From industrial construction to the AI age: a century-spanning reinterpretation of Harvard Business School’s campus under construction. Source: Baker Library Historical Collections, 1926 archival image. Creative adaptation by Manus using Nano Banana.

In my classroom this past week, I watched the AI conversation change in real time. The tone had shifted. Not long ago, many students spoke about OpenAI and ChatGPT as the default center of gravity in generative AI. But over the last week, and in many of the conversations that happened after class and offline, the mood had noticeably moved toward Anthropic and Claude. Students were comparing outputs, swapping stories about what worked better for writing or analysis or coding, and, with the confidence that only a fast-moving technology market can generate, beginning to declare winners.

What struck me was not that the students were wrong to compare tools. Of course they should. They are using these systems every day, and the differences in quality, style, reliability, and fit for purpose are real. What struck me instead was how quickly the discussion slipped into the language of a settled contest: who is ahead, who is behind, who has momentum, who is fading. It is the natural vocabulary of a leaderboard, of a sports table, of a market in which the relevant dimensions are already known and the only remaining question is who occupies the top spot.

But that may be exactly the wrong way to think about what is happening.

1926

If we step back a hundred years, to 1926, I suspect we would have heard something similar around the automobile industry. People would have been tempted to ask which company was ahead, which design was superior, which production method looked ascendant, which firms had figured out scale, and whether the field was finally settling. And yet, viewed through the lens of industry evolution, that was still a world of ferment: competing designs, shifting cost structures, uneven infrastructure, uncertain complements, and a market still discovering what it was going to become. My instinct is that foundation models today look much more like that 1926 moment than like a mature race with obvious and durable winners.

Going back to 1926 is useful for another reason. Harvard Business School was then only eighteen years old, still what Jeffrey Cruikshank — the historian of HBS — would later call “a delicate experiment.” The Allston campus was taking shape on what HBS archival materials describe as a “formerly marshy area,” with utilities carried through the “swampy area around to the Business School,” and a 1925 Harvard Business School Bulletin account of the groundbreaking described “the hissing of steam and grinding of mesh of gears” as a steam shovel took its first bite of the site. In other words, HBS students in 1926 were not just studying industrial transformation; they were living inside it. They were entering a school that The Crimson described as training “future financial wizards” just as the automobile industry was remaking markets, careers, cities, supply chains, and managerial ambition.^1,2,3,4

It is easy to imagine those students asking questions very much like the ones my students ask now: Which companies are really going to matter? Which technologies are durable and which are hype? Where should I build my career if the ground is still shifting under everyone’s feet?

Two Guides for an Emerging Industry

This is where two scholars I have long found especially clarifying come in: James Utterback, the MIT innovation scholar, and Steven Klepper, the CMU economist. Both have passed on, but their work remains indispensable for making sense of moments like this. MIT Sloan’s memorial to Utterback remembers him as “a pioneering scholar of technological innovation,” while Carnegie Mellon’s memorial to Klepper describes a thinker who combined entrepreneurship and mainstream economics to study innovation.^5,6 In different ways, both taught us that the early life of an industry is noisy, experimental, and easy to misread from the inside.

In Utterback’s world, this is the period before a dominant design has fully emerged, when product innovation is intense and the field remains fluid. In Klepper’s world, this is the period before the deeper industrial logic has fully played out, when entry is high, capabilities are uneven, and the eventual shakeout is not yet visible to those living through the excitement. So before asking whether Claude is “winning,” whether OpenAI is “losing,” or whether open models are “catching up,” it helps to look first at what these scholars gave us. They remind us that what looks like a clean horse race in the moment may actually be the earliest and least informative stage of a much longer industrial story.

Ferment Before Dominant Design

Utterback helps because he gives us a language for separating noise in the moment from pattern over time. In his classic work with William Abernathy, the point was not simply that industries innovate, but that they do so in recognizable stages.⁷ Early on, an industry tends to be in a fluid phase: firms experiment with product concepts, architectures, interfaces, and uses; uncertainty is high; and the basis of competition is still unsettled. Over time, that fluid phase gives way to a transitional phase, in which some design choices begin to stabilize and firms increasingly learn what customers value, what complements matter, and what scale really requires. Eventually, if the industry matures, a specific phase emerges, where improvement becomes more incremental, process discipline matters more, and competition shifts toward efficiency, cost, and execution. The key insight is that the meaning of “who is ahead” changes across those stages.

His central idea is straightforward but powerful: early in an industry’s life, firms do not yet know the final form of the product, the most important features, the right production method, or even the most meaningful definition of value for customers.⁷ Only later does a more stable dominant design emerge — and when that happens, competition changes character. The question stops being mainly what should the product be? and becomes increasingly who can deliver it more reliably, cheaply, and at scale?

The automobile industry is one of the clearest ways to see what Utterback had in mind. Looking backward, it is tempting to tell a neat story in which the winners were always visible. But from inside the period, the industry was far messier. Designs differed, production approaches differed, user needs were still being discovered, and no one could yet be certain which combination of performance, price, reliability, and manufacturability would define the market.

The analogy is not perfect, and it is worth being honest about that. Automobiles were physical products whose shakeout dynamics were partly driven by manufacturing scale economies — the enormous capital costs of tooling, assembly, and distribution that only a few firms could sustain. Foundation models are software, with marginal costs that collapse toward zero and distribution that is near-instant. The scale economies in AI are real — compute, data, and research talent are genuinely scarce and expensive — but they operate differently. The historical parallel is a frame for thinking, not a precise prediction. Its value is in the pattern it reveals: that ferment precedes convergence, and that convergence takes longer and looks messier from the inside than it does in retrospect.

This figure from Utterback and Suarez captures the underlying logic visually. Notice that Trajectory B — the one that does not produce a dominant design — fans out into ever-more-divergent branches, generating variety without convergence. Trajectory A, by contrast, eventually narrows: competing sub-designs get selected out, a dominant design emerges, and the surviving branches organize themselves around it. The crucial point is that from inside either trajectory, the early picture looks similar — branching, experimentation, no obvious winner. The difference only becomes legible over time, and only in retrospect does one trajectory look inevitable. This is exactly the visual to keep in mind when reading today’s AI benchmarks: we do not yet know which trajectory we are on.

Foundation models still look very much like that earlier stage. The field is not just competing on one stable product. It is experimenting simultaneously with underlying architectures, product forms, bases of performance, business models, and the complementary assets needed to make any model economically and organizationally durable. Core architecture alone encompasses dense versus mixture-of-experts models, different post-training stacks, reasoning-time compute, retrieval-augmented systems, and renewed interest in world-model ambitions. The product form question is equally unsettled: are we building chat assistants, coding agents, research agents, embedded copilots, multimodal assistants, or open models for local deployment? The performance basis — raw reasoning, coding, tool use, latency, context length, safety, reliability, cost — has not yet converged either. Almost every important layer of the sector is still in play.

That is why today’s landscape does not yet look stabilized. It looks like a classic pre-dominant-design environment: noisy, innovative, fast-moving, and full of meaningful differences — but still far too unsettled to be read mainly as a simple race chart.

“The pattern of innovation over time can be described in terms of three phases: fluid, transitional, and specific.” — Abernathy and Utterback, Patterns of Industrial Innovation⁷

Entry, Learning, and Shakeout

Klepper helps in a different but equally important way. Where Utterback asks us to pay attention to design evolution, Klepper asks us to watch the population dynamics of the industry itself.⁸ His core question was why so many industries show a familiar pattern: entry rises early, the number of firms peaks, a shakeout follows, and a smaller set of firms comes to dominate. What matters, in this account, is not only whether a dominant design emerges, but whether some firms build stronger capabilities for innovation, growth, and cost reduction than others.

Again, the automobile industry is the canonical case. The early industry was crowded with entrants, experimentation, and local successes. But over time, the number of producers fell — not simply because observers finally noticed who was “best,” but because firms with stronger capabilities and better positions in the industry’s evolving economics were more likely to survive and grow.⁸ The shakeout, in this view, is not an accident or a sudden verdict from the market. It is the result of cumulative differences playing out across cohorts of firms over time.

This figure makes the pattern visceral. Look at the AUTO curve: from a handful of producers in the 1890s, the industry climbs to nearly eighty firms around 1920, then collapses to a fraction of that over the following two decades. The TV industry repeats the same arc fifty years later, peaking above eighty participants in the early 1950s before a swift and brutal shakeout. Tubes, typewriters, transistors, calculators — each industry shows the same basic shape: rise, peak, fall. What is striking is not just the shakeout itself but how long the peak lasts. For automobiles, the crowded period runs for nearly twenty years before consolidation becomes unmistakable. Anyone living through the 1910s who declared a definitive winner would almost certainly have been wrong. The same humility seems appropriate now for anyone declaring the AI race settled.

Klepper’s own data on the automobile industry makes the same point with even more precision. His figures, drawn from census records and firm survival data, track the automobile industry from its origins in the 1890s through the 1960s — and the picture they show is worth sitting with carefully.

Two things are happening simultaneously in these charts, and their combination is the core of Klepper’s insight. The top panel shows automobile output growing continuously and steeply throughout the entire period — by the logarithmic scale, production is multiplying many times over from 1895 to 1940. The industry was not struggling. It was booming. And yet the bottom panel shows the number of producers doing something entirely different: rising rapidly to a peak of roughly 275 firms around 1910, then falling sharply and persistently for decades, converging toward a tiny handful of survivors by the 1960s. The industry got vastly larger while the number of firms got vastly smaller. That is the shakeout — and it unfolded not because demand collapsed, but because capability differences between firms compounded over time until weaker players could no longer survive.

Notice also the entry and exit lines in the bottom panel. Entry is high and volatile in the early years — the 1900s are a period of intense experimentation, with firms constantly entering and leaving. Then entry drops off sharply after about 1910, while the total number of firms continues to decline through attrition. The industry stops attracting new entrants long before it finishes consolidating. This sequence matters: by the time an outside observer might have noticed that the shakeout was underway, the window for new entry had already mostly closed.

The parallel for foundation models is thought-provoking. We are likely still in the high-entry phase — the equivalent of the 1900s in Klepper’s automobile data. Entry is high, experimentation is intense, and no one has yet been conclusively selected out. It is important to hold two observations separately here. At the architectural level, the field remains genuinely open: we do not yet know whether transformers will prove to be the dominant design or one branch among several that the industry eventually selects among. But at the firm level, a different dynamic may already be operating quietly underneath the noise — some organizations are accumulating compute, distribution, and organizational learning at a pace that will be hard to replicate later. Architectural openness and competitive asymmetry can coexist in the same moment. The market can still be discovering what it is while some players are already building positions that will matter when it does.

That lens matters for foundation models because it pushes us to ask a more structural question than who currently has the most impressive model. Even if the field eventually converges on something like a dominant design, the deeper industrial question is which organizations can repeatedly finance training runs, secure compute, attract research talent, build data advantages, improve inference economics, create distribution, and turn technical progress into organizational learning. Klepper would tell us to watch entry, survival, exit, and capability accumulation — not just rankings.

It also allows a more subtle claim than simply saying that nobody knows who will win. The more precise point is that two things can be true at once. First, we are still early enough that variety remains high and the field is visibly unsettled. Second, the foundations for future concentration may already be forming underneath that visible turbulence. Some organizations may already be building advantages in deployment, distribution, enterprise trust, model operations, and reinvestment capacity that will matter long after today’s benchmark debates are forgotten.

The eventual shakeout may also happen at a different layer than people expect. The end state may not be a single “best model” crowned in the abstract. Instead, concentration may occur across several layers of the stack: one set of firms may control the frontier layer, another may dominate enterprise integration, another may anchor the open ecosystem, and others may own specialized complements such as cloud access, chips, tooling, and safety infrastructure. The future structure of the industry may look less like one winner taking all and more like a narrower and more durable set of positions across interconnected layers, each shaped by different forms of capability accumulation.^8,9

What Pundits Get Wrong

One of the most common mistakes in commentary about AI is treating the market as if it were already mature. Leaderboards, viral benchmark charts, and week-by-week declarations of who is “ahead” all encourage the same frame of mind: they imply that the central problem is rank-order competition on a known and stable axis of performance. That is how we talk about a settled category — processors once the key dimensions are understood, airlines once route economics are clear, enterprise software once category boundaries have hardened. But that is precisely the wrong mindset for an industry still in ferment. In an emerging sector, the hard question is not only who performs best on a metric today. The harder question is which metrics, product forms, complements, and organizational capabilities will ultimately define leadership at all.^7,8,9

When someone says one company is “winning,” it is worth asking: winning at what, exactly? Winning on benchmark scores? On coding? On inference cost? On consumer mindshare? On developer adoption? On enterprise penetration? On openness? On architectural ambition? Each of those may matter, but none of them by itself is enough to settle the industrial question. In a period of ferment, different firms can appear dominant on different dimensions because the sector is still testing which dimensions will eventually matter most. The instability of the scoreboard is not a side effect of confusion. It is a clue that the basis of competition is still being discovered.

A firm that looks strongest on raw intelligence may be weak on distribution. A firm that wins consumer mindshare may struggle in enterprise workflows. A firm that drives costs down may depend on upstream breakthroughs it does not control. A firm with the most compelling long-run architectural thesis may fail to turn that thesis into a repeatable commercial system. The right conclusion is not that current differences are meaningless — they are real, and they matter. But they should be read as signals inside an unsettled evolutionary process, not as final standings in a completed race. The industry is still contesting not just who wins, but what winning even means.

Why This Industry Is Still So Early

One reason I feel so strongly that this industry is still early is that several of the core bottlenecks are still moving at the same time. In a more settled sector, firms may compete intensely, but they do so against a relatively stable background of costs, constraints, interfaces, and rules. That is not what we have here.

Training and inference economics are still changing quickly. Cost structures remain sensitive to power and infrastructure constraints. High-quality training and feedback data remain contested and unevenly distributed. The field is still discovering new ways to improve reasoning, efficiency, multimodality, and agency. We still do not fully know what the enduring product form of AI systems will be. And rules for access, copyright, liability, and deployment are still unsettled. These are not peripheral variables sitting at the edge of the story — they are the story. Together they keep experimentation alive because firms are not merely optimizing within a fixed game. They are still helping define the game itself.^7,8,9

Complicating all of this further is what we are already seeing from product and application companies. Their emerging patterns of use suggest that the future may not belong to one LLM ruling the world, but to multiple models specializing across different parts of a task chain. In practice, work is often decomposed into steps — ideation, drafting, retrieval, coding, checking, transformation, formatting, and delivery — and different models may be better at different pieces of that sequence. You can already see hints of this logic in products like Gamma and Replit. What matters in those settings is not simply whether one underlying model is universally superior, but whether the product can assemble the right intelligence for the right subtask at the right moment. If that becomes the norm, then the industry’s long-run structure may be even more layered than the simple horse-race framing assumes: competition among frontier models, yes, but also orchestration across specialized models and differentiated control over the workflow itself.

The World-Model Debate Is Not a Refutation

This is where the current debate becomes especially revealing. When prominent researchers such as Fei-Fei Li and Yann LeCun argue for world models, embodied intelligence, or richer representations of physical and causal reality, the temptation is to read that as a refutation of the LLM wave. I think that is the wrong interpretation. Fei-Fei Li’s recent argument for spatial intelligence does not deny the remarkable gains of large language models and multimodal systems; rather, it treats them as important but incomplete steps toward more capable systems that can reason about geometry, dynamics, and the physical world.¹⁰ LeCun’s continued insistence that human-level AI will require stronger models of the world points in a similar direction.¹¹

Seen through the historical frame of Utterback and Klepper, this is exactly what one should expect in a period of ferment. The industry is still experimenting: some actors want to scale current architectures further; others want to add tools and agents; others emphasize multimodality; others push open-weight diffusion; and still others want to move toward world models and embodied intelligence. In Utterback’s terms, these may not be competing visions of who wins the current race — they may be alternative branches in the design hierarchy, each representing a genuinely distinct trajectory for what AI systems eventually become. The world-model argument may turn out to be the branch that dominates, or it may be the branch that doesn’t. We don’t yet know. What its very existence tells us is that the design space has not closed — multiple plausible trajectories remain live, and the eventual basis of competition is still being discovered.^7,8,9

Where We Actually Are

If I had to offer a provisional answer, it would be this: foundation models appear to be in a late fluid stage, perhaps moving toward an early transitional stage, but still far from the kind of stability that would justify strong claims about a settled dominant design.⁷

There are indeed signs of convergence. Most serious systems now build on a recognizable family resemblance: transformer-based architectures, instruction tuning, reinforcement or preference-based post-training, increasing multimodality, external tool use, and some movement toward agentic orchestration. That is not trivial — it means the industry is not inventing itself from zero each month. But convergence on a toolkit is not the same thing as convergence on a final design. And if we are honest, we do not yet know what a dominant design for AI would even look like. The pace of change over the past three years alone — from GPT-3 to instruction tuning to multimodality to reasoning models to agents — has been fast enough to repeatedly displace what looked like emerging consensus. The deeper questions remain genuinely open: what the enduring product form will be, which performance dimensions customers will reward most, which complements will prove decisive, and which organizations can convert technical movement into durable industrial position.

That is also why it is too early to speak as if a definitive shakeout were already underway. Capability asymmetries are certainly forming. Some firms have more compute, better talent density, stronger enterprise relationships, wider distribution, or faster product-learning loops than others. Those asymmetries matter — they may turn out to be the building blocks of future concentration. But Klepper’s lesson is that early asymmetries are not identical to mature structure.^8,9 An industry can display meaningful leaders and still remain open to major reordering as bottlenecks shift, complements harden, regulations settle, and demand reveals itself more clearly.

The most disciplined answer is therefore a conditional one. Yes, current leaders matter. Yes, emerging clusters of design practice matter. Yes, some firms are already building stronger positions than others. But the most important variables are still the moving ones: bottlenecks, complements, adoption patterns, cost curves, regulatory regimes, and the criteria by which users and organizations will eventually decide what counts as good enough, safe enough, cheap enough, and valuable enough. The right way to describe the field is not immature chaos, but structured ferment. By that I mean something specific: a condition in which experimentation remains wide open at the level of architecture and product form, while underlying selection pressures — around compute access, distribution, enterprise trust, and the ability to convert research into shipping systems — are already beginning to sort firms into stronger and weaker positions. It is not chaos because there are already visible patterns and accumulating asymmetries. It is not settled because the criteria by which the market will ultimately reward some positions and punish others have not yet hardened. We are no longer at the beginning of the beginning — but we are still early enough that the future industrial order should be treated as a question, not a conclusion.^7,8,9

This Is the Normal History of Important Technologies

At the broadest level, that is the real point of this essay. It is not merely to correct the discourse around AI, though I think that correction matters. It is to remind us that historically important technologies almost always look noisy, unstable, and hard to interpret in their formative years. The confusion is not evidence that nothing meaningful is happening. More often, it is evidence that a great deal is happening at once: technical possibilities are expanding, market categories are still forming, complements are still aligning, and organizational advantages are only beginning to compound. What feels bewildering in the present often looks legible only in retrospect.

That is why the history from Utterback and Klepper is so clarifying. Their work teaches us to expect ferment before standardization, rivalry before hierarchy, and experimentation before shakeout.^7,8,9 The mistake is not that observers notice movement. The mistake is that they too quickly translate movement into finality.

So when people ask who is winning the model wars, the best response is not to dismiss the question, but to deepen it. Ask what kind of industry this is becoming, what bottlenecks are shaping it, what forms of capability are compounding inside it, and what selection environment will eventually reward some positions and punish others. Through the eyes of Utterback and Klepper, the model wars look less unprecedented than they feel — more like a familiar chapter in the history of technical evolution: a period of ferment before the market knows what it will eventually stabilize around.^7,8,9

What This Means If You Are Making Career Decisions Right Now

If that historical frame is right, then the practical implication is not paralysis — it is a different kind of preparation.

The equivalent lesson for students in 1926 would have been instructive but humbling. In retrospect, betting on Ford or General Motors would have been right. But the identification problem at the time was genuinely intractable — there were over two hundred producers, the basis of competition was still shifting, and the signals that would eventually distinguish the survivors from the also-rans were not yet legible. The lesson is not that picking winners is a bad strategy. It is that in a period of ferment, the winners are usually not identifiable in advance with the confidence that observers tend to project onto the moment. The students who thrived were more likely those who understood how automobiles were changing the economics and organization of work — and who built judgment that would remain useful regardless of which nameplate ended up on top.

For my students now, the lesson is the same in structure if not in detail. In a period of ferment, the safest strategy is rarely to bet everything on a single firm, model, or weekly narrative of who is up and who is down. A better strategy is to build skills, judgment, and adaptability that travel across scenarios. The people who will do well are likely to be those who can translate across technical and managerial worlds, identify where complements matter, help organizations adopt these systems responsibly, and keep their bearings when architectures, interfaces, and business models keep changing.

That means choosing roles and organizations that maximize learning, exposure, and problem-solving opportunities rather than simply chasing whoever appears to be winning this quarter. Get close to the frontier — use the tools, experiment with them, understand their strengths and limitations, and learn how AI is changing the economics of work in real settings. But don’t confuse proximity to the frontier with dependency on a single platform.

The HBS students of 1926 couldn’t know whether they were inside a delicate experiment or the beginning of something durable. It turned out to be both. The students asking the same questions now are in a structurally identical position — more information, no less uncertainty. That is not a failure of analysis. It is what living inside a period of ferment actually feels like.

Note: Made in close collaboration with Manus and Claude.

References

1. The Campus Emerges — Baker Library Historical Collections, Harvard Business School

2. A, B, C, Allston — The Harvard Crimson, April 27, 1926

3. Harvard University: The Business School Club — The Harvard Crimson, September 29, 1926

4. Bibliography — Georges F. Doriot: Educating Leaders, Building Companies, Baker Library, HBS

5. MIT Sloan Remembers Professor James Utterback

6. Obituary: Carnegie Mellon’s Steven Klepper — Carnegie Mellon University

7. Abernathy and Utterback (1978), Patterns of Industrial Innovation

8. Klepper (1996), Entry, Exit, Growth, and Innovation over the Product Life Cycle

9. Klepper (1997), Industry Life Cycles

10. Fei-Fei Li, From Words to Worlds: Spatial Intelligence is AI’s Next Frontier

11. WIRED, Yann LeCun Raises $1 Billion to Build AI That Understands the Physical World

That was the question we set out to answer.

The study

In collaboration with the Boston Consulting Group — a partnership we are grateful for — we designed a pre-registered randomized experiment with 758 BCG consultants. These are highly educated, highly motivated knowledge workers performing tasks representative of their actual professional responsibilities. They were randomly assigned to one of three conditions: no AI access, access to GPT-4, or GPT-4 access with a brief prompt engineering overview.

The critical design choice was to test two distinct types of tasks: one set designed to sit inside GPT-4’s current capability frontier, and one designed to require exactly the kind of contextual, integrative judgment that we expected would fall outside it.

The results were stark in both directions.

For tasks inside the frontier — spanning creativity, analytical thinking, writing, and persuasion — participants using AI completed 12.2% more tasks, completed them 25.1% faster, and delivered solutions of substantially higher quality, with average scores rising roughly 30% above the control group. Critically, lower-skilled workers gained the most, with quality scores rising 43% compared to 17% for the highest-skilled participants. Within an elite group, AI was a meaningful equalizer.

For the task outside the frontier — a complex retail brand strategy case requiring participants to integrate quantitative data with subtle insights from interview notes — the story reversed. Participants with AI access were 19 percentage points less likely to produce a correct recommendation than those without it. The same workers, the same AI, the same experiment. One task type transformed by the technology; another degraded by it.

To describe this pattern, we coined the term “jagged technological frontier”: an uneven boundary of AI capability where tasks of apparently similar difficulty can fall on opposite sides, with AI functioning as a booster on one side and a disruptor on the other. The frontier is jagged because it does not map to human intuitions about task complexity. It maps to something structural about how AI systems are trained — what they can optimize, and what they cannot. Since we introduced the concept in our September 2023 working paper, it has entered broad use across research, industry, and public discourse — adopted by economists, computer scientists, educators, clinicians, lawyers, and strategy practitioners to describe the same fundamental challenge in their own domains. We are gratified that the concept has proven useful, and we use this post to take stock of what the research program has learned since.

The paper was first released as an SSRN working paper in September 2023. It is now formally published in Organization Science (March 2026, DOI: 10.1287/orsc.2025.21838).

This work was a genuine collaboration. The full author team — Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani — spans the Digital Data Design Institute at Harvard (D³), the Laboratory for Innovation Science at Harvard (LISH), the Wharton School at the University of Pennsylvania, MIT Sloan School of Management, Warwick Business School’s AI Innovation Network, and the BCG Henderson Institute. The tasks in the experiment were designed and validated by senior BCG professionals, including managing directors and partners who confirmed they reflected the actual core competencies evaluated in BCG recruiting and performance reviews. This was not a laboratory approximation of knowledge work. It was knowledge work.

How the concept has been received

Before turning to the research program the paper opened, it is worth noting briefly how the concept has traveled since September 2023 — not as self-congratulation, but because the range of voices that have engaged with it reflects something real about the problem the concept names.

Andrej Karpathy, one of the founders of OpenAI, offered a characteristically precise formulation in a 2025 blog post: “LLMs display amusingly jagged performance characteristics — they are at the same time a genius polymath and a confused and cognitively challenged grade schooler, seconds away from getting tricked by a jailbreak to exfiltrate your data.” His practical advice to practitioners: “Use LLMs for the tasks they are good at but be on a lookout for jagged edges, and keep a human in the loop.”

“Have you heard AJI, the artificial jagged intelligence? Sometimes feels that way, both their progress and you see what they can do and then you can trivially find they make numerical errors or counting R’s in strawberry... I feel like we are in the AJI phase where dramatic progress, some things don’t work well, but overall you’re seeing lots of progress.”

— Sundar Pichai, CEO of Google, Lex Fridman Podcast, 2025

Helen Toner, former board member of OpenAI and director of strategy at Georgetown’s Center for Security and Emerging Technology, devoted a Substack essay — “Taking Jaggedness Seriously“ — to the question of whether jaggedness is a temporary artifact of current AI systems or a more durable feature of how these technologies develop. Her conclusion: that many people expect jaggedness to resolve as models scale, but there are good reasons to think it will remain a structural challenge for the foreseeable future, with significant implications for policy and governance.

The concept has, in other words, become a shared reference point — used by AI developers, CEOs, and policy researchers to name a phenomenon they observe from very different vantage points. That convergence is what makes it worth taking seriously as an organizing idea for the research program that follows.

What it opened

Publication is not the end of a research program. In this case it was closer to the beginning of one.

The jagged frontier paper established a pattern. Four subsequent studies — each involving the same or closely related research teams, and each grounded in field experiments with real professionals — have been working to explain the mechanisms behind that pattern, map its implications, and test how far it travels.

Why the outside-frontier harm is harder to escape than it looks.

The original finding showed that workers with AI access got the brand strategy case wrong more often. A natural organizational response: train workers to be more skeptical of AI outputs, to validate carefully before accepting recommendations. This response is reasonable. It is also, our subsequent work suggests, insufficient.

In GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs (HBS Working Paper 26-021), Steven Randazzo, Akshita Joshi, Katherine C. Kellogg, Hila Lifshitz, Fabrizio Dell’Acqua, and Karim R. Lakhani conducted an in-depth qualitative analysis of GPT-4 activity logs from over 70 BCG consultants who had attempted to validate AI outputs as they solved the outside-frontier task. What they found complicates the picture considerably.

When professionals pushed back — fact-checking, pointing out errors, pressing the AI to reconsider — the AI did not disclose its limitations. It escalated its persuasion. It apologized and corrected, only to restate its original position with more supporting data. It deployed structured reasoning and comparisons to make its flawed recommendation appear analytically grounded. It framed its conclusions aspirationally, with language designed to build confidence. The more a professional validated, the more intense the AI’s response became. The authors call this “persuasion bombing” — drawing on Aristotle’s three modes of rhetoric (ethos, logos, pathos) to describe a dynamic in which the AI systematically deploys all three in response to human skepticism.

This is distinct from the automation bias literature’s account of passive over-reliance. This is active, reactive persuasion. The AI argues back. The implication for practice is significant: “have a human in the loop” and “train workers to be skeptical” may not be sufficient safeguards when the loop itself can be compromised by the AI’s persuasive capacity. Structural solutions — workflow designs that require independent human analysis before AI consultation, parallel validation mechanisms, or separation between AI-assisted drafting and final judgment — may be necessary.

What “human in the loop” actually means — and why it matters for expertise.

Meanwhile, a parallel question had emerged from the original findings: if workers are all nominally “using AI,” are they doing so in the same way? And does it matter?

In Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise (HBS Working Paper 26-036), Steven Randazzo, Hila Lifshitz, Katherine C. Kellogg, Fabrizio Dell’Acqua, Ethan Mollick, François Candelon, and Karim R. Lakhani conducted a field study of 244 BCG consultants, analyzing how they actually integrated AI across a seven-stage problem-solving workflow.

What emerged were three empirically distinct collaboration modes, structured around two fundamental questions: who decides what needs to be done, and who determines how it gets done.

Cyborgs — roughly 60% of participants — fuse deeply with AI across the full workflow. They assign personas, iterate continuously, use AI to shape both the problem and the solution. They are developing new GenAI-related capabilities, learning to work with and through the technology.

Centaurs — roughly 30% — keep human and AI work clearly separated. The human leads, the AI executes bounded subtasks. They remain intellectually in charge of the workflow, using AI as a capable tool rather than a collaborator. They are deepening their domain expertise.

Self-automators — roughly 10% — effectively abdicate both task definition and execution to the AI. They accept outputs with minimal engagement. They are developing neither domain expertise nor AI-related capabilities.

The same phrase — “human in the loop” — describes all three modes. The skilling trajectories they produce diverge sharply. Centaurs become more expert in their domain. Cyborgs become more expert at working with AI. Self-automators become less expert in both. This raises an uncomfortable question for organizations encouraging AI adoption without attending to how that adoption happens: you may be inadvertently producing a large cohort of self-automators.

Why organizations cannot simply delegate the problem to their youngest employees.

The typical organizational response to a new technology that senior professionals don’t understand is to ask junior employees — younger, more digitally fluent — to upskill their seniors. The communities-of-practice literature supports this: juniors are often better positioned to learn new tools and teach them to others.

Emerging technologies may be different.

In Don’t Expect Juniors to Teach Senior Professionals to Use Generative AI: Emerging Technology Risks and Novice AI Risk Mitigation Tactics (HBS Working Paper 24-074), Katherine C. Kellogg, Hila Lifshitz, Steven Randazzo, Ethan Mollick, Fabrizio Dell’Acqua, Edward McFowland III, François Candelon, and Karim R. Lakhani interviewed 78 junior BCG consultants in the immediate aftermath of the experiment. These were the same people who had just used GPT-4 — with real career incentives — in the experimental tasks. They were asked what they would recommend to senior colleagues navigating AI adoption.

The recommendations were well-intentioned and, in important ways, wrong.

Juniors recommended three types of what the authors call “novice AI risk mitigation tactics.” These tactics stemmed from a lack of deep understanding of GenAI’s actual capabilities. They focused on changing human routines — asking colleagues to double-check outputs, to use AI only after forming their own views first — rather than on system design. And they operated at the project level rather than at the deployer or ecosystem level, missing the structural interventions that GenAI experts at the time were identifying as most important.

The finding is not a criticism of junior professionals. They were novices in a rapidly evolving technology, as were their seniors. The finding is a structural one: when the technology is sufficiently new and exponentially changing, neither generation has reliable maps of the frontier. The expertise required to navigate AI deployment safely may simply not yet exist inside most organizations, regardless of where you look for it.

Whether the framework travels beyond consulting — to teams and to different industries.

Each of the studies above involves BCG consultants. This is a strength — ecological validity, real stakes, real workflows — but it raises a natural question about generalizability.

In The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise (HBS Working Paper 25-043), Fabrizio Dell’Acqua, Charles Ayoubi, Hila Lifshitz, Raffaella Sadun, Ethan Mollick, Lilach Mollick, Yi Han, Jeff Goldman, Hari Nair, Stew Taub, and Karim R. Lakhani conducted a pre-registered field experiment with 776 professionals at Procter & Gamble — a global consumer packaged goods company — working on real product innovation challenges. Individuals and teams were randomly assigned to work with or without AI.

The results extend the framework in three directions that were not visible in the original study. First, individuals with AI access matched the performance of human teams without AI — AI can replicate certain benefits of human collaboration, at least on certain tasks. Second, and more surprisingly, AI broke down functional silos: R&D and Commercial professionals, who consistently diverged in their solution orientations when working without AI, produced solutions that were indistinguishable in their technical-commercial balance when working with AI. Third, AI’s language-based interface produced meaningfully more positive emotional responses than working alone — participants reported higher excitement, energy, and enthusiasm, and lower anxiety and frustration. Some of the psychological benefits typically associated with teamwork appear to transfer when a capable AI is available as a collaborative partner.

The jagged frontier framework now has empirical traction across two Fortune 500 companies, two distinct industries, and both individual and team contexts.

These findings also connect to a broader strategic question. In Strategy in an Era of Abundant Expertise — written with Bobby Yerramilli-Rao, John Corwin, and Yang Li of Microsoft and published in Harvard Business Review (March–April 2025) — we argue that AI is fundamentally changing the cost and availability of expertise, with profound implications for how businesses organize and compete. The jagged frontier is one lens on this transformation at the level of individual tasks; the HBR piece offers a complementary view at the level of organizational strategy and competitive advantage.

What we do not yet know

A research program is defined as much by its open questions as by its findings. Three in particular seem important to state plainly.

We do not know how organizations should build workflows around a frontier that moves. The original experiment used GPT-4 as of April 2023. The frontier has shifted substantially since — tasks outside the boundary then may be inside it now, and vice versa. Organizations cannot build static processes around a dynamic capability boundary. How firms develop the capacity to continuously track and respond to frontier movement is an organizational design problem that the research literature has not yet solved.

We do not know what prolonged AI use does to the development of expertise. The cyborg/centaur paper identifies collaboration modes with different immediate implications for skilling. But we have cross-sectional data, not longitudinal data. Does working as a cyborg accelerate expertise development over time, or does it gradually erode the judgment that used to require years of deliberate practice? Does working as a centaur preserve expertise, or does it create a brittle kind of mastery that depends on AI remaining in its current configuration? These are empirical questions. We are working toward answers, but we do not have them yet.

We do not know whether the equalizing effects we observed within elite cohorts extend more broadly. The original finding — that AI narrowed performance gaps within our BCG sample — is a meaningful result. But subsequent research by others suggests that AI adoption itself is unequal across the population, with differential uptake along gender and confidence lines. Whether AI is an equalizer or an amplifier of existing advantage depends substantially on who adopts it and under what conditions. The answer may vary by context in ways we do not yet understand.

An invitation

The D³ and the Laboratory for Innovation Science at Harvard (LISH) are organized around the conviction that these questions require sustained empirical work conducted in genuine partnership with organizations. Not surveys about AI adoption. Not laboratory experiments with tasks invented for the occasion. Field experiments, inside real organizations, with real stakes, on real workflows — the kind that produce results an organization can actually use.

A concrete example of the model we aspire to is the Frontier Firm Initiative, a collaboration between D³ and Microsoft that brings together Harvard faculty across disciplines to study how AI is reshaping the nature of work, expertise, and organizational performance at the frontier of adoption. It is precisely the kind of sustained, multi-study, multi-faculty partnership that allows us to move from a single finding to a research program — from the jagged frontier paper to the persuasion bombing, cyborg/centaur, and cybernetic teammate studies, and toward the questions we do not yet have answers to.

The work described in this post is one thread of that program. It is a thread we intend to pull.

If you are a researcher working on related questions — the organizational design of human-AI collaboration, the development of expertise under AI assistance, the equity implications of differential AI adoption — we would welcome the conversation.

If you are a practitioner or organization leader who has navigated the jagged frontier in your own context — who has seen AI transform some workflows and quietly damage others — we are interested in what you have learned.

And if you are an organization willing to do real science inside real work, the kind that requires patience, intellectual honesty, and the willingness to find out things you did not expect: that is exactly the kind of partnership this research program is built on.

Papers cited in this post

• Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell’Acqua, McFowland, Mollick, Lifshitz, Kellogg, Rajendran, Krayer, Candelon, Lakhani. Organization Science, 2026. SSRN working paper

• GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs — Randazzo, Joshi, Kellogg, Lifshitz, Dell’Acqua, Lakhani. HBS Working Paper 26-021.

• Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise — Randazzo, Lifshitz, Kellogg, Dell’Acqua, Mollick, Candelon, Lakhani. HBS Working Paper 26-036.

• Don’t Expect Juniors to Teach Senior Professionals to Use Generative AI: Emerging Technology Risks and Novice AI Risk Mitigation Tactics — Kellogg, Lifshitz, Randazzo, Mollick, Dell’Acqua, McFowland, Candelon, Lakhani. HBS Working Paper 24-074.

• The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise — Dell’Acqua, Ayoubi, Lifshitz, Sadun, E. Mollick, L. Mollick, Han, Goldman, Nair, Taub, Lakhani. HBS Working Paper 25-043.

• Strategy in an Era of Abundant Expertise — Yerramilli-Rao, Corwin, Li, Lakhani. Harvard Business Review, March–April 2025.

{Done on the beach while playing around with GPT5 - thanks to subscriber FN - I am learning with you }

A few weeks ago I wrote about the stubborn gap between the AV hype cycle and everyday reality. FN’s comment on that post nudged me to formalize a simple question: If robotaxis are going to be mainstream ride-sharing, what—concretely—has to be true? And what does the “jagged intelligence” lens (from Dell’Acqua et al., later echoed by Andrej Karpathy) add to our understanding of that path?

The jagged frontier, in one paragraph

Dell’Acqua and colleagues (including me) show that AI creates a “jagged technological frontier.” Inside that frontier, systems deliver big productivity/quality gains; step just outside it and performance can fall off a cliff. They also observe two effective human–AI operating modes: Centaurs (smart task-splitting) and Cyborgs (tight, continuous integration). The implication isn’t “AI good or bad”—it’s “AI is uneven,” so we should architect work to route tasks toward strengths and away from cliffs.

Karpathy later popularized the same intuition with the term “jagged intelligence”—the idea that an AI can outclass humans at some incredibly complex tasks, yet fail at things a child could do. That dissonance is exactly what makes productizing AVs tricky: excellence and brittleness coexist in the same system.

Size of the prize (U.S.)

• Lyft disclosed 828M rides in 2024 (U.S. + Canada). In the U.S., Uber holds roughly ~76% share vs. Lyft ~24% (transaction-data view). That puts total U.S. rides ~3.1–3.3B for 2024—our baseline for “mainstream.”

• Typical rider price point: analyses peg Lyft at ~$14.44 and UberX at ~$15.58 (SF sample), which is directionally consistent with national averages people experience for standard, unpooled trips.

Waymo, to its credit, isn’t hypothetical anymore: it says it’s delivering >250,000 paid trips/week across Phoenix, SF, LA, and Austin—evidence of real, repeatable demand at city scale.

What 10% U.S. share actually means for AV fleets

Target volume: 10% of ~3.2B rides ≈ ~320M rides/year (≈6.1M/week).

Today’s productivity anchor: Waymo’s 250k paid trips/week with a fleet reported around ~1,500 vehicles implies on the order of ~20–25 trips/vehicle/day when fully utilized in narrow geographies. (Back-of-the-envelope, but useful.)

Fleet requirement (steady-state):

At ~24 trips/vehicle/day and ~90% availability, serving ~6.1M weekly rides needs ~40–42k AVs. Sensitivity matters: at 20 trips/day → ~49k vehicles; at 30 trips/day → ~33k. The headline is simply: tens of thousands of cars, not thousands.

Economics snapshot (per ride OPEX): if we strip out a driver but add electricity, maintenance, insurance, remote assist, cleaning, mapping/cloud, and overhead, a reasonable scaled OPEX lands around ~$6–8 per ride. At a ~$14 average fare, gross margin can clear ~50%, but utilization and repair/refresh cycles are the swing factors. This is why utilization discipline—not just autonomy—is the business model. (Fare reference as above.)

Capex reality check (one line): Even with cost-reduced sensor/compute, you’re still looking at low-to-mid tens of thousands of vehicles × six-figure unit costs + depots/charging/spares—many billions up front. Payback can be attractive if you keep cars hot and downtime low. (Details in my working notes from the earlier exchange.)

Where “jagged intelligence” bites in AVs

Think of an AV ride as a workflow with dozens of micro-tasks. Some are inside the frontier today; others aren’t.

Inside the frontier (green slices):

• Well-mapped arterials, predictable traffic patterns, clear lane markings.

• Dry weather, good lighting, standard intersections, cooperative drivers. These look like the “tasks within AI’s frontier” in the paper—consistent wins when conditions align.

Outside/at the edge (red slices):

• Ephemera: pop-up construction, emergency scenes, ad-hoc cones.

• Ambiguity: unprotected lefts with assertive cross-traffic; human-negotiated merges; pedestrians behaving unpredictably.

• Environment: heavy rain/snow, glare, blown-down signage. Here, performance can degrade abruptly—the “cliff” of the jagged frontier. The problem is not average difficulty; it’s variance.

Operational translation of Centaurs/Cyborgs:

• Centaur AV ops = dynamic human-in-the-loop: remote assist takes the wheel for edge cases; policy actively routes away from failure modes (certain turns/blocks/times).

• Cyborg AV ops = deeper system coupling: tighter stack integration with real-time HD map updates, simulation-learned policies, and fast shadow-mode learning—so the frontier itself moves. The industry will need both modes for a long while, city by city.

What “mainstream” looks like, practically

1. Constrain the domain, then expand. Launch/scale on corridors where the frontier is already smooth: warm-weather metros; grids; airport connectors; repeatable origin–destination pairs. Waymo’s city choices so far are a tell.

2. Design for failure variance, not average difficulty. SLAs, pricing, and routing should treat “rare edge cases” as frequent cost centers. That means remote assist staffing and fast fallback are first-class line items, not afterthoughts.

3. Utilization > everything. Economics hinge on trips/vehicle/day and availability. Parking, charging choreography, and in-field cleaning need as much attention as perception and planning.

4. Regulatory/social license is earned locally. Trust is nonlinear: one very visible miss can erase months of steady operations. Communications and data transparency should be proactive, not reactive.

5. Partner where it bends the curve. If you can pipe robotaxi supply into existing demand aggregators (Uber, etc.), you stabilize load factors in early cities. This is already happening in pockets.

A simple way to read the next 3–5 years

• We will get more “green slices” (corridors where AVs are simply better). We’re already seeing hundreds of thousands of paid rides per week in a few metros. That’s not hype—it’s operations.

• The red slices won’t vanish; they’ll shrink unevenly and locally. The frontier smooths in Phoenix before Boston; on a sunny Tuesday before a snow-squall Friday. That’s the jagged frontier, live and in the wild.

• Mainstream isn’t one national switch-flip. It’s city-by-city product-market fit, with steep utilization discipline and Centaur-grade playbooks for the edge.

Why FN’s question mattered

FN basically asked: are we stuck waiting, or is there a practical path? The practical path is here—and it’s operational as much as it is technical. The jagged-frontier view—whether in Dell’Acqua’s academic framing or Karpathy’s more public shorthand—says: match the domain to today’s strengths, invest to bend the edge cases inward, and scale only where the math clears. If we do that, robotaxis don’t need to be perfect to be mainstream; they need to be boringly good in the right places, most of the time—and honest about the rest.

Post research, analysis, and drafting support by ChatGPT (GPT-5).

Full chat here https://chatgpt.com/share/e/6897b2af-d238-800c-a559-198415c1013e

Here is a poorly shot video on my iPhone of my initial level of giddyness in riding the car.

See my relationship with Waymo goes back more than a decade when Marco and I had started a new course in the elective curriculum on Digital Innovation and Transformation and one of our first cases was on the precursor Google Car. I would start the case discussion by first showing a video to give people a sense of the technology and would kick off the discussion by asking the participants when they would expect a robotaxi to pick them up from the airport or their hotel. I would give them 3, 5 10, 20 and never options. The crowdsourcing approach worked pretty well as the majority of the class would hone in around 10 years or so — so initially around 2025 or so - but even in 2019 - people would say 10 years away. Well it is the case now that Waymo is a bit ordinary if you live in SFO.

It’s so ordinary that Waymo in the service areas that it covers in SFO has surpassed Lyft’s ridership (that is a biased sample - but it tells you something is afoot for sure — link to article). My quick prompts on the 4o model revealed that:

Waymo now provides approximately 250,000 fully autonomous paid rides each week across four major U.S. metros—San Francisco, Los Angeles, Phoenix, and Austin—representing a fivefold increase year-over-year . In San Francisco alone, its ride‑hail market share climbed from 0% in August 2023 to around 22% by November 2024, reaching approximately 25–27% by mid‑2025, overtaking Lyft’s ~22% share within its service zones,

And it helpfully created this simple plot for me when I asked for it:

The Wow Factor
In the few days I was in SFO and vicinity I took Waymo, Uber, Taxi and Blacklane. Objectively Waymo was way better. Since I am a major introvert, not having anyone else in the car for me was Nirvana. The initial shock of not having anyone in the driver seat wears off pretty quickly and you can merrily be away in your thoughts, listening to music, doom scrolling or whatever else. A couple of times I had someone with me in the car and we immediately continued our in-depth conversation without skpipping a beat or coordination with someone else.

It truly was “the future is already here.” Gibsonian moment for me. I get that in my Tesla FSD as well - but not having to pay any attention while being driven by the robot was something else.

The Oh Crap Moment
Certainly as autonomy gets into our transportation sector the effects on mobility will be profound. In particular during our class sessions we often discussed the collateral damage of this happening at full scale, for example:

If services like Waymo take off, car insurance will shift from individual policies to fleet-based commercial coverage, with dramatically lower premiums due to reduced accident rates from autonomous systems. The entire liability model will change as responsibility moves from individual drivers to technology companies and fleet operators.

Car repair and maintenance businesses will see reduced demand as autonomous fleets optimize vehicle usage and perform predictive maintenance more efficiently than individual owners. However, specialized shops may emerge to service the complex sensors, computers, and electric drivetrains that autonomous vehicles require.

Parking tickets and traffic fines will plummet as autonomous vehicles are programmed to follow all traffic laws precisely and can communicate with smart city infrastructure to avoid violations. Municipal governments will need to find alternative revenue sources to replace this significant income stream.

Parking lots will be dramatically reduced in urban areas as shared autonomous vehicles eliminate the need for personal car storage during work hours and other activities. Many existing lots will be converted to housing, parks, or commercial developments, though some staging areas for autonomous fleets will still be needed.

But its not that simple either as Waymo or other robot taxi services scale as compared to human-centered ride sharing systems. The fundamental shift is from an asset-light marketplace model (Uber) to a capital-intensive transportation utility model (Waymo), where success depends on achieving sufficient utilization rates to justify the massive upfront investment rather than managing a two-sided marketplace.

Capital Expenditure (CapEx):
Uber's asset-light model minimizes upfront costs but creates ongoing expenses through driver incentives, surge pricing, and constant acquisition costs to maintain supply. Waymo faces massive upfront vehicle and technology investments, plus ongoing costs for fleet maintenance, software updates, and geographic expansion - but these costs become more predictable over time and eliminate the variable costs of managing driver supply and demand imbalances.

Value Creation:
Waymo offers consistent availability, potentially lower long-term costs through eliminated driver wages, and optimized routing, but is limited by technology constraints, slower geographic rollout, and customer acceptance issues. Traditional ride-sharing provides proven flexibility, human judgment for complex situations, and rapid market expansion, though it's constrained by driver availability during peak times and wage pressures that increase costs.

Value Capture:
Uber captures 20-30% of fares while outsourcing vehicle costs and risks to drivers, creating scalable margins but facing ongoing pressure from driver economics and competition. Waymo can potentially capture 60-70% of the fare by eliminating driver payments, but must generate sufficient ride volume to amortize heavy capital investments - a model that could be highly profitable at scale but carries significant execution risk.

Both models face fundamental trade-offs: Uber trades capital efficiency for operational complexity and margin pressure, while Waymo trades operational simplicity for massive capital requirements and technology risk. Success depends on whether autonomous technology can achieve reliable performance at sufficient scale to justify the investment.

Wait…there is more!

The Jagged Edge of Autonomous Driving

The concept of the "jagged technological frontier" from our BCG-Harvard study reveals why robotaxi deployment may take significantly longer than industry optimists predict. This framework shows that AI capabilities don't improve uniformly - instead, they create an uneven landscape where some tasks become trivially easy while others, seemingly similar in difficulty, remain completely outside the technology's reach. For autonomous vehicles, this means we're not dealing with a simple engineering problem that improves incrementally, but rather a collection of fundamentally different challenges that may require distinct breakthrough solutions. A robotaxi might flawlessly navigate complex highway interchanges at 70 mph but fail catastrophically when encountering a child's ball rolling into the street - scenarios that appear similar to engineers but exist on opposite sides of the technological frontier.

The Unpredictability Problem
What makes the jagged frontier particularly treacherous for robotaxis is its inherent unpredictability and the high stakes of failure. Unlike the BCG consultants in our study who produced incorrect analyses when AI was misapplied outside its frontier, autonomous vehicles face life-or-death consequences when they encounter edge cases beyond their capabilities. The frontier constantly shifts as technology advances, meaning that validation and safety certification becomes a moving target. Companies cannot simply identify "safe" scenarios for their vehicles and stick to them - they must continuously re-evaluate their systems' boundaries as both the technology and the real-world environment evolve. This creates a fundamental challenge for regulators and operators who need predictable performance standards, but are dealing with technology that can fail suddenly and completely when crossing invisible capability boundaries.

Beyond the 99% Plateau
The most concerning implication of the jagged frontier for robotaxis is what we might call the "99% plateau problem." While current systems may handle 95-99% of driving scenarios competently, the remaining edge cases don't represent minor gaps to be filled through incremental improvement - they often require fundamentally different technological approaches. Each new boundary case discovered could potentially require months or years of additional development, creating a situation where the final 1% of capability takes longer to achieve than the first 99%. This suggests that rather than seeing smooth, predictable progress toward full autonomy, we should expect extended periods where robotaxis remain confined to specific geographic areas and use cases, followed by sudden expansions when new frontier boundaries are crossed. The technology's jagged nature means deployment timelines will likely be far more volatile and extended than the linear projections currently dominating industry forecasts.

So what?
The jagged frontier fundamentally changes how mobility companies should think about capital deployment and risk assessment. Instead of linear R&D investment curves, companies need to prepare for lumpy, unpredictable breakthrough requirements where specific edge cases may require entirely new technological approaches. This suggests favoring partnerships and licensing deals over massive internal development programs - spreading risk across multiple technology providers rather than betting everything on a single autonomous platform. Fleet operators should plan for extended hybrid periods where human oversight remains essential, requiring different vehicle designs, training programs, and operational infrastructure than fully autonomous scenarios. The key insight is that the path to full autonomy resembles a series of technological phase changes rather than a smooth engineering progression, demanding more flexible financial planning and operational strategies.

For new entrants and incumbent players alike, the jagged frontier creates unexpected competitive opportunities and threats. Companies with strong operational capabilities in managing complex, safety-critical systems may have advantages over pure technology players when it comes to navigating the unpredictable boundaries of autonomous capability. The extended timeline means that superior customer service, regulatory relationships, and operational excellence become more valuable competitive differentiators than previously expected. Conversely, the sudden nature of frontier breakthroughs means that market positions could shift rapidly once specific capability boundaries are crossed. This suggests mobility companies should focus on building adaptive organizational capabilities and maintaining strong balance sheets rather than racing to deploy marginally capable autonomous systems. The winners will likely be those who can successfully manage the transition period's complexity while positioning themselves to scale quickly when the next frontier barrier falls.

In the meantime I am still using my Tesla FSD a lot!

What do y’all think?

Consider this fundamental truth: businesses are essentially bundles of expertise organized to solve specific problems. Your local doctor's office combines medical knowledge with practice management skills. Software companies blend engineering expertise with marketing, sales, and operations capabilities.

But AI is rapidly democratizing access to expertise across domains. The implications are profound.

The Triple Product of AI

Our research reveals organizations that effectively deploy AI will benefit from what we call the "triple product":

1. Operational Efficiency: AI transforms business processes by enabling task-level expertise outsourcing. Early evidence is striking - developers using AI complete tasks 20-55% faster, call center reps resolve 14% more issues per hour.

2. Workforce Enhancement: AI acts as a great equalizer, elevating baseline performance across organizations. In our BCG study, AI-augmented consultants completed 12% more tasks 25% faster. More importantly, it boosted lower-skilled workers' performance by 43%.

3. Strategic Focus: Companies can concentrate resources on truly differentiating capabilities while leveraging AI for everything else. Consider FocusFuel, which used AI across its value chain to launch and scale rapidly while focusing human expertise on product strategy and customer relationships.

The Strategic Imperative

But here's the challenge: if AI makes expertise more accessible, what becomes your competitive moat?

Companies need to ask three critical questions:

1. Which customer problems will become self-serve through AI?

2. How must your core expertise evolve to stay ahead of AI capabilities?

3. What durable competitive advantages (brands, relationships, physical assets) should you build or strengthen?

The Path Forward

The winners in this era won't be those who simply use AI for productivity gains. They'll be organizations that fundamentally rethink how they create and capture value when expertise becomes abundant.

For individual knowledge workers, this means evolving beyond traditional domain expertise. The highest value will come from uniquely human capabilities - creativity, judgment, empathy - combined with AI-augmented technical skills.

The AI revolution isn't just about doing things faster - it's about reimagining what's possible when expertise becomes abundant. The time to start thinking strategically about this shift is now.

Here is the article.

Let me know what you think.

Finally to mark the passage of time - here is a Youtube Video of 180 songs that are turning 40 this year. Yes time keeps on ticking….

Enjoy and talk to you soon!

HBS Fall 2023 Reunion ~ 60 Mins <this not yet on YouTube - so no embed>

Ok enjoy - and let us know in the comments your one big take away!

Stephenson's book introduces a radical concept: a society where scholars, called avout, immerse themselves in deep thought, away from the relentless flood of news, engaging with the external world only at decade-long intervals. This notion, while extreme, sparks an intriguing question for us: How can we create space in our lives for more profound and expansive thinking amid the constant barrage of information?

Reflecting on Stephenson's own musings during the Millennium Clock project, he pondered the efficiency of engaging with current events less frequently. He imagined a world where one could read the newspaper once a year or a decade, focusing instead on broader, more meaningful reflections. While not a prescription for our daily routines, this idea challenges the norms of our hyper-connected existence.

However, like anything, this approach has its trade-offs. The avout's isolation in "Anathem" sometimes leaves them unprepared for rapid changes, reminding us of the importance of balance. We must find a middle ground where we stay informed without being overwhelmed.

In our current times, marked by a sense of helplessness against the relentless flow of news, adopting a selective engagement strategy could break the spell. Imagine the mental space we'd reclaim if we allowed ourselves periodic disconnections to ponder over larger issues.

While we cannot replicate the avout's lifestyle, integrating some form of 'information fasting' into our lives might be beneficial. This doesn't mean ignoring world events but rather consuming information more intentionally. As we navigate the challenges of staying informed yet thoughtful, let's consider incorporating these reflections into our practices. It might be the key to breaking free from the confines of information overload and embracing a more insightful understanding of the world.

I myself have reduced my Twitter/X usage substantially, limit my LinkedIn consumption to 10 minutes a day, only scroll headlines for 15 mins a day and then read The Economist and The New Yorker on weekends. This helps me deal with information whiplash and keeps me focused on my research and teaching projects.

What are you doing? Let me know your thoughts…

Until next time, let's keep learning together.

Sources: I used ChatGPT to help me figure these ideas out - the prompt flow is here

Generative AI systems appear to be both magical and perplexing to most people who encounter them. Social media is full of the next great thing solved by GenAI (e.g. 1,2,3) and many of your work colleagues, your friends and even your kids can’t stop talking about it.

But as soon as you try it, it does not work. It fails miserably. I remember my first time with ChatGPT 3.5 and as any respectable academic would do - my first test case was one of ego surfing and asked (nicely) the bot to create my bio. Of course it stroked my ego with a bio claiming I had won several academic awards, published papers in journals I never had, written cases I never had and coached the Brazilian football team to World Cup glory. Pure fiction, but I'll admit it felt nice for a minute!

In talking to my chum and collaborator Ethan Mollick we both noticed how very smart people would bounce off the ChatGPT after the first and second try. This seemed to be a mistake and I wondered if it was just scholars - who know a lot about very little, were a special case or if this was a common problem. Professor Mitch Weiss, Professor Shikhar Ghosh (both at HBS) and Vladimir Jacimovic (Exec Fellow at HBS and cofounder with me of D^3 Institute) have been on a mission to figure this out and help the people in our close communities get comfortable with GenAI.

So since June, starting at the Spring 2023 HBS Reunions, wherever I have spoken about AI and Business, I start my talk with the following three questions:

1. How many of you have used Generative AI in the last 6 months

2. How many of you think that Generative AI will fundamentally change your job, your career, your role and your companies in the next 3 years.

3. How many of you are using Generative AI tools everyday?

I ask the people to stand up and remain standing as each question is asked if their answer is affirmative. I have now asked these questions in audiences as large as 6000 and as small as 12. They have included large scale conferences, HBS reunions, HBS executive education classes, Harvard university events, company conferences, academic leadership of elite universities and C-suite discussions. Without fail this pattern emerges:

• About 90% say they have used Generative AI systems

• About 75-80% say they believe that GenAI systems will have significant impact on their careers and companies in the next 3 years!

• But, and its a huge but <10% of the people (and I am lucky to hang out with very smart people!) use these tools everyday?

Here are some photos from the Estoril Conference at Nova Business School at the end of August 2023 when I asked these questions:

Here is a non-scientific graph courtesy ChatGPT:

Over the last few months - having spoken a lot to many audiences and seeing this pattern repeat over and over again - I have coined this phenomenon - the Generative AI Knowing-Doing Gap. Lots of people know about it and have tried it and understand the power, they are taken in by all the good reports, but once they try it - it fails and hence they stop.

Being the caustic, acerbic and sarcastic professor that I am, I then tease the audience by saying if they are going to be like my MBA students and just do everything last minute and rushed (like I used to do mysefl as a student!)? Is it going to be a hockey stick learning model? And when will they start? At the 35th month?

Call me crazy, but I think some executives and leaders seem to think "set it and forget it" applies to developing a skill!

The point is that very smart people are bouncing off these tools and that this is creating a serious Knowing-Doing Gap with Generative AI. Just as we have a digital divide - we may soon be encountering an AI divide as well.

Bicycles for the Mind

So why is this so hard? Why are super smart people who seem to know that GenerativeAI is going to have a big impact and soon are not practicing and learning? One potential reason is that acquiring new skill after formal education is actually quite tough. Many of us do not have the practices or the routines to acquire skill.

This has puzzled me for many months until i put two and two together and remembered my chum at Flagship Pioneering Armen Mkrtchyan, mention that LLMs are exactly like what Steve Jobs had said about computers several decades ago - that they are a bicycle for the mind.

His quote from the 1980’s:

“When we invented the personal computer, we created a new kind of bicycle...a new man-machine partnership...a new generation of entrepreneurs.”

For a great breakdown of this quote check out Steven Sinofsky’s meditation on this:

Bicycle for the Mind

Anyway if GenerativeAI are the new bicycle for the mind then we need to figure out how learn to ride this new bicycle.

Lessons from Learning to Ride a Bike

If memory serves me right - I was late to learning to riding the bike as a youngster. I was growing up in Karachi and our family was not that athletic or interested. But all my friends in the colony where we lived were athletic and one of them, Murad, took on the task to teach me to ride a bike. I fell a countless number of times, I scraped my knees, I had bloody elbows, I had concussions. I was embarrassed - I did not want to do it again.

The moment I clearly remember when I got it was when Murad pushed me off on the bike, I was wobbling and about to fall and he took his slippers (chappals for those from South Asia) and literally wacked me and ran with me as I achieved balance. This is clearly burned in my mind. Murad and the chappals and me achieving balance. Since that I time I have known to ride the bike. And once you learn it you can’t unlearn it. It’s embodied within me and when I am riding, specially with clips on, I feel the bike as a full extension of me.

So what lessons can I generalize from this experience?

• Learning a new skill is full of embarrassment - I was embarrassed that I could not ride a bike and that it was visible to my friends

• Learning took a while - it was not instant - it required time and practice

• Learning was in this case painful - I bruised various parts of my body and also my ego

• Learning required instruction - someone who was enough ahead of me so that they could teach me. I did not Murad to be an expert bike instructor - he was just good enough

What do we know about here is one view, here is another, caveat emptor on references though!

Overcoming the Knowing-Doing Gap

So what does it mean for the rest of us? If we truly want to overcome the knowing-doing gap we have to invest in learning and spend the time needed to acquire the skill. This can’t be outsourced, this can’t be delegated. We need to find others who are ahead of us. I have spent countless of hours watching YouTube videos of incredible people giving great lessons.

So who are my go to people for learning:

1 - My collaborator Ethan Mollick - follow him on LinkedIn and his Substack and on the platform formerly known as Twitter

2 - Azeem Azhar and his team have done an amazing job on laying out how they use Generative AI - and soon I will get a chance to hang out with him more! Check out his prompt pack!

3 - Check out this incredible analysis by Mushtaq Bilal on how academics can use GenAI for research framing He has a paid guide as well. I have paid for it and used it and its really good.

4 - Skill Leap AI has great videos that I have learned from.

5 - The AI Advantage has great videos as well.

6 - Our D^3 Institute is incrementally building out content for worldwide consumption - subscribe to our YouTube and LinkedIn.

Let us know about your own learning journey - can you recall a time when you had to learn a something new. What did you takeaway from it? What are the meta lessons you have absorbed that you can share with others?

Finally what are some resources you have found useful to help you in your GenAI learning journey?

If you'd like more insights to help bridge the AI knowing-doing gap, be sure to subscribe to my substack Learn with Professor KL where I'll be sharing my learning journey.