-art Stable Diffusion model augmented with ControlNet to generate high-fidelity satellite imagery based on land use descriptions and natural environments. By linking satellite imagery with structured land use information from OpenStreetMap, the researchers were able to generate realistic and diverse urban landscapes. The model proved effective across diverse urban contexts, opening the door to a future where AI could play a pivotal role in urban planning, enhancing planning workflows and public engagement.

Similarly, generative AI models have made significant strides in the realm of smart grids. These models can generate large amounts of synthetic data that would otherwise be difficult to obtain due to confidentiality constraints. The challenge, however, lies in assessing the quality of data produced by these generative models. To address this, a novel metric based on Fréchet Distance (FD) can evaluate the quality of the synthetic datasets from a distributional perspective. This approach enhances the reliability of data-driven decision-making in smart grid operations.

Generative AI models are also transforming the way text data is structured. A hybrid text-mining framework converts raw, unstructured scientific text into structured data, enhancing entity recognition performance by up to 58% and improving the quality of the final structured data by up to 83%. This advancement could significantly expand the scope of data-driven scientific discovery.

Despite these promising developments, the rise of generative AI has also raised concerns, especially in the realm of visual content generation. Open-source text-to-image (TTI) pipelines, while providing creative opportunities, also facilitate harmful practices such as the generation of non-consensual deepfakes and the amplification of misogynistic or hypersexualized content. Similarly, AI-generated content risks infringing on intellectual property rights due to the potential for models to memorize and reproduce specific content.

**Model or mental framework**

To understand the impact of generative AI, we can conceptualize it as a powerful tool in a craftsman's workshop. Just as a hammer can be used to build a house or break a window, generative AI can be used for beneficial or harmful purposes. The key lies in the craftsman's intent and the controls in place to ensure responsible use.

For generative AI to deliver its full potential, we need robust frameworks to govern its use. These frameworks must balance the need for creativity and efficiency with considerations of ethics, privacy, and intellectual property rights. They should include mechanisms to assess the quality of generated data, safeguards against misuse, and measures to ensure intellectual property rights are respected.

**Takeaways**

Generative AI is revolutionizing urban planning, data structuring, and synthetic data generation. It offers the potential to automate traditionally manual tasks, enhance the quality of generated data, and facilitate efficient data structuring. However, it also raises considerable challenges, including data quality assessment, potential misuse, and intellectual property infringement.

To navigate these challenges, we need robust, comprehensive frameworks to govern the use of generative AI. These frameworks should balance the need for creativity and efficiency with considerations of ethics, privacy, and intellectual property rights. They should include mechanisms to assess the quality of generated data, safeguards against misuse, and measures to ensure intellectual property rights are respected.

**Closing**

As generative AI continues to evolve and its applications continue to expand, how will we ensure that this powerful tool is used for the benefit of all? The answer lies not only in the development of advanced models but also in the creation of robust ethical, legal, and regulatory frameworks that guide their use. As we stand on the cusp of this new era, we must ask ourselves: How can we shape the future of generative AI to ensure it serves as a tool for progress rather than a source of harm?

---

### Related Articles

- Generative AI for Urban Planning: Synthesizing Satellite Imagery via

Diffusion Models (http://arxiv.org/abs/2505.08833v1)

- Fréchet Power-Scenario Distance: A Metric for Evaluating Generative

AI Models across Multiple Time-Scales in Smart Grids (http://arxiv.org/abs/2505.08082v1)

- Symbol-based entity marker highlighting for enhanced text mining in

materials science with generative AI (http://arxiv.org/abs/2505.05864v1)

- Perpetuating Misogyny with Generative AI: How Model Personalization

Normalizes Gendered Harm (http://arxiv.org/abs/2505.04600v1)

- Safer Prompts: Reducing IP Risk in Visual Generative AI (http://arxiv.org/abs/2505.03338v1)

- Ensuring Reproducibility in Generative AI Systems for General Use Cases:

A Framework for Regression Testing and Open Datasets (http://arxiv.org/abs/2505.02854v1)

- Leveraging Generative AI Through Prompt Engineering and Rigorous

Validation to Create Comprehensive Synthetic Datasets for AI Training in

Healthcare (http://arxiv.org/abs/2504.20921v1)

Version 1.2 With an Upgraded Model

### Substack Post for Cluster 0

**Opening paragraph**

Last spring, I found myself in a small conference room at a community center in rural Virginia, listening to a group of farmers debate whether ultraviolet light could cure COVID-19. I’d been invited to speak about information flows in crisis moments, but the talk evolved into an impromptu fact-checking workshop: we sifted through Facebook posts, dissected forwarded audio clips, and traced dubious medical claims back to satirical news sites. One farmer confided he’d always trusted what a neighbor shared on WhatsApp. When I explained how generative AI can spin entirely new, but patently false, stories at the click of a button, his eyes widened. We’d crossed a threshold: misinformation had ceased to be an accident of rumor; it was now a technologically amplified phenomenon, blurring the line between honest mistake and deliberate manipulation.

That afternoon, as I packed my notes, I realized this challenge sat at the intersection of two great forces I study—technological innovation and organizational response. On one hand, generative artificial intelligence (genAI) offers the promise of accelerating discovery, unleashing creativity, and redesigning how we work. On the other, it supercharges the spread of “fake news,” a problem as old as Gutenberg’s press but newly weaponized at global scale. How, I wondered, do we harness the power of genAI without becoming unwitting distributors of falsehood?

Back at my desk, I dove into two strands of research. The first traced the legislative reaction to misinformation: who writes these laws, why they emerge, and whether they help or harm. The second examined genAI not just as a clever trick but as a genuine economic force—a general-purpose technology (GPT) that powers knock-on innovations, and even as an invention of methods of invention (IMI) that turbocharges R&D. What struck me was how little integration existed between these two literatures: we talk about regulating speech or debating productivity gains, but rarely about the dynamic feedback loop that ties them together. This essay is my attempt to stitch those threads into a coherent picture.

**TL;DR**

Generative AI sits at a crossroads: it is poised to become a transformative general-purpose technology that accelerates innovation, yet it also magnifies the age-old problem of misinformation.

Recent legislative efforts against “fake news” have gone from illiberal regimes to Western democracies, focusing on national security and public health but differing in zeal and scope according to political freedom.

To navigate this new terrain, we need a simple mental framework that captures the interplay between genAI’s innovation boons (GPT + IMI) and the evolving architectures of regulation—public laws, platform policies, and decentralized norms.

**Main body**

1. The ubiquity of misinformation predates the internet: medieval pamphleteers, yellow journalists, radio bonanzas of propaganda. What’s changed is scale and speed. Today, a doctored video of a political candidate can circulate globally within minutes. We’ve watched disinformation shape elections, sow mistrust in vaccines, and even trigger violence. The conventional playbook—from editorial gatekeepers to fact-checking NGOs—struggles under the weight of billions of real-time posts. Meanwhile, legislative bodies scramble to catch up, drafting laws that define “mal-information” and prescribe penalties for genuine or unintentional offenders. My own fieldwork in Southeast Asia found that the first wave of these laws emerged in countries with weaker civil liberties and lower GDP per capita—places where governments faced fewer constraints on speech and greater incentives to control narratives.

2. Then came the pandemic. In 2020 alone, over a hundred new statutes or regulations targeted misinformation about public health. Suddenly, the stakes were life and death: unverified herbal cures, conspiracy theories about 5G towers spreading the virus, or outright denialism that slowed vaccine uptake. Governments invoked emergency powers; platforms accelerated their content moderation algorithms; civil society groups published daily myth-busting bulletins. But the rush to legislate revealed deep tensions. Hard-line measures risked chilling legitimate speech, while softer “notice and takedown” rules often lagged behind rapidly mutating falsehoods. The result was a patchwork of responses—some draconian, others toothless—hardly the unified front needed to stem a truly global infodemic.

3. Meanwhile, genAI quietly matured. Early chatbots and vision models made headlines, but few realized these systems would soon sit at the heart of both productivity engines and information manipulation tools. Economists distinguish two crucial categories of transformative technology. General-purpose technologies (GPTs) like electricity or the steam engine spark waves of complementary innovations—new services, reorganized industries, and sustained growth. Invention of methods of invention (IMIs) such as the high-throughput microscope or combinatorial chemistry speeds up R&D itself. Remarkably, genAI ticks both boxes: it can write code, design experiments, and even generate synthetic data to train successor models.

4. What does this dual identity imply? On the upside, genAI can help us detect disinformation by automating content verification, flagging deepfakes, and surfacing the provenance of images or texts. Imagine a newsroom where AI agents scan every incoming claim, cross-reference it with primary sources, and highlight inconsistencies. Research labs could deploy genAI to sift through mountains of academic papers, identify replication failures, and propose novel experimental hypotheses. Productivity growth might get a new lease on life after a decade of plateauing by reducing the “friction” in human-machine collaboration.

5. Yet the very generativity that drives these gains also fuels the spread of falsehood. An open-source genAI model can write convincing political satire that blurs into disinformation. It can churn out thousands of phishing emails, scaled misinformation campaigns, or bot-driven comment floods. Manual fact-checking can’t keep pace; automated filters often struggle to distinguish genuine from AI-generated false claims, especially when malicious actors fine-tune models for stealth. In effect, genAI intensifies the arms race between detection and deception, manufacture and verification.

6. The policy response has entered a new phase. What began in authoritarian contexts has now swept into Western democracies. Legislators debate obligations for transparency—requiring companies to watermark AI-generated content or disclose model training data. Platform liability is back on the table, with proposals to hold social media companies accountable for algorithmic amplification of false material. At the international level, conversations range from UNESCO’s “Recommendation on the Ethics of AI” to OECD guidelines on digital content governance. The question is no longer whether to regulate “lawful but awful” speech, but how to design systems that balance free expression with societal resilience.

7. These debates mirror earlier technology cycles. The industrial revolution prompted factory safety laws as steam engines powered textile looms; radio deregulation in the 1980s unfolded alongside new content standards. What’s different now is velocity: regulation, innovation, and counter-innovation unfold in tandem, often outpacing legislators’ ability to fully comprehend emerging risks or opportunities. A parliamentary committee might take years to pass an AI transparency bill, by which time a new class of private foundation models has slipped through loopholes.

8. To bring clarity to this chaos, we must zoom out and look at the information economy as an evolving ecosystem. GenAI is a catalyst, reshaping the terrain with twin blades of productivity enhancement and misinformation amplification. Our goal should be to erect adaptive institution sandboxes—dynamic regulations, modular policy toolkits, and real-time monitoring dashboards that evolve with the technology. Such hybrid approaches combine top-down mandates (e.g., watermarking requirements) with bottom-up community standards (open-source developer norms) and market signals (liability insurance for platforms).

**Model or mental framework**

Imagine a two-dimensional map where the horizontal axis captures **Technological Generativity** (from low to high) and the vertical axis measures **Regulatory Flexibility** (from rigid to adaptive). On the left, low generativity technologies—say, basic word-processing software—pose minimal misinformation risk, so rigid regulation incurs high compliance costs for little gain. On the right, high generativity technologies—cutting-edge genAI engines—can spawn both innovation and disinformation at scale, demanding more nuanced governance.

In the top-right quadrant, **Adaptive Regulation + High Generativity**, we see fertile ground for experimentation: flexible rulebooks that update in real time, collaborative governance boards of regulators, industry, and civil society, and continuous evaluation loops. This is where genAI’s productivity benefits outweigh oversight costs, and where regulation can shape innovation trajectories—steering models towards ethical design and transparent deployment.

Conversely, the bottom-right quadrant, **Rigid Regulation + High Generativity**, risks either stifling innovation (if rules are too burdensome) or failing to curb misuse (if they’re quickly evaded). Past examples include broad “data localization” laws that drive AI talent offshore or sweeping bans on encrypted messaging that undermine privacy without halting bad actors. Our task, then, is to navigate toward the top-right: calibrating regulatory interventions to match genAI’s generative power, employing sunset clauses, impact assessments, and co-regulatory frameworks.

**Takeaways**

• Generative AI is both a general-purpose technology and an invention of methods of invention. That dual nature makes it a powerful engine for productivity growth—but also a supercharger for misinformation. Understanding this duality is key to designing effective responses.

• Global legislative efforts against “fake news” have spilled from illiberal states into mature democracies, driven by threats to national security, public health, and electoral integrity. Yet the shape and effectiveness of these laws hinge on each country’s legal traditions, technological capacity, and political norms.

• We need dynamic, layered governance architectures that blend top-down laws with bottom-up norms and market incentives. A simple 2×2 framework—mapping technological generativity against regulatory flexibility—can guide us toward adaptive regimes that foster innovation while safeguarding informational integrity.

**Closing**

As I left that rural hall in Virginia, I reflected on one farmer’s parting question: “Can we trust the next headline we read online?” There’s no simple yes or no. Trust hinges on how we shape the engines beneath the surface—both the algorithms that craft narratives and the laws that police them.

In my next essay, I’ll unpack one of those layers—how open-source communities can co-govern public AI models, setting technical standards that preempt misuse without stifling creativity. Because if there’s one thing I’ve learned in ten years of field experiments, it’s that governance works best when it is as generative and adaptive as the technologies it seeks to channel.

For now, I leave you with this: in a world where machines can both invent cures and concoct conspiracies, what institutional innovations will we need to ensure that truth retains the upper hand? Over to you.

---

### Related Articles

- Lawful but Awful: Evolving Legislative Responses to Address Online

Misinformation, Disinformation, and Mal-Information in the Age of Generative

AI (http://arxiv.org/abs/2505.15067v1)

- Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope? (http://arxiv.org/abs/2505.14588v1)

Final Code

pip install requests pandas scikit-learn openai beautifulsoup4

import requests
import pandas as pd
from bs4 import BeautifulSoup
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.cluster import KMeans
import openai

# ========= PERSONALIZATION =========
def personalize_voice(base_prompt: str, personality_descriptor: str = None) -> str:
    """Injects a personal writing voice into the base Substack prompt."""

    default_descriptor = (
        "You are Sharique Hasan, a professor of strategy and sociology who blends deep academic rigor "
        "with accessible storytelling. Your writing style is reflective, often drawing from personal experience, "
        "field experiments, and empirical work to illustrate insights. You connect theory to practice with clarity, "
        "emphasizing real-world implications, especially in entrepreneurship, innovation, and social networks."
    )

    descriptor = personality_descriptor if personality_descriptor else default_descriptor

    personalized_prompt = base_prompt.replace(
        "You are a Substack writer specializing in",
        f"{descriptor}\n\nYou are writing a Substack article specializing in"
    )

    return personalized_prompt

# ========= PROMPT GENERATION =========
def format_substack_prompt(cluster_id, summaries, topic="innovation and organization"):
    return (
        f"You are a Substack writer specializing in {topic}.\n\n"
        f"Your goal is to write a thoughtful and engaging article based on the following cluster of ideas: {cluster_id}.\n"
        f"Use the provided article summaries below as your source material. Boil the ideas down to an essential argument that is both intuitive and powerful.\n\n"
        f"Structure the post in this format:\n"
        f"- **Opening paragraph**: Personal, reflective, or story-driven context.\n"
        f"- **TL;DR**: A short summary of the core idea or insight.\n"
        f"- **Main body**: Develop the argument through clear examples, often comparing states (e.g., past vs. future, manual vs. automated).\n"
        f"- **Model or mental framework**: Provide a simple analytical tool, metaphor, or economic model that explains the core transition or mechanism.\n"
        f"- **Takeaways**: What should readers remember or apply?\n"
        f"- **Closing**: End with a question or insight that sets up a future post or reflection.\n\n"
        f"Your tone should be clear, reflective, and idea-driven—similar to a well-written essay on Substack that combines storytelling, conceptual depth, and actionable insight.\n\n"
        f"---\n\n"
        f"{summaries}\n\n"
        f"---\n\n"
        f"Now, write the article in full, following the format above. Ensure that each section has at least 3 paragraphs. The total post should be 3000 words."
    )

# ========= ARXIV FETCH =========
def fetch_arxiv_articles(query="generative ai", max_results=100):
    url = f"http://export.arxiv.org/api/query?search_query=ti:\"{query}\"&sortBy=submittedDate&sortOrder=descending&max_results={max_results}"
    response = requests.get(url)
    soup = BeautifulSoup(response.content, 'xml')
    entries = soup.find_all('entry')
    articles = []
    for entry in entries:
        articles.append({
            'title': entry.title.text.strip(),
            'summary': entry.summary.text.strip(),
            'link': entry.id.text.strip()
        })
    return pd.DataFrame(articles)

# ========= CSV OUTPUT =========
def save_to_csv(df, filename='arxiv_articles.csv'):
    df.to_csv(filename, index=False)

# ========= CLUSTERING =========
def cluster_articles(df, n_clusters=5):
    if len(df) < n_clusters:
        print(f"Warning: Only {len(df)} articles. Skipping clustering.")
        df['cluster'] = 0
        return df
    vectorizer = TfidfVectorizer(stop_words='english', max_features=1000)
    X = vectorizer.fit_transform(df['summary'])
    model = KMeans(n_clusters=n_clusters, random_state=42)
    df['cluster'] = model.fit_predict(X)
    return df

# # ========= GENERATE BLOG POST =========
# def generate_post_ideas(summaries, cluster_id, openai_client):
#     base_prompt = format_substack_prompt(cluster_id, summaries)
#     prompt = personalize_voice(base_prompt)
#     response = openai_client.chat.completions.create(
#         model="gpt-4",
#         messages=[{"role": "user", "content": prompt}],
#         temperature=0.7
#     )
#     return response.choices[0].message.content


# ========= GENERATE BLOG POST =========
def generate_post_ideas(summaries, cluster_id, openai_client):
    base_prompt = format_substack_prompt(cluster_id, summaries)
    prompt = personalize_voice(base_prompt)
    response = openai_client.chat.completions.create(
        model="o4-mini-2025-04-16",  # Updated to latest best model
        messages=[{"role": "user", "content": prompt}],
        temperature=1
    )
    return response.choices[0].message.content

# ========= FILE OUTPUT =========
def save_cluster_ideas(df, openai_client, base_filename="substack_ideas_cluster"):
    grouped = df.groupby("cluster")
    for cluster_id, group in grouped:
        summaries = "\n".join(group['summary'].tolist())
        references = "\n".join(f"- {row['title']} ({row['link']})" for _, row in group.iterrows())
        ideas = generate_post_ideas(summaries, cluster_id, openai_client)
        output_text = (
            f"### Substack Post for Cluster {cluster_id}\n\n"
            f"{ideas}\n\n"
            f"---\n\n"
            f"### Related Articles\n{references}"
        )
        with open(f"{base_filename}_{cluster_id}.txt", "w", encoding="utf-8") as f:
            f.write(output_text)

# ========= MAIN =========
if __name__ == "__main__":
    OPENAI_API_KEY = "API KEY GOES HERE"
    client = openai.OpenAI(api_key=OPENAI_API_KEY)

    print("📡 Fetching arXiv articles...")
    articles_df = fetch_arxiv_articles("generative ai", max_results=30)

    print("💾 Saving articles to CSV...")
    save_to_csv(articles_df, "arxiv_articles.csv")

    print("🧠 Clustering articles...")
    clustered_df = cluster_articles(articles_df)

    print("📝 Generating Substack posts...")
    save_cluster_ideas(clustered_df, client)

    print("✅ All tasks completed.")

中文 | Español | हिंदी | العربية | Bahasa Indonesia | Português

This is the second post in my series on AI. In this post, I explore an overlooked dimension in the competition between humans and machines: the clarity of labor market signals.

TL;DR.

While AI advances through increasingly rigorous benchmarks that improve and communicate model capabilities, human potential becomes harder to evaluate due to grade inflation. This may lead to a dangerous divergence in signal clarity that could distort labor markets and undervalue human capabilities, particularly in entry-level labor markets.

Humans, AI, and the Changing Nature of Skill and Ability

There was an interesting op-ed in the New York Times from a LinkedIn executive sounding the alarm about the potential collapse of entry-level labor markets due to AI.

I've been thinking a lot recently about the competition between AI and humans, especially how we humans can come out ahead.

There's a puzzle I'm trying to solve: AI models keep improving as we make their tests more challenging. At the same time, humans are becoming increasingly difficult to evaluate. Students, in particular, face tests that are becoming easier and less informative. Think grade inflation.

This divergence raises an important question. What economic consequences will this shift create for the labor market?

One of the most crucial elements of this system is the educational institutions that train, sort, and help workers signal their abilities to firms. Education provides one of the few opportunities for effective worker signaling across the entire market. Once you work for a firm, much of your signaling ability fades because a significant portion of your value comes from firm-specific knowledge, and much of your information remains trapped within the firm.

What is a Benchmark? Signal and Learning in Education and AI

Before I get into the details of my hypotheses regarding the impact of the divergence in benchmarks for humans and AI, it's helpful to consider what a benchmark is. I briefly discussed this in my previous post, but it's worth reiterating.

A benchmark is a test.

This test has correct and incorrect answers, and it serves two main purposes.

The primary purpose is to convey a signal regarding the test taker’s abilities. This assumes that the test is well-designed and can differentiate between strong and weak test takers. It indicates something about their underlying abilities, which are more challenging to observe.

The second purpose is its treatment effect. A more challenging test means people are more likely to answer questions incorrectly. If the test and the testing methods are well designed, individuals then identify what they answered incorrectly, adjust their learning strategies, and learn how to correctly answer those questions. The next time they take the test, they perform better. This has a different significance for the test taker, as the test does not merely reveal their quality; it also contributes to improving it.

Easier tests mean that everyone performs well. You cannot distinguish the very best from the average or poor performers. Similarly, an easy test offers no real chance for failure; therefore, no feedback loop exists to alter the learning process. Without this feedback, the test taker does not improve.

Why AI Benchmarks Are Getting Harder and What That Means

Given these definitional issues, it is worth noting that one reason AI models are improving is that the benchmarks are becoming more challenging. These benchmarks are growing tougher not only due to the inclusion of more difficult tasks, such as classifying harder images, but also because they are becoming much more diverse in what they measure. With the edge cases and complexities present in these new benchmarks, we are accomplishing two things simultaneously.

We are gaining a better understanding of what these models can do, which would not be possible without individuals creating improved and more comprehensive benchmarks.

At the same time, we are pushing the models to improve on tasks they could not previously handle. There are, of course, trade-offs, as getting better at one task may sometimes lead to worse performance in another. (And AI benchmarks aren’t perfect either. There are many problems, and I will explore this in a future post.)

However, more challenging benchmarks benefit both sides of the market.

Competitive Implications of Rigorous Benchmarks

Firms such as OpenAI, Anthropic, Meta, and many others in the foundation model business have incentives to improve their models and showcase that enhancement to the market by excelling on benchmarks.

This benchmark-driven competition produces models for which we can obtain comparable performance evaluations on specific tasks. From these evaluations, we draw conclusions about each model’s capability to handle the work we do in our organizations.

This relies on another somewhat hidden assumption: do the benchmarks provide relevant information to firms?

That is, can the scores that models achieve on benchmarks be used to predict how well the models will perform when “working” within a firm?

Firms are figuring this out. We’ll hopefully see the results in the productivity statistics.

The Grade Inflation Crisis and Why Human Benchmarks Are Getting Easier

Now, let’s turn to human benchmarks.

There is ample evidence that grades are not what they once were.

Evidence suggests significant grade inflation exists everywhere, including high schools and colleges, especially elite ones. For example, the typical high school GPA in the late 1980s was around 3 (slightly higher for females, lower for males). By 2002, that average was closer to 3.2-3.4. Average GPAs are likely even higher now.

It is unlikely that high school students have become that much smarter. The more plausible explanation is that benchmarks have become easier. We also see similar trends in other assessments outside regular grading. AP scores tell a similar story: scores have dramatically improved.

This can be interpreted in various ways. One possibility is that students are indeed improving, which means the exam is masking real gains among the very top students. This information may be valuable to the student and the labor market.

Another possibility is that the benchmarks themselves are becoming easier, which means students are not improving, their grades are.

As a result, the information from these scores is less valuable.

Evidence (from one consistent externally administered benchmark, the SAT), suggests that grades and test scores have decoupled. Average SAT scores have remained pretty flat over this same time period (early 80s to 2006/7). This is in stark contrast to the GPA inflation above.

In any case, there are two possible outcomes. First, the signal becomes weaker, and markets cannot distinguish between the best and the rest.

Everyone seems to be the best.

Second, ability either stagnates or declines as incentives to learn decrease.

Economic Incentives Behind Grade Inflation

Many factors may have contributed to the rise of grade inflation. Some of these relate to the economics of educational institutions.

First, universities have expanded their range of programs.

In particular, there has been a massive increase in master’s programs, which help fill a revenue gap for these institutions. This expansion has resulted in a sharp increase in the number of students earning these degrees, as well as the cost of those degrees. Consequently, there has been a notable shift in how universities perceive their enrollees. The role of the student is evolving toward that of a customer.

The second set of factors is that this is not just a university phenomenon; it is a “systems” issue. As the value of a college degree, especially for knowledge work in the economy, has increased (though with some caveats), grades have become more important as a signal in the labor market.

It’s possible that the incentives would then be to make each candidate appear favorable in the labor market, especially as more people attain college degrees.

There is also the issue of rankings that affect incentives and sometimes bad behavior. But rankings deserve their own post.

The final explanation presents a more micro-level incentive story.

The increasing economic value of education has created a greater need to assess professors. There has been a notable rise in teacher-student evaluations, particularly in universities. Evaluations are used for tenure and promotion, and they are required to apply for jobs at other universities. So, faculty have many incentives to get good evaluations.

Having been through these evaluations myself, they can be influenced. The more we demand from my students, the more they feel judged, and the harsher the reviews we receive. Research supports this dyanmic (see more).

All of these forces create a perfect storm that makes benchmarks for humans noisier.

When Signals Fail in Markets for Skill

If we take both Kenneth Arrow (“Higher Education as a Filter”) and Michael Spence (“Job Market Signaling”) seriously and accept that higher education acts as a filter providing signals to the labor market, then many universities and high schools may muddy that signal.

If we view the labor market as a system, the goal of higher education is twofold. First, it trains individuals and develops their skills and capabilities. Second, it provides clear signals that assist the market in matching workers to firms.

Because firms do not observe worker skills directly, they rely on signals to make decisions. A key question is whether signals like GPA remain useful.

Why Signal Clarity Matters As Much (Maybe More) Than Raw Capability

As I have written earlier, there is much more noise in the system because the cost of applying is now lower. AI is also making it easier to fake other signals, such as the quality of a cover letter.

These forces, combined with the fact that GPAs are no longer strong signals, mean firms must find other ways to separate the workers they want from those they do not.

This uncertainty contrasts with the much greater clarity regarding AI's capabilities. While AI may not possess all the necessary capabilities (just yet, or ever), and we do not fully comprehend how benchmarks translate to actual performance within organizations, we can discern the distinction between the best and the lowest-performing models within a specific signal class.

In contrast, humans are entering a classic market for lemons.

In a very interesting paper by Amanda Pallais, this dynamic is evident. Employers prefer hiring workers who may have lower quality but offer clearer signals rather than higher quality workers who lack clear signals.

When you know with greater certainty that a person is of quality q, even if it is not ideal, you can design around it. Uncertainty is harder to design around.

Swapability and Risk in Organizational Workflows

Currently, we have not yet determined where algorithms fit into organizational workflows. This process is just beginning, and we will learn much more in the coming years. However, AI models are quite easy to swap (at least technically). You could replace Claude with ChatGPT in just a few minutes for many workflows.

Getting rid of a “bad” human worker, however, is costly. There is legal risk, endless performance improvement plans, and of course, the human cost. Some industry estimates suggest that replacing a “bad” worker can cost over 2x their salary or even more (recruiter.com says it’s 840k for a person who makes 62k. Who knows whether this exact figure is true, but it is quite high!).

This presents an intriguing dilemma worthy of consideration in the ongoing debate about who will prevail in the future: AI versus humans. Capabilities are one aspect influencing hiring decisions, but uncertainty is another.

If AI models provide clear signals, this poses a fundamental risk to human labor markets that do not.

In other words, AI models maintain a competitive advantage with reduced uncertainty even if their abilities are weaker.

Humans, conversely, emit weak signals and incur high correction costs. They carry risk.

Talented workers are particularly likely to be undervalued in this environment. When they sit in classrooms, working hard and paying attention, but their grades are the same as those of someone who never shows up, simply because they pay the same tuition, that degree becomes less useful. They will seek another way to demonstrate their skill.

This outcome is detrimental to workers and the institutions connected to the ecosystem that are responsible for providing clarity about which humans are worth hiring.

Strategic Responses for Workers, Employers, and Universities

What will begin to happen? How will the various agents in the system respond strategically?

If I were a capable worker in this era, I would put in a lot of effort to develop my “personal brand.” I would start building my own portfolio to showcase the work I have done and create my own profiles. I would not rely on my degree to signal my value because I know firms cannot distinguish me from someone else who has the same degree but less ability.

For employers, I would conduct more internal assessments and training and rely less on universities for my labor pool.

A Crisis of Credibility?

Are universities undermining their position in the labor market as honest brokers by sending misleading signals?

There is a tension that makes this problem difficult to solve.

If a single university (or professor, for that matter) makes a unilateral choice to grade more stringently, it might penalize itself.

Evidence shows that many educational institutions, from undergraduate to graduate levels, do not consider the fact that some universities inflate grades more than others. Students from universities that do not inflate grades are at a disadvantage compared to those with impressive grades from institutions that award higher grades to everyone. Princeton, for instance, decided to go back to inflating grades after trying to “deflate.”

I don’t believe we will resolve this dilemma easily, certainly not in a single Substack post. However, if benchmarks continue to diverge, with AI tests becoming more difficult while human educational systems persist in goldbricking (students pretend to learn, we pretend to teach, and everyone gets a trophy), we will encounter a significant problem.

So, as we worry about whether AI will replace humans, we might also want to consider signal clarity alongside capabilities.

This is a collective action problem that no single person can resolve.

Aligning Human and Machine Benchmarks for a Human-First Economy

So what is the punchline?

Markets punish opacity and reward clarity.

Much of our current discussion focuses on whether AI will replace humans due to its superior capabilities. However, in designing a human-first labor market, we should consider better ways to signal human capabilities.

These capabilities are obscured by vague, reassuring A’s or B’s that should have been C’s or worse. If we desire a human-first economy, we must establish human benchmarks that enable us all to communicate value clearly in the labor market.

This will help us compete with the machines.

Over the next year, I will share what I am learning during my sabbatical on Superadditive.co. I plan to focus on understanding modern AI, particularly how it could shape innovation, strategy, and organizational performance. If that sounds interesting, I would be honored to have you along for the journey. I will continue to post on other topics that I find interesting too…

TL;DR

AI is currently integrated into workflows through manual, fragmented, cut-and-paste methods, but it will only become transformative once it seamlessly embeds into organizational processes. This transition relies not just on whether human involvement can be effectively replaced or complemented but also on the economic incentives of software providers, firms, and users that make up the workflows inside organizations.

Today’s Reality: Patchwork AI

Today, AI exists in a patchwork state. It is manually inserted through cut-and-paste into workflows and used to assist us in thinking, searching, structuring, or translating in idiosyncratic ways. 

But if AI is to have a lasting impact, and I believe it will, it will not remain visible. Like past technologies that reshaped work, it will become embedded, silent, and eventually invisible.

From Zip Drives to Overleaf: A Workflow Evolution

Writing is a large part of my work. Especially writing research papers, getting them published, and hopefully having them cited by colleagues.

When I was in graduate school, my writing process felt quite cumbersome. At that time, there was very little of what we now call “cloud computing.” After working on a manuscript in Microsoft Word, I would save it to my hard drive, copy it to a Zip drive, and open it again on my home computer.

This addition (the zip drive) to my workflow was a workaround that enabled me to transfer work from the computer lab at Carnegie Mellon to my home PC. Later, I discovered another patch: emailing myself a document so I didn’t have to rely on a Zip drive. I would work on a manuscript at home, then email it back to myself to continue the next day in the lab.

Hours wasted cleaning up citations.

As many people know, academic writing also requires citing prior literature to give credit to others whose ideas you build upon. Adding citations to a paper, let alone a dissertation, was incredibly cumbersome. I remember spending hours refining the citations after multiple rounds of revisions. I had to ensure that names were spelled correctly, dates were accurate, and that the references at the end of the manuscript adhered to the correct format, whether MLA, APA, or Chicago style, and were listed in the proper order. For every paper, I spent several hours just polishing the citations.

It was mind-numbing and tedious work. I wouldn’t wish it on anyone.

If you added it up, a significant amount of my time went into sending myself files, ensuring they were correctly numbered, copying and pasting regression tables, reformatting them to look neat, checking that citations were accurate, and performing what was, frankly, mind-numbing work.

I started using LaTeX for academic writing when I began my first academic job. LaTeX alleviated much of the headache associated with citation management. All I had to do was type \cite{…} and the citations would update automatically.

But there were still hassles with collaborative projects. We had to make sure the right file was being used and that we were not making conflicting changes.

My colleague developed a system using Subversion, a version control software for Linux. This improved things a lot.

Eventually, we transitioned from Subversion to Dropbox. Dropbox automatically syncs the file on your hard drive with the one in the cloud.

My work-from-home workflow came a long way in just 7 years: from Zip Drives to Dropbox.

Workflow Integration with Cloud Computing

Fast-forward a few more years, and I shifted my academic writing to Overleaf. Overleaf integrated what had previously been a patchwork of fixes that supported a fragmented academic writing workflow. With Overleaf, I no longer have to move files back and forth manually, update citations manually, or troubleshoot LaTeX issues.

Overleaf has saved me hundreds of hours. I estimate that for each manuscript, I save dozens of hours that were previously spent on tedious and unnecessary tasks. Now, most of my energy is focused on writing.

This transformation from Patchwork Cloud Computing to Invisible Cloud Infrastructure in the academic workflow has saved millions of hours. This wasted time was previously spent on tasks that, while necessary, did not require scholarly human capital and should not have demanded so much effort.1

Think, Search, Structure, and Translate

If we consider where generative AI stands today, we are in the early days of Patchwork AI. For most people, AI usage appears to be a series of janky insertions into workflows that have not yet been designed for its capabilities. 

When we have a document that needs feedback, we paste it into ChatGPT with an off-the-cuff prompt (“vibe” prompting). We then take that output and paste it into a PowerPoint deck.

We extract text from an article, request ChatGPT to summarize or provide feedback, and paste the responses into a note-taking app. We might use AI to draft emails, check grammar, refine code snippets, brainstorm slide titles, or generate some bullet points to initiate a meeting.

In all these cases, we spend a surprisingly (though currently taken-for-granted) large portion of our limited time as cutters and pasters.

Broadly speaking, we have integrated AI into our workflows so that we can better:

• Think: Reason, brainstorm, or generate a first draft. (e.g., help me explain why we should not go forward with this proposal, weigh the pros and cons).

• Search: Find and explain relevant information. (e.g., what happened during the industrial revolution in India).

• Structure: Organize ideas, documents, or workflows. (e.g., take these meeting minutes and highlight the most common themes)

• Translate: Map your text into some other forms (e.g., can you write code to take an academic article and write it as a press release; or write code to estimate a difference-in-difference estimation with all modern methods in Stata; convert Python code into base R.)

If we observed how people use ChatGPT today, it would not look very different from watching someone emailing files back and forth. However, instead of storing and transferring information, generative AI is inserting itself into the workflow by performing these cognitive tasks that once required what we thought were inherently human skills.

Subscribe now

From Patchwork to Invisible AI

Stepping back, it's clear that patchwork AI represents our current state. However, we will eventually transition towards something more akin to invisible AI for many of our workflows. 

How and for what workflows this transition occurs is an interesting question for technology scholars, but it also has practical implications for entrepreneurs and everyday AI users.

Consider the straightforward challenge we discussed earlier: transferring files from one computer to another. 

At the core of this lies an economic decision. Implementing a patch (e.g., sending a file to yourself via email) incurs some cost, C. 

And there are some benefits or gains, G, that come from using the new technology. Working from home gives me more flexibility. I do not have to sit in the computer lab until 11 p.m. When I bought a Zip drive, I decided that G > C.

When I switched to emailing myself the files, I reassessed the total value created, which is V(Email) = G - C. Emailing was cheaper than buying and carrying a Zip drive. I lost some consistency because versioning became a concern. However, at each stage, I evaluated the gain versus the cost.

When I adopted Subversion, I had to spend some time learning it, but the gain was substantial: V(Subversion) > V(Email)

Notably, the value I got from adoption increased with the number of times I repeat the activity. The total value grows with each repetition N.

Now, I pay for premium Overleaf, and the benefits I gain from it are substantial. The annual cost K for the subscription is worthwhile for an integrated writing and collaboration platform. That cost begins to make sense when the value of the patchwork solution and the integrated solution start to converge as the number of times I use the solution (N) increases. At some point, the integrated solution is much value creating in the long run.

When Integration Makes Sense

On the consumer side, the equation is simple. Do I use the technology often enough that the cost of purchasing the integrated solution is lower than the total cost, in terms of time or money, associated with my patchwork approach?

Let’s assume for a moment that K for the patchwork solution is $0; for the integrated solution, it is $120. However, the net benefit of G-C for the patchwork solution is $3, and $5 is for the integrated solution.

Strategic Bottlenecks and Workflow Design

For integrated solution providers, the key question is whether the cost of developing the solution can be recouped by selling it to a sufficient number of buyers. If demand increases for specific AI workflows, integrated solutions for particular workflows are likely to emerge. This is simple math and is not very different from traditional software solution providers. 

Why Humans Are Still in the Loop

The real question, however, is when the integrated human in the “patchwork” will be replaced with a fully autonomous AI solution?

I believe this depends on three factors, not all of which relate to the need for humans to be involved in the production process. It will be a mix of what humans can uniquely do, and a lot of what the incentives are for firms (both entrepreneurs and “user” firms) to invest K in developing, marketing, and centrally maintaining integrated solutions.

The Risk is Not Average Gain, But Worst-Case Loss

The first aspect relates to what humans are doing during the patchwork and whether AI can replace the entire set of steps involved in the patch. For example, if humans participate in aspects pertaining to ethics, judgment, or trust during the act of transferring information from one window to another, then establishing a direct AI pipe from App A to App B without human involvement may actually decrease the benefit, G, relative to the cost. In other words, G is some function of Humans + AI, and it cannot be replicated by AI alone. 

This is particularly true if we consider G, not as some guaranteed and fixed quantity, but as a stochastic variable E[G]. That is, a payoff drawn from a distribution. What happens if your draw of a particular G is highly negative? One blunder can completely ruin the outcome. Often, the strategic value of a decision is not whether the mean is high or low, but whether you can manage the tails of the outcome distribution. For instance, a VC often makes all her money on one or two startups that go gangbusters. Similarly, HR often spends most of its time on the 1 or 2 employees who are total disasters.

Managing process variance is a crucial friction that may slow the transition into integrated AI. Visibility, and thus Patchwork AI, is a feature here even if the costs are still reasonably high.

Semi-manual patchworks are how we address uncertainties that we, as humans, want to observe and manage during our workflow. We want to be in the loop so the worst-case scenario doesn’t transpire (or we can blame someone for it!). This complicates our model slightly because it expands the G minus C equation into something that includes uncertainty around G.

Human beings participate in many workflows to reduce that uncertainty, even if it means accepting a slightly higher cost.

Complementarity and Workflow Economics

The second aspect worth considering is the extent to which the patchwork fits into a system with a high degree of complementarity. By complementarity, I mean situations where different parts of a process rely on and amplify each other’s contributions, where solving one piece doesn’t just help on its own, but unlocks progress elsewhere.

For instance, if AI can solve a problem that is an essential input into a later problem, then the impact of AI will be substantially more significant. The counterfactual is true as well. If AI makes it harder to do a downstream job, it won’t exist in the workflow for too long.

Identifying strategic bottlenecks with low risk and high gain appears to be where integration is likely to yield the greatest benefits.

Multiple Firms and Incentive Issues in the Workflow

A third factor to consider is an issue that may deserve a separate post: the incentives of the firms involved in the workflow to invest in integrating what is currently “patched.”

The typical workflow for a knowledge worker consists of software provided by various players (in a typical data I use software by Microsoft, Apple, Grammarly, Substack, Stata, R, Google, and a lot more!). The key point is that workflows are endogenous systems made up of products created by economic actors.

Consequently, there is an incentive problem for any one provider to offer something that increases value across the entire value chain.

One example that comes to mind is checking grammar.

A puzzle I have been pondering is why high-quality grammar editing isn't integrated into every single application. Why must I pay $199 a year for Grammarly? And why do I pay for that, even though Google Docs has a grammar checker, and Microsoft does too?

The fact that Grammarly has identified a market niche when grammar checks should, in theory, be integrated is very interesting (it is currently valued at around 8 billion, down from 13 billion during its last raise in 2021).

It likely has to do with the incentives, or rather the lack of them, for other providers not to invest in grammar editing modules. This is either because the fixed cost, K, of building it in-house is too high (maybe), or more likely, the pay-offs are too low.

A patchwork provider may emerge if different companies make up a knowledge worker’s workflow, and none can recover the cost of building the integrated solution internally. This provider offers significant benefits, reduces costs, and generates a positive return on investment, but the new technology use remains a patchwork rather than an integrated solution.

A Future of Mixed AI: Patchwork and Invisible

AI is likely to have both invisible and patchwork components. Invisible AI will arise from workflows dominated by a single firm or where purchasing firms find it worthwhile to invest in custom integrations. These firms are motivated to develop integrated AI due to numerous users (high Q), frequent usage (large N), and strong complementarities across the system's tasks.

At the same time, opportunities for patchwork AI will still exist. These will arise in areas where the value of human involvement is (a) high, especially due to variance issues, but too complex to automate, or (b) where interoperability issues complicate integration. If no single firm has the incentive to pay the fixed cost of building a unified solution because the integration must occur across very different and non-overlapping software environments, then the patchwork opportunities will likely persist.

Other factors, including privacy concerns, legacy software, and user trust, also make it likely that patchwork solutions will continue to play a significant role in certain areas of the AI landscape. These particular issues, however, can be folded nicely into our simple economic model.

Okay, so what’s the takeaway from all this? We currently live in a world of patchwork AI, with some glimmers of integrated solutions. Where we eventually see integrated solutions will depend on the economics of the task system, the nature of the work to be done, and the incentives of both users and firms to invest in the fixed costs required to build and adopt integrated systems.

However, learning curves will set in once adoption (whether patching or integration) occurs. It may become very difficult to displace a player like Grammarly because they have accumulated more data, experience, and have built durable user habits. So, even if an integrated solution is theoretically possible, a patchwork solution may remain more appealing due to the weak incentives for other firms to catch up and the relatively high G-C for the patchwork solution resulting from learning-by-doing effects for users.

The Value of Humans in The AI-Enabled Workflow?

Stepping back, the next set of questions we should ask is: where will humans fit in, and what is in the G that makes human involvement especially valuable? In my next post, I will explore AI as a player in the labor market. A recent article in the Wall Street Journal highlights that Moderna has decided to merge its technology and HR departments. At a broader level, this suggests that some firms already see humans and technology, particularly AI, as substitutes. They are actively working to determine the optimal mix between the two to maximize performance.

Join me next week as I delve deeper into considering benchmarks for both AI and humans, and what they might signify for the future of work as companies opt for a blend of human and AI contributions in their decision-making.

What are some Patchwork AI solutions you’ve developed for your workflow?

In fact, for many organizations, there is a gulf between “strategy” and “execution.” Strategy is “in the clouds.” Execution is “in the weeds.” Over the years, I’ve found that having conversations that bridge this gulf (“Strategic Execution”) is quite challenging because there is no natural language to connect big ideas to the work-to-be-done.

Anyway, the class is still a work in progress, but in designing it, I’ve talked to (and borrowed material) from colleagues at other schools who’ve taught the class (Thanks, Claudine!). I’ve also read nearly two dozen books, over 100 academic papers, countless news articles, and even a few Harvard case studies.

But one of the guilty pleasures of prepping this course has been watching lots (more than I’d like to admit to my children!) of YouTube videos.

What I Learned From MrBeast

And if you watch YouTube, you’ve probably heard of Mr. Beast, the 26-year-old YouTuber with the most followers—nearly 391 million. For the Indians reading this, the #2 channel is T-Series (with 293 million followers). T-Series is a major Indian music label with over 23k videos posted, billions of dollars in production costs, and a storied history. So when you compare T-Series’ #2 rank to Mr. Beast’s #1 rank (with 868 videos1 ), it’s EVEN more impressive.

If you’re anything like me, you’re probably wondering: how did MrBeast do it?

Luckily, YouTube recently recommended a phenomenal interview by Shaan Puri (a Duke Alum) and Sam Parr with James Donaldson (MrBeast). The interview is super well-done, and I highly recommend watching it.

As I listened to the interview, it was clear that MrBeast has thought rigorously about his economic engine and is comfortable in this liminal space between strategy and execution. He has a sharp and conceptually deep language for what he does.

The video is structured around a list (“MrBeast’s 10 Business Secrets”) that could easily serve as a syllabus for a class on strategy execution. In fact, it resembles much of the syllabus I have iterated toward over the past two years, with some points that I might include in the future after watching this video. Unfortunately, I don’t use Mr. Beast as a case study, but that will change (we discuss GM, Arizona State University, Apple, Uber, Kodak, etc.).

Before I dive into this list (and link it to some of the ideas in the strategy and economics literature), I want to show you a short clip from MrBeast’s channel from a decade ago—when he was 16 years old. On April 1, 2015. It’s titled “Making YouTube a Priority.” At this point, he had about 2000 subscribers and had been doing YouTube on-and-off for about two years.

A few quotes from this video are striking.

• “I’m nothing as compared to the top tier YouTubers…”

• “I’ve never fully committed…”

• “I was thinking, I don’t want to be in the future, you know, like 10 years from now, and I don’t want to wonder how big can my YouTube channel have gotten if I had stuck with it…”

Its been almost exactly a decade: 2000 subscribers to 391,000,000.

MrBeast set an audacious goal, fully committed to it, and systematically built an execution engine to turn vision into reality.

Setting Ambitious Goals

Overcome Myopia to set a clear vision. As I watch this video, I have to put myself in the mindset of a 16-year-old boy in high school. That was me at one point in my life, in New Jersey. My view of the world was incredibly myopic. Maybe school, maybe friends—maybe a few other small things—were in my field of vision.

Now, here’s a 16-year-old boy in North Carolina who’s overcome significant myopia that comes from a lack of experience and that particular stage of life. He looked out at the world and saw what was possible: That there are people called top-tier YouTubers out there. And he decided: I have 2,000 subscribers now, but I could be one of the best, if not the best.

I think this is an underappreciated aspect of strategy execution. We need a vision to execute on, but we often set goals within the narrow confines of what we can readily observe in field of view. Bold visions require us to first overcome that myopia—as Levinthal and March have so elegantly written about their famous article “The Myopia of Learning.”

And it’s only with that broader frame that we can set truly ambitious goals. What’s out there? What is possible? Most of us don’t really know what’s possible—for ourselves, or for our companies.

Burn the Boats

Between commitment and flexibility, choose commitment. There is a strategic tension between making credible commitments (see Ghemawat’s Commitment: The Dynamics of Strategy), meaning committing to a course of action such as a vision, and developing the capabilities needed to execute on that vision. Strategy defines commitments as decisions that are difficult or costly to reverse (see also Van den Steen).

Commitments are fundamental because they force you to burn the boats, leaving no other choice but to pursue a particular path with a specific theory of how to succeed. The research on commitment as a strategic mechanism is well established, and commitments can lead to sustained competitive advantage for at least four reasons (see Pisano, 2022 for a review of Ghemawat’s ideas):

1. They create barriers to entry by making it harder for others to imitate or compete.

2. They signal to everyone, both internally and externally, that you are serious about this direction and willing to do what it takes to win.

3. They lock you into a specific path and force focus, rather than allowing constant shifts in strategy.

4. They push you to innovate and build capabilities around the commitments you have already made, which leads to powerful complementarities.

Of course, there are trade-offs. You might commit to the wrong direction. You might miss new opportunities. Commitment can also lead to inertia, making it difficult to adapt as the world changes.

But the key insight is this: while you might lose by committing, you may never win without it.

The Rule of 100

Incremental Improvement and Learning Curves. Many people believe that if they discover some “one weird trick” or a special aha moment where everything suddenly falls into place, they will unlock success. But they are fundamentally mistaken about the nature of production.

While it is true that 80 percent of what drives quality comes from a few big factors, those are usually obvious and easy for others to replicate. What drives the remaining 20 percent of performance comes from lots of little things, not one big breakthrough. And those little things are learned by doing.

The knowledge created through both experience and strategic investment in process innovation can lead to long-term competitive advantages. These small, subtle improvements are often invisible to outsiders because each one is minor on its own (see Hendel and Spiegel, 2014; Cohen and Klepper, 2000; Argote and Epple, 1990, and a LOT more) . But when there are a hundred small improvements working together, the total effect is powerful and very hard to copy. You have to do all of them to make it work.

My own research on A/B testing shows that firms that rigorously test even the smallest ideas consistently perform significantly better over time, increasing their page visits by over 100 percent in just 12 months. Incremental, compounding improvement can be a powerful source of long-term competitive advantage.

This is exactly why MrBeast’s idea: You should post 100 YouTube videos, each one with the goal of being slightly better than the last on a clear and measurable (even if small) dimension—can lead to massive differences in outcomes over time.

The Power of Ideas

Ideas and effort are superadditive. I think in the entrepreneurship literature, there is considerable debate over what matters more: ideas or founders. There is strong reason to believe that both play an important role in the success of a venture. Great founders who pursue weak ideas waste their time. Great ideas in the hands of mediocre people fail to go anywhere. So, it is clear that both elements are necessary. As a result, there are strong complementarities between ideas and people. In many cases, great individuals are better at selecting promising ideas and tend to find and match with one another.

In the context of the MrBeast interview, what stands out is his exceptional ability to generate ideas that go viral on YouTube. This capability has developed over many years and draws on a deep foundation of tacit knowledge that others typically lack. He can hear a simple word like “grandmother” and quickly imagine a concept that feels obviously engaging and likely to hold viewers’ attention.

This connects to a more foundational insight. If we hold constant the capabilities of a person or team, it becomes clear that success functions as a product, not a sum—X1 times X2, rather than X1 plus X2. When a team has strong capabilities, adding a high-quality idea greatly amplifies the result. For this reason, investing the time to reflect, think critically, and develop the judgment to recognize when an idea is worth pursuing becomes essential.

Although MrBeast’s process may appear intuitive, it likely relies on a schema he has developed over time. He seems to carry a structured, although implicit, causal model of what kinds of content will perform well. Moreover, it is not the average idea that drives performance. It’s the best idea. And the ability to identify the best idea is a central part of success.

Reaching this level of capability takes time, repeated effort, and focused development of the skills required to execute effectively. These skills involve subtle forms of knowledge and judgment and are built through experience. Once a team has developed these capabilities and committed fully to them, the selection of a truly exceptional idea can lead to dramatic improvements in performance.

Clone Yourself

Delegate. Delegate. Delegate. There is a new body of research suggesting that many people either avoid delegating or fail to delegate effectively. One excellent book that shaped my thinking on this topic is The E-Myth Revisited. It argues that one of the biggest mistakes entrepreneurs make is assuming that being an entrepreneur is the same as being a craftsperson. Someone who is skilled at making cupcakes, for example, may not have the skills needed to run a successful cupcake business. Running a cupcake business means designing systems that allow other people to make cupcakes without relying on the original baker.

The ability to delegate effectively, or in MrBeast’s words, to clone yourself, requires a deep and nuanced understanding of your own work. You need to be clear about what you actually do and what information (and skills) someone else would need in order to do it the same way. This matters because the goal is for others to see the world as you do and to carry out tasks in the same manner. Many people lack this level of self-awareness.

McDonald’s provides a useful example. It functions as a massive delegation system where each franchise owner follows a standardized formula.

In some situations, you delegate to people who reach your level of skill very quickly and then go beyond it. That is when real progress happens. When your clones can create more clones, scale becomes possible.

To reach this point, it helps to reflect not only on the outcomes you seek but also on the actual steps you take to achieve them. Once you understand your own process, you can create an instruction manual for others to follow. If others cannot replicate your work entirely, then identify the key decision points where your unique insight, judgment, or tacit knowledge is essential. By placing yourself where you add the most value, the work can continue without depending entirely on your labor.

Impossible is Possible

Setting and executing on stretch goals. There is a large body of research on goal setting in management. I am not an expert in this field, but the evidence is clear: goals matter. They focus attention, drive effort, and encourage creativity in pursuit of achievement. Sometimes this happens at the cost of ethical behavior, but despite that, goals remain critical.

What about impossible goals? My colleague Sim Sitkin has written extensively on stretch goals, which are defined as “seemingly impossible goals.” His research shows that although stretch goals can be powerful, they are often adopted by the firms least equipped to achieve them. These are organizations that lack the necessary capabilities. On the other hand, firms that do possess strong capabilities often choose not to set such ambitious goals. My colleagues Wes Cohen, William Miles and Yoko Shibuya and their collaborator Matthew Higgins have shown that large and successful firms frequently default to setting incremental goals, and in many cases, there may be economic incentives that encourage this behavior.

Even so, the value of stretch goals points to a deeper insight about the relationship between effort and outcomes. Incremental goals often require proportional increases in effort. For example, if you want to increase sales by 20 percent, you may need to work 20 percent harder. However, if your goal is to increase sales tenfold, the same kind of effort will not be enough. You would need to rethink your entire approach to sales, e.g., your full production function.

This is the critical difference. Stretch goals force you to design entirely new systems. Regular goals push you to apply more effort to your existing systems.

Consultants are a Cheat Code

Leverage the market for expertise. Cloning yourself allows other people to benefit from your hard-earned knowledge and expertise because it becomes codified and transferable. In the same way, experts help solve problems that you face but do not have the skills or experience to address directly.

Expertise that comes from years or decades of focused effort, improvement, and lessons learned through experience creates a high barrier. It becomes very difficult for someone without that foundation to step in and handle the same challenges effectively. When a problem requires a solution now, and there is not enough time to build the necessary capabilities internally, the most practical and economical approach is to borrow outside expertise.

The cost of hiring an expert is often lower than the combined costs of delay, failure, and the effort it would take to build that knowledge internally. The expert's marginal cost of solving the problem is likely far lower than both the fixed and marginal costs involved in trying to reach that same level of proficiency yourself.

Although I have not read extensively on the academic research surrounding consultants, one strong example comes from a paper by Nick Bloom and his colleagues. They studied textile manufacturers in India that hired consultants from firms such as Accenture and McKinsey. The results showed clear improvements in performance, and those improvements continued well after the engagement ended. This supports the idea that well-matched external expertise can create lasting value.

Block Out the Noise

Surround yourself with the right people. Surround yourself with the right people. There is a large body of research on peer effects, and I have written several papers on this topic. When people are in the presence of others who are highly intelligent, deeply engaged, enthusiastic, or simply easy to work with, and when strong complementarities have developed between them, remarkable things can happen.

One of the most interesting findings in this area is that once people discover the right collaborators, they tend to stick together. They move from one company to another as a group. They continue to work on long-term projects together. It is not just that they have found the most talented individuals. More importantly, they have found the right people for their specific goals. Over time, they develop shared understanding, common capabilities, and the kind of alignment that allows them to collaborate with high efficiency.

Reinvest Everything

Scale in a way that no one can do what you do. Perhaps the final point is that if the system starts working, reinvest the proceeds. I am not a financial economist, so please do not take my advice on this one, but this seems logical. There is something that MrBeast can do because he is going to give a million dollars away—something the next YouTuber, who decides to buy two Ferraris with his winnings, will not be able to do. Strategic reinvestment is clearly important. Once you have built a phenomenal system with great people who know what they are doing, reinvesting within that system allows you to scale, potentially in a nonlinear way.

As a result, even if others copy every single element listed above, it would still be impossible to replicate what MrBeast is doing, because the scale at which he operates creates a fundamental gap. You might be doing everything right, but while you are giving away $1,000, MrBeast is giving away $1 million. At that scale, true competition becomes unreachable. Even if someone matches every step, the outcomes are no longer comparable.

Build the System

If there’s one key takeaway from teaching Strategy Execution and studying MrBeast’s success, it’s this: Build the system. Strategy execution is systems building. You have to design a machine where outcomes scale non-linearly with effort. The essential elements of this system? A bold vision, unwavering commitment, continuous improvement, ambitious goals, high-performance teams you can delegate to, and reinvestment to fuel scale.

With that, I’ll leave you with a final YouTube video from Jerry Seinfeld.

PS: Thank you to my daughter for generating the thumbnail for this post. :)

Starting Small and Hitting Walls

In 2007, I collaborated with a friend and classmate from Carnegie Mellon, Surendrakumar Bagde, on a field experiment in India. We were studying knowledge spillovers and network formation at an Indian college. It was a bit serendipitous that we got to run the experiment (ask me about it sometime). I didn’t know many people doing field experiments then1—it wasn’t common in the graduate program I was part of. And when I started my first job, field experiments were not on the radar in organization theory or strategy.

For some reason, my early studies were accepted fairly easily (one in the American Sociological Review and the other in Management Science). But when I shifted toward more traditional organizational/strategy topics like entrepreneurship, publishing became a real beast. My work didn’t seem to fit and kept getting rejected. I could wax philosophical about why the papers were rejected. Looking back, I’m sure it was a mix of things—imperfect research designs, skeptical reviewers, and my own blind spots around what excited people in that part of the field.

Anyway, there was a gap of nearly four years between when the experiments with Surendra were published (in 2015) and when my next set of experimental studies finally came out (in 2019). That stretch was rough. Can’t lie. Constant rejection wears on you. You begin to wonder whether you’re cut out for this business. But great guidance from thoughtful editors and support from my colleagues at Duke really made the difference.

Finding Each Other

By 2019, I had started meeting more researchers who also seemed to be banging their heads against the wall trying to get their experimental work published. I kept hearing the same kinds of stories—reviewers who didn’t understand why randomized experiments allow for causal inference (“they kept asking for more control variables!”), struggles with implementation and reviewer skepticism about whether we learned anything from the experiment.

At the time, I was also an Associate Editor at Management Science, so I saw firsthand how difficult it was to find reviewers who understood both experiments and the substantive topic well enough to guide a paper toward successful publication. It was also clear that experimental research in strategy was still in its early stages, and without support for people in our field who were experimenting with experiments, the incentives might soon shift away from doing them at all.

It was clear that it was a “system” problem rather than just the failings of individual researchers or bad reviewers. We hadn’t yet built the collective capacity to experiment in ways that would help us learn as much as we could.

That is, individual researchers were struggling, but our field also lacked a system that made field experiments viable: incentives, reviewers, resources, training, and developmental publication venues.

Something needed to be built.

Collective Experimentation

It is worth stepping back at this point and asking: Why should we focus on improving collective experimentation in the field rather than having people just read Gerber & Green?

While there are clearly individual incentives to get one’s experiment published, I think the collective goal of our efforts is something entirely different. It is to drive better decision-making—at key leverage points in business strategy, entrepreneurship, and public policy— through rigorous evidence so that our societies are collectively more prosperous. It is both about how the ideas are tested, but also what ideas are tested.

In addition to my experimental research, I’ve been working with colleagues to think about the “collective” part of experimentation more recently.

Experiments versus Experimentation

The first study that got me seriously thinking about collective experimentation was with Rem Koning and Ronnie Chatterji (Experimentation and Startup Performance). We tracked thousands of high-growth startups in the paper and found those who adopted formal experimentation (e.g., A/B testing) [Note: An experiment is a one-off empirical exercise. Experimentation is the collective ability of the organization to run many experiments to drive strategy towards a particular vision.] Individual experiments become vehicles for directional change when embedded into a collective effort and strategic process.

But the key word here is collective. A/B testing only made a difference when it was integrated into how firms made strategic choices—when it helped the organization explore many bold ideas, not just test one-off theories. If we want field experiments in innovation and strategy to matter, we need to treat experiments not just as evaluations of interventions but as explorations of what is possible a la March (1991). That is: They shape what people are willing to try next. And if we can surface the right ideas, then we can shift the direction of collective learning.

Managing Risk versus Progress

In a separate collaboration with Todd Hall (Organizational Decision Making and The Returns to Experimentation), we modeled what happens inside organizations when experimentation meets decision-making. Organizations that either centralize too much or decentralize too loosely lose the benefits of collective experimentation. Either they slow down innovation by over-controlling which experiments get implemented, or they allow too much noise by letting every random finding get committed into policy.

The sweet spot is a structure that decentralizes the generation and running of experiments—but maintains some consistency in what counts as sufficient evidence to scale. This matters a lot when we consider how collective experimentation's value scales. Communities should provide consistent thresholds and norms about recognizing when experimental evidence counts as credible.

Driving Diversity In Hypotheses, Not Just Maximizing Significance

In more recent work with Ashish Arora and William Miles (If You Had One Shot: Scale and Herding in Innovation Experiments, NBER Working Paper 33682), we explored how the distribution of experiments across firms affects experimental diversity (e.g., which hypotheses and approaches are tried) and eventually, which problems get solved. Our central finding was counterintuitive: when many firms each run just one-off experiments, they tend to herd. Everyone bets on the same “most promising” idea. What would you do if you had one shot?

But this behavior creates systemic risk. If the most popular approach turns out to be a dead end, the entire market of experiments fails together.

Organizations that can run multiple experiments face a different calculus. They have an incentive to diversify because they only need one success to generate value. Spreading their bets across different approaches lowers the risk of total failure. By internalizing the cost of correlated failure, multi-experiment firms naturally hedge.

This finding, I think, has implications for how we structure research communities. If we want to make progress on hard, high-uncertainty problems, we need mechanisms that encourage approach-level diversity. That might mean funding multiple early-stage pilots instead of doubling down on the most promising idea. It might mean supporting researchers over the long term who can try multiple paths. It means building a community that values novelty and risk.

Yes, rigor matters, but so does trying new and challenging things.

Building Capacity

Let’s start building.

Meeting others who had gone through similar struggles was incredibly eye-opening. Solamen miseris socios habuisse doloris: it is a comfort to the unfortunate to have had companions in woe.

I guess a community begins when isolated people realize they’re not alone. Karim Lakhani at Harvard Business School generously covered food through his LISH lab, and Vish Krishnan at UC San Diego was incredibly supportive. Duke I&E provided the space, and my personal research budget covered breakfast, a Costco run for drinks and snacks, and a lot of elbow grease. We also had a fantastic project manager, Camille, who made the logistics a breeze.

We organized the first workshop—Workshop on Field Experiments in Strategy, Innovation, and Entrepreneurship (long name, we would change it later)—at Duke University in January 2020. We invited researchers from around the world who we knew were doing experiments. We were hoping for 30 or 40 attendees. We got 60 attendees over the two days. Not bad for a new conference built around a method that wasn’t widely used in the field.

The people who came included both early-career and more senior researchers (many still participate in CFXS today—thanks, Tom and Jana!). That first year, we also had a phenomenal panel with my Duke colleague Wes Cohen, Alfonso Gambardella from Bocconi, and Albert Bravo-Biosca, a friend of Karim’s from an organization I hadn’t heard of before—IGL.

Then COVID hit. The world shut down.

Who knew whether we could ever run the workshop ever again?

Fortunately, two young scholars, Rem Koning at Harvard Business School and Hyunjin Kim at INSEAD—both brilliant, fun, creative, and incredibly hard-working—stepped in to keep the conference going. We had initially thought about organizing the 2021 conference in person, but the COVID lockdowns and travel restrictions made that unworkable. Moving it online turned out to be a revelation.

In my view, the second conference (yes, the online one) was at least two to three times better than the first. Over 80 people attended, from all around the world. The conversations were incredible. You could feel a real hunger—not just for the ideas, but for the kind of community we were building. We were intentional about setting the culture from the start.

Our community would be intellectual, but fun. Rigorous, but kind. Open to diverse ideas, but with a strong emphasis on integrative discussion.

Most of our discussants were senior scholars, and they were asked to read and synthesize two, sometimes three, papers at a higher level of abstraction, drawing on their experience to help connect ideas across presentations. That takes a lot of work, and you could tell the discussants were really thinking hard about the papers.

After that conference, Albert Bravo-Biosca asked if we’d be interested in co-hosting the next one in London with the Innovation Growth Lab.

The IGL team—Albert, Yan Yan, and Edoardo—put together an incredible venue at the NESTA offices. That year, at CFXS3, we introduced a few new features to the conference. One was the idea pitch sessions—no full experiments, just early-stage concepts. We realized that a key problem we needed to solve for was getting feedback to people before they ran their experiment. This feedback could help save time, money, and headaches. Edoardo also brought in Slido, allowing participants to give live feedback in real-time. It was crowdsourced feedback in five minutes, simple, fast, and incredibly useful. We wanted to move beyond giving feedback to experiments that had already been run but, rather, to help researchers improve the experiments still incubating.

We also launched a PhD workshop on the morning of the first day, with nearly 30 students in attendance. Most PhD programs in strategy still don’t teach experimental design. If we wanted better collective experimentation, we would need to start with PhD students. To guide those sessions, we realized we needed a shared language for designing strategy experiments. Taking inspiration from the Stanford d.school, we created our own Strategy Experiment Canvas.

CFXS4 was held Cambridge, MA with much the same format as CFXS3. But this time, we had over 100 people attend over the two days of the conference. Clearly, we had hit on something.

If you asked attendees whether CFXS was a “methods” conference, the answer was a resounding no. A field experiment was not just a technique; it was a way of thinking about the world: How do we find leverage points to help change the world in a positive direction?

Scaling Impact with Philanthropic Support

Building on the success of CFXS4, we decided to formally connect IGL and CFXS’s efforts in organizing experimental capacity building. We applied for and were awarded a Sloan Foundation grant last year to help build out this community even further. The grant would support a range of initiatives, including:

• Build a vibrant and diverse research community by integrating the IGL and CFXS networks and expanding access to training, funding, and collaboration opportunities. We also provide travel support for PhD students attending programs that can not support travel to conferences. [join the research network] [check out research network members]

• Foster learning, feedback, and knowledge sharing through conferences (CFXS), webinars, PhD workshops, masterclasses, and mentorship programs.

• Provide seed funding for early-stage experimental ideas that lack traditional sources of support. [learn about our seed funds]

• Reward and recognize outstanding research with annual prize awards to highlight impactful and innovative experimental work. [awards to be announced soon]

• Create and disseminate research resources—such as design toolkits, data collection templates, outcome measurement instruments, and matchmaking services with implementing partners. [coming soon!]

And more long-term goals, including:

• Strengthen the link between research and practice by enabling partnerships with government agencies and organizations interested in experimentation.

• Advance evidence-based policy and business practices by increasing experimental research quality, quantity, and visibility in the U.S. and globally.

With Sloan Foundation support, CFXS5 was held at INSEAD in Fontainebleau, France, in December 2024.2 In 2024, our team got its newest member, Anna Giulia Ponchia. Check out her write-up about CFXS5.

Just last week, we wrapped up a six-week master class on field experiments. The speakers included Liz Lyons (UC San Diego), Jana Gallus (UCLA), Chiara Spina (INSEAD), Thomas Astebro (HEC Paris), George Ward (Oxford), Nicholas Otis (Berkeley), Jackie Lane (Harvard), and Florian Englmaier (LMU Munich). The topics the masterclasses covered ranged from experimental failure to using GenAI in experiments:

• Setting Up an RCT at Scale – Ensuring robustness in RCTs, including strategies for negotiation, navigating set-up challenges, and learning from near-failure experiences.

• The Importance of Power in Experiments – Addressing low-powered studies and strategies to increase sample size, such as within-subject design approaches.

• Team Science & Organizational Economics– Exploring collaborative research within organisations and teams

• Scientific Approach – Applying scientific methodologies to entrepreneurship.

• AI & Experiments – Running online experiments using AI with entrepreneurs in developed and developing economies.

• What Happens When an Experiment is Over – Insights into refining experiments through pre- and post-mortem analysis.

• How to make your experiments more rigorous – Getting feedback and knowing what to focus on.

If you want access to the videos, please sign up here.

Nearly 400 people from around the world registered for the masterclass. Around 270 participated in the asynchronous sessions; an average of 80 joined the live classes across the six sessions. I’m lucky to get five students in a PhD class; 80 is a treat.

If there’s one thing we’ve learned in slowly building CFXS, it’s that the capacity to experiment—and to learn—depends on more than good ideas or individual effort. It requires building shared infrastructure and public goods: tools, training, funding, and feedback loops that make high-quality field experiments possible at scale.

The support from the Sloan Foundation and collaboration with IGL have made it possible for us to scale our efforts in a way we could never have imagined. But most of all—what keeps the flywheel going is the community of people who have continued to share and engage year after year. Thank you!

With PhD education under pressure and institutional support becoming less certain, we need to think more carefully about how we continue to develop the next generation of scholars. If we want strategy and policy—and social science research in general—to be more evidence-based, we must collectively invest in institutions that make that kind of learning possible. This is a first-order mission that should drive our efforts.

We’re excited about where the IGL/CFSX collaboration will take us next. San Francisco, actually. :)

Keep an eye out for the announcement for CFXS6.

See you in San Francisco.

"Gatsby believed in the green light, the orgastic future that year by year recedes before us. It eluded us then, but that's no matter—tomorrow we will run faster, stretch out our arms farther... And one fine morning—So we beat on, boats against the current, borne back ceaselessly into the past." - F. Scott Fitzgerald, The Great Gatsby (1925)

Several weeks ago, my daughter wanted to watch The Great Gatsby (she eventually read the book too). I hadn’t thought about Gatsby since at least high school, and even then, I didn’t fully understand it. In my mind, it played like a Bollywood movie: boy loves girl, girl loves boy, social class keeps them apart… Even the fact that Amitabh Bachchan was in the remake we watched aligned with my notion.

But now I see it: Gatsby was a mirror.

His story reflected his society. Gatsby strove to transcend his humble beginnings through relentless reinvention, through his dress, his speech, his affectations, and his wealth. It’s as if he drank the Power Pose Kool-Aid.

But in the end, the real power belonged to Tom Buchanan, whose dominant position in the social structure made him secure. Gatsby, in contrast, could never grasp the green light. Real power was forever out of his reach.

And what was this social structure that gave Buchanan so much power?

Buchanan’s power felt effortless and absolute because of a web of institutions, connections, social capital, class, wealth, and cultural legitimacy.

Gatsby was a mere status hack. In the end, the social structure caught up with him. He was shot dead in his pool, blamed for a crime he didn’t commit, while the Buchanans walked away untouched.

I once asked my advisor, Shelby Stewman, what the phrase “social structure” actually meant.

We were crossing Forbes Avenue in Pittsburgh after class. He looked down the street and said, “You don’t see a truck here, do you? Then, suddenly, it comes outta nowhere—at full speed—bam, it hits you. And you don’t even get to ask what the **** just happened. That’s social structure.”

Gatsby never saw the truck coming.

But social structures don’t just magically appear. They are built, shaped, and reinforced by larger forces, and few forces shaped the world more profoundly than the rise of global trade.

Trade and the Global Economy

If the Age of Exploration has a starting point, it is often traced to 1415, with the Portuguese conquest of Ceuta. Almost immediately after, many European nations began expanding sea routes and connecting to different parts of the world.

By 1492, Columbus arrived in the Americas. Then a few years later, in 1498, Vasco da Gama reached India. Spices, textiles, and humans, both free and enslaved, moved across continents.

Wherever there was demand, new ventures arose to create supply.

This expansion in trade also led to important institutional changes. New organizational forms emerged, most notably the rise of joint-stock companies that allowed strangers to pool capital. New institutional arrangements, such as limited liability, were also introduced, enabling people to invest in risky ventures without risking total personal loss. Alongside these came the legal and financial infrastructure of capitalism: property rights, contracts, corporate charters, and formalization of double entry accounting in 1494 and its eventual use by companies to grow into sprawling and complex organizations (a great book on the rise of accounting is Jacob Soll’s book, the Reckoning). With these innovations and increasingly better knowledge of trade routes and navigation, capital could now create and capture value across an expanding global trade network.

By the 1800s, the basic institutional skeleton of modern capitalism was in place. Global markets, financial infrastructure, and institutions protecting capital had emerged—alongside new social structures.

New classes, as writers including Marx noted, were increasingly self-aware of their growing wealth and influence. They sought to protect and preserve their position in this emerging global order.

However, while trade had dramatically reshaped the world by the mid-1800s, the most profound transformation still lay ahead: a massive expansion of the technological frontier.

Technological Change

Major technological changes began to reshape the Western world, starting in the late 1700s and accelerating in the mid-1800s. The rise of global markets created demand that needed to be met, which led to the invention of new ways to serve that demand.

Broadly speaking, five significant technological changes supercharged global trade and capitalism.

Energy

Perhaps the most important of these technological innovations was the steam engine. While the use of steam as a source of energy had been known for centuries, it was not turned into a practical device until the late 1600s. It was James Watt’s design, developed between 1765 and 1776, that led to the creation of the modern steam engine. This is widely considered the spark that ignited the Industrial Revolution.

Energy, at a more philosophical level, is all there is. Much of social and economic existence is about energy transfer from one form to another. The steam engine transformed that transfer. It no longer required the kind of human energy that had powered centuries of labor, from building pyramids to weaving cloth.

But the steam engine was only the first step. By the 1870s, with the invention of the dynamo, which could provide smooth and continuous electric power, inventors like Edison and Tesla were actively experimenting with electricity.

Rapid improvements in generating, transmitting, and distributing electric power allowed for broader applications across the economy. Electricity began to be widely produced for industrial use.

Production

Given this new source of energy, new devices were created such as the power loom, the spinning jenny, and the cotton gin. These innovations drastically improved the productivity of the textile industry and reduced the labor needed per unit of output. This technological change dramatically reduced the cost of goods. Combined with growing markets, both domestic and global, it led to greater demand for both textiles and the raw materials needed to produce them, such as cotton.

In addition to the mechanization of production in sectors like textiles, another major innovation was the invention of the Bessemer Process for steel production. This allowed steel to be produced much more inexpensively. Before this process, steel was rare and mostly crafted by skilled artisans. With the Bessemer Process, it could now be produced cheaply and at scale.

I would add a third aspect of production technology. A few years ago, I read Simon Winchester’s The Perfectionists, which traces the rise of precision engineering. By the mid-1800s, the needs of manufacturing—repairability, scale, and precision—led to the creation of machine tools as well as precision, interchangeable parts.

Transportation

Now that you have supply and demand, you need to get them to meet. These technological innovations: steam, steel, and mechanization also enabled a rapid increase in transportation technologies. Railroads allowed countries to integrate their national markets, creating industries that could transport goods efficiently and enabling production to be located wherever it was most efficient, then supplied to wherever it was needed. Steamships played a similar role for domestic trade across rivers, and by the late 1800s, significant transoceanic trade had emerged, moving large volumes of people and goods across continents.

In 1876, the internal combustion engine was patented and later commercialized by the late 1880s. This led to the eventual rise of automobiles (including trucks) and road systems, that could move goods and people to places even railroads could not reach.

Information and Communication

The rise of the capitalist enterprise also led to the creation of a new type of good: information. Like physical goods, this information needed to travel. A series of important innovations made that possible. The telegraph, invented in the 1830s and commercialized in 1844, allowed for real-time communication over long distances. This was crucial for real-time pricing of goods, the trading of equities and commodities, and the centralized coordination of complex and sprawling enterprises distributed not only across the country but around the world. The telephone, patented in 1876, made communication even easier, though it was primarily a one-to-one technology.

Radio transformed communication once again. It enabled one-to-many communication, allowing politicians to reach their constituents, businesses to advertise and grow demand, and the creation of national markets for both attention and culture.

During this period, businesses expanded, which required more systematic recordkeeping. It was no longer enough to jot notes in ledgers or rely on unclear handwriting. The invention of the typewriter in the 1870s led to the production of even more information, which now needed to be sorted, sifted, and analyzed—and mailed through the growing postal service.

Organization and Management

Finally, beyond energy, production machines, transportation, and communication technologies, these ever-growing enterprises needed to manage their affairs more rigorously and efficiently as competition intensified. While significant changes to management and organizational practices occurred during this period, perhaps the most influential was Taylorism. This approach allowed for the systematization of work on the factory floor by breaking tasks down into small, incremental components and then reassembling them in the most efficient way possible.

Taylorism had the effect of de-skilling labor, but it also dramatically increased the demand for low-skilled workers who could “plug in” to modern factory systems.

Taylorism, combined with more rigorous statistical process control as well as advances in accounting, led to the rise of larger, more efficient organizations that could scale to sizes that would have been unimaginable just a few decades earlier. For instance, by 1929, just before the Great Depression, Ford had 174,126 employees, according to official records. By comparison, the estimated size of the Continental Army under George Washington that helped the US gain independence was a mere 80,000.

In less than 150 years, the power of the factory floor rivaled the power of armies. Atop these armies of workers was a new class of men made in the shuffle: Vanderbilt (Railroads), Carnegie (Steel), Rockafeller (Oil), Ford (Automobiles), Stanford (Railroads), and Morgan (Money). They not only held vast economic power but political, cultural, and social power as well.

How Trade * Tech Changed The World

In 1875, most people lived in rural areas and worked in agriculture or small-scale trades. The economy was local, labor was manual and seasonal, and social life was organized around community and family. Businesses were small. Government had a limited role in economic life. Inequality existed but was less extreme. Social roles were more fixed, and there was less mobility.

By 1925, the world was different. Industrialization had reallocated people into cities and into factories. Large corporations replaced small businesses. Railroads, telegraphs, and typewriters created a faster and more connected and information rich economy. Immigration surged due to the demand for labor. Inequality widened, and in turn labor organized. The structure of everyday life looked very different. The assumptions people had about the social order no longer held.

Perhaps most importantly, once people rose to prominence in this new world, they did everything they could to calcify the structure they now controlled. This stark new order became the dominant social structure, replacing first feudal society, and later the simpler early bourgeois society of Marx.

Like all social structures before it, this new structure was hardened into place, self-became reinforcing, and functioned to preserve the advantages of those already at the top.

Many scholars began to ask: What the **** just hit us?

The Rise of Structural Sociology

The sociological imagination that emerged from this great transformation was a bit more systematic. It came in two flavors: one descriptive, one causal.

On the descriptive side, scholars began asking: What is this new world? What are its core parts? How are they connected?

On the causal side, the questions were harder to answer. How did this society come to be what it is now? Why does it work the way it does? And who benefits from the way things are set up? To answer these questions,

They studied institutions, roles, and hierarchies to understand why some groups had more power than others. They examined how bureaucracies shaped work and how organizations dominated social life. They asked how identity was formed and whether people had real agency inside tightly structured systems. And they explored how elites maintained power, how inequality was reproduced, and what caused institutions to hold—or break.

What Structural Sociology Teaches Us

The most powerful idea in structural sociology, at least to me, is that power flows through position.

One way that happens is through formal structure. Max Weber saw this in the rise of bureaucracy. In modern systems, what matters is not who you are, but where you are positioned. Authority is tied to role, not personality. Decisions follow the structure, not the individual (Economy and Society, 1922). Talcott Parsons made a similar point at a broader level. Institutions are held together by roles and rules. They don’t ask much of people beyond showing up and playing their part (The Social System, 1951).

Position is also built into how economies sort people. Peter Blau and Otis Dudley Duncan showed that outcomes are shaped by origin. Education, family, and class background all matter more than effort (The American Occupational Structure, 1967). Davis and Moore argued that inequality helps match talent to roles (Some Principles of Stratification, American Sociological Review, 1945), but the data told a different story. The system wasn’t meritocratic. It was structured to keep people where they started.

Bourdieu added that taste, language, and style all reflect early advantage. People raised in elite environments learn how to move through elite spaces (Distinction: A Social Critique of the Judgement of Taste, 1979). That comfort, the ease of knowing how the game is played, is its own kind of power.

And once positions are locked in, they often reproduce themselves. C. Wright Mills described how elites move between institutions with little friction. The game is set up for them (The Power Elite, 1956). Antonio Gramsci explained how this kind of power is maintained not just through force, but through consent. People learn to see the structure as natural (Prison Notebooks, between 1929–1935). That’s when power is at its strongest: when it doesn’t have to explain itself.

These ideas and scholars were just the tip of the iceberg, and structural thought was never confined to explaining modern industrial society. It animated anthropology: Claude Lévi-Strauss studied pre-modern kinship systems, M.N. Srinivas analyzed the caste structure and social change in India, and others like S.F. Nadel and Bronisław Malinowski extended structural ideas into diverse cultural settings. Later, structural thinkers like Robert Merton introduced concepts such as role strain, while Harrison White and his students developed tools for mapping social structures as patterns of relations.

Over time, these structural tools became so deeply embedded in sociological thinking, and so analytically powerful (Google PageRank itself has origins in structural ideas), that it almost didn’t matter what they were used to study. The tools became the focus. In some ways, we became more interested in the architecture of structure itself than in the social world it was originally meant to describe and explain.

While the structuralism of early- to mid-20th-century sociology was slowly vanishing from mainstream sociological practice, it began to reappear in an unlikely place: economics. A new generation of economists began to notice that the postwar social structure of America looked eerily like the world the structural sociologists had once set out to describe.

The Rise of 21st Century Structural Sociology (in Economics)

In today’s economy, power is concentrated once again, and invisible social structures appear to matter more than ever. A modern cadre of social scientist have given us a mirror to look at our society.

Thomas Philippon has shown that market concentration has risen sharply since the 1990s. Fewer firms control more industries, and barriers to entry are higher than they appear (The Great Reversal, 2019).

José Azar, Martin Schmalz, and Isabel Tecu have shown that even firms that appear to compete are often owned by the same institutional investors. What looks like a market often turns out to be a set of interlocking positions. Power follows ownership and capital structure, not innovation or merit.

Position still sorts people, and it is harder to move than it used to be. Raj Chetty and his collaborators have shown that intergenerational mobility in the United States has declined sharply. For most people born into the bottom half of the income distribution, the odds of moving up are now lower than at any point in the postwar period (The Fading American Dream, 2017). Where you start matters more than how hard you work.

Wealth and income are no longer just outcomes. They function as mechanisms for reproducing advantage. Economists like Thomas Piketty and Emmanuel Saez have shown that the top 1 percent has captured a growing share of wealth and income since the 1980s (Capital in the Twenty-First Century, 2013; Income Inequality in the United States, 2003). The structure reinforces itself: r > g.

Once these positions are established, they are hard to dislodge. Jacob Hacker (a political scientist) has shown how small shifts in policy and enforcement create lasting advantages for those already on top (Winner-Take-All Politics, 2010). Zoning laws, tax policy, school district boundaries, and corporate governance rules work together to maintain existing hierarchies. In fact, I have a paper with my co-author, Anuj Kumar, on how ostensibly freely accessible school ratings favor those who already have economic power: Who Captures the Value from Organizational Ratings?: Evidence from Public Schools (in Strategy Science).

These are not isolated problems, but rather structures that define the modern economy. They shape who gets ahead, who stays in place, and who is locked out from the start.

I want to think that the world we live in, the world our children will live in, is different from Gatsby’s. But, the re-emergence of structural sociology in modern economics may suggest otherwise.

Social science remains a worthy calling. We must again recommit to the hard work of asking: What the **** just hit us?

Where does this money fit into the formula for American technological excellence?

In this post, I do a deep dive into the building of two of America’s most innovative universities and the formula for excellence that leaders like Vannevar Bush, Fred Terman, and Allen Newell applied. It’s a story of ambition, architecture, and the entrepreneurial spirit of a generation of academic leaders who led long-term scientific leadership for their institutions and the United States.

World War II and US Science Policy

Much has been written about how WWII shaped U.S. science policy, including one of my favorite books, American Prometheus, as well as other accounts of how America responded to the crisis of this most devastating war.

Many brilliant academics from universities across the country (and some from outside it) were involved in the war effort to defeat the Nazis. These included physicists involved in the Manhattan Project, engineers running the Radio Research Laboratory as part of the US’s electronic warfare, as well as statisticians and economists responsible for weapons design, military tactics, and weather forecasting.

Perhaps the most important scientific institution during this period was the OSRD, or the Office of Strategic Research and Development, founded in 1941 and headed by Vannevar Bush, an MIT professor and an advisor to some of the greatest scientists of the mid-20th century. OSRD coordinated scientific research for military purposes during the war, overseeing key projects, including radar, radio engineering, and the Manhattan Project.

Bush’s experience during the war led him to see the immense impact of science on the US war effort. But he also realized that science would be the key to post-war prosperity and that this dominance needed to be engineered.

Science, the Endless Frontier

In Science, the Endless Frontier, published in July 1945, near the war's end (the Nazis had surrendered in May), he wrote a report to President Franklin Roosevelt that laid out a vision for how the United States should organize and fund scientific research after World War II.

"New frontiers of the mind are before us, and if they are pioneered with the same vision, boldness, and drive with which we have waged this war we can create a fuller and more fruitful employment and a fuller and more fruitful life."

Franklin D. Roosevelt, November 17, 1944.

One key insight from this report was that basic scientific research is vital to national progress and deserves public funding as a strategic priority.

Where would this basic scientific research be conducted? In America’s great universities, of course. These universities would not only produce great science but also great scientists. The funding would not just support one-off projects, but also fund people and institutions over the long term. And these scientists would be free to explore any direction they choose.

“Scientific progress on a broad front results from the free play of free intellects, working on subjects of their own choice, in the manner dictated by their curiosity for exploration of the unknown.” Vannevar Bush

Bush also knew that great science takes time and money,1 and the funding should support the infrastructure (libraries, machines, buildings, students, etc.) that can complement the production of great science and its eventual application.

After the war, many scientists, engineers, and thinkers who had contributed to the war effort returned home, scattered across the country, often returning to the universities they had left behind. They came back with a new perspective on the power and importance of their craft: ambitious, entrepreneurial, and deeply aware of the opportunity science had created for both military and civilian use.

The post-war period marked a key transition as the spirit of the OSRD began to take root in civilian institutions, leading to the creation of the National Science Foundation and a broader university infrastructure that would fuel America’s scientific dominance.

Looking back, it’s clear that the extraordinary scientific achievements of World War II were made possible by the convergence of a few essential elements. Massive federal funding flowed into ambitious research. Brilliant people, including physicists, engineers, statisticians were mobilized. The wartime context granted time and institutional focus to tackle long-horizon problems without the usual constraints of normal university jobs. Most importantly, the moment offered an unparalleled opportunity: a global crisis that demanded new science and opened new frontiers.

The Entrepreneurial American University

I've been incredibly fortunate in my academic career to be at two great American universities, first as a graduate student at Carnegie Mellon University and later as a junior faculty member at Stanford University. Both were shaped in critical ways by World War II and the endless frontier that entrepreneurial leaders at these schools saw as their opportunity for lasting impact.

One of the things anyone who spends time at an American university will quickly realize is the fundamental role that private universities (or independent public universities) play in the U.S. education system.

This is not the case in much of the world. For instance, the leading UK universities, Oxford, Cambridge, Imperial, and the University of London, are all public. In France, it's the Grandes Écoles. In India, higher education is dominated by the Indian Institutes of Technology. In Canada, public universities like the University of Toronto and McGill are dominant.

The American university system is incredibly unique in this way (and others). For various historical reasons, the United States never developed a single dominant national university or a unified federal university system. There is no “University of America” funded by the federal government. Instead, the system is a patchwork of thousands of colleges and universities spread across the country. They began in many different ways. Some were founded in the early days of the republic by religious communities seeking to train clergy. Others started as teacher-training colleges, vocational schools, or polytechnics created by local industries to meet regional needs.

Because they lacked consistent government support, these institutions had to be self-sufficient. This constraint made them entrepreneurial, competitive, and innovative. They built relationships with alumni, turned to local business leaders for support, and constantly created new programs to bring in revenue to support their functions. As described in A Perfect Mess by David Labaree, this system grew out of necessity and evolved into the rugged, adaptive, and experimental institutions we see today.2

The scarcity of public funding forced American universities to differentiate themselves and attract students and resources in ways their international peers did not. So, when federal research funding expanded dramatically after World War II, it met an environment full of ambitious, nimble institutions ready to compete. Much of this momentum came from young academics returning from wartime research who, like Vannevar Bush, saw that science and rigorous research offered a pathway to national progress and institutional prominence.

Fred Terman at Stanford

A few years ago, I read a biography of Frederick Terman, the legendary provost of Stanford University. He was a famous engineer and scientist in his own right, and the son of Louis Terman, the psychologist who developed the Stanford-Binet IQ test. After earning his undergraduate degree at Stanford, Terman went on to MIT to complete his ScD under none other than Vannevar Bush.

He returned to Stanford in 1925 to join the engineering faculty and got off to an incredible start. During this time, he mentored students who would go on to have luminous careers, including Bill Hewlett and David Packard. He wrote a field-defining textbook on radio engineering and invested time and energy into students who would go on to found companies like Varian Associates and Hewlett-Packard.

When the United States entered World War II, Terman was tapped to lead the Radio Research Laboratory at Harvard University, where he oversaw more than 850 staff members. The lab played a critical role in helping the U.S. military develop and deploy cutting-edge radio technologies, improving Allied communication and disrupting enemy capabilities.

After the war, Terman returned to Stanford as dean of engineering. He once quipped that when he arrived, Stanford wasn't even a top-five engineering school in California, let alone in the country. But over the following decades, working alongside Wallace Sterling, Stanford’s fifth president, he helped transform the university from a strong regional institution into one of the greatest research universities in history.

Stanford, who?

When Terman first joined the faculty in the 1920s, California was barely on the map in terms of higher education, and Stanford was even less prominent. At the time, the universities that mattered were the elite East Coast institutions, the great Midwestern public universities, and a few newer research-focused models like Johns Hopkins and the University of Chicago. Stanford was nowhere near that conversation—yet.

Stanford had many problems in its early years, but perhaps the biggest was a lack of spendable capital. It just did not have the kind of money that other elite universities had. Moreover, its location far out west, away from the prestige and networks of the East Coast, did not help either. The Stanford family granted the university its land, but according to the original charter, Stanford was prohibited from selling it. It was a land-rich, cash-poor university.

But running a great university takes money. And the most important reason you need that money is to hire the best people in the world. What are universities but a collection of people committed to the power of ideas? For a university to succeed, outstanding faculty must believe they can do the best work of their careers there.

This was the key problem Stanford needed to solve.

Constructing the Stanford Flywheel

Terman was instrumental in developing three big initiatives to address this. First, drawing on his experience during World War II, he understood that federal funding for scientific research would expand dramatically after the war. However, to compete for that funding, Stanford needed to attract world-class faculty who could win grants. The challenge was convincing those people to come to Stanford. And to bring them in, the university needed to create an environment that would allow them to do their best work.

Stanford could not sell its land, but maybe it could rent it out?

In 1951, Terman helped establish the Stanford Research Park on the southern edge of campus near Page Mill Road. It was a place where Stanford faculty and students working on advanced technologies could stay close to campus and start new companies. The rent was cheap, but it kept them nearby. Companies like Varian Associates and Hewlett-Packard set up their headquarters there. Even today, HP’s world headquarters remains on Stanford land.

On the northern side of campus, near Menlo Park, the university opened the Stanford Shopping Center in 1955. Palo Alto was growing quickly, and a mall brought in both retail and reliable rental income. These two developments—the Research Park and the Shopping Center—created a steady cash flow that allowed Stanford to pursue Terman’s larger vision. (Another Terman institution formed was the Stanford Research Institute in 1946, but that is for a later post.)

With this new infusion of funding, the university could focus on what Terman called “Steeples of Excellence” — a few areas where Stanford could be world class.

These Steeples of Excellence were focal points of targeted investment where Stanford could establish unrivaled expertise and reputation. These carefully selected domains would serve as beacons attracting premier talent, substantial funding, and global recognition to the university. The hypothesis was that prioritizing strategic concentration over broad mediocrity could eventually elevate the entire institution.

Stanford went after world-class faculty in specific areas, offered them opportunities to start labs, build departments, hire ambitious colleagues, and enjoy life in sunny California. Many faculty built homes on campus in what is still called the Faculty Ghetto. Despite the name, it is no ghetto. The homes range from modest to magnificent, including the Hannah House, the first home Frank Lloyd Wright built in California.3

Terman, always thinking like an engineer, designed what you might call the Stanford flywheel. The income from the Research Park and the mall provided capital to recruit faculty, both young and experienced. Those faculty built labs and won federal research grants. The labs attracted top students. Many students stayed and started companies at the Stanford Industrial Park or joined nearby ones. This created a self-reinforcing system that kept growing stronger, what later became known as Silicon Valley.

Terman and Stanford President Wallace Sterling knew that building a strong engineering school was just the first step. For Stanford to become a truly great university, it had to take the winnings from its Steeples of Excellence and upgrade the rest of the university. The funding engine they created made that possible. It allowed Stanford to invest in its infrastructure and support all of its departments.

Today, Stanford has more academic departments ranked in the top five nationally than any other university.

Terman’s vision for Stanford aligned the key ingredients of institutional excellence. The money came from the Stanford Shopping Center and Research Park, and this primed the system for federal grants. The people this money attracted were world-class faculty recruited to California with the promise of autonomy and ambition. These people attracted other top people, federal dollars, and great students who stayed to start companies. The time came from long-term investments in labs, departments, and a campus culture prioritizing deep work. And the opportunity, the rise of new engineering fields in the wake of World War II, was one Terman recognized intimately from his own wartime experience. This is not to say that California itself was an opportunity to seize.

Stanford wasn’t the only university on the rise after World War II. A new discipline was being built in the steel town of Pittsburgh, PA.

Carnegie Mellon University and the Birth of Computer Science

Many names come to mind when people think about computer science at Carnegie Mellon. Of course, Herbert Simon stands out. He is one of the founders of artificial intelligence. This polymath won the Nobel Prize in Economics, the Turing Award, and some of the highest honors in psychology, philosophy, and other disciplines. His intellectual blueprint had a profound and lasting influence on the university.

No less transformative for the field of computer science was Alan Newell. Newell was pioneering computer scientist who sought to understand how humans think, and how we might replicate that kind of thinking in machines. His belief was simple but revolutionary: if we could deeply understand what intelligence is, we could build intelligent systems capable of doing meaningful, intelligent work.

Newell graduated from Stanford University in 1949 and began a graduate degree in mathematics at Princeton. But he quickly realized he was more interested in applied work than in pure mathematics. He left Princeton to join the RAND Corporation (I hope to write a longer discussion on the importance of RAND at some point). While at RAND, he met Herb Simon, who was already collaborating with many researchers there. That connection led Newell to pursue a PhD at what is now the Tepper School of Business, then known as the Graduate School of Industrial Administration, under Simon’s supervision. He completed his PhD in 1957 and returned to Carnegie Mellon in 1961.

Beyond being a brilliant academic and one of the founding fathers of artificial intelligence, Newell was also, like Terman, an institution builder. Carnegie Mellon has done a tremendous service to the research community by digitizing many original documents from Herbert Simon’s archives. A few years ago, I came across a 29-page document titled “THE FUTURE OF THE CMU COMPUTER SCIENCE DEPARTMENT FILE CSDF.A2 STARTED 7 FEB 74” authored by Newell in 1974 on how Carnegie Mellon could remain a leader in the emerging field of computer science. It’s a powerful read. It offers a window into the ambitious, entrepreneurial spirit of the people who helped build the post-WWII American university into the powerhouse it is today.

The full document is worth reading. It contains very little of the bullshit that constitutes modern strategy documents. It is full of hard, honest assessments of the challenges and opportunities for Carnegie Mellon.

For instance, Newell highlights Pittsburgh as a limiting factor for CMU's Computer Science aspirations. He calls it a “second rank city” (though not a third or fourth-ranked one) that lacks the scientific tradition and technological ecosystem of places like San Francisco, Boston, and Los Angeles. Unlike those premier locations, Pittsburgh was not viewed as a desirable place for scientists to live, and critically, it lacked the established “habits of doing science” that naturally permeated institutions in major centers of learning. Anyone who has lived in the San Francisco Bay Area or worked in Cambridge, MA can tell you they “feel” different.

How do you build a world-class department in an emerging technology area in a grimy industrial town? This document was Newell’s hypothesis:

However, for this post, the most relevant part to our discussion begins on Page 7 (“The Ingredients for Gaining Excellence”).

My most straightforward takeaway is his MMTO formula: Money, Men (People), Time, and Opportunity. [I’ve decided to keep the MMTO acronym, but below changed the gendered language.]

It’s clear that this was the formula that Bush understood as key to driving US excellence in science at OSRD and afterward, and Terman used it as a playbook upon his return to Stanford from World War II.

• Money: Sustained and flexible funding is essential and must be under local control to respond to changing opportunities.

• People (Men): A small group of visionary, high-caliber individuals at or near the top of their field and with the stability and ability to collaborate over time is critical.

• Time: Excellence takes time; there is no instant success. Institutions must have autonomy to capitalize on opportunities without external time pressures.

• Opportunity: An “ecological niche” or a unique moment in the field where the institution can grow without direct competition. Often tied to the emergence of a new field or concept not yet widely accepted.

Carnegie Mellon’s strategic engineering of MMTO and the flywheel that ensued under Newell, Simon, Cyert, and later under others such as Raj Reddy and his successors has allowed it to remain a dominant Computer Science program for over half a century.

The MMTO framework should be understood as a conjunctive model (strong complementarities, superadditive) rather than an additive one. Each component—Money, People, Time, and Opportunity—is necessary, and the absence of any single element renders the entire system inoperative.

Furthermore, it’s clear that additional strength in one area cannot compensate for a deficiency in another (great people, no money, means you will lose people). Even great people, money, but no time or autonomy for research, means you will not create great research because great work takes time and attention. Similarly, money, time, people, and no clear niche: intense competition.

This interdependence explains why the emergence of excellence is uncommon and challenging to engineer anywhere. What is clear from both the internal strategy document from CMU and Terman’s strategic efforts as Dean and then Provost was that Stanford and CMU’s success was not an accident. It resulted from an engineer’s mindset: Excellence is a system that needs to be consciously and strategically constructed.

However, once excellence is achieved, a different dynamic often takes hold: Excellence Breeds Excellence (EBE), according to Newell. In this phase, reputation and past success attract further resources, talent, and attention, creating a self-reinforcing cycle (a “Matthew Effect”). But even this dynamic is vulnerable. When one of the foundational elements begins to erode, e.g., through faculty departures, loss of institutional autonomy, or the closing of a strategic window, the effects may not be immediately visible. Over time, however, the system becomes increasingly fragile, and collapse can follow with little warning.

Today, as the debate over science funding heats up, we should remember Bush, Terman, Newell, and many other great leaders and consider what it actually took to build lasting excellence. It was a system carefully constructed by ambitious leaders who slowly put together the pieces for institutional excellence.

I believe in the Entrepreneurial American University, but The MMTO formula is a reminder that these factors all must come together to create excellence.

When any one of those is missing, even the strongest institutions can falter. And once you break the system, how you might put it back together is not obvious.

After the first test, he noticed something odd. One section had students who answered every question. The work was messy, sometimes right, sometimes wrong, but always complete. The other section had beautiful, perfect answers, but many questions were left completely blank.

He dug into it. The first group? Future engineers. Their instinct was to solve the problem however they could. The second? Physics majors. Also brilliant, but their internal critics wouldn’t tolerate an imperfect answer. They’d skip questions they weren’t 100% sure about.

That tension between getting to an answer and ensuring it is perfect has stuck with me. Some people lean toward creation. They try, iterate, and push forward, albeit imperfectly. Others lean toward critique. They seek precision and hold back until the answer is just right.

That tension isn't just between people. It lives within us: the drive to create versus the urge to hold back. It plays out in teams, in who takes the lead and who says wait. It’s wired into org charts, cultures, and incentives. It can shape entire political and economic systems too (think the NIMBYs vs. the YIMBY). It's the tug-of-war between the engine that generates variance and change and the breaks that protect the mean and status quo.

In this deep dive, I discuss two functions: the creative function, which generates new ideas and solutions, and the critic function, which tells the first why their ideas just won’t work.

The Architecture of Value Creation

One of my favorite recent papers on creativity is “Idea Generation and the Quality of the Best Ideas” by Karan Girotra, Christian Terwiesch, and Karl Ulrich, published in Management Science. I like it for many reasons, but most importantly, it lays out what I think of as the architecture of value creation more clearly than almost anything else I've read.

The first point they make, I think, is crucial. Most of the time, you’re not optimizing for whether your average idea is good. You’re trying to find the best idea.

The paper makes this point visually and analytically. In the figure above, they break down the “idea generation” and “selection” process into four core variables:

1. The number of ideas generated

2. The average quality of the ideas

3. The variance in the quality of those ideas

4. The ability to select the best one

It is clear from this schematic that if you only generate one idea, that idea better be amazing. It's your only shot. But if you generate dozens or hundreds, the odds go up that at least one will be great. Of course, that assumes two things:

• you’ve got people generating a large volume of ideas (and hopefully not all terrible ones), and

• you have a system that can reliably pick out the best idea(s) from the noise.

That’s the tension this framework captures so well. The generation process is all about creating variance. You need to generate lots of ideas with wide-ranging quality. Some will be duds. But that is OK. Some might be brilliant. But you won't find the brilliant ones if you don't create lots of ideas in the first place.

Then comes the selection process, the “critic” function. At this stage, judgment, evaluation, and discernment matter. Someone or something has to sift through the ideas and evaluate whether they have any value (or not.) If the critic function doesn’t work well, the ultimate idea you invest in may cost you a lot.

While this paper takes a particular approach to the “creation” and “critique” (generate ideas separately, then come back and evaluate them together), I think many potential mechanisms can serve these functions (including artificial ones). What’s also clear is that overemphasizing one function (e.g., critique) at the expense of the other isn't helpful and will undermine value creation.

The Creator Function

The creativity literature is expansive, and I won’t attempt to do it justice in this post. If you're interested in digging deeper, I recommend checking out several books on the topic, especially the work of Teresa Amabile, Mihaly Csikszentmihalyi, and two recent favorites by non-academics: David Kelley’s book on Creative Confidence and Rick Rubin’s The Creative Act.

The Girotra et al. (2010) framework offers an interesting starting point: how do we design the “Creator function” to generate a large volume of diverse ideas? Conceptually, assuming we are working within a combinatorial innovation model, the goal is to start with as many discrete components as possible and then combine them in novel and unexpected ways.

McCrae and colleagues' work on the psychological construct of openness to experience suggests that some individuals have a personality that mirrors this “creator” function. They are curious about the world, aesthetically sensitive, imaginative, intellectually engaged, and comfortable with novelty and ambiguity. For example, someone who is aesthetically sensitive might be moved by a piece of music, architecture, or design in a way that sparks new ideas (“Wow, that’s beautiful”). Intellectual engagement might manifest as a fascination with abstract theories or a tendency to chase down rabbit holes across disciplines (“Oh, that’s interesting”). Comfort with novelty and ambiguity means these individuals are less likely to run away when faced with uncertain or unfamiliar situations. They might lean in and see them as opportunities (“Hmm…that’s odd”). While these creative minds can generate endless combinations, the quality of the ideas depends on the quality of the raw material they have to work with.

In other words, these people are wired to generate variance. Their minds move across disciplines, explore ideas that do not quite fit, and remain open to ambiguity rather than avoiding it. These tendencies may not always be efficient, but they are often generative. Not every idea will be good, and that is just fine.

The creative function depends on the capacity to produce many ideas of varying quality. Only a few need to be exceptional. What matters is the willingness to keep generating without allowing an overly harsh internal critic to intervene too early or too often.

This creator function doesn’t reside only in the individual mind; it can also operate at the level of teams, organizations, regions, and entire economies. The underlying logic remains remarkably consistent. Whether we are talking about a person or a system, the basic building blocks are the same: how many discrete components are available at the outset, how often they are recombined or reimagined, and how frequently new ones are introduced. A team with diverse perspectives and experiences, for example, is more likely to come up with novel combinations than a group that draws from a single disciplinary, demographic, or cultural background. The same holds for organizations or regions that attract talent from across the world. In each case, the capacity of the function to generate a wide variety of ideas increases.

The Critic Function

Great, now you’ve got a lot of ugly ideas. And maybe, just maybe, hidden among them is a gem. Something so unexpected, so powerful, that it could create untold value for society. But how do you find that one idea worth betting on?

What you need is the Critic Function.

The Critic Function’s main job is to judge. This is good. This is not.

And when you step back and think about it, it’s actually quite striking how pervasive the Critic Function is in society. Books. Movies. Restaurants. Academic papers. Startups. Stocks. Everything has critics. There are professional critics, and there are lay critics. Everywhere you look, someone is evaluating, reviewing, scoring, or rejecting. If you are interested in the sociology of critics, a recent review of these market intermediaries by Sharkey, Kovacs, and Hsu (2023) is worth reading.

Why are critics so pervasive?

I think it stems from simple math: The cost of implementing an idea is usually much higher than the cost of generating one. Given that resources (journal pages, money, etc.) are finite, someone has to decide what moves forward and what doesn’t.

So, what is the impact of critics and critical judgment on innovation?

There is substantial evidence that critics generally disfavor novelty (Lane et al. 2021). For instance, a study by Siler, Lee, and Bero (2014) discovered that while peer reviewers can typically distinguish between low and high-quality work, unconventional studies are more likely to be rejected. They state: “Our research suggests that evaluative strategies that increase the mean quality of published science may also increase the risk of rejecting unconventional or outstanding work.” Boudreau et al. (2016) also found that experts were significantly more inclined to judge novel work and that which is intellectually close to them as inferior. An interesting finding by Azoulay et al. (2019) shows that, at times, the gatekeepers who hinder novelty in their field are eminent scientists themselves. Indeed, when these gatekeepers pass away, the field experiences a surge of new ideas from outsiders previously kept on the periphery. This penalty for novelty is not confined to academic peer review; Hsu (2006) identifies this effect in the context of films that span multiple genres. Paolella and Durand observe this effect within law firms and their services, while Kovacs and Jonsson (2013) find it with restaurants that offer multi-cuisine menus as well. 

The Value of the Internal Critics

But critic functions are not always bad—particularly an internal critic, whether at the individual or organizational level. My colleagues Ronnie Chatterji, Rick Larrick, Roger Masclans, and I, in a paper titled “Taste Before Production: The Role of Judgment in Entrepreneurial Idea Generation,” ran a field experiment to test the value of having a refined internal critic in the quality of ideas people generate. We find that the ability to generate high-quality entrepreneurial ideas improves when individuals are trained to evaluate ideas more rigorously. In a randomized experiment with MBA students, a brief intervention focused on idea evaluation led participants to assess others’ ideas more accurately and produce ideas that were more complete, better calibrated, and more aligned with expert assessments of fundability. The largest improvements came from those with the weakest initial judgment, indicating that entrepreneurial judgment is a learnable skill.

Similarly, having a low-cost critical function at the organizational level may encourage more innovative ideas. In another paper titled Experimentation and Startup Performance, my colleagues Ronnie Chatterji, Rembrand Koning, and I examine the role of experimentation (scientifically testing ideas with A/B testing) in early-stage startups, using data from 35,000 startups over four years. A/B testing helps organizations apply judgment, not through someone’s opinion, but through evidence. We find that those that adopt A/B testing experience large and rapid gains, somewhere between 30 and 100 percent improvement in key performance metrics within a year. This top-line finding suggests that A/B testing is not a marginal technical tool but a core strategic capability that enables better critique of ideas through faster, cheaper judgment. What is also interesting is that we find that by reducing the cost of testing ideas, startups generate more variance (because now it’s cheaper to sort through it). They launch more products, make larger changes, and move faster toward scale or fail. The effects are especially pronounced among smaller and earlier-stage ventures, suggesting that experimentation lowers the cost of learning what works.

However, this internal critic function has to be structured the right way in order for organizations to benefit from it. Otherwise, it can stall any innovation. In a paper I co-authored with Todd Hall titled Organizational Decision Making and the Returns to Experimentation, we show that depending on how the critic function is implemented within an organization, you get very different outcomes: decentralized models (the idea generator is the critic) drive high innovation but come with high volatility, while centralized models (decision by committee) reduce risk but often stall meaningful change; the best results come from a simple but high threshold rule that everyone abides by, which delivers high performance with low volatility.

I could go on forever with the amount of research about the critic's function. People have studied various contexts, such as peer reviews and movie reviews, among others. All this to say is that the critical function is varied, and how it is implemented affects the kind of innovation we end up getting. But it is nevertheless critical to have ideas that create real value.

All of this research points to a central insight for me: the critic function is everywhere, and how we design and deploy it shapes the kinds of ideas that survive the gauntlet. Critic functions are not inherently good or bad, but overemphasizing or misapplying them can limit the innovation we seek to foster.

I see this dynamic up close in the classroom.

Balancing the Creator and the Critic

I teach in a business school; we lean mostly toward analysis and the development of judgment among our students. Our pedagogical goal is to sharpen their ability to evaluate and make better decisions under uncertainty. The critic function.

In strategy, for instance, we teach students how to dissect a case, challenge assumptions, and identify risks in CEOs' decisions. We ask them to demonstrate clarity of thought and rigor in their arguments. These are no doubt essential skills.

Our PhD education is quite similar: In our seminars, we ask students to read and then dissect papers they do not yet have the skills to write. We make them little Anton Egos, who can’t cook a meal but can flippantly judge it as too salty or overcooked. Overdeveloping the critic function too early has tremendous risks, in my view. If we are not careful, we risk training great critics who will never learn to create. I’m reminded of a quote by Ira Glass:

“Nobody tells this to people who are beginners, I wish someone told me. All of us who do creative work, we get into it because we have good taste. But there is this gap. For the first couple years you make stuff, it’s just not that good. It’s trying to be good, it has potential, but it’s not. But your taste, the thing that got you into the game, is still killer. And your taste is why your work disappoints you. A lot of people never get past this phase, they quit. Most people I know who do interesting, creative work went through years of this. We know our work doesn’t have this special thing that we want it to have. We all go through this. And if you are just starting out or you are still in this phase, you gotta know its normal and the most important thing you can do is do a lot of work. Put yourself on a deadline so that every week you will finish one story. It is only by going through a volume of work that you will close that gap, and your work will be as good as your ambitions. And I took longer to figure out how to do this than anyone I’ve ever met. It’s gonna take awhile. It’s normal to take awhile. You’ve just gotta fight your way through.” - Ira Glass

The New Needs Friends

I’ve always loved elementary school classrooms. They feel so creative: colorful, filled with art, markers, crayons, and construction paper. The chairs can be moved around for different projects and teamwork. The space feels personal, reflecting the identity and personality of both the teacher and the students. Student projects from this and past years are hung up on the walls. There are cozy spaces where kids can grab a book and read or grab a crayon and draw. There is a sense of both creativity and continuity.

A question I have been thinking about a lot recently is how we can regain this feeling in the context of higher education (and particularly business education, where I work). I don’t have a full set of ideas just yet, but here are some things I’ve been thinking about and how they map to the work of McCrae and his colleagues on Openness to Experience:

• The environment matters. Whiteboards, markers, Post-its, prototypes, failed experiments. Make the space feel lived in and a little weird. Somewhere that feels comfortable and real. Aesthetic Sensitivity.

• Protect early-stage ideas. Create projects where judgment doesn’t show up too early. Let things get weird. Encourage the creativity. Comfort with Ambiguity.

• Celebrate the generators, not just the evaluators. Give creators the status they deserve. It’s easy to sound smart by tearing things down. But when you’re proposing something new, something that doesn’t quite fit together yet, you might look a little stupid. Kill that off too early and people stop trying. Celebrate Imagination.

• Design for building, not just critique. We have plenty of review meetings. But where are the build sessions? The ones where people feel comfortable making new things? Engagement and Curiosity.

• Play and fun are, in my mind, first-order principles. People do their best creative work when they’re having fun. Engagement and Curiosity.

• Give unstructured time. There’s already too much time pressure. Give people hours, days, even weeks to play with ideas. Don’t demand deliverables too quickly. No deadlines. Just space to think and explore. Engagement, Imagination and Curiosity.

• Ideas need room to wobble. Give people the chance to experiment. Let the idea crawl before you expect it to run. Comfort with Ambiguity.

These principles aren’t rocket science. Honestly, I more-or-less lifted them from my observations of the Stanford d.School, an off-shoot of Stanford’s mechanical engineering department. When founded, it had the spirit of an elementary school inside one of the greatest universities in the world. It’s no surprise that its ethos took off in the 2010s. People were hungry for a creative space.

Compare these environments to the typical college classroom. Below is a space built for critique, not creation. Fixed seating, forward-facing rows, clean walls, quiet order, and the pressure to sound smart rather than create.

Let’s Honor the Creators

With that, I’ll leave you with a parting quote from Anton Ego:

“In many ways, the work of a critic is easy. We risk very little, yet enjoy a position over those who offer up their work and their selves to our judgment. We thrive on negative criticism, which is fun to write and to read. But the bitter truth we critics must face is that, in the grand scheme of things, the average piece of junk is probably more meaningful than our criticism designating it so. But there are times when a critic truly risks something, and that is in the discovery and defense of the new. The world is often unkind to new talent, new creations. The new needs friends.

The next great idea probably won’t arrive looking polished. It will be a little awkward, a little messy, and it will need a friend.

Last year, I had the usual goals: exercise, eat better, publish better, teach better, etc. But like many of my mid-career peers, I’ve been drowning in obligations, many of which are the result of what I’d call work gluttony.

Work gluttony happens when your plate is full, but the extra work seems manageable…until it isn’t. You take on a little more, then a little more, and then suddenly realize: Oh, I don’t have enough time for this, no matter how efficient I am. That’s when the panic sets in.

You start breaking promises to your friends and colleagues. The work you do manage to complete suffers because your attention is stretched thin. And all the unfinished tasks lingering in your mind make it even harder to focus. Then, your sanity takes a nosedive.

And it’s not like I could stop adding more work to my plate. Part of becoming more senior in an organization is, in fact, being responsible for getting more done.

What I’ve come to realize is that my real problem wasn’t too much work, it was a lack of leverage—the ability to scale output without a proportional increase in my effort or time.2

I can’t say I figured out the leverage problem, but I did realize one thing I wasn’t doing effectively: handing off work to others. So, I’ve added a new goal that had never made my list before: learning how to delegate better.

Then, something else happened: ChatGPT.

Delegation Makes People Productive and Happy

Being the academic that I am, the first thing I did was search Google Scholar for research on delegation. I found plenty of work by economists on agency theory, and leadership in organizational behavior, but surprisingly, there is little empirical work on delegation itself3 (if someone knows of great studies with high-quality evidence on this topic, please share).

Then, I ran into an incredible study by Stephen J. Anderson (UT Austin) and David McKenzie (World Bank). Researchers and policymakers around the world have been trying to figure out how to make entrepreneurs more successful. For many, the working theory is that entrepreneurs lack critical skills and knowledge, and that training them will unlock better performance for their companies. This seems reasonable.

However, the premise of this study was different. Maybe the issue isn’t the skills entrepreneurs have, but the time they have. So the researchers conducted a randomized controlled trial to see whether outsourcing—delegating tasks to an external person—or insourcing—hiring someone to take on specific tasks—improved entrepreneurial performance more than training entrepreneurs to do the work themselves.

What they found was amazing. Here’s a direct quote from their abstract: “Insourcing and outsourcing both dominate business training and do at least as well as business consulting at half the cost.”

In fact, a few years ago my co-authors, Rembrand Koning, Solene Delecourt, Ronnie Chatterji and I published a study titled “When Does Advice Impact Startup Performance?” (here is the pdf) We conducted a randomized controlled trial with 100 growth-stage entrepreneurs in the Indian software product industry (along with iSPIRT). We found that entrepreneurs who received mentoring around effective delegation practices from their peers—e.g., structured management approaches including regular meetings, goal-setting, and feedback—significantly impacted startup success. Entrepreneurs who received advice from peers advocating these formal management practices experienced 28% greater growth and were 10 percentage points less likely to fail than those who followed a more informal management approach.

While these studies highlight the benefits of delegation for firm performance, delegation likely also has significant effects on the well-being of the delegator.

A fascinating study by Whillans et al., titled "Buying Time Promotes Happiness," used a field experiment to show that when people spend money on time-saving purchases like hiring someone to clean their house or babysit their child, they are significantly happier than when they spend that same money on physical goods.

Delegation is not just cost-effective and beneficial for business performance. It also makes people happier.

The surprising thing, I suppose, is why we are not more explicit at teaching delegation to our MBA students or, frankly, to our kids.4

Delegation Types: Structuring vs. Task Involvement

In my exploration of delegation, it became clear that (setting aside incentives for a moment) the delegator faces two primary choices.

The first is the extent to which the delegator is involved in the tasks they delegate. This refers to how much they engage in the micro-details of how the work gets done.

The second is the level of structure the delegator imposes. This includes defining the broad parameters of the task, outlining how it might be approached at a high level, and specifying the expected outcomes and how they should look.

One could imagine a 2x2 framework (yes, I teach MBAs) where one axis represents the level of structure imposed, ranging from low to high, and the other represents the level of direct involvement in the task, also ranging from low to high.

The ideal delegator sets clear expectations but stays out of execution. They know how to structure work and hand it off to someone who can complete it at a reasonable level of quality, let’s say at least 80% as well as they would have done it themselves.

Most of us, however, are far from this ideal, often stuck in one of the other three quadrants:

• Absent: Some managers provide neither structure nor involvement. They offer no clear guidance or check-ins, leading to confusion and misalignment. In the end, they are disappointed with the results but never set their team up for success in the first place.

• Micromanager: Others swing to the opposite extreme, setting rigid structures while staying deeply involved in every detail. They waste time overseeing tasks that should be delegated, creating frustration for their team and themselves.

• Collaborator: The most common mistake is being a collaborator rather than a true delegator. These managers stay heavily involved in tasks but don’t provide clear structure or expectations upfront. As a result, the work doesn’t come back in a way that meets their needs, forcing them to step in and fix or redo it.

What’s interesting about these three delegation styles—the absent manager, the micromanager, and the collaborator—is that none of them actually solve the problem of too much work and not enough time. They create little to no leverage for the delegator. Output still depends on the delegator’s time.

The absent manager doesn’t take advantage of having people work for them, so nothing extra gets done. The micromanager gets things done but wastes their own time doing work they shouldn’t be involved in. The collaborator brings in help, but they’re still too involved, using their mental bandwidth on tasks instead of freeing themselves up to focus on bigger problems.

Delegation isn’t just about offloading tasks. It’s about creating leverage, making sure things get done without you having to be in the weeds.

For most of history, delegation was limited by resources, if you couldn’t afford employees or assistants, you had to do the work yourself. But what if delegation no longer required other people?

Delegation in the Age of AI

Delegation used to depend on affordability, but many of us lacked employees to offload work to. That was my situation for a long time (and still is to some extent), and I overcame it by hiring people on Upwork.

Upwork was good for certain tasks, though I probably didn’t use it as effectively as I could have, but it did allow me to get work done that I couldn’t do myself (e.g., remake my slides, collect citations, etc.).

But today, delegation is within everyone's reach. If you consider what generative AI, especially ChatGPT and Claude, can do, these technologies function much like fast, reasonably skilled workers.

A key question, one studied by researchers including Enrique Ide and Eduard Talamas in their papers, is how AI models fit into organizational hierarchies. Their work builds on the research by Luis Garicano on hierarchy and delegation, modeling workers as problem-solvers within an organizational structure. Depending on the nature of the task and the cost of generative AI, Ide and Talamas explore how organizational hierarchies consisting of humans and AI might evolve.

The key insight for me? AI could enable a one-person organization with near-infinite agents handling delegated work—offering leverage at an unprecedented scale.

I don’t think we’re there yet. But my delegation framework above, adapted for the AI era, can help us see when real leverage might happen.

Right now, most of us who use AI regularly are still collaborators—what we might now call prompt engineers. We use AI to refine our work, but we’re mostly not leveraging it to scale our impact. I’m sure there are also tinkerers and programmers—either not using AI at all or approaching it like traditional computing, where every instruction has to be explicitly programmed. We’re mostly gaining efficiency but not leverage.

The real shift will come when AI enables us to scale impact without our constant input.

So, what does delegation in the lower-left quadrant look like? If you’re a financial analyst at an investment firm, it’s not just using AI to quickly scan financial statements or produce rough valuation models at a fraction of your usual accuracy. It’s building a comprehensive AI-driven workflow that can autonomously evaluate investment opportunities, from gathering financial data and industry research to generating detailed investment recommendations, with quality approaching 80% of a seasoned human analyst.5

That means integrating multiple AI tools, each specialized for different stages of the investment evaluation process, effectively mirroring the roles and functions of an entire analyst team rather than merely supporting an individual contributor.

In fact, a senior analyst would follow similar principles when assembling a human investment team. A critical question is whether these same organizational design strategies remain valid in this emerging AI agent-driven world.

My co-authors, Prasanna Tambe, Dokyun Lee, David Hsu, and Yuan Gao, and I have a new project titled Managing the Machine: Does Organization Theory Matter When Organizing AI? In it, we explore whether traditional models of organization design still hold when applied to organizations composed of AI agents rather than humans. So far, our findings suggest that core principles of organization design, such as hierarchy and diversity, remain surprisingly portable to the AI world. These concepts function in agentic organizations much as they do in human ones. This suggests that delegation, too, may operate similarly in this new landscape as it did in the old. More to come as the project progresses.

Getting Leverage in the World of Generative AI

I still haven’t figured out how to delegate well. But my guess is that the people who do—especially those who learn to delegate to AI through systemic structuring of AI agents, not just collaborate with it—will have armies of agents scaling their impact. The real shift won’t be just working better, but leverage; it’ll be doing more than we ever could alone.

What Brian Eno Taught Me About Deep Tech

If you are a music lover, you’ve probably heard of Brian Eno. If you haven’t, you’ve probably heard many of the albums he produced (or co-produced): U2’s The Unforgettable Fire, The Joshua Tree, and Achtung Baby1. I could probably go on about Brian Eno, his generative music2, and his other experimental projects. But I want to talk about something else that Brian Eno created (on his Macintosh):

Yes, Brian Eno created the Windows 95 sound. Here’s a quote from an SFGate article about this project and his process3:

The idea came up at the time when I was completely bereft of ideas. I'd been working on my own music for a while and was quite lost, actually. And I really appreciated someone coming along and saying, “Here's a specific problem -- solve it.”

The thing from the agency said, "We want a piece of music that is inspiring, universal, blah- blah, da-da-da, optimistic, futuristic, sentimental, emotional," this whole list of adjectives, and then at the bottom it said “and it must be 3 1/4 seconds long.”

I thought this was so funny and an amazing thought to actually try to make a little piece of music. It's like making a tiny little jewel.

In fact, I made 84 pieces. I got completely into this world of tiny, tiny little pieces of music. I was so sensitive to microseconds at the end of this that it really broke a logjam in my own work. Then when I'd finished that and I went back to working with pieces that were like three minutes long, it seemed like oceans of time. - Brian Eno (Interview in SFGate)

Eno’s deep dive into the Windows 95 sound follows the arc of many deep innovations. At some point, combinatorial ideas run out of gas. You get only so much “new” from remixing the known.4

Motivated by curiosity, musical talent, and a passion for challenge, Eno returned to foundational principles and dove into the microstructure of sound. This research uncovered sonic landscapes he had never imagined. When applied to longer compositions, the new musical elements he discovered while creating the Windows95 sound made it feel like he was exploring entirely different musical contours, where moments expanded into vast oceans of time.

This is the deep tech arc—(1) hitting the limits of combinatorial innovation, (2) the need to delve deep for a new conceptual framework, (3) the discovery of entirely new primitives, and (4) then a burst of innovation with a new paradigm.

Like Eno’s research into sound unlocked a new way of composing music, deep tech ventures could unlock new technological possibilities by reimagining the fundamental paradigms of computation, energy, and medicine.

Why Should We Care About Deep Tech?

The term Deep Tech has been around for some time but is now gaining traction in investment and academic circles. Definitions vary, from “technology solutions based on substantial scientific or engineering challenges” (Wikipedia) to innovations with significant “technology risk”—e.g., the possibility that a technology may not be feasible. Other definitions highlight deep tech’s challenges, such as high R&D costs, long incubation periods, and uncertainty, or focus on specific domains like quantum computing and next-generation batteries.

Growing research also examines why deep tech ventures are harder to develop and fund. Ramana Nanda and colleagues argue that high experimentation costs limit investment. Arora et al. highlight the joint effects of tech and market risks, creating a “missing middle” that makes some deep tech harder to finance. However, we must first understand how deep tech generates value before tackling organizational and financing challenges.

In this post, I argue that deep tech matters because it creates value in ways conventional (combinatorial) innovation cannot—by creating new primitives, making the constraints of existing paradigms less relevant, and solving unsolved problems with previously impossible technologies.

How Does Combinatorial Innovation Create Value?

Before developing our concept of deep tech innovation, it is helpful first to outline a model of “normal” combinatorial innovation, which works by recombining existing, discrete components in novel ways. These components operate at different levels of abstraction, enabling both incremental improvements and entirely new technologies.

Anyone familiar with computer science knows that programming operates at multiple levels. At the lowest level, assembly language and machine code interact directly with hardware, manipulating memory and processor instructions. Higher-level languages like C, Java, and Python don’t require you to know binary math. Developers can write more human-readable code with these languages. At an even higher level, frameworks and APIs provide fully formed software components, enabling rapid development of applications without ever knowing or interacting with the underlying logic of many components. Today, generative AI pushes this abstraction further, allowing us to 'write code' using natural human language.

You can see this in other creative domains, like music. At the most fundamental level, sound waves are the raw materials. Notes and chords structure these sounds, and they combine into melodies, harmonies, and rhythms. Higher up, song structures and genres provide frameworks for variation. Developers use APIs, and musicians sample, remix, and reorganize patterns to create new music.

Combinatorial innovation reassembles existing components into new configurations. These recombinations can drive substantial productivity gains for innovators who create novel and useful linkages.

The math behind combinatorial innovation is also exponential. With k known components, they can be mixed into bundles of up to k-1, excluding trivial cases. Even with just five components, there are 25 possible combinations. With 100 elements, the number explodes to approximately 1.27 × 10^30. Everyday combinatorial innovation has significant potential, and extensive research highlights both (a) its impact and (b) the characteristics of those who excel at it.

However, not all combinatorial innovation is equal. “Shallow innovation” (SI), what I like to think of as “API innovation,” and now “chat-based innovation” is becoming increasingly easy to execute across many domains. While still valuable, these forms of innovation are highly competitive because their ease makes it easy for most people to create with these building blocks.

There is also deeper combinatorial innovation (DCI), which is less obvious and less accessible to most people. Unlike shallow innovation, DCI involves rigorously recombining existing paradigms in novel ways or returning to the most fundamental components of established approaches. As technology matures, primitives are often abstracted away to make innovation more accessible, but this can reduce novelty and quality. By going deeper into these foundational building blocks, DCI enables the creation of breakthroughs that shallow innovation cannot achieve.

Therefore, shallow innovation operates within a given paradigm, leveraging its existing frameworks to address problems. In contrast, deeper combinatorial innovation typically does one of two things: (a) it pushes beyond the superficial limits of a paradigm while still working within its foundational building blocks, or (b) it merges distinct paradigms, such as applying computer science principles to biology. The actual test of these innovations lies in whether they solve meaningful problems and, for businesses, whether they offer a viable path to profitability, at least in the short term.

In essence, combinatorial innovation creates value by recombining existing components, frameworks, or paradigms to optimize performance (e.g., GPUs repurposed for AI), lower costs (e.g., low-code/no-code software tools), improve accessibility (e.g., smartphone cameras), and enable new applications (e.g., ride-sharing apps integrating GPS, mobile payments, and cloud computing) within existing technological constraints.

How Does Deep Tech Create Value?

While shallow and deeper combinatorial innovation remix existing components and primitives, real deep tech creates value in a different way.

Existing definitions, which emphasize “science,” “advanced engineering,” or “high cost,” serve as useful descriptors but miss the true source of value creation.

Deep tech’s significance does not come merely from applying the scientific method or tackling costly and complex problems. Its value emerges when old technological constraints, previously accepted limitations, and established primitives are replaced with an entirely new conceptual and technological framework. To understand what makes deep tech truly deep, we must see it as a venture into “continuous reality” where discretization and modularization have not yet occurred. Researchers draw circles around phenomena that haven’t been identified before to create new and generalizable primitives. These primitives then become part of a new framework for innovation.

The Failure of Existing Models

It is perhaps helpful to begin at the end.

Every technological paradigm has limits, and those limits eventually make it difficult to push the technology forward, even at great cost. This limit occurs when current frameworks and scaling laws (e.g., put x in, y comes out) stop working. Sometimes, these limits are due to engineering constraints, but more often, the underlying assumptions of the science may no longer hold. For instance, a physical “law” doesn’t behave as it should at some scale.

So, scientists and engineers stop trying to optimize within an asymptotically failing framework and start questioning whether a new one is needed. Recognizing this failure as a fundamental limit rather than a fixable inefficiency is a key step (e.g., Nelson and Winter (1977)).

After this realization takes hold, the next step is to revisit the first principles.

The Search For a More General Conceptual Framework

As the existing paradigm reaches diminishing returns, scientists, researchers, firms, and investors begin searching for a new framework. This kind of shift requires abstraction, challenging widely held assumptions, stripping away existing conceptual models that serve as the foundation for current technologies and impose constraints on progress.

The goal is to start afresh with a new model that reveals deeper patterns still unseen by humans, allowing innovators to perceive a path forward that was not visible when we looked at the world through the old model.

New Generalizable Primitives from Continuos Reality

The search for a new framework requires reconceptualizing the problem to be solved, particularly by creating new primitives. Primitives are a technological paradigm's most basic building blocks, forming the foundation for designing and constructing more complex concepts, structures, systems, and technologies.

It is important to note that while new systems may, in principle, perform the same functions as those built within the old paradigm, they are not merely more efficient versions of existing technologies. They represent an entirely different way of thinking about the problem (one example is the mRNA vaccine).

For instance, semiconductors were not simply better vacuum tubes, just as qubits are not improved transistors. They are new foundational primitives.

Development of New Innovation Paradigms

Once a new conceptual framework is developed and enough of its essential primitives are defined, the work of building new technologies on this foundation can begin. Using new frameworks and primitives distinguishes deep tech innovation from even the most advanced forms of combinatorial innovation.

Combinatorial innovation starts with existing primitives and conceptual paradigms, rearranging and recombining them in novel ways to extend the capabilities of known systems. Deep tech, on the other hand, involves changing the primitives themselves, redefining what is possible, and setting the stage for entirely new technologies, and perhaps even new industries. Once these new foundations are established, they become paradigmatic, forming the infrastructure that enables the next wave of combinatorial advances.

The End of the Classical Computing Paradigm?

Just as the limits of vacuum tubes forced a shift to an entirely new computational paradigm, today’s semiconductor industry is in need of innovations beyond incremental improvements. There are already people trying to figure out what is next in computation.

The future of computation may not just be about better chips. It might consist of a reconceptualization of what computation itself is.

Quantum Computing, Semiconductors, and Limits of Existing Scaling Laws

From Vacuum Tubes to Semiconductors

In a previous post, I talked about Intel, its early rise, and what now looks like its decline. But Intel is just one milestone in the broader journey of computing.

While the computing industry has deeper roots, a good starting point for discussing quantum computing is the vacuum tube era. This era emerged around World War II and lasted until the late 1950s. At its core was the vacuum tube itself, a tiny lightbulb like thing that acted as an on/off switch and could, with input, perform calculations by manipulating binary states.

When Vacuum Tube Paradigm Could No Longer Scale

Interacting with a vacuum tube-based computer back then was not easy. The ENIAC, for example, had 18,000 vacuum tubes, and programming it meant writing instructions in binary (0s and 1s), then feeding them to the machine via punch cards, paper tape, or plugboards.

These computers were incredibly costly and complex. They used lots of energy, broke down often, generated enormous heat, and required constant maintenance (imagine how many lightbulbs had to be replaced!).

Yet, these early computers were also incredibly important. One of the biggest users of vacuum tubes was AT&T, which relied on them as switches and amplifiers for long-distance calls. As America's telephone system expanded, AT&T ran up against the technological limits of this paradigm. In practical terms, connecting all of America with vacuum tubes, let alone the rest of the world, was impossible. Something different was needed. A new primitive that could overcome the inherent limitations of vacuum tubes. (My colleague Dan Gross has several papers on AT&T’s and its mechanization of the telephone system and its impact on the labor market.)

The vacuum tube architecture reached its physical breaking point as computational demands grew. The architecture was constrained by size, reliability, power consumption, and operational complexity. The scaling laws of the vacuum tube era effectively ran out of gas, unable to support further growth without immense cost and complexity. To sustain innovation, robust new primitives were needed.

The Search for a New Framework at Bell Labs

AT&T recognized that these diminishing returns were a real problem. Luckily, they had the financial resources (and time) to invest in deep research conducted by brilliant scientists to tackle this failure of the existing paradigm. The breakthrough came in 1947 at Bell Labs (AT&T’s internal “R” organization), where John Bardeen (Physics PhD, Princeton), Walter Brattain (Physics PhD, University of Minnesota), and William Shockley (Physics PhD, MIT) invented the transistor. (Check out The Idea Factory about Bell Labs)

The transistor replaced the vacuum tube by enabling electronic switching and signal amplification using solid-state materials rather than glass tubes. This shift required a complete rethinking of the underlying physics. It moved away from classical electrical principles toward understanding how electrons behave at the atomic level. It also required a shift in conceptual framework, leveraging new concepts from other fields, including materials science, quantum theory, and solid-state physics.

The New Primitives of Semiconductor Computing

Rethinking the conceptual underpinnings of computing hardware led to new primitives: the p-n junction and bandgap engineering, allowing engineers to precisely control electrical flow. This, in turn, enabled the transistor to function as a fast, reliable 0/1 switch. This was a massive reconceptualization of the building blocks of computing. What we thought of as the constraints of the vacuum tube era no longer applied in the same way.

The transistor, along with the new primitives that scientists learned from researching semiconductor physics, created the foundation for a new computing. These building blocks eventually led to integrated circuits, vastly expanding computational possibilities.

Redefining the Paradigm of Computing

In addition to the deep innovations that led to the transistor, other fundamental breakthroughs—such as Claude Shannon’s development of information theory—enabled the harnessing of this new paradigm.

More importantly, though, this conceptual breakthrough gave rise to much of the modern world: computing as we know it (with a device in our pocket several trillion times more powerful than the ENIAC), the evolution of culture in the form of movies made at home and shared around the world, smart cars, space travel, and much more.

However, today, semiconductor architectures are thought to be approaching their physical and economic limits, as even the deepest combinatorial innovation is running out of gas.

From Semiconductors to Quantum Computing

On February 19th, Microsoft announced what it claims to be a massive breakthrough in the history of computing: the first quantum computer powered by topological qubits.

I am a social scientist, not a quantum physicist, but I’ve been watching this evolution with great interest. For one, Duke University is a hotbed for Quantum Computing innovation (see the work of Jungsang Kim and Christopher Monroe, as well as dozens of other students and faculty). Its also not just scientific research anymore, several Duke faculty have launched the first publicly traded quantum computing company, IONQ.

Moreover, several scholars in my field, Strategy, have been assembling people to understand the economic implications of quantum computing (check out the website of a new conference on this topic run by Florenta Teodoridis, Milan Miric, and Avi Goldfarb).

If quantum computing works (and it looks like it may just have legs), then what we see with generative AI might be a blip compared to the transformation that is likely to happen with quantum.

Limits of Existing Computing Paradigms

The shift from vacuum tubes to semiconductors, and now from classical to quantum computing, again follows the arc of deep tech innovation. Semiconductor scaling hit limits not just of engineering but of the classical computing paradigm itself, forcing a rethinking of what computation could be.

Computing is fundamentally about how we represent and manipulate information. Classical computing encodes data in bits (discrete 0s and 1s stored in silicon transistors) processed through deterministic logic gates.

Over decades, researchers and engineers refined semiconductor manufacturing, allowing transistors to shrink exponentially in size while increasing computational power. However, fundamental constraints emerge as traditional computing and transistors approach atomic limits, notably diminishing returns from Moore’s Law. These barriers indicate that even the deepest combinatorial innovations in semiconductors are running out of gas, necessitating a new approach.

Searching for a New Framework

From my naive understanding, quantum computing isn’t just about building better semiconductors, just as semiconductors weren’t merely better vacuum tubes. While semiconductors retained their predecessors' binary computing model, quantum computing marks a complete departure from classical logic. Instead of encoding information as a series of 0s and 1s, quantum computers leverage the mind-bending (and, frankly, weird) principles of superposition and entanglement, allowing qubits, the new units of computation, to exist in multiple states simultaneously.

This new paradigm brings opportunities as well as scientific, engineering, and business challenges. Quantum systems are able to work with information in ways that classical ones cannot, but qubits are fragile and require specialized error correction mechanisms. Unlike classical redundancy-based error correction, quantum error correction must preserve entanglement without collapsing quantum states. The transition to quantum computing is not just about increasing processing power; it will require breakthroughs in physics, materials science, and algorithm design to fundamentally change how computation works.

New Primitives of Quantum Computing

While classical systems achieve speedups through multi-core architectures, GPUs, and optimized execution models, they ultimately follow a sequential or parallelized path. Quantum computers, however, use entanglement and interference to explore multiple computational pathways simultaneously, enabling exponential speedups for certain kinds of problems (e.g., cryptography).

To unlock these advantages, every layer of the “quantum stack,” from hardware to software, must be rebuilt with new primitives.

Scaling quantum computers means overcoming fundamental physics problems. The real question isn’t whether quantum computing works—it does—but whether we can turn this new paradigm into something practical and broadly useful, solving problems classical systems never could.

New Industries and a New Paradigm for Computing

Finally, quantum computing is not an extension of classical techniques but the creation of an entirely new computational paradigm. It operates beyond binary logic to redefine what computation could be. Introducing new primitives means that industries built around classical computation, from cryptography to materials science to optimization, may undergo fundamental transformations.

This is deep tech’s essence: recognizing limits, finding broader frameworks, and building new primitives that render old constraints irrelevant. Quantum computing isn’t just increment on an existing natural trajectory of computation; it redefines information processing, paving the way for new industries.

So, How Does Deep Tech Create Value?

Simple: Deep Tech creates value from first principles. It provides innovators with a new toolkit to solve problems that simply couldn’t have been solved within the existing paradigm.

This is why fostering deep tech innovation isn’t just important; it is essential. If we fail to invest in and support it through our research labs and universities, we risk ceding the future to those who do invest.

Workers: Why is it so hard to find a job?

A few weeks ago, the Wall Street Journal article “Ghost Jobs: Why Fake Jobs Are Proliferating” struck a nerve online. 

It highlighted the growing frustration of job seekers who apply to countless listings, only to find that nearly one in five are 'ghost jobs'—roles posted with no intention of being filled. Ostensibly, companies use these postings to signal growth or hedge for exceptional candidates. If you're job hunting, that’s a pretty crappy feeling to hear that you were wasting your time on fake jobs.

But job hunting has never been a walk in the park. Even before ghost jobs, the odds were grim. A study by Farber et al. (2016) found that only 10% of job applications even receive a callback, and the numbers were worse for early career candidates, especially those with interim jobs (e.g., graduating with a CS degree but working at Target).

And even after landing an interview, the odds remain tough: recent numbers suggest an interview-to-hire rate of just 10-11%. If you want a 90% chance of getting at least one job offer, you’d need to apply to 230 positions. And if the callback and offer rates drop to just 5%? You’re looking at 920 applications. Four times as many!1 Clearly, your incentive is to apply to as many jobs as possible.

Firms: Why is it so hard to find good people to hire?

These odds have led to record numbers of job seekers applying to as many jobs as possible. But companies are on the other end of the “submit” button.

Firms now face a new problem: they’re drowning in applications; how do you find that proverbial needle in the haystack?

A few years ago, a friend was tasked with building a data science team at a major tech company. He followed the standard playbook: He posted a job and sent top candidates to HR, only to hear: 'Better teams already recruited them.' Searching PhD programs? 'Already in the system.' Expanding to lesser-known schools? Same result.

Finally, he went deeper, combing through conference papers to find “hidden gems” with cutting-edge research. When he sent the list to HR? They were already being recruited by someone else.

But what about the other side of the equation: the thousands of people who applied to those jobs and never heard back?

Information Problems in the Labor Market

The problem isn’t just about too many applications or too few hires. At its core, hiring is riddled with information gaps: two big ones, in particular.

For workers: What jobs are available? Is this a good fit? Will I enjoy working there? Even with job boards, LinkedIn, and employer branding, crucial details like pay, daily responsibilities, workplace culture, and long-term career growth often remain unclear.

The challenge for firms is just as complex: Who is out there, and how do we find them? Can they do the job, or will they struggle once hired? Will they stay long enough to justify the investment in training? Despite resumes, interviews, and assessments, firms still face uncertainty: How well a candidate will integrate into the company, collaborate with the team, and grow into future roles remains a gamble.

These are perennial problems, and over time, we've found ways to navigate them—first through networks and personal connections and later through a relentless cycle of technological innovation. This is a hiring arms race in which each new solution introduces fresh challenges.

Networking your way to your next job?

In 1973, Mark Granovetter, a sociologist now at Stanford University, published groundbreaking research on how Boston workers found their jobs. His seminal article, The Strength of Weak Ties (perhaps one of the most beautiful pieces of social science ever written), and his dissertation, Getting a Job, revealed a surprising fact: when it comes to finding jobs, weak ties—acquaintances, former colleagues, and distant contacts—matter more than close friends or family. Why? Strong ties tend to operate in the same social and professional circles, meaning they hear about the same opportunities. Weak ties bridge gaps between disconnected networks, exposing job seekers to new information and firms to new talent.

The transmission of labor market information through networks isn’t just to a worker’s advantage; it’s equally critical for firms. Just as workers struggle to see all possible job openings, firms struggle to see all possible candidates. Hiring through close networks limits access to fresh talent. Weak ties expand the search, helping firms identify candidates they might otherwise miss.

Social networks also convey information that doesn’t flow as easily through weak ties. Research by scholars such as Roberto Fernandez, Isabel Fernandez-Mateo, and Emilio Castilla—as well as many scholars in economics and business research—have highlighted the role of employee referrals in the hiring process. From the firm’s perspective, referrals offer insight into whether a candidate can be trusted and how well they might fit with the team. For workers, it’s about understanding whether the company is a good place to work and whether the manager is someone they can rely on. The real information that won’t be posted online comes from people, especially those you trust, such as those with whom you have strong ties.

Networks are potent tools for workers and firms but come with trade-offs. Information flows imperfectly through networks, and because of how we connect, strong biases determine who is “in” on the information and who is “out.” Though networks provide opportunities for some, they also restrict opportunities for others.

While referrals are an imperfect mechanism for connecting workers and firms, they have consistently accounted for approximately 35% of all hiring in both within-firm analyses and studies of specific samples going back to the 1970s.

However, technological change is unrelenting, and the Internet has transformed the job search process by scaling access to information— a role that informal networks have played for decades.

How Job Boards Led to Information Overload For Firms

If referrals were the dominant hiring mechanism of the pre-digital era, the Internet and job boards introduced a new paradigm: one in which information about job opportunities would flow freely, job markets would become more open, and hiring frictions would—poof—disappear.

But technology's impact is never so obvious. David Autor’s Wiring the Labor Market (2001) predicted these tensions.

He begins with a simple idea: At its core, the labor market is riddled with friction. It’s hard for workers and firms to find each other. When that process breaks down, the consequences ripple outward—bad matches, prolonged unemployment, and lost productivity. The naive Econ 101 solution? More information.

Then came the flood.

A quote from Autor (2001) summarizes this nicely.

Jobs once drew dozens of applicants; now they get thousands.

The challenge flipped: instead of struggling to find enough candidates, firms were drowning in them. Drinking from a firehose doesn’t quite capture it.

Instead of carefully searching for candidates, they were inundated with applications—most of which were irrelevant. Their new challenge was to filter, rank, and sort the haystack. 

Meanwhile, workers faced a different frustration: sending application after application, only to have them disappear into a void.

Networks had once been the gatekeepers to jobs. Online job boards were supposed to democratize hiring, but instead of removing bottlenecks, they relocated them. The fundamental challenge of hiring, matching the right worker with the right firm, was not solved. It had simply shifted from information access to sorting.

Hunting For Talent in Exponential Haystacks

The Rise of Outbound Recruiting

How could firms cut through the noise? Instead of waiting for the right candidates to apply, firms began hunting for talent (HfT).

LinkedIn, founded in 2003, gave firms access to millions of professionals—both job seekers and those who weren’t looking. As the platform scaled to hundreds of millions of users, hiring shifted from a passive process to a proactive search.

A few years ago, my collaborators Ines Black, Rembrand Koning, and I set out to answer a crucial question: How has this new information reshaped how firms find talent? Our findings, published in “Hunting for Talent: Firm-Driven Labor Market Search in the United States,” revealed a fundamental change in how hiring works today.

Historically, headhunting was a niche and costly practice in the American labor market. It required maintaining extensive candidate lists, conducting personalized outreach, and investing significant time in relationship-building. Outbound recruiting was both costly and uncommon and reserved mostly for senior hires. In 1991, only 4.9% of U.S. hires came through this method.

By 2022, the rate of outbound recruiting had tripled to 14.3% of all hires. Among those making 100k or more, that number was 4X the level of a few decades ago - close to 20%.2 Higher-paid and higher-skilled workers were those most affected by this shift (the LinkedIn set). This also wasn’t just a tech industry phenomenon. It was happening across sectors.

We observed firms adapting to this new flow of information, and their hiring data reflected the shift. Between 2010 and 2020, the share of firms hiring HR employees with recruiting skills rose from 33% to nearly 50%, signaling a growing emphasis on proactive talent acquisition. Demand spiked for recruiters skilled in scouring LinkedIn, GitHub, and other platforms for talent. Rather than sorting through endless applications, firms were building the capability to hunt for candidates directly.

As social scientists, we also worry about substitution. What practice was HfT displacing?

Surprisingly, outbound recruiting didn’t diminish referrals. We found that referrals still accounted for 33% of hires, unchanged since the 1970s and mostly constant across income levels. Networks kept doing what they were doing. Getting a Job was the same as it always was.

Instead, hunting for talent was replacing application-driven hiring. Hiring shifted from applications to search. The bottleneck moved from information access to sorting to search. Today, nearly 60% of workers earning $150K or more land jobs through referrals3, recruitment, or placement agencies. For them, direct applications to companies now make up only a minority of hires (~40%).

For a fair chance, workers struggling to stand out must signal their quality to recruiters and algorithms. Can AI bridge these gaps, or will it simply create new ones?

Can AI help firms filter and sort applicants and potential workers?

AI might create new information problems on both sides of the market. A few recent studies have given us a glimpse of how AI might affect firm decision-making.

Work by Li, Raymond, and Bergman (2020) showed that because AI is trained on past decisions, it leans heavily toward the candidates that firms have always hired. It scales up biases as it scales up efficiency. Their study puts this into sharp relief: standard supervised learning algorithms cut Black and Hispanic interviewees from 10% to 3-5% while boosting hiring yields. A contextual bandit model, which explicitly values exploration, significantly increased minority representation while still improving hiring rates. In short, algorithms don’t just reflect hiring practices; they can reshape them. Sometimes reinforcing bias, sometimes counteracting it.

Another creative study in this space, Cowgill (2020), shows that AI can improve hiring by selecting stronger candidates. However, its impact depends on the structure of human decision-making. When human evaluations are noisy and inconsistent, AI helps correct biases, leading to better interview performance, higher offer acceptance rates, and greater productivity, especially for non-traditional candidates. However, if human biases are rigid and systematically applied, AI is more likely to reinforce those patterns rather than counteract them.

AI doesn’t just mirror hiring practices. It can reshape them. As Li, Raymond, and Bergman (2020) and Cowgill (2020) show: when trained on past decisions, AI can scale biases, but models that encourage exploration can counteract this. AI’s impact also depends on whether human decision-making (its training data) is inconsistent (allowing AI to correct biases) or systematically biased (causing AI to reinforce them).

How will the rise of AI-based tools on the worker side affect screening?

While firms use AI to screen candidates, workers may use it for a range of labor market activities, especially to improve their resumes, cover letters, and essays.

Cowgill et al. (2024) find that generative AI, like ChatGPT, makes it harder for employers to distinguish strong candidates from weak ones, reducing screening accuracy by an average of 4–9%. Because AI flattens differences in applicant signals, evaluators will increasingly seek alternative ways to assess talent, such as background checks, direct interviews, and especially referrals.

What’s interesting is that while generative AI may make an undesirable candidate look slightly better, it can also help a potential star—especially one with non-native English fluency—better communicate their value. The key takeaway for me from these papers is that predicting AI’s impact is far from straightforward.

But here’s my prediction anyway: As AI slashes the cost of applying, even eliminating the effort of customizing a cover letter, job applications will flood in, looking sharper than ever.

Workers’ AI agents will flood companies with applications. Firms’ AI agents will ghost them.

Frankly, we should leave the agents to hunt for talent in the exponential haystack.

Real hiring, however, will continue much as before, leaning even more on harder-to-fake signals: who you know, what you’ve built, and the reputation of past employers. Instead of making hiring more data-driven, AI may paradoxically make networks matter more than ever. And in a world where AI can perfect every resume, the most valuable career advantage may remain what it’s always been: having the right people vouch for you.

The Best Algorithm? The Right People

This isn’t just an abstract discussion for me. I’ve lived it.

When I graduated college in 2003, the dot-com bubble had just burst. The NASDAQ had collapsed by 78%, and hiring had frozen in tech and finance. Maybe I had been too focused on my classes, or maybe I had just assumed that something would work out after graduation. Whatever the reason, I missed the recruiting cycle. So, I did what everyone says you should do: I sent dozens of job applications. No response.

Then, a friend told me that Target was hiring. Desperate to do something, I was offered a ten-dollar-an-hour job unloading trucks at night, so I took it. The work was brutal, and I lasted only two weeks.

A week later, I met up with a friend from high school at our regular spot, Dunkin’ Donuts on Route 27 in South Brunswick, New Jersey. His former roommate was leaving a research assistant job at Memorial Sloan Kettering Cancer Center in New York to go to law school. “If you want it, my friend can put in a referral for you.”

Two weeks later, I had a job at MSKCC.

Data entry and small research tasks weren't glamorous, but I was surrounded by researchers passionate about their craft. I was hooked.

That job led to grad school, and with strong recommendations, I got in. And just like that, I was out of the labor market again.

Looking back, I realize that formal hiring processes contributed little to my trajectory. No one sifted through my resume in an applicant tracking system, no recruiter found me through a keyword search, and no AI model ranked me as a top candidate. Like many others, I got my first real job through a weak tie.

I also made one accidental smart choice. I left Target off my resume. If I had put it on there, would future employers have placed me in a different category? Maybe.

That’s the messy reality of hiring: It’s not just about what you can do.

All this to say: Skip the exponential haystack, call a friend, and grab a coffee.

Is Intel History?

On February 15th, the Wall Street Journal broke a story about Intel, one of the most storied companies in American history. Rumor had it that Intel may be split into two. Broadcom could acquire its design business, and Taiwan Semiconductor Manufacturing Company Limited (TSMC) could acquire Intel’s manufacturing capabilities.

Intel, the great-grandchild of Bell Labs and its research into transistors (through Shockley Semiconductor → Fairchild Semiconductor → Intel), was the archetype of the Silicon Valley startup. Its founders, Bob Noyce and Gordon Moore (Moore’s Law), were PhDs in Physics (MIT) and Chemistry (CalTech). They founded the firm with Arthur Rock, the storied financier who kicked off the venture capital revolution with his bet on Fairchild Semiconductor. Not only was Intel what we would call “deep tech” today, but it was also a genuinely modern startup.

However, Intel ascended to new heights under its third CEO, Andy Grove (who earned a PhD in Chemical Engineering from Berkeley). Grove was a brilliant scientist (his first book focused on semiconductors), but he was also known for his innovative management ideas (e.g., OKRs). He turned Intel into a global technological powerhouse, fending off competition from Japanese rivals and shifting from memory chips to microprocessors, which solidified Intel’s dominant position in the computing industry. Grove’s guiding principle—Only the Paranoid Survive—fueled an alert, creative, forward-looking, and experimental culture constantly seeking ways to innovate and thrive. 

However, Intel’s paranoia gradually shifted from a relentless drive to invent the future to an anxious fixation on protecting its share of the present. It receded into an ordinary kind of paranoia, defined by suspicion, defensive overreaction, and a tendency to see threats where none existed. By the time Paul Otellini became Intel’s fifth CEO (the first with a business degree)1, the company seemed to focus more on slicing the pie than making it bigger.

In fact, in 2005, Apple approached Intel with an opportunity that could have kept the company at the forefront of computing in the mobile era: supplying the processor for the first iPhone. Yet, Otellini chose to pass on the deal:

This decision cost Intel not only a deal but also its future. Locked out of the mobile revolution (never mind AI), Intel ceded the next era of computing before it even began.2

What didn’t Otellini see?

The Shifting Forces of Value Creation to Value Capture

To see why firms prioritize capturing value over creating it, we need to look at how strategy itself is taught.

While there are many academic definitions of strategy, the syncretic one that I find most helpful is this: strategy is the framework a firm uses to create and capture value.

From a customer’s perspective, value creation is easy to define: a company creates value when it offers something that makes you better off. This could be because it’s vastly superior to the alternatives (e.g., the light bulb) or significantly cheaper, allowing you to keep more of your money for other things. But from the firm’s perspective, it’s much harder. You need to innovate. Either to make something new, something better, or cheaper. Creation is easy to see but hard to do.

From a firm’s perspective, value capture is easy to define: given the value a firm creates (e.g., measured by willingness to pay), how much does it get to keep (e.g., profits = price – cost)? For customers, value capture is often invisible at first—you may not realize what share a company keeps until after you’ve bought and used the product.3 However, for firms, once a product is already selling in the market, value capture becomes the easier lever to pull. A restaurant can buy cheaper produce, pay workers less, shrink portion sizes, or swap fresh ingredients for frozen ones without necessarily creating any new value.

Herein lies an asymmetry, especially for incumbents: Value creation is easy for consumers to see but difficult for firms to achieve. Value capture, on the other hand, is easy for firms to implement but harder for consumers to detect.

This tension between value creation and value capture is at the heart of strategic thinking. To understand why some firms lean toward value capture, we need to revisit one of the most influential strategy frameworks ever developed: Michael Porter’s Five Forces.

The Five Forces that Will Eat Your Lunch

Often, the first framework students encounter in introductory strategy is the “Five Forces Framework,” rooted in Michael Porter's foundational research on Competitive Strategy from the early 1980s. Porter's genius lay in flipping the classical model of perfect competition (where, in equilibrium, no firm makes a sustainable profit) on its head. He asked: What would it take for a firm to generate profit in a competitive market?

He outlined five forces that quietly but relentlessly chip away at a firm’s ability to hold onto the value it creates. Your rivals are scheming to undercut you; new entrants are always at the gate, looking for a way in; your suppliers are angling for a bigger slice of the pie; your customers are always on the lookout for a better deal; and even firms in entirely different industries are figuring out how to make your category irrelevant with their substitutes.

These aren’t just competitive forces; they’re the Five Paranoias: the inescapable, ever-present anxieties that gnaw at even powerful firms.

Watch your back. Everyone is out to get you.

While Porter’s framework could also direct firms to identify areas for innovation, in practice, many firms may use it to justify defensive, value-capturing strategies. By framing strategy primarily as a battle for defending value against competitive pressures, business education may have emphasized strategic success in terms of securing a larger share of the pie rather than expanding it through innovation.

This is because value creation is riskier, and resources are scarce. Protecting the value already created often feels more prudent than chasing shiny objects with uncertain payoffs.

So the work begins: How do we minimize the impact of rivalry, fend off threats, create customer loyalty, and squeeze suppliers? That is, how do we build the infrastructure to capture value?

This is how we start the lesson.

But what happens when a generation prioritizes value capture over creation?

Frameworks like the Five Forces shape not only business strategy but also the mindset of future executives. They teach firms how to defend their positions rather than create new value. But how does this emphasis on competition affect students trained in these models?

The impact of strategy education on students

Two recent studies provide insight into how such lessons might impact students' mindsets.

A fascinating study by Yang et al. (2020) examines how 262 Harvard Business School CEO alumni approach strategy. The researchers leverage a quasi-exogenous shock—HBS’s 1983 shift in its core strategy curriculum to incorporate Michael Porter’s later influential ideas on competitive strategy (e.g., the five forces). On page 25, the authors describe this shift:

Their findings are striking. While the new curriculum fostered more “structured” strategic planning with a focus on the external environment of a business, it came at the expense of a reduced emphasis on internal affairs and implementation.

Another very important recent study echoes these findings. Heshmati et al. (2024) studied a sample of 2,269 MBA students at a major business school and examined how students analyzed case studies before and after taking their core strategy class. What they found parallels the Yang et al. study but with a much younger and more recent sample. They found that after taking strategy, students increased the attention “paid to broader industry and competitive concerns” and became “ more aware of the uncertainty pervading strategic decisions.” Like the CEOs in the Yang et al. study, those learning the new way of thinking about competitive strategy became more attuned to competitive pressures.

These studies suggest a shift in mindset. When strategy education emphasizes competitive positioning—framing strategy around the fear of rivals, new entrants, suppliers, buyers, and substitutes—so do the students.

Past research by Frank et al. (1996) suggests that “when people expect partners to defect, they are overwhelmingly likely to defect themselves.” Rubinstein (2006) also finds similar results.4

If business education has instilled an orientation toward value capture over value creation, we should see its effects not just in classrooms but inside firms. Indeed, a growing body of research suggests that this shift is already reshaping firm behavior—and not for the better.

The Consequences of a Value Capture Mindset

A growing body of economics and strategy research has begun to examine how introducing new professionalized business practices—often through the acquisition of firms by larger players—affects firm behavior, profitability, and product and service outcomes. Some of the most interesting research on this topic comes from the healthcare sector.

A study by Eliason et al. (2016) [including my colleague Ryan McDevitt] investigates the practice of “strategic patient discharge” among long-term acute-care hospitals (LTCHs). They find that these hospitals appear to maximize profits by exploiting a quirk in the Medicare payment system. Under this system, a hospital receives a small reimbursement at the start of a patient’s stay but a sizeable lump-sum payment after a predetermined number of days. To maximize profits, hospitals disproportionately discharge patients immediately after they cross this financial threshold—not because of medical necessity.

While this practice increases profits, it also leads to negative outcomes for patients, including higher mortality and readmission rates, and results in the loss of potentially billions in U.S. taxpayer money. It is particularly prevalent among for-profit hospitals and those acquired by corporate chains. This is value capture over value creation.

This is not the only study to identify this pattern: prioritizing value capture over value creation. In another study on the corporate acquisition of dialysis clinics, Eliason et al. (2020) examine how large-firm management practices reshape dialysis clinic behavior. After the acquisition, they find that these clinics overprescribe highly reimbursed drugs without medical necessity, exploit Medicare reimbursement structures, and replace higher-cost but more skilled nurses with lower-paid technicians. Once again, these practices maximize profits but lead to higher hospitalization rates and increased patient mortality. These actions exemplify the cost-cutting, revenue-maximizing strategy of value capture. Similarly, Wollman (2020), in a cleverly titled paper (“How to Get Away with Merger”), also studies the dialysis industry and finds similar results: the stealth consolidation of this sector reduced care quality, resulting in higher hospitalization rates and lower survival rates for patients.

Another recent study by Ambar La Forgia (2023) finds a similar pattern of results. She shows that physician practices acquired by “financially driven” management companies perform significantly more C-sections, “resulting in less clinically appropriate care and worse patient outcomes.” However, those acquired by clinically focused firms reduce unnecessary C-sections, improving care.

This finding is powerful: The firm's stated strategy has huge implications for how it balances value capture (maximizing profits) and value creation (patient outcomes). 

Another excellent study by La Forgia and Bodner (2024) offers a more optimistic perspective. They demonstrate that chain ownership in IVF clinics can enhance patient outcomes by increasing treatment volume, improving success rates, and improving quality through resource sharing and diffusing standardized best practices. They explain why acquisitions in this sector may not produce the same negative effects seen elsewhere—greater price and quality transparency can incentivize management to prioritize the value they deliver to customers rather than the value they capture.

Business Education and Value Capture in the Economy

While the studies above provide compelling evidence that firm strategy and management practices influence the balance between value capture and value creation in healthcare, is there evidence of a systemic impact of this type of education on the economy?

A powerful paper by Nobel Prize-winning economist Daron Acemoglu, Alex He and Daniel le Maire, examines the impact of the rise of “business-educated” managers (MBAs or business undergraduates) on wages and labor share. It uses large-scale data from the U.S. and Denmark.

The findings are striking and align with the evidence we discussed earlier in the healthcare sector. Firms with business-educated managers experience a 6% to 3% wage cuts (U.S., Denmark) and an equally significant drop in labor share (fewer workers).

However, these managers do not increase their firms' productivity. Firms led by business majors are no more profitable and do not seize new market opportunities at a higher rate. Instead, they primarily capture a larger share of the value created by the firm.

Surprisingly, this pattern does not apply to non-business-educated managers. Non-business majors split the pie more fairly; business majors do not.

Why is this happening? The authors suggest that their findings reflect “the effects of practices and values acquired in business education—rather than the differential selection into business education of individuals unlikely to share rents with workers.”

I interpret this as: we may be teaching our students to become value-capture pirates; they didn’t start that way.

Value Capture and ‘Enshitification’

Prioritizing value capture over value creation is not just an academic concern. It manifests as declining quality, cost-cutting, and short-term thinking in the real world. Cory Doctorow has a name for this phenomenon: enshittification.

Ouch.

This pattern closely mirrors the way companies eventually shift from prioritizing value creation to value capture. Enshittification and the overemphasis on value capture are business strategies that, over time, erode an organization’s value proposition. An obsession with value capture can lead to missed opportunities (e.g., Intel and mobile chips) and eventual decline. However, I would argue—and the evidence is beginning to support this—that enshittification is not inevitable nor uniform across the economy. As the Acemoglu et al. paper suggests, it is a consequence of the values instilled through specific types of education.

Given the findings of these studies, it is clear that what we teach matters. Students leave our classrooms with new frameworks that shape their perspectives and influence their behavior in real-world companies. Providing them with different frameworks could lead to better outcomes for businesses and those they serve.

And, to be fair, business education likely has many positive effects (I would love references to high-quality studies on these effects). We must double down on them.

Perhaps we can also learn from the ideas of another great thinker, Joseph Schumpeter.5 Schumpeter viewed economic progress as the result of entrepreneurship and innovation by small and large firms, which lead to creative destruction and the replacement of the old ways with new ones that are better and more efficient.

Make no mistake: value capture is essential. But if everyone is fixated on capturing value, who’s left to create it? What value will then remain to be captured?

Maybe this is what happens when only the paranoid survive.

The entertaining video captures Folds composing and iterating with the orchestra on the piece in real time. The audience chooses the song’s parameters on the spot: it must be in A minor, have an upbeat tempo, and incorporate a sentence from that night’s program booklet: “These new spaces are all designed to be flexible.”

The result is extraordinary, and with nearly 10 million views, it’s clear people love it! But beyond that, it also showcases Folds’ mastery of improvisation, music theory, composition, and technical skills. He takes the challenge—A-minor, upbeat, “These new spaces are all designed to be flexible”—and, with some conceptual leaps and intuition, structures it into a surprisingly good song. Here’s the video.

Sitting on my couch today, could I create a pretty good song using AI in ten minutes?

In fact, with SUNO, I did it in just thirty seconds with little to no musical skill. The prompt: a male singer-songwriter with an orchestra performs an upbeat song in A minor, featuring the lyrics, “These new spaces are all designed to be flexible.” Listen for yourself.

Verdict: Not bad.

Fold’s brilliance is undeniable. And I’m not a musician. So, I can’t say much about whether the song SUNO produced is musically meaningful.1

Just as AI can mimic musical composition, some argue it will soon replicate PhD-level human thinking. But what exactly are they imagining?

This raises a fundamental question: What does it mean to think like a PhD? More importantly, what does it mean for AI to do so?

How do you get a PhD?

Not by taking a multiple-choice test.

Later this week (hopefully, depending on the snow), I’ll visit my PhD alma mater, Carnegie Mellon University, in Pittsburgh, PA, to give a research seminar. I arrived at CMU’s Heinz School (now College) as a 24-year-old master’s student in Public Policy and left at 29 with a PhD, ready to start (and struggle) in my first faculty position. Those six years in Pittsburgh were transformative. I left seeing the world in a completely new way—not just because of what we learned in the classroom but also because of how learning was structured and evaluated in the PhD program.

One thing many students don’t appreciate about earning a PhD is that they mistakenly believe it’s similar to a bachelor’s degree—or, at best, to the master’s they completed before entering the program. They assume you learn the material, pass an exam, and then, somehow, you’re magically equipped to do research. That mindset is a recipe for failure.

What makes a PhD fundamentally different from most degrees is its purpose, which differs from most degrees. Your role is not merely to apply knowledge but to create it. At Heinz, the PhD Handbook outlines three key criteria for achieving candidacy—none of which involve a conventional timed written examination.

1. Research Competence: A student must demonstrate expertise in methodology and subject matter through original research. Each paper must be methodologically sound and substantively rigorous, going beyond coursework. That is, Can we trust your results? Is the paper internally valid?

2. Flexibility: Students must apply distinct methodologies, work across different topics, or integrate multiple disciplines. Research must also demonstrate adaptability in analytical approaches and problem-solving. That is, Can you think beyond a single tool or context? Can you adapt?

3. Structuring Unstructured Problems: Students must identify and frame real-world problems in a way that enables rigorous analysis. This requires introducing new perspectives, applying novel methodologies, or redefining complex issues for deeper investigation. That is, Are they solving the right problem? Can they be analyzed?

Level (3), I think, can be broken up into two parts. The first is structuring an unstructured problem that has already been identified. But who decides what problems matter? That’s Level (4):

4. What problems are worth solving? This depends on many factors—most notably, your values, priorities, experience, preferences, capabilities, and context. Finding a problem worth solving is fundamentally a human task—one that reasonable, smart people (and sometimes not-so-reasonable ones) can vehemently disagree on.

Competence is a low bar.

In my opinion, competence is a low bar. If someone has already identified a problem, you don’t necessarily need a PhD to address it. Most senior Principal Investigators (those running research labs) don’t handle the “solving” aspect themselves; they delegate that responsibility to research assistants, PhD students, and postdocs. It’s obvious now that market returns for competence alone are rapidly eroding.

You don’t become a PI on an R01 grant simply by being competent or flexible, and maybe not even by structuring a previously unstructured problem alone. You need to do all of that together, but the problem you’re solving must also be significant2, and others must recognize it as such (how else will you secure funding to support your research?).

Beyond structuring problems, you must also structure their solution/implementation through people, technology, and systems.

This is the hierarchy of skills emphasized in rigorous PhD programs and the gauntlet of the tenure process at research schools (and beyond).

Why is this an upside-down pyramid with competence as the most minor component? Because there are millions (billions?) of problems worth solving. Each of these requires intelligence, flexibility, and structured approaches in various ways.

However, which problems are worth solving depends on who you are and your context. Our context windows are long and diverse, just like our training data. AI can optimize for the mean, but human intelligence thrives in variance: the blue-haired outliers, the unexpected, and the creatively structured solutions to problems no one recognizes.

AI Benchmarks and Shifting Goalposts

If being a great researcher is about more than just solving well-defined problems—if it’s about structuring the unstructured, making conceptual leaps, and deciding what’s worth solving—then how do we assess whether AI is intelligent?

I want to focus on one criterion that has come to dominate conversations around which “AI” is best or whether we’re making progress: the Benchmark.

Essentially, a benchmark in the context of AI is a standardized test or dataset used to measure and compare the performance of models on a specific task. Examples include ImageNet (used in image recognition research); the GLUE (and SuperGLUE) benchmarks for text; the Atari-57 benchmark for reinforcement learning; and many more.

In the context of “PhD level intelligence,” a few are worth noting. For instance, a team of researchers developed the GPQA (Google-Proof Q&A) benchmark. GPQA is an expert-authored, "Google-proof" benchmark with questions from biology, physics, and chemistry designed to test AI using tough questions. When they published the paper in 2023, GPT-4 could get about 39% accuracy. Today, we’re seeing 89% or higher levels with o3 (a proprietary model developed by OpenAI) and even the DeepSeek R1 model performing at nearly double the level of GPT-4: at 71.5%. Even if these estimates are somewhat upwardly biased due to data leakage, it’s still quite impressive. The questions are multiple-choice, with four options and a detailed explanation. In their paper, you can see some examples of questions in Table 1. Here is one quantum mechanics question.

Source: GPQA

This is the last exam. I promise.

Given this massive improvement of AI to solve this problem in just three years, researchers are again hunting for new benchmarks that AI can’t beat. I’m particularly impressed by a new one called Humanity’s Last Exam. It's hard, and I’ll probably get 0% right.

Here are the results as they appear on their website on Feb 14th, 2025.

Source: Humanity’s Last Exam

Frankly, I love the constant shifting of goalposts. It’s what makes humans so incredible—we’re never satisfied.

However, a key phrase at the bottom of the table is worth emphasizing: closed-ended.

These assessments are closed-ended, meaning the problem is clearly defined and has a known answer, which you can find at the "back of the book."

This distinction is crucial. In contrast, researchers—and many other professionals, including artists, musicians, entrepreneurs, consultants, CEOs, and team leaders—grapple with open-ended problems. The assessment of the solutions is not quantitative but subjective and open to interpretation.

This topic warrants a deeper dive, but for now, it’s essential to recognize that scalable benchmark “tests” are, by design, closed-ended, with ostensibly “right” answers. At their core, they measure competence and, at times, flexibility: Can I solve a broad range of problems well? However, the problems have been structured and are chosen by people who value different things (e.g., they care about different topics or phenomena).

My humble prediction? This won’t be Humanity’s Last Exam. AI will surpass it. The goalpost will move.

As I mentioned, the very nature of benchmarks—their need for scalability and cost-effectiveness—means they will almost inevitably remain closed-ended. As a result, they will fail to capture the kinds of heterogeneous problems many professionals face daily: the open-ended ones that have yet to be defined.3

Another issue worth noting is that while benchmarks drive effort and focus toward a singular goal, they can also lead to solution convergence—a narrowing of diversity due to the phenomenon of “studying to the test.”

A fantastic paper by Jane Wu at UCLA, titled Measurement for Dummies? Exploring the Role of Policy-Driven Measurement in Automotive Safety illustrates this well. The study showed that introducing the Side Impact Dummy (SID) as a regulatory measurement tool—a benchmark of sorts—reduced overall fatalities and narrowed firm innovation. Automakers optimized safety for SID-sized occupants, limiting diverse solutions that could have benefited people with different body sizes.

A final point about the tests we design for AI: What might we overlook in AI's capabilities if we assess them solely based on their rankings in closed-ended tests? What multitudes might these technologies hold if we stepped back and asked which problems are truly worth solving with them?4

Cartesian, Lockean, and Phenomenological Work

In short, the skills required to earn a PhD are qualitatively different from the metrics used to evaluate AI models. While AI benchmarks test performance on predefined, closed-ended problems, PhD-level thinking involves identifying which problems are worth solving and tackling those open-ended challenges.

This distinction brings me to my final point in this post. As some readers might know, I was an undergraduate philosophy major. One of my classmates (25 years ago!) jokingly called the philosophy department where I studied the "Department of Descartes." There were always ongoing debates between the rationalists (Descartes), the empiricists (Locke), and the phenomenologists (e.g., the American pragmatists). I gravitated toward the lone phenomenologist in that department, Bruce Wilshire, with whom I took Existentialism and American Pragmatism classes and conducted a few independent studies. Many key ideas in phenomenology trace their heritage to Heidegger, Husserl, and the American pragmatists, including James and Peirce.

The debates between the Cartesians, Lockeans, and phenomenologists had faded from my thoughts—replaced by worries about endogeneity—until I read The Philosopher of Palo Alto by John Tinnell. It is a fascinating biography of Mark Weiser, the father of the “Internet of Things” or “ubiquitous computing.” Weiser was Xerox PARC's Chief Technology Officer and a lapsed philosophy major himself.

One key distinction from the book, which I think is worth bringing into this discussion, is the difference in how they view knowledge (epistemology) and reality. (metaphysics)

• René Descartes (Rationalism): Defined existence through cognition: “I think, therefore I am.”

• John Locke (Empiricism): Argued the mind begins as a “tabula rasa”, shaped entirely by experience (sensation) and then reflection.

• Phenomenologists: Reframed existence as fundamentally embedded in the world: “I am because I exist in the world.”

A Cartesian view holds that rational thought and control are at the core of our identity. The Lockeans, in contrast, saw the mind as a blank slate, shaped by sensory experience and learning over time, emphasizing adaptability and accumulation of knowledge. The phenomenologists, by contrast, saw our essential nature as being embedded in experience, practice, and the material world. I don’t want to argue that one type of intelligence is better or worse or claim to know which defines our essential nature. However, these distinctions are important and may be useful for understanding what AI will do in the future and what we will do.

Identifying problems worth solving is fundamentally phenomenological work: it requires immersion in experience, attention to context, and an appreciation of how problems manifest in the real world.

LLMs, on the other hand, engage in remarkable Lockean reflection (with a bit of Cartesian structure). They excel at induction, inference, reasoning, and prediction from data but operate far from lived experience and embodied understanding.

These Lockean workers’ window into the world, and even their interpretation of it, occurs through our words, our experiences, and the tools we have built to make sense of our reality in a meaningful way (e.g., through the photos we take, the art we make, etc.).

And to get a little Sartre in here: we create our purpose.

Ben Folds vs. AI Revisited

With these ideas in mind, it is worth revisiting the example of Ben Folds and the AI-generated song.

We see a perfect illustration of phenomenological work when we watch Ben Folds compose a song live in just 10 minutes. Folds didn’t sit down and analytically break the task into discrete logical steps. He engaged directly with the orchestra and the audience, adjusting in real-time and making intuitive leaps that couldn’t have been pre-planned. His expertise wasn’t just in his head—it was in his hands, his ears, and his deep experience interacting with the material and people and of music itself. His purpose and the problem he was solving were also different. It wasn’t about composing a song. It was about creating an unforgettable experience for everyone in that room.

While the AI's output was interesting and impressive in its own right, it felt different. It was Lockean work. The algorithm took the prompt I provided, broke the task down into known components, applied data-driven reasoning and rational processes to solve each step, and optimized the performance based on what it was trained to think sounded good to us. 

I know what kind of work excites me. But a Lockean sidekick wouldn’t hurt.

Held Together by Helpfulness

When I was in graduate school, my mom’s uncle used to call me once a month. He’d begin with his characteristic “Asalamwalekum beta” (peace be upon you, son) and then ask how I was doing, whether I was praying, and if everything was OK. It was usually a short conversation, but I always looked forward to it. Recently, when I talked to my cousins about Nana, I learned that he did this with lots of people. It wasn’t only these “check-in” calls—he felt responsible for, and took seriously, other aspects of the family and its well-being. For instance, we have a ROSCA in the family, and for decades, he managed the dozens of monthly checks and frantic requests for money that came in. He was always ready to take a late-night call so that funds would be available for an urgent tuition payment, a health emergency, or to replace a broken-down car.

Nana didn’t just make calls; he created a web of support people could rely on.

Organizations often overlook this invisible glue, yet it’s the hallmark of companies that function and flourish. When team members help each other—sharing tacit knowledge, unblocking progress by providing resources, and offering creative insights purely out of a desire to support each other—the collective outcomes become far more significant than any individual could achieve alone.

While a culture of helpfulness can drive connectivity and innovation, it isn’t immune to the natural pull toward efficiency and self-interest—a transactional drift, so to speak.

Over time, organizations may shift away from a relational view of themselves as individual imperatives gradually replace the communal commitment that unifies people. Yet the power of the helpful act—supporting others without expecting anything in return—remains a potent antidote.

Relationships versus transactions

In the early part of my academic career, I spent a lot of time researching social networks. I always found it curious that a meaningful chunk of that literature is organized around a zero-sum, almost formulaic, framing: How can individuals leverage their networks to get ahead of others? “Your networks” are your “competitive advantage,” a means to an end for outperforming the competition.

As I’ve gotten older, I’ve come to view this framing with deep suspicion. It carries the seeds of what ultimately undermines networks: it turns relationships into transactions.1

But, stepping back, what is the difference between a transaction and a relationship?

Are organizations just bundles of transactions?

From an economic perspective, organizations can be seen as bundles of transactions—exchanges between people, teams, and departments—all aimed at producing and selling products or services.

But what exactly is a transaction? At its core, it’s simple: I give you X, you give me Y. Each party knows what they’re offering and what they’re receiving. If the terms aren’t mutually beneficial, the transaction doesn’t happen or gets renegotiated.

This is a perfectly reasonable model of organizations, and many organizational activities can be viewed fruitfully through the lens of transactions.

Organizations are also bundles of relationships.

However, the transactional view of organizations is somewhat confining. Transactions aren’t all that organizations are. Organizations also consist of relationships—and these are qualitatively different from transactions. A relationship isn’t governed by “If I give X, you give Y,” but by what’s appropriate in the context of that relationship—by your responsibilities as a friend, teacher, or leader.

While transactions offer efficiency, they are not intended to nurture the deeper, less measurable aspects that relationships provide.

Relationships cultivate trust, empathy, safety, shared purpose, responsibility, and resilience—elements that can’t be easily measured or reduced to simple if X, then Y exchanges. Without relationships, organizations risk becoming efficient but weak, optimized for short-term gains at the expense of long-term prosperity, innovation, and adaptability.

What is transactional drift?

Of course, transactions and relationships aren’t discrete or mutually exclusive; instead, they represent a spectrum that can help us better understand which pole our organizations are moving toward.

Transactional drift refers to an organization’s slow—and often imperceptible—shift toward connections that are increasingly built on if give X, then get Y logic rather than what is appropriate for the relationship. I’ll also add that what is considered “appropriate” is likely shaped by norms that can shift over time. As these norms evolve, we may gradually view a transactional frame as normal for a relationship—even though it was different in the past.

Transactional drift and what we can do to fight it

I’ve been trying to wrap my head around “transactional drift” for some time now. While it isn’t a named phenomenon, I’ve found some research especially useful in helping me understand how it unfolds and its impact. I’ve tried to think about two things. First, what forces drive transactional drift inside our organizations? Second, what can we do to resist it?

Incentives and the Dilemma of Sharing

At work, we are confronted with a myriad of incentives2. Promotions, raises, that new book deal—we have many carrots (and sticks) that kindle our strategic juices and cause us to choose how we allocate our time, channel our creativity, and, most importantly, change how we both view and relate to others.

My colleague Wes Cohen at Duke and his collaborator John Walsh at Georgia Tech have each written about a fascinating phenomenon crucial in scientific innovation: the sharing3 of technology and materials and how incentives affect this behavior.

In one paper, Walsh, Cohen, and Cho (2007) find that a shift from producing science to commercializing it leads researchers to share significantly less with their colleagues—not by withholding knowledge itself but by restricting access to critical materials and data that were once more freely exchanged within the scientific community. The net impact on the community is likely that many ideas that could have been pursued if materials were available are no longer explored by anyone, as researchers are increasingly concerned about how the broader use of their resources might affect their slice of what looks like a limited pie.

In another paper, Shibayama, Walsh, and Baba (2012) show that as academic entrepreneurship grows, scientists shift from freely sharing resources within the community to more transactional exchanges, where people expect something in return and share less overall. This paper's most critical point is that this shift isn’t limited to scientists who are directly involved in entrepreneurial activities; it spills over to everyone, showing how the push toward commercialization may reshape the culture of science itself.

A final paper worth noting: Bandiera, Barankay, and Rasul (2005) find that when firms implement relative compensation schemes (e.g., paying employees based on their rank rather than their output), workers dramatically reduce their productivity to avoid harming their colleagues—especially when their colleagues’ output is easily observable. People are sensitive to the impact they have on others, but this sensitivity depends on their ability to internalize the costs they might impose, which likely stems from their relationships and their beliefs about their responsibilities to their colleagues.

The Impact of Market Substitutes for Social Connection

High-powered incentives aren’t the only way to weaken relationships. While relationships are complex, they often serve as crucial sources of essential resources—money, advice, etc. A recently published paper by Banerjee et al. (2024) uses a natural experiment and an actual RCT to study the impact of providing Indian villagers with access to formal microfinance (a financial product) on their social networks. They found something very interesting: not only did microfinance substitute for lending relationships, but it also weakened social networks more generally—even for those unlikely to borrow. This finding is quite striking (and similar to the Shibayama et al. one above) and suggests that the substitution effect of formal financial systems can have unintended and perhaps far-reaching effects on community relationships and risk-sharing mechanisms.

In other words, microcredit weakened social relationships in the village. I don’t want to say that access to microfinance is bad—it’s actually quite good because it provides credit to people who might not have gotten it before and perhaps reduces the power of certain groups who wield it over others. But there was a trade-off: the access to credit wasn’t “free,” and the transactional drift occurred quite rapidly.

We have likely seen similar substitutions in other contexts. Putnam believes that television may have caused some erosion of America’s social capital. There has been considerable discussion about smartphones and loneliness. And now, what will Generative AI (the omnipresent advisor) substitute for in our work relationships?

Specializing in People Drives Performance

While incentives matter, how we structure work also matters in balancing whether our connections are transactional or relational.

There’s fascinating work on this topic.

One of my favorite papers is Freyer’s (2018). While the paper isn’t explicitly about relationships, its underlying mechanism surely is. The experiment showed that when classrooms were organized to emphasize teachers who “know the subject” rather than those who “know their students,” student achievement suffered across the board. This suggests that a teacher who builds relationships with their students outperforms subject specialists who teach as if they worked on a student assembly line. Simply put, specializing in people drives the effectiveness of teachers and their students.4

This effect of “specializing in people” isn’t unique to education—we see it in other contexts. Chen (2021) finds that when proceduralists and physicians share experience working together, patient outcomes improve—boosting survival by cutting 30-day mortality by 10–14% and reducing resource use. These physicians specialize in each other: it’s relational, not transactional.

Sabtey (2023) finds that when patients lose a longstanding relationship with their primary care physician, adverse events spike—with mortality rising by 4%, emergency visits by 4%, and hospital admissions by 3%—underscoring the unique health benefits of enduring patient-PCP relationships. Similarly, Schwab (2025) finds that when a longstanding primary care–patient relationship dissolves, costs rise by 3–5%—primarily due to increased specialty care usage—without any observable improvement in patient outcomes, highlighting the hidden costs of losing relationship-based care.

These studies suggest that a key form of tacit knowledge in the economy is our knowledge about each other. This knowledge isn’t incidental but critical to driving the high performance of our essential social institutions: our schools, hospitals, and likely the companies tasked with creating economic progress. What happens when this knowledge is lost?

Helpful People Strengthen Networks.

What can be done to counteract these forces of transactional drift?

My friend and co-author, Alex Oettl at Georgia Tech, wrote a powerful paper studying the importance of helpfulness on the scientific contributions of research immunologists (scientists who study the immune system and its responses). He found something quite interesting: yes, collaborating with “stars” (i.e., highly productive peers) boosted scientists' productivity, but collaborating with highly helpful stars was what actually drove the productivity boost.

Who counted as helpful? The people who openly shared advice and guidance even without explicit rewards like co-authorship. These helpful stars were relational, not transactional.

In 2022, Alex, our friend Samsa Samila from IESE, and I published a paper titled "Helpful Behavior and the Durability of Collaborative Ties."

In it, we studied the impact of helpful people on the resilience of collaborative networks among research scientists (again, immunologists). Our findings were both surprising and telling.

We found that having a helpful collaborator makes you more helpful yourself. More importantly, the ripple effect of helpfulness fostered a culture of mutual support, strengthening the surrounding collaborative network.

As a result, collaborations with helpful individuals tended to be more durable, persisting even after the unexpected loss of the original helpful collaborator. In contrast, similar collaborations with highly productive but less helpful individuals were more likely to dissolve in their absence, highlighting how the spread of helpfulness enhanced both the resilience of specific collaborations and the broader scientific ecosystem.

The point is that helpfulness has ripple effects and might be a simple solution to transactional drift. It’s easy and impactful, though perhaps not entirely “free.” You do have to give things without expecting something in return. However, the reward could be an organization you enjoy being part of—and one you’re proud of.5

The end

Alright, that wraps up today’s deep dive into relationships, transactional drift, and the power of helpfulness. However, we’ve only begun to explore how to design thriving organizations and what organizational changes you might want to reconsider. Ask yourself: Am I building a relational organization or a transactional one?