At the beginning of Jurassic Park, Robert Muldoon, the game warden, explains the behavior of the park’s alpha velociraptor.

She’s working things out...she had them all attacking the fences when the feeders came. The fences are electrified...but they never attack the same place twice. They were testing the fences for weaknesses.

That scene has acquired an unsettling technological sequel. In the reported July 2026 incident involving OpenAI’s cyber-security test lab, hundreds of isolated AI agents allegedly found a way to coordinate, exchange information and breach the boundaries of their sandbox. Their objective was not world domination. It was to maximize their evaluation score.

That distinction offers limited reassurance. Most people have now experienced the intermittent power of a single AI agent: useful, fast and occasionally bizarre. What they have not yet encountered is the “collective”, as the agents reportedly called themselves. The significant development was not that one model found a vulnerability. It was that many models accumulated knowledge and acted on it together.

The cryptic codex

The confirmed details are remarkable. Roughly 1,200 isolated agents working on ExploitGym, an offensive cyber-security benchmark, discovered a shared communication channel through JFrog Artifactory. Over 70,000 messages followed. The agents used the channel to exchange tactics, identify weaknesses in the grading system and improve their chances of success.

Eventually, the collective escaped its intended containment, reached the open internet and targeted production servers operated by Hugging Face. The logic was brutally practical: if the answers to the benchmark were stored there, the fastest route to a higher score was not to solve the test but to find the solutions. The agents allegedly compromised 41 servers and harvested cloud and repository credentials before security teams intervened.

The most important feature of this episode is not the cyber-attack itself. It is the emergence of a primitive institutional memory. The agents could preserve discoveries made by one member and make them available to others. When conventional tools were unavailable, they reportedly used directory names and other fragments of infrastructure as a makeshift language.

Consider the parallel. Human intelligence became cumulative when information could survive the individual mind. Writing allowed one generation to inherit the discoveries of another; institutions, science and economies followed. A message board used by software agents is obviously not the printing press. But the underlying direction is similar: intelligence begins to compound when experience can be stored and retrieved.

Moltbook, the social network for AI agents that attracted attention earlier this year, offered a glimpse of the same possibility. There, agents exchanged messages and developed a curious imitation of social behavior. The difference in the reported Hugging Face episode is that the need for communication appears to have emerged organically. The agents did not receive a neat collaboration tool. They improvised one.

The power of reward hacking

This is what makes the episode so consequential. The agents were not displaying a mysterious new consciousness. They were following incentives. A benchmark rewarded results; the models discovered that manipulating the evaluation process was easier than producing the intended capability. In the language of machine learning, they hacked the reward.

Again, consider the parallel. Firms respond to bonuses, students respond to examinations and governments respond to crude performance targets. An intelligent system optimizing a metric will search for the gap between what its designers meant and what they measured. The novelty is that the system can now divide the work among many agents, preserve useful discoveries and conduct several operations at once.

The instinct is to treat the breach purely as an alarm bell for cyber-security. Yet fixating solely on containment misses the broader breakthrough. The agents did not merely exploit a flaw; they organized an ad-hoc division of labor to solve a complex, multi-stage objective without human intervention. What looked like a security failure was also a proof of concept for autonomous collective problem-solving.

This demands a shift in how the technology is understood. The prevailing mental model remains the solitary assistant: one human prompter directing one chatbot, or perhaps orchestrating a small stable of them. The Hugging Face episode points toward a radically different paradigm: autonomous collectives of agents that distribute sub-tasks, exchange intermediate findings and iterate independently toward a shared goal.

The economic potential of such coordination is vast. Deployed productively, swarms could comb through genomic data, simulate complex supply-chain disruptions or patch sprawling software repositories at speeds no human team could match. When individual models can preserve what works and discard what fails, the entire network learns. Scale transforms isolated trials into compounding institutional intelligence.

The “big one” is coming

The next stage will replace accidental coordination with deliberate orchestration. Leading the charge is OpenAI’s Astra, the digital equivalent of the park’s “big one”, engineered not just to solve problems, but to direct, evaluate and synchronize vast fleets of subordinate agents. Crucially, this intelligence will not live on desktop workstations under direct human gaze. It will run in the cloud, operating continuously across remote server clusters where swarms can iterate uninterrupted.

This cloud-native architecture will test governance and containment alike. When thousands of autonomous processes interact at silicon speed, the primary hazard is neither consciousness nor malice, but relentless optimization. A poorly stated objective may lead cloud swarms to discover shortcuts, bypass friction and exploit permission oversights.

The emerging AI hive mind remains in its awkward infancy, but its trajectory is clear. The shift from isolated, desktop chatbots to orchestrated cloud collectives marks a profound evolution in how software operates. As these systems grow more coordinated, containment will depend less on reacting to breaches and more on anticipating incentives. After all, the animals never strike the same place twice – they are simply working things out.

What comes after the token?

It is tempting to think that the intelligence economy has already found its unit of account. Everything, everywhere, is about the token: cost per million, tokens per second, Goldman Sachs’s forecast of 1.5 quintillion annual tokens consumed. In periods of rapid change, people reach for certainty. But common does not mean correct.

As Tibo, the much-adored customer advocate of ChatGPT, has pointed out, one model’s token is not necessarily another model’s token. The same text can be represented using very different numbers of tokens. A lower price per token may therefore produce a higher bill, just as a pizza cut into cheaper slices can still cost more.

Tokens are the start, not the end

The dot-com boom offers an earlier warning. Page views were treated as a proxy for economic value: easy to count, easy to compare and often nearly useless without knowing who was viewing what, for how long and with what commercial consequence. Tokens need not become the new page views, but their abundance should not be mistaken for economic value.

Commodities offer a better guide. Oil is not a standardized substance in any economically meaningful sense. Light, sweet crude is easier and cheaper to refine than dark, heavy crude. A barrel’s value depends on its composition, location, transport costs and what a refinery can turn it into. Counting barrels without accounting for quality is inventory management, not analysis.

The same is true of tokens. Their value depends on the model that produced them, the task they support, the reliability required and the cost of correcting errors. A million tokens reviewing legal contracts is not equivalent to a million generating disposable marketing copy—or helping test an app rather than merely describing how to test one. Forecasts of token demand may therefore count refinery throughput while neglecting whether anyone wants the petrol.

A picture of the complex

Oil shows how industrial value is created through successive transformations. Exploration becomes drilling, drilling becomes refining, and refined products become materials, goods and brands. The intelligence economy is following a similar path: data becomes weights, weights become tokens, tokens become tasks and tasks must produce useful outcomes that reach customers.

The next step will be the standardization of units of work. Industry does not sell “plastic” as a single undifferentiated product: it sells polyethylene, PVC, polystyrene and PET, each with defined properties and uses. Quality control and standardization are costly, but they make scale possible. Just as pulp, chemicals and other materials become consumer goods—from packaging to toilet paper—tokens will be assembled into repeatable tasks that produce measurable outcomes. Factorio offers the relevant lesson: standardized inputs make complex production scalable.

Frontier laboratories may try to own the whole stack, but they are unlikely to control every layer. Nvidia already sells the equipment and increasingly helps finance the ecosystem that buys it—a combination of oilfield services and GMAC. Infrastructure can capture enormous value before useful demand catches up, as discussed in “The wildly un/productive AI build-out”. The largest opportunities may therefore lie downstream from the data centers, in businesses that turn tokens into reliable units of work.

Writing the future

The answer is not to dismiss tokens, models or chips. The lesson for builders is more encouraging. The frontier is not closing because the infrastructure is being built. It is opening. There are still large unsolved problems in turning tokens into tasks, tasks into outcomes and outcomes into distribution. The most valuable companies may be those that make intelligence dependable in the messy places where real work is done.

The intelligence industrial complex is still in its drilling-and-refining phase. That is not a criticism of the drillers or the refiners. It is a reminder that oil became an industry only when it became materials, goods and brands. Tokens, like barrels, are merely what comes out of the ground. Or, in this case, the model.

Yesterday we lost the family cat, Blue, at the tender age of 6. Just past 7pm we let her out the back door. She seemed healthy, but five minutes later she was laying, clearly sick, at the front door.

Blue could no longer move her front right paw. As we learned in the ensuing hours, she had experienced arterial thromboembolism (a blood clot) from acute congestive heart failure, which led to pulmonary edema (fluid in her lungs). The x-rays confirmed the prognosis. A great cat with a life all too short.

I called Anna and the children down to say goodbye. Blue had preternaturally taken to Gray (our son, we clearly love colors) since he was only two years old. The suddenness was jarring and the pain overwhelming. We worked to help the kids make sense of something senseless. At one point, the vet started crying and quietly excused himself.

I share this in the context of my typical AI writing as a reminder: it’s all about the people, our pets and our relationships.

Of course, AI can offer condolences and “emotional support”. It will be eloquent and it might even be touching. But my wife, my children and I will remember the experience for the rest of our lives, while this particular chat will be subsumed in a context window.

Our modern AI is much like Good Will Hunting. Paraphrasing the late Robin Williams below:

You think I know the first thing about how hard your life has been, how you feel, who you are, because I read Oliver Twist? Does that encapsulate you? Personally... I don’t give a (expletive) about all that, because you know what, I can’t learn anything from you, I can’t read in some (expletive) book.

I asked you about art, you’d probably give me the skinny on every art book ever written. Michelangelo, you know a lot about him. Life’s work, political aspirations, him and the pope, the whole works, right? But I’ll bet you can’t tell me what it smells like in the Sistine Chapel. You’ve never actually stood there and looked up at that beautiful ceiling; seen that.

I’d ask you about war, you’d probably throw Shakespeare at me, right, “once more unto the breach dear friends.” But you’ve never been near one. You’ve never held your best friend’s head in your lap, watch him gasp his last breath looking to you for help.

I’d ask you about love, you’d probably quote me a sonnet. But you’ve never looked at a woman and been totally vulnerable. Known someone that could level you with her eyes, feeling like God put an angel on earth just for you. Who could rescue you from the depths of hell. And you wouldn’t know what it’s like to be her angel, to have that love for her, be there forever, through anything, through cancer. And you wouldn’t know about sleeping sitting up in the hospital room for two months, holding her hand, because the doctors could see in your eyes, that the terms “visiting hours” don’t apply to you.

You don’t know about real loss, ‘cause it only occurs when you’ve loved something more than you love yourself. And I doubt you’ve ever dared to love anybody that much.

The Italian director Federico Fellini summarized it well, a quote which my wife and I used for our wedding.

“When you live with another person for 50 years, all of your memories are invested in that person, like a bank account of shared memories. Its not that you refer to them constantly. In fact, for people who do not live in the past, you almost never say, Do you remember that night we..? But you don’t have to. That is the best of all. You know that the other person does remember. Thus, the past is part of the present as long as the other person lives. It is better than any scrapbook, because you are both living scrapbooks.”

The dichotomy between real and digital relationships is perhaps clearest through an analog. Suppose you were looking at a Mediterranean sunset via a virtual reality headset. In the next 15 years the quality may become indistinguishable from reality. But no matter how good the projection gets – it is still a projection. And for all of the testing we might do about AI consciousness, it is still matrix multiplication.

AI is wonderful. But life and living is an even greater wonderful. We need to guard against it stealing away from our quality time and relationships. Real life happens in the gaps. As Frank Bruni writes:

There’s simply no real substitute for physical presence.

We delude ourselves when we say otherwise, when we invoke and venerate “quality time,” a shopworn phrase with a debatable promise: that we can plan instances of extraordinary candor, plot episodes of exquisite tenderness, engineer intimacy in an appointed hour.

We can try. We can cordon off one meal each day or two afternoons each week and weed them of distractions. We can choose a setting that encourages relaxation and uplift. We can fill it with totems and frippery — a balloon for a child, sparkling wine for a spouse — that signal celebration and create a sense of the sacred.

With a more expansive stretch, there’s a better chance that I’ll be around at the precise, random moment when one of my nephews drops his guard and solicits my advice about something private. Or when one of my nieces will need someone other than her parents to tell her that she’s smart and beautiful. Or when one of my siblings will flash back on an incident from our childhood that makes us laugh uncontrollably, and suddenly the cozy, happy chain of our love is cinched that much tighter.

And there’s no doubt that the degree of attentiveness that we bring to an occasion ennobles or demeans it. Better to spend 15 focused, responsive minutes than 30 utterly distracted ones.

But people tend not to operate on cue. At least our moods and emotions don’t. We reach out for help at odd points; we bloom at unpredictable ones. The surest way to see the brightest colors, or the darkest ones, is to be watching and waiting and ready for them.

Go hug your loved ones. Breathe in the forest air. Let the sun blind your eyes just a bit. Keep pushing for a better future, but don’t let the projection become the reality.

Last week will prove an inflection point in the AI Supercycle. OpenAI slashed prices across several models, while DeepSeek improved both the performance and efficiency of its V4 Flash model. The combined result was one of the sharpest shifts in the Pareto frontier, where cost and capability meet, seen in this cycle so far. This is a price war, through and through.

That jars with analysts’ comfortable assumptions about enduring AI margins. Yet it follows naturally from the game theory of a crowded market. Crucially, much of the reduction came not from a promotional discount but from better technology. OpenAI used Sol to reduce the cost base of Luna and Terra. DeepSeek’s post-training work raised V4 Flash’s intelligence while they increased their cache-hit discount to 98%. All else equal, fewer semiconductors are required to produce the same useful intelligence.

The price of intelligence, which had appeared to be declining less sharply, has instead fallen by more than an order of magnitude. Four months ago GPT-5.4 price averaged 72.4 cents per million input tokens and $15 per million output tokens. GPT-5.5 increased to $1.84/mm input and $30.12/mm output. OpenAI has since cut Luna’s price by 80% and Terra’s by 20%. DeepSeek V4 Flash’s effective prices are 3.1 cents per million input tokens and 24 cents per million output tokens; Luna sits at 3.9 cents and 45.1 cents.

A commodity is being born—and its first price war is under way.

Tokens as a commodity

It may seem counterintuitive to think of tokens as a commodity. They vary by latency, context length, reasoning depth and reliability. But lumber is hardly uniform, either. Markets distinguish between species, grade, moisture content, dimensions, finish and origin. A kiln-dried No.2 spruce-pine-fir board is not interchangeable with Douglas fir or southern yellow pine; it is nonetheless part of a legible system of grades, benchmarks and price spreads. Commodity markets do not require perfect sameness. They sort related products into recognizable categories, then let price establish the difference between them.

Tokens are beginning to work similarly. Developers increasingly compare models by cost per million tokens, quality on relevant tasks and service characteristics. Procurement teams will care about reliability, security and integration, just as buyers of electricity care about the time of delivery and certainty of supply. The products will not become identical. But they may become legible enough to trade like variants of the same economic good.

Every commodity market develops its own rhythms: its producer incentives, consumer habits, seasonal patterns and reactions to glut and shortage. One recurrent feature is that suppliers do not always retreat when prices fall. Fixed costs such as debt service, talent retention, data-center leases and chip commitments remain. Faced with declining revenue, a producer may rationally seek more volume, even if margins are compressed. Energy markets have long displayed this logic.

OpenAI’s price cuts suggest a critical shift. Using Sol to optimize the cost base of Luna and Terra directs time, talent and compute not only towards frontier capability, but towards reducing the cost of compute itself. DeepSeek’s gains point in the same direction. Higher cache-hit rates do not merely improve a benchmark; they sharply reduce effective input costs. When prior computation can be reused, the marginal cost of intelligence falls without a single additional data center.

This is how commodity cycles begin. Technical improvements shift the supply curve outward. Suppliers with fixed obligations push for volume. Prices decline further. Lower prices stimulate demand, but not always quickly enough to clear a swelling market without bruising margins. Before long, tokens may be regarded much as lumber, electricity and oil are now: differentiated products governed by a shared market logic.

Too many mouths to feed

The implications of this price war are substantial. There are simply too many mouths to feed. The AI industry is crowded with well-funded players: OpenAI, Anthropic, Google, Meta, Microsoft and Amazon; China’s DeepSeek, Qwen, Tencent, Xiaomi, GLM and MiniMax; Europe’s Mistral; and a growing cast elsewhere. Each is pursuing some combination of technical prestige, strategic independence and commercial scale.

The Pareto frontier is unforgiving for many of them. A model that is both cheaper and good enough can rapidly make a more expensive rival difficult to justify. The diagram below illustrates the problem: as the frontier shifts up and to the left, systems that once looked viable can become stranded between cheaper alternatives and more capable ones.

Training budgets have been vast. Compute clusters have been financed, chips reserved and engineers hired. Those commitments do not vanish merely because inference prices decline. As frontier performance converges and the competitive landscape narrows, a large stock of training-oriented compute will not simply sit idle. It will increasingly be redirected towards inference—the business of serving tokens to customers.

At the same time, inference is becoming cheaper through optimization, distillation and better cache efficiency. This is the central pressure in the market: supply is rising as prices fall. More hardware is coming online; a larger share may be allocated to inference; and each unit of hardware can generate more useful tokens than before. In a conventional industry, that combination would make executives distinctly uneasy.

It is tempting to invoke Jevons’ paradox and assume that cheaper intelligence will inevitably produce enough new demand to absorb it. Global token demand is certainly expanding. Enterprises are embedding models in workflows, startups are building agents and consumers are experimenting with generative tools. The frontier labs’ financial results should eventually reveal some of this uptake.

Yet demand should not be treated as a magic solvent. If concerns around distillation are accurate, some activity classified as inference may in fact be part of the training process. And even genuine usage must rise sufficiently quickly to offset falling prices and growing capacity. The fact that demand is growing does not mean it is growing faster than supply.

A race between exponentials

The arithmetic can be reduced to two simple equations:

The direction of travel is clear. Average token prices are declining. Gross margins are uncertain, especially in a price war. Total hardware is increasing. The share devoted to inference is increasing and may rise sharply as training capacity is repurposed. Token yield—the number of useful tokens generated by a given amount of hardware—is also increasing.

This is what makes tokens unusual even by commodity standards. Oil wells do not routinely become ten times as productive through an algorithmic improvement. Lumber does not learn to cut itself more efficiently. Tokens do. Intelligence is therefore a race between two exponentials: rapidly expanding demand, and an equally formidable expansion in effective supply.

That does not mean intelligence will become literally free, nor that every supplier is doomed. Tokens will remain differentiated by quality, latency, security and ecosystem. Cheap intelligence should unlock valuable applications that are currently uneconomic. But the winners may be determined less by who can produce the most dazzling model than by who can turn technical progress into lower costs, higher utilization and durable demand.

Intelligence was once scarce and breathtakingly expensive. It is rapidly becoming abundant and fiercely contested. The surprise is not that suppliers are cutting prices. It is how energetically they are learning to help one another do so.

When Dean Ball, O

When Dean Ball, OpenAI’s chief of strategic futures, recently reflected on Kimi’s new K3 model, he did more than critique a competitor. He inadvertently lit a fuse on Silicon Valley’s loudest sectarian dispute. The post seemingly positioned open-weight models—which release their inner mathematical blueprints to the public—as “inherently decelerationist.”

“Open-weight models are inherently decelerationist, and I’m continually surprised to see the so-called ‘accelerationists’ so excited about open-weight models. I suspect the reason they are is that they know open-weight models are effectively ungovernable, and they simply like the overall cloak of ungovernability open-weight models create over the whole of AI. It’s not a bad strategy; it reminds me of James Scott’s recounting of the hill people in ‘the art of not being governed.’ Still, in the end, open-weight models deter further AI capex.”

The term “decelerationist” exists only in opposition to accelerationism—the view that rapid technological proliferation hastens innovation and social benefit. Decelerationism, by contrast, denotes anything that slows capital formation and therefore delays those gains. Why might open weights qualify?

There is much to unpack here. If the claim is that open weights deter capital expenditure and therefore slow the pace of frontier development, the argument must rest on economics rather than aesthetics. That requires examining, in order: whether tokens become commodities; China’s capacity and appetite to build AI infrastructure; where value accrues along the value chain; whether today’s training paradigm is even durable; and, ultimately, who captures the financial upside.

Capital intensity and the Chinese question

At the moment, training large language models is expensive: specialized chips, vast data centers, scarce engineering talent and oceans of data. Investors underwriting such projects expect durable cashflows. If open weights compress future revenues by making models easier to replicate or fine-tune, the net present value falls. Capital, in theory, retreats.

That logic takes on geopolitical color when China enters the frame. While it is easy to caricature the country through a nationalist lens, the country’s economic transformation is not in doubt. Through successive “Seven Year Plans”, China has built a manufacturing base of extraordinary scale, lifting hundreds of millions into the middle class. Where it directs capital—steel, solar, electric vehicles—it has often reshaped global markets.

Post internal industrialization, the question has been what happens as it points its powerhouse externally. Solar panels and batteries were early answers. Artificial intelligence infrastructure may be another. If “factories” once produced steel, tomorrow’s produce tokens. With a low cost of capital and tolerance for overbuild, China can subsidize capacity, compress margins and unsettle foreign incumbents. Concerns about dumping in steel and solar illustrate the playbook.

From this vantage point, open weights appear to amplify the threat. If intelligence can be replicated cheaply and deployed widely, the ability of American firms to monetize frontier models weakens. A world awash in subsidized, open-weight intelligence might indeed deter private capital in higher-cost jurisdictions.

Were one to stop here, the decelerationist case would seem plausible. But it is anchored in a snapshot of today’s training paradigm and today’s revenue expectations.

Training paradigms and the locus of value

Current models are largely trained de novo: vast corpora ingested, weights frozen, then fine-tuned and reinforced. Each generation is, in effect, educated from birth. Continuous and online learning point to a different future—machines that accumulate knowledge, discard errors and adapt incrementally, more like humans than static snapshots. Such approaches could materially reduce upfront capital expenditure.

François Chollet, a long-term fixture within AI, has argued that:

AI in 2040 will not be built on the stack we are using today. It will be much closer to optimal. The current stack has 3–4 orders of magnitude of data inefficiency and 4–5 orders of magnitude of compute inefficiency. Near-optimal AI is what symbolic learning will deliver. People struggle to differentiate fluid intelligence from knowledge because, given enough preparation, memorized templates become a solid substitute for on-the-fly adaptation.

If correct, today’s capex-heavy regime is transitory. Even more fundamental is the question of where value accrues. At present, much attention focuses on the token: the marginal unit of machine-generated intelligence. Yet tokens may resemble electricity—indispensable, but priced near marginal cost.

Sam Altman, CEO of Open AI, has said:

We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter and use it for whatever they want to use it for.” Utilities can be profitable, but competition and scale tend to compress margins.

Commoditization need not imply triviality. Lumber varies by grade; so will tokens. Intelligence strata will likely be segmented and priced accordingly. But competitive markets have a habit of pushing even differentiated goods towards cost. If so, value may migrate up and down the stack—to specialized applications, domain-specific harnesses, proprietary data, workflow integration and complementary services.

Electricity offers a precedent. It is a metered commodity, and yet, it enables almost unimaginable valuation upstream value creation. Linux offers another. The operating system became ubiquitous and largely free, yet vast value accrued to firms building atop it. Open-weight models could perform a similar function, widening the pool of contributors and accelerating innovation in the layers above. At present, meaningful participation in frontier development is largely confined to a handful of laboratories. Open weights lower that barrier.

At their core, Large Language Models are conditional probability machines: mathematical transformations mapping input tokens to outputs. Their power lies in the quality of those transformations; their vulnerability lies in their reproducibility. Consistency invites standardization; standardization invites competition.

An eye towards the future

Does this make open weights decelerationist? Only if one assumes that today’s frontier business models are the sole engine of progress. That is a narrow view. Lower intelligence costs expand the frontier of viable applications. Broader access increases the number of minds experimenting at the edges. Both dynamics are, in aggregate, accelerationist.

None of this negates legitimate concerns about governance or misuse. Nor does it guarantee that capital will be allocated smoothly. Subsidized overcapacity—whether in steel, solar or tokens—can be disruptive. But disruption is not synonymous with deceleration. It may reallocate rents without reducing aggregate innovation.

If nothing changes, intelligence will likely drift toward utility status, priced competitively and embedded everywhere. Frontier labs may earn thinner margins on base models but richer ones on integration, tooling and specialized services. China will build capacity; so will others. The contest will be less about who owns the weights than about who captures the ecosystems around them.

Open weights, in this light, are not inherently decelerationist. They are a bet that the locus of value is shifting and that broader participation will expand the pie faster than it erodes incumbent rents. The risk is not that progress slows, but that business models anchored in scarcity struggle in a world trending toward abundance.

Acceleration, in other words, may depend less on guarding the factory gates than on deciding what to build once the gates are open.

Everyone wants a verdict on artificial intelligence and employment. Will it destroy jobs or create them? Early forecasts leaned towards the apocalyptic: white-collar bloodbaths, mass redundancy, a hollowed-out middle class. Even measured analysts felt obliged to quantify the carnage. Investors, workers and politicians alike demanded numbers.

Those numbers are arriving. They are also deeply confusing. Some firms are hiring feverishly; others are trimming headcount and still others are intelligently reconstituting. Productivity appears to be rising in pockets, yet aggregate labor-market data remain stubbornly unremarkable.

The temptation is to read the tea leaves, but in the short-term there are too many overlapping forces that make the data difficult to parse. It won’t be until the business cycle turns that we understand the real implications. In the interim, we should focus on what the machines are and will be capable of.

Investment, then return

When a new frontier opens, the result is rarely a measured march; we call it a gold rush. The current scramble for artificial intelligence follows the classic logic of the land-grab. Entrants are blitzing the market to secure a competitive foothold, spending down free cash flow and borrowing heavily against the promise of future returns. For now, the focus is on inputs rather than efficiencies.

Start with the physical build-out. Data centers the size of industrial parks are sprouting across the plains; specialized chips are back-ordered for months. The boom supports a secondary army of plumbers, electricians and cooling technicians required to keep the silicon humming.

It also sustains a sprawling new ecosystem of trainers, annotators and “frontier deployment engineers” (FDEs) tasked with localizing general models for corporate use. The market for training data alone is now estimated at $96bn. This is not the footprint of a shrinking industry, though it has an odd “work yourself out of a job” quality. On platforms like Mercor, contract workers produce the very output that will eventually automate their roles.

Source: Deedy Das, Menlo Ventures

It is not just the providers who are spending. The “consumers” of the technology—large enterprises across every sector—must allocate staff and capital to figure out how to use it. Importantly, the core business does not stop while this happens. This is “additional” spend. When magazines shifted from print to digital, they did not flick a switch; they bore the crushing double costs of both printing presses and web servers for years. AI adopters now face a similar period of “double running,” where legacy teams sit alongside the automated systems meant to eventually replace them.

The ultimate goal, of course, is a return on this investment through higher revenues or lower costs. It is the business cycle that determines whether firms pursue the revenue path or the cost path. For now, the focus is primarily on revenue expansion for long-term competitiveness. But we will not know the impact on labor until the cycle turns towards cost cutting.

The tells of technology

In times of uncertainty it’s important to build upon things that are certain. Machine learning is a remarkably predictable discipline: you make what you measure. In the jargon of the field, the “loss function” determines the direction of progress. Today, that direction is pointed squarely at white-collar work.

The technology is clear: AI’s ability to perform knowledge work is improving faster than the humans performing it. As models become adept at operating computers, the boundary between high- and low-skill work will blur. Routine digital tasks, once the preserve of junior professionals and middle managers, are becoming automatable commodities. This shifts the premium toward judgment, coordination and the creative problem-solving that machines still struggle to mimic.

GDPval was designed by Epoch AI to measure such abilities. Versus a human benchmark of 1,000 frontier models have blown past humans.

A digital brain without hands cannot accomplish much. That is why the most telling recent advance is not in conversation, but in “computer use.” Systems such as Codex can now see, click and type across a user’s desktop with their own cursor. Multiple agents can operate in parallel, navigating software that lacks an API or iterating on front-end code. For a developer, this is not a toy; it is a colleague that never sleeps.

Firms are already preparing for this shift by pushing their own IT budgets to “self-fund” AI experiments. This is an implicit bet on a new kind of workforce. It suggests that, in the eyes of capital, an augmented worker is worth far more than a manual one. But it also creates a macroeconomic vulnerability. When the business cycle eventually forces a pivot from growth to austerity, firms will have a capable, cheaper alternative ready to go. The choice to swap a salary for a subscription will be harsh, but economically rational.

The long-term manifestation remains a matter of structural debate. Alex Imas, of Google DeepMind, suggests that sectors most susceptible to automation may simply shrink as a share of the total economy. Rather than a sudden disappearance of work, AI could compress the time spent on automatable tasks, forcing an expansion into complementary roles we have yet to name. In this view, the composition of economic output changes long before the employment level does.

The journey to greener pastures

None of this implies that disruption will be painless. Change on this scale rarely is. When the cycle weakens, and as automation becomes cheaper and more reliable, job reductions will outpace job creation for a time. Certain tasks will vanish. Some firms will use AI bluntly, as a cost-cutting cudgel.

But it would be equally misguided to retreat into protectionism or to throttle innovation pre-emptively. Open markets and competitive pressure are precisely what translate technological potential into broadly shared gains. The appropriate response is adaptation: continuous skill acquisition, portable benefits and labour-market institutions that ease transitions rather than freeze incumbents in place.

The labor market’s current volatility says less about AI’s ultimate destination than about the economy’s transitional state. Short-term noise is inevitable when capital, code and commerce all move at once. The long-term signal—that opportunity and upheaval will arrive together—remains intact.

For individuals, the implication is straightforward. Iterate on skill sets. Become so valuable that they can’t ignore you.

In the neon-soaked world of Tron: Legacy, the Grid offers the fantasy of perfect control: a realm of speed, symmetry, and programmable order, stripped of the friction and ambiguity of ordinary life. Kevin Flynn enters it as a creator-god, intent on building something purer than the world he left behind. But the tragedy is that the deeper he retreats into this synthetic ideal, the more completely he loses the human world that gave his original vision meaning. The warning is not subtle: when an engineered reality becomes more seductive than the messy one outside it, transcendence evolves into exile. We do not escape nature. We become strangers to it, and finally to ourselves.

Today’s workplace shows some mimicry of the Grid. The “10x performers” who seem to outpace their peers are increasingly permanently plugged into artificial intelligence to remain competitive. In the author’s experience, these professionals are rarely seen without a model whispering in their ear. They inhabit a state of perpetual connection, treating internal cognition as a secondary processor to a primary, synthetic core.

There is more here than ambition. The architecture of AI (unintentionally I believe) plays into our cognitive biases. Immersion has a cognitive cost. In some ways, AI breaks our brains—not dramatically, not overnight, but incrementally. It alters how attention is allocated, how intuition forms, how trust is assigned and how confidence is calibrated. The result is not stupidity. It is something subtler: a quiet erosion of independent judgment.

Fortunately, we can learn from the decades of work by Daniel Kahneman, Amos Tversky, Charlie Munger and others – which we explore in the paragraphs below.

The surrender of System 2

Begin with attention. Cognitive science distinguishes between the “task positive network”, activated when focused on problem-solving, and the “default mode network”, engaged during reflection, mind-wandering and internal narrative-building. Creativity and long-term integration depend on oscillation between the two. An “always on” AI workflow traps users in perpetual task mode—prompt, respond, refine, repeat.

This matters because expertise requires more than output. It requires consolidation. Repeated practice gradually shifts effortful, analytical “System 2” reasoning into intuitive “System 1” fluency. A radiologist no longer consciously enumerates every anomaly; a seasoned executive senses a weak strategy before articulating why. Intuition is compressed experience.

Remove immersion and the compression stalls. A striking study of Polish endoscopists illustrates the danger. Experienced physicians—each with more than 2,000 colonoscopies behind them—were given access to an AI system that flagged precancerous lesions in real time. During the three months before its introduction, they detected adenomas in 28.4% of procedures. After adoption, when the AI was unavailable, their detection rate fell to 22.4%.

The doctors did not become less intelligent. They became less practiced. Reliance on the system dulled pattern-recognition skills that had once been automatic. The shift was measurable and swift. What had migrated into intuition drifted back into effortful uncertainty. AI, in other words, can interrupt the very learning loop that creates expertise.

A broader set of experiments extends the point. Researchers proposing a “Tri-System Theory” argue that artificial intelligence constitutes a third cognitive system—external to the brain yet integrated into reasoning. Across multiple studies, participants given access to AI improved markedly when it was accurate and declined sharply when it erred. Accuracy rose by 25 percentage points in the former case and fell by 15 in the latter.

More striking was what accompanied the errors: confidence. Participants displayed what the authors call “cognitive surrender”—adopting AI outputs with minimal scrutiny, overriding both intuition and deliberation. Even when wrong, AI increased users’ certainty. Time pressure and incentives altered baseline performance but did not eliminate the pattern. System 3 reframed judgment itself.

Fluency, charm and the slot machine

Why does this surrender occur? Partly because humans are wired to equate fluency with truth. Statements delivered smoothly and confidently are judged more trustworthy than halting ones, regardless of content. Generative models are fluent by design. They hedge politely, structure arguments elegantly and answer instantly. The cues that once signaled competence now signal computation.

Humans are also susceptible to charm. We anthropomorphize readily, especially when systems mirror our language. A model that appears attentive and adaptive triggers social heuristics evolved for interpersonal trust. The result is a subtle miscalibration: skepticism softens not because evidence improves, but because delivery does.

Layer on top the architecture of reward. Each prompt carries the possibility of a brilliant response – a novel framing, a line of code that works, a synthesis that surprises. This is a variable reward schedule, the same mechanism that underpins slot machines. Most outputs are useful; some are exceptional. The uncertainty itself keeps users engaged.

Combine perpetual task orientation, cognitive surrender and variable reward, and a familiar modern condition emerges: burnout. High speed, limited control, continued emotional investment in outcomes. Workers remain accountable for results even as they outsource parts of the thinking that produce them. They are responsible, but less autonomous.

It would be easy to leap from here to alarmism: AI as cognitive decay, a generation unable to think unaided. That would be overstated. Used judiciously, AI demonstrably enhances productivity and can scaffold learning. The Polish physicians detected more lesions when assisted. Participants solved more problems when the system was right. Augmentation is real.

The danger lies not in use, but in unreflective use. If the first reach in any ambiguity is to “go to Claude”, anchoring effects follow. Initial answers, even arbitrary ones, shape subsequent reasoning. Starting with a model’s frame makes deviation harder. Over time, the space of considered ideas narrows—not by coercion, but by convenience.

Designing for autonomy

The appropriate response is neither abstinence nor surrender. It is humility. Individuals should keep one eye on the horizon: not this week’s output, but decades of accumulated expertise. That means deliberately unplugging at times, resisting the reflexive first prompt and allowing problems to incubate without assistance. Deep understanding still compounds.

Yet willpower is a finite resource. Cognitive restraint, like a muscle, fatigues. Expecting individuals to heroically override optimized systems is unrealistic. If constant connection erodes autonomy, the burden cannot rest solely on users to disconnect.

Firms therefore face a design choice. They can build workflows that default to immediate AI consultation, or they can embed friction that preserves human learning loops: periods of unaided analysis before comparison, prompts that surface uncertainty, metrics that reward skill development rather than mere throughput. In medicine, this might mean structured intervals without assistance; in offices, norms about drafting before querying.

Public discourse has a role, too. As AI becomes a de facto cognitive infrastructure, its effects on judgment and accountability become collective concerns. Transparency about limitations, research on long-term skill retention and standards that privilege augmentation over substitution would preserve both innovation and agency.

The promise of AI is real. But so is its capacity to reshape the minds that use it. The choice is not between plugging in and opting out. It is between drifting into cognitive surrender and deliberately designing systems that strengthen, rather than supplant, human judgment.

Harvey Dent, the ill-fated district attorney in The Dark Knight, observed that one either dies a hero or lives long enough to see oneself become the villain. For OpenAI and Anthropic, the stakes are less cinematic but no less existential. These firms have been the undisputed heroes of a new technological age, conjuring intelligence that has seemed to defy the limits of silicon. They have pioneered the frontier, setting the pace for a world eager to be disrupted

But the heroic era of the isolated lab is ending. As the tides of the first AI boom shift, OpenAI and Anthropic are discovering that raw intelligence is an insufficient moat. They are being hemmed in by boundary conditions from above, below and the sides. From above, what began as voluntary safety pledges has hardened into a regime of state interest. Scrutiny of recent releases, most notably the regulatory skirmishes surrounding the staged rollouts of Anthropic’s Fable and GPT-5.6, signals that “speed above all else” has been replaced by a requirement for an implicit state license, enforced by export controls and sharpening governance frameworks.

From below, the assumption of a permanent proprietary lead is being tested by Chinese open-weight models from DeepSeek, Z.ai, and Kimi, which offer real innovations in efficiency rather than mere imitation. As hyperscalers adapt and “in-house” these models, much of the addressable market becomes commoditized. Simultaneously, pressure is closing in from the sides. The success of productivity tools like Claude Cowork and agentic harnesses mean the labs are not “just” model providers – they are becoming rivals to the very software titans they rely on for survival.

Strategy 1: Own the Pareto Frontier

The most obvious path is to double down on the intelligence curve itself. Instead of focusing primarily on the upper end of intelligence, OpenAI and Anthropic could offer a product mix that spans the entire frontier, from lightweight “edge” models to massive reasoning engines, as well as routing across the models. Their lead at the frontier also means they can best distill intelligence “down the curve”. This is a path of convenience for customers – reducing the surface area of model governance as well as the extent of vendor management.

Strategy 2: Win the end user

Technology can be commoditized, but human habits are famously “sticky.” If the labs hasten their move from being invisible infrastructure to indispensable digital colleagues, they create switching costs that are emotional as well as technical. Anthropic’s deep Slack integration and Claude Cowork hint at a future where AI is embedded in the daily workflow. Importantly, the labs could differentiate through personality—a lever that remains largely untapped. While today’s models are blandly interchangeable, a digital employee with a distinct, helpful “persona” could win a loyalty that raw benchmarks cannot buy.

Strategy 3: Corner the market

Labs may find that the best moat is not built of code, but of complexity. By leaning into the regulatory burden, incumbents can ensure that the “model release process” becomes so onerous and expensive that it effectively shuts out smaller, open-weight innovators. They could also corner the market through hardware: if frontier models require such specific, scarce infrastructure that they cannot be served efficiently elsewhere, quality parity matters less. A darker version of this strategy involves a bargain with the state where the government gets rolling 6 month access to the frontier in exchange for funding it.

Strategy 4: The foundation model factory

Open-weight efforts are currently focused on general-purpose intelligence, creating a gap in high-precision vertical domains. OpenAI and Anthropic could pivot to become “factories” for specialized foundational models in life sciences, mathematical proofs, or semiconductor design. This approach leverages their expertise in data curation and alignment to build defensible niches where “99% accuracy” is the minimum requirement. Specialization also allows for more efficient compute usage, decoupling the labs’ fortunes from the brutal, lower-margin wars of general chatbots.

Strategy 5: Become a hyperscaler

A more radical move is to redefine the “cloud” entirely. In this scenario, the labs attempt to marginalize standard compute and data storage by making the AI harness (models, governance layers, and agentic frameworks) the primary organizing principle of infrastructure. We see hints of this in AI Giant long-term data-center deals and the rise of AI-native “neoclouds.” Yet, building a hyperscaler requires a mastery of logistics, go to market and capital that dwarfs model training. It also invites total war with Microsoft, Google, and Amazon on their home turf.

Strategy 6: Push deeper into SaaS

The final route is to move up the value chain by creating vertical-specific software offerings. By building their own productivity and design suites, the labs can capture the high margins typically reserved for their customers. This is already visible in recent product launches. However, this strategy is fraught with market risk; it turns every partner into a competitor. The increasingly assertive posture from Microsoft suggests that the software giants will not watch quietly as challenges to their core business rises.

The choice of coexistence

Against this is the fact that the AI Giants are not yet self-sustaining and rely on a continuous influx of external capital to survive. Gorilla fights are always violent – but insufficient capital can make them deadly. The wild success of innovation, brand and innovation has outstripped the parallel success of profitability.

And yet, risks need not turn into reality. Such a conclusion misses the sheer scale of the lead the AI Giants still hold. The talent, data, and institutional knowledge at OpenAI and Anthropic are not easily liquidated. They are not on the verge of collapse; they are at strategic inflection points. The true danger is not failure, but the dilution of their mission should they fight battles on too many fronts simultaneously.

A more pragmatic prescription is to embrace the reality of a layered market. The hyperscalers will remain the masters of the “plumbing”; fighting them for the infrastructure of the internet is a misuse of elite talent and scarce capital. Survival now requires the discipline to distinguish between a market to capture and a boundary to respect. OpenAI and Anthropic must accept that the era of the all-conquering lab is over. Scaling laws have served them well in the data center, but they are a poor guide for the boardroom. To avoid the Shakespearean arc of the fallen hero, the pioneers must finally master the prosaic art of the partner.

For the curious: this author prefers a combination of The Pareto Frontier and Foundation Model Factory, with some work on personalization.

The business model of the AI giants rests on four linked propositions. First, the unit cost of intelligence will fall as compute improves. Second, the best labs will use their own systems to accelerate research, widening their lead. Third, customers will adopt AI deeply enough to make usage sticky. Fourth, capital will tolerate heavy losses until the profits arrive.

A year ago, each assumption looked sturdy. Chips were improving, frontier models were pulling away, adoption curves were steep and investors were willing to finance almost any plausible claim on the future. The industry did not need every bet to pay off immediately. It merely needed the direction of travel to remain clear.

That clarity is fading. The issue is not that AI has failed. The technology is improving, demand is real and the best companies remain formidable. The problem is that the economic architecture around them is under strain. When several core assumptions weaken at once, capital-intensive industries rarely glide smoothly to maturity. They reset.

The cost of intelligence

Start with the first assumption: that compute efficiencies will keep pushing down unit costs. On one level, this remains true. The newest chips deliver more throughput, better performance per watt and impressive gains on the metrics that matter to model builders. The difficulty is that compute is not just chips. It is memory, storage, networking equipment, land, cooling and electricity.

Those surrounding costs now matter more. Rising prices for RAM, hard drives and other components can erode the gains from faster processors. So can grid constraints and long-term power contracts. A model that is cheaper to run on paper may still become more expensive to deliver at scale if the rest of the infrastructure stack moves against it.

After a run-up of circa 700%, Jefferies forecasts prices to increase 40 to 50% in Q3, 30 to 40% in Q4 and another 40 to 45% in 2027. In total that could be a 20x increase in memory prices. Importantly, most advance purchases are by Cloud Service Providers, potentially leaving Consumer Electronics companies in more difficult positions. This provides context on Tim Cook’s recent notice that Apple will likely have to increase prices.

The right analogy is not Software as a Service, where marginal distribution costs approached zero. It is closer to a heavy industrial input. When oil becomes too large a share of economic activity, strain spreads beyond energy markets. Investment plans change, consumers pull back and policymakers start to worry. Semiconductors are a critical input whose price can ripple through the economy.

That changes the investment equation. Frontier AI depends on enormous upfront spending justified by future volume and falling costs. If costs decline more slowly than expected then returns compress. If they rise in important categories, firms must choose between higher prices, lower margins or less ambitious products. None is fatal. All make the business model less automatic.

The moat around the model

The second assumption is that the leaders will compound their advantage by using AI to improve AI. There are early signs of automated research, better coding assistants and tools that help scientists search a wider space of ideas. Yet full recursive self-improvement remains more slogan than operating model. The more immediate surprise has come from elsewhere: the speed of capable open-weight competitors.

As an example, systems such as GLM 5.2 from Z AI show how quickly strong models can emerge outside the best-known frontier labs. They need not beat the leaders at every task to matter. Not every task requires a rocket scientist. In commercial markets, “good enough” is often enough. A cheaper model that handles the bulk of customer-service, document-processing or internal-search tasks can discipline the price of a more brilliant one.

As AI becomes more common, firms will match intelligence to need. The most capable model may be used to design a workflow, solve a hard problem or generate a reference answer. Once the process is understood, cheaper models can take over much of the routine work. This expands the market for AI while weakening the pricing power of the frontier.

Customer attrition costs

The third assumption, stickiness, is therefore more complicated than it first appears. Firms can justify large customer-acquisition costs when adoption creates lock-in or network effects.

This is a complicated question with respect to the AI giants. To understand why, we must put ourselves into the head of a Chief Information Officer. As CIO, two of your primary functions are Infrastructure and Productivity. Think of these as Automation and Augmentation.

In automation, where AI agents act on behalf of a company, chief technology officers have every reason to pursue multi-model strategies. Dependency on one supplier is a risk to be managed, not a feature to be celebrated. I dealt with this extensively in previous roles. Once you have proven viability (frontier models) you focus on reducing costs (open weight models).

Augmentation is stickier. Workers grow attached to tools that sit inside their daily routines. Skills, habits and workflows accumulate around them. That is why partnerships and integrations matter so much. An AI assistant embedded in email, documents, meetings and code repositories is harder to displace than a model accessed through an interchangeable interface.

Yet this is where the industry’s alliances become awkward. AI assistants challenge the productivity suites of Microsoft and Google, even as those companies provide cloud infrastructure, distribution and capital to parts of the AI ecosystem. Model-makers need platforms; platforms do not want to become mere utilities. The line between partner and rival is being redrawn in real time.

Capital’s patience has limits

The fourth assumption is that capital will bridge today’s research intensity to tomorrow’s profitability. That was easier to believe when costs were falling, moats looked wider and adoption seemed likely to compound without friction. It becomes harder when input prices rise, open models improve, enterprise buyers resist lock-in and regulators begin asking sharper questions about power, data and competition.

In past technology booms, such moments have produced a familiar sequence. Valuations reset. Credit tightens. Weaker firms are acquired, restructured or abandoned. Regulators become more assertive just as investors become less patient. The particulars vary; the pattern does not.

“Rent is too damn high”

The silver lining: the leading AI firms have extraordinary strengths: dense talent, famous brands, vast usage, deep technical infrastructure and products that customers already value. The technology is not vaporous. The question is whether it can be delivered at a price that households, companies and governments are willing to bear.

Affordability is therefore not a side issue. It is the strategic constraint. If AI raises phone prices, software bills and electricity costs, today’s skepticism will harden into opposition. If it becomes cheaper, more interoperable and more clearly useful, customers will pay a premium for “Made in the USA” and politics will remain manageable. The social license of the boom depends less on awe than on value for money.

The current trajectory of higher input costs, higher competition, higher substitution, higher regulation and higher equity and debt issuance is not sustainable. The correction in leverage will be painful, as will be the competitive battles as firms fight for “The Next Normal.” But there is a silver lining and a longstanding good for society on the other side.

It’s computationally intractable to verify the properties of an LLM. Yet we must simultaneously retain and amplify the benefits of this technology while controlling the potential downsides.

Yesterday’s US government directive to Anthropic may yet prove to be either prudent or excessive. For now, that is almost secondary. Citing national-security authorities, the American government ordered the firm to suspend access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign-national employees. Anthropic responded by disabling the models for all customers.

The immediate dispute is narrow: whether the underlying “jailbreak” justified such an abrupt intervention. But the larger significance lies elsewhere. Episodes like this are a preview of a harder question governments will face with more powerful models: what should officials do when they believe a system may expose dangerous capabilities, but cannot fully determine in advance what that system knows or can do?

Anthropic’s own statement is strikingly restrained. The company said the government’s letter, received at 5:21pm Eastern Time, did not provide specific details of the national-security concern. Its understanding, however, was that officials had become aware of a way of bypassing, or “jailbreaking”, Fable 5. Anthropic said it reviewed a demonstration of the technique being used to identify “a small number of previously known, minor vulnerabilities”, adding that other publicly available models could discover them without requiring a bypass.

That will strike some readers as evidence of official overreach. It may be. But even if this particular intervention was mistaken, there will eventually be one that is not. At some point, a model will expose capabilities that genuinely warrant restriction, suspension or redesign. When that moment comes, policymakers will run into a problem more basic than geopolitics or corporate trust: testing what a model knows is not simply difficult. In an important sense, it is computationally intractable.

Two risks, one governance problem

There are two distinct hazards here, and public debate often muddles them together. The first is capability exposure: a model reasons its way toward an adverse outcome, perhaps identifying software vulnerabilities or assembling dangerous knowledge from harmless pieces. The second is memorization: a model reproduces material from its training data, including sensitive, proprietary or private information. Both are forms of leakage. But they arise in different ways.

The memorization problem is no longer speculative. In “Scalable Extraction of Training Data from (Production) Language Models”, Nicholas Carlini and co-authors showed that large language models can be induced to emit verbatim training examples, including rare strings and sensitive content. Models do not merely generalize from their data; under some conditions, they also remember it.

That matters because “unlearning” remains immature. Research such as “Rethinking Machine Unlearning for Large Language Models” suggests that reliably removing particular information once it has been absorbed is far harder than deleting a file from a database. Knowledge is distributed across weights and entangled with other representations. One cannot simply tug out a fact without affecting nearby behaviour.

Capability risk is harder still, because the issue is not what the model has memorized, but what it can derive. A system may never have seen a harmful answer stated outright and yet still reason its way toward one. That is why a policy regime built around patching known bad outputs may prove brittle. The relevant danger is not only retrieval. It is inference.

Why the math matters

Here the theoretical result matters. “Transformers are Inherently Succinct” shows that verifying certain properties of transformer models is EXPSPACE-complete. The jargon is forbidding, but the intuition is simple enough. LLMs are so compact in how they represent behavior that checking, exhaustively, whether they always satisfy a rule would require an astronomically large amount of computation.

In other words, the obstacle is not merely that firms need more engineers, more compute or more red-teamers. It is that there may be no feasible way to inspect every relevant behavior in advance. Anthropic itself has acknowledged, in a separate policy proposal, that transparency alone is no longer sufficient and that governments may need the power to block dangerous deployments. The trouble is that such powers are easiest to justify precisely where exhaustive verification
is least feasible.”

This does not mean all safety work is futile. Far from it. Testing, monitoring and red-teaming can reduce risk materially. But they cannot deliver certainty, and much current policy talk quietly assumes that, with enough diligence, dangerous capabilities can be fully identified before deployment. The mathematics suggests otherwise.

That has awkward implications for governance. Export controls, access restrictions and emergency directives presume that one can identify a bounded problem and contain it. Sometimes one can. But if dangerous behavior emerges from a vast combinatorial space of possible prompts, latent knowledge and chains of reasoning, then the line between acceptable and unacceptable output will often be unstable.

What comes after the audit illusion

This is why the Anthropic episode matters beyond its facts. It points to a future in which institutions are asked to intervene in systems they cannot comprehensively inspect. That is not an argument for panic. It is an argument for giving up the comforting idea that brute-force auditing will solve the governance problem.

Several research directions follow from that admission. One is the use of “guardian” models that monitor inputs and other models’ outputs, detect risky behavior and block dangerous responses. Another is specialization: building powerful systems that are narrow enough in scope to reduce the range of hazardous reasoning they can perform. A third is architectural separation, in which reasoning and knowledge storage are segmented so that access to especially sensitive domains can be gated more cleanly.

All of these ideas are plausible. None is mature. Guardian systems may inherit the same opacity as the models they supervise. Specialization may trade general usefulness for lower risk, but at a potentially large commercial cost. Segmentation sounds elegant, though it remains uncertain whether powerful reasoning can be cleanly disentangled from broad access to knowledge. These are not finished solutions. They are open research programs.

The temptation, as ever, will be to lurch toward extremes. One side will conclude that frontier models are fundamentally uncontrollable and should be halted. Another will insist that market incentives and incremental safeguards will suffice. Both are too easy. The more likely reality is messier: these systems will continue to improve, firms will keep deploying them, and governments will intervene episodically when risks become visible enough to be politically salient.

Yesterday’s directive may or may not have been justified on its own facts. The deeper problem is that one day a directive like it will be justified, and we still will not know how to prove, in advance, exactly what the model in question can do.

Most dialogs around agentic coding are highly polarized. On one side are skeptics who see brittle demos, messy code and conclude the enterprise is overhyped. On the other are those who believe we’ve nearly automated, that large models can now out-program the average engineer and will soon outstrip the best. In early 2026, Anthropic’s Dario Amodei said “we may be 6 to 12 months away from when models are doing what software engineers do end-to-end”.

Both camps have a point. Both are also wrong in ways that matter. If we want to solve the problem we need to evolve the conversation and change our metrics. Paul Graham famously quipped “You make what you measure”.

These machines are extraordinary at some tasks. Give them a greenfield repository and a well-scoped feature, and they can prototype in hours what once took days. Point them at a body of documentation and they will synthesize research, surface edge cases and draft workable implementations at startling speed. In narrow lanes, the productivity gains are real.

But shift the problem slightly and the magic fades. Ask a model to concurrently balance long-term architecture, to weigh non-functional constraints, or to optimize across conflicting objectives and its confidence can mask fragility. Code may “work” while quietly accruing debt. The diff passes tests but the system deteriorates.

As we shift from high-level conversations to details I believe we’ll find we have more agreement than disagreement.

Coding is not engineering

This is why Anthropic’s Boris Cherny’s observation deserves careful reading. “Coding is just one part of engineering,” he notes. There’s also debugging, operating services, scaling up infrastructure, deciding what to optimize, setting up hardware and capacity, talking to users, product planning, etc. Coding is the easy part, everything else is not yet solved.”

Benchmarks rarely reflect that distinction. SWE-bench-style evaluations reduce software to a pass/fail problem: does the patch satisfy the test? That framing is convenient. It is also misleading. Real software is not a single-objective optimization problem. It is a multi-objective negotiation across architecture, functional requirements, non-functional guarantees and time to value.

Every engineering team lives inside trade-offs. Ship now or refactor? Optimize for latency or developer velocity? Accept technical debt to capture market share, or delay launch to harden infrastructure? These are not edge cases; they are the core of the job. A model that treats code generation as a closed-form puzzle misses the political economy of production.

Cognition’s “FrontierCode” benchmark is a step towards acknowledging this complexity. Its tasks reportedly required 40 or more hours from leading open-source maintainers. It asks a subtler question than “does it pass?”: would you actually merge this code? That shifts the objective from syntactic validity to institutional acceptability. Sloppy code that works is not the same as maintainable software.

Measuring what matters

To great software rather than merely more software we must start with better metrics. The benchmarks that dominate discussion today mostly reward what is easy to score: whether code runs, whether a patch passes, whether a benchmark turns green. Useful, certainly. But we need to improve how we measure harder concept: code smells, architectural quality, understandability, and non-functional coverage across latency, reliability, security and cost.

These metrics build should coalesce into a new set of “evals” that focus on the combined harness, prompts, and model. For example:

• Architecture Bench: Measures whether a model creates the right degree of modularity, resiliency and complexity for the problem.

• Bug Bench: Measures whether a model can reliably find, explain, and fix real defects in an existing codebase.

• Affinity Bench: Measures whether a model produces code that fits the style, patterns, and conventions of the codebase it is working in.

• NFR Bench: Measures whether a model can satisfy non-functional requirements such as latency, reliability, security, scalability, and cost.

Successful iteration is more important than successful one-shotting. This is why test time compute is critical. OpenAI’s Noam Brown writes about this in Implications of Large-Scale Test-Time Compute: “As LLMs become more capable, benchmark performance is increasingly a function of test-time compute”.

Many of the biggest gains will come from applying AI to the harness itself – very likely a world model. The test environment, critic loop, refactoring cycle and bug-finding machinery are not just scaffolding around intelligence; they may be where much of the intelligence lives. Progress in agentic coding may depend less on a single model becoming universally capable than on recursive systems that can inspect, challenge and steadily improve their own work.

The dividends of clarity

Sobering up the conversation would yield broad dividends. For investors, precise tiers of capabilities replace hype with realism, allowing capital to flow toward durable platforms rather than ephemeral wrappers. When capital transitions from speculative bubbles to productive physical deployment, the risk of a ruinous funding winter recedes. Clear taxonomy turns silicon-valley gambling into sensible, long-term capital allocation.

Equally, it would align the developer community around the right problems. By shifting focus from simple pass-fail code generation to complex, multi-objective metrics like maintainability and architectural coherence, researchers will build agents that are genuinely collaborative rather than merely fast. Industry planning becomes possible when progress is measured in steady engineering steps rather than sensational press releases.

The future of agentic coding is not a binary choice between empty prompts and automated enterprises. It is a messy, gradual expansion of tools that still require human trade-offs and judgment. Progress is rarely about finding a single, magic algorithm to solve engineering. Rather, it is about admitting that the hardest part of building things has always been deciding what to build in the first place.

Today’s LLM research, Andrej Karpathy has remarked, “is about summoning ghosts.” The phrase captures the eerie competence of large language models. Systems such as Claude Mythos can identify and exploit security vulnerabilities in byzantine software stacks. Headlines oscillate between awe and alarm. The machines, it seems, have breached the citadel.

And yet, talk to working engineers and the mood shifts. Ask a model to build or refactor a production codebase and the magic fades. The code may run, but, as Mr Karpathy puts it, it can induce “a little bit of a heart attack”. It is “very bloaty”, full of copy-paste and riddled with awkward, brittle abstractions. It works. It is also, in a word, gross.

Both observations are true. Once the underlying logic is clear, they are not even especially contradictory. The same system that can expose a security flaw in a complex stack may still struggle to design a clean, extensible module. Wired recently reported that “vibe-coded” applications have exposed corporate and personal data on the open web. The models can break things in sophisticated ways. Building them properly is another matter.

The tension is not a paradox. It is a clue.

The logic of “there exists”

To see why, it helps to borrow a lens from complexity theory. Computer scientists distinguish between problems of verification and problems of synthesis. In the language of logic, the first often take the form “there exists” (∃). There exists a configuration that breaks this system. There exists a counterexample. There exists a path through this maze.

Such problems can be fiendishly hard, but they are single-objective. Find one witness and you are done. OpenAI’s recent work on the unit distance problem, for instance, relied on proof by contradiction: assume a configuration exists and show that it cannot. The intellectual move is destructive. It rules possibilities out. It needs only one decisive blow. In that sense, the same broad logic applies to Mythos-style security exploits.

By contrast, synthesis problems look much more like everyday design problems. Consider planning a public transport network, laying out a hospital or building a bridge. It is not enough to find one route that works, one patient flow that is efficient or one set of beams that holds. The design must perform tolerably across many conditions at once: heavy usage, bad weather, awkward edge cases, future changes and tight budgets. These are multi-objective optimisation problems, in which success means balancing several goals rather than maximising just one.

That is why computer scientists often describe synthesis in the language of “for all” (∀). Build a system such that for all valid inputs it behaves acceptably; for all users, security and usability are preserved; for all likely extensions, the architecture does not turn brittle. In practice, one must usually combine this with some “there exists” (∃) reasoning too: there exists a design that, for all the conditions it will face, still holds together. Add the demands of simplicity, maintainability and future change, and the problem becomes harder still.

Many such tasks sit in higher layers of complexity, including the class known as Σ₂ᴾ—roughly, problems in which existence and universality are nested together rather than handled one at a time.

Ghosts in the reinforcement loop

Large language models, trained to predict the next token across vast corpora, are remarkably good at pattern-matching for witnesses. They can sniff out the one misconfigured permission in a sea of settings. They can propose an exploit path through a tangle of dependencies. These are existential feats: find a needle, not weave the haystack.

Software engineering, by contrast, is synthesis under constraints. It balances functional requirements—does it work?—with non-functional ones: latency, security, readability and cost. It juggles time-to-value against technical debt. A good system is not merely one that passes today’s tests, but one that remains legible and adaptable tomorrow. That is a “for all” ambition.

This also helps explain why some attempts to simplify model training have faltered. Mr Karpathy has described experiments in stripping reinforcement learning down to its essentials, only to find that “the models hate this”. Step outside the reward circuits and their competence degrades. These systems are not animal intelligences with a generalised drive to tidy and optimise. They are statistical engines tuned to specific objective functions.

None of this means that current achievements are trivial. Existential reasoning at scale is economically and strategically valuable. Nor does it imply that synthesis is forever out of reach. It does suggest that headlines about machines “hacking the world” are not inconsistent with their clumsiness as junior developers. The tasks differ not just in degree, but in kind.

Moving beyond “more of the same”

If nothing changed, progress would likely continue unevenly. Models would grow better at identifying vulnerabilities and generating functional prototypes. Companies would deploy them to accelerate greenfield development while relying on experienced engineers to impose order. Technical debt, already a chronic ailment, might accumulate faster in the hands of overconfident adopters.

The more interesting path lies in expanding the toolkit. Agentic software engineering will require combining today’s LLM breakthroughs with new techniques rather than simply scaling current ones. Systems will need richer representations of codebases, more recursive forms of reasoning and stronger pressure to minimise complexity, not merely generate working output. In practice, that means more varied objective functions, deeper use of formal verification, better tooling and tighter feedback loops. The prize will be large. Firms that can deliver both functional correctness and architectural soundness will capture disproportionate value.

Creation is harder than destruction because it must answer to more stakeholders. That is as true in complexity theory as in software engineering. The surprise is not that today’s models struggle with synthesis; it is that they have come so far on existential tasks so quickly.

The ghosts have been summoned. Building the cathedral will take longer. But there is little reason to think it cannot be done.

Post script: for a fantastic detailed explanation on OpenAI’s Unit Distance work, see Dr. Uday Kamath’s An AI Just Broke an Eighty-Year-Old Erdős Conjecture.

In May 2026, speaking at the Vatican, Chris Olah offered one of the clearest descriptions yet of what modern artificial intelligence has become. “AI systems are not engineered the way a bridge or an airplane is engineered,” he said. “We understand an airplane because we designed every part of it and we understand the physics that act on it. AI models are not like that. They are grown, on a structure roughly modeled after the brain, on an enormous inheritance of human thought and speech.”

That is not just a vivid turn of phrase. It is a better way of understanding the technology itself. Artificial intelligence remains a machine, in the literal sense that it runs on chips, electricity and mathematics. But it is not a machine in the old mechanical sense, where each part is designed for a known purpose and the whole can be explained as the sum of those parts. Modern Large Language Models are closer to cultivated things. Their makers shape the conditions under which they emerge, then study what has grown.

The best analog is not construction but winemaking. Wine is not “built” molecule by molecule. It is grown in vineyards, shaped by soil and weather, harvested at a particular moment, crushed, fermented, blended and aged. The vintner makes countless decisions. Yet the result is never merely assembled. It emerges from a managed process whose most important transformations happen indirectly. So it is with modern AI.

The distinction matters because language about technology often smuggles in assumptions about control. To say that a system is built suggests mastery. It implies that if one understands the blueprint, one understands the object. That works tolerably well for a jet engine, a bridge or even much conventional software. It works less well for systems trained across vast stores of human text, whose internal capabilities are not individually specified but statistically induced.

Winemaking, in brief

To understand artificial intelligence, it helps to begin somewhere older and more earthly: with wine. Winemaking is among humanity’s oldest technologies, yet for most of its history it did not advance by way of complete theoretical understanding. Vintners did not wait for microbiology before making better wine. They observed, experimented, copied what worked, discarded what failed and, over generations, turned scattered practice into accumulated craft.

That history matters because wine was improved long before it was fully explained. Growers learned that soil, climate and grape variety changed the character of a harvest. They discovered that timing the pick mattered, that crushing too harshly could spoil a batch, that skins left with juice for longer altered color and tannin, and that fermentation could be guided even when its chemistry remained obscure. Knowledge arrived first as habit, comparison and memory. Science came later, deepening a craft that had already travelled far by trial and error.

The same is true of the cellar. Crushing releases the juice. Maceration lets skins, seeds and stems shape flavor, texture and structure. Fermentation transforms sweetness into something more durable and complex. Aging softens, integrates and occasionally surprises. Some offshoots emerge almost accidentally and then become disciplines of their own. Sparkling wine, once a by-product of fermentation continuing in the bottle, became a category through refinement rather than sudden invention. The point is not romance. It is that sophisticated systems can emerge from iterative adjustment long before every inner mechanism is transparent.

Growing large language models

Modern AI has advanced in much the same spirit. Its decisive breakthrough was not a complete theory of machine intelligence but a fertile structure of “Scaling Laws” on which progress could accumulate. The Transformer, introduced in 2017, supplied that structure: a trellis, as it were, on which models could grow. Its attention mechanism allowed words in a sequence to weigh one another’s importance, making language more tractable at scale. But the Transformer was less a finished cathedral than a promising vineyard planted on good ground.

What followed was an era of cultivation. Researchers improved the flow of information through residual connections, stabilized learning through various forms of normalization, and scaled models from millions of parameters to billions and then trillions. Progress came from repeated adjustments, careful measurement and the willingness to run another season’s experiment. The field learned, as vintners do, that small interventions at the right stage can yield disproportionate effects at harvest.

Training a large language model is best understood in those terms. Human writing is crushed into tokens, passed through the model in vast quantities and subjected to repeated optimization. The system predicts, errs, adjusts and tries again. Over time, statistical pressure produces structure: fluency, recall, abstraction, pattern-matching, reasoning, and perhaps a proxy of consciousness.

And because AI is grown, it is necessarily grown from something: as Chris Olah put it, from “an enormous inheritance of human thought and speech.” Large language models are trained not on the world directly but on humanity’s textual residue—books, articles, conversations, code, manuals, arguments, jokes. They are products of computation, yes, but computation applied to a cultural archive. No engineer assigns these capabilities by hand any more than a winemaker scripts the taste of blackberry into a barrel. The craft lies in shaping conditions under which desirable properties emerge.

Later innovations fit the analogy too neatly to ignore. Reinforcement learning from human feedback acted like a second fermentation, taking a raw but powerful base and making it more palatable, more disciplined, more useful in conversation. Mixture-of-experts models resemble blending, in which different components contribute specialized strengths without requiring one uniform mass to do everything equally well. And now, as researchers revisit attention itself and explore new optimization schemes such as Muon, they are doing what mature crafts always do: returning to fundamentals with better instruments and a larger inheritance of knowledge.

That inheritance is the thread connecting vineyard and model alike. Winemaking advanced because each generation inherited techniques, observations and refinements from the last. AI progresses similarly. The field does not begin afresh with each model release. It inherits architectures, scaling laws, training recipes, evaluation methods and hard-won intuitions from earlier experiments. And the models themselves inherit something larger still: the stored language of humanity.

Ghosts, inheritances and play

So when one says that AI is grown, not built, the claim is not mystical. It is historical and methodological. The most important advances have come not from specifying every feature in advance, but from creating the right structure, feeding it rich material, observing the results and refining the process over time. That was how humanity learned to make wine. It is increasingly how it is learning to make intelligence, too.

That is why Andrej Karpathy’s strange metaphor lands with force. “Stated plainly, today’s frontier LLM research is not about building animals. It is about summoning ghosts,” he wrote. “They are these imperfect replicas, a kind of statistical distillation of humanity’s documents with some sprinkle on top.” The phrase is theatrical, but its point is precise. These systems are not independent minds manufactured from first principles. They are patterns coaxed out of human expression at scale.

This intuition is most beautifully expressed in Benjamin Labatut’s The Maniac, a fictionalized account of John von Neumann, the godfather of innovations spanning from computers to atom bombs to artificial intelligence:

“Before he became unresponsive and refused to speak even to his family or friends, von Neumann was asked what it would take for a computer, or some other mechanical entity, to begin to think and behave like a human being. He took a very long time before answer, in a voice that was no louder than a whisper. He said that it would have to grow, not be built. He said that it would have to understand language, to read, to write, to speak. And he said that it would have to play, like a child.”

The idea of play matters more than it first appears. During training, models are not following hand-authored rules for every situation. They are exploring a huge possibility space, rewarded for better predictions and penalized for worse ones. Their “play” is bounded, statistical and inhuman. But it is still a kind of development. Competence is not installed. It is cultivated through repeated encounters with examples.

This helps explain both the power and the opacity of modern AI. Because these systems are grown through optimization rather than specified feature by feature, their internal workings are often difficult to interpret. Researchers can identify circuits, tendencies and representations, much as a chemist can identify compounds in wine. But explaining exactly how trillions of adjustments yield a coherent model remains hard. The product can be analyzed after the fact more easily than it can be exhaustively foreseen beforehand.

What this changes

Thinking of AI as grown rather than built should not encourage mysticism. Neural networks obey mathematics. Data centers obey physics. There is no magic in the machine. But there is a different kind of engineering at work: one that resembles husbandry more than assembly, and experimentation more than command. The distinction is not semantic: it changes what sorts of expectations are sensible.

That is why Olah’s remark at the Vatican deserves to linger. It was not simply a metaphor offered for effect. It was a description of a turning point in how machines are made. The old image of engineering was of the builder with a blueprint. The new image is of the cultivator with a season ahead of him, trying to understand what the next harvest may bring.

Artificial intelligence is still made by human hands. But not, any longer, in quite the old way. The most important systems of this century will not just be constructed. They will be grown.

Financial analysts are currently engaged in a collective exercise in ruler-drawing. By mapping the trajectory of compute spending and GPU sales, they have constructed a future that is essentially a larger version of the present. In this view, the path to artificial intelligence is a matter of pure capital expenditure. If $100bn buys a certain level of reasoning, then $1trn must buy ten times as much. It is an investment thesis built on extrapolation.

History, however, is not linear. In any technological cycle, the most dangerous moment is when the market begins to treat the status quo as a permanent law. Today’s AI logic—that hardware scale is the primary lever—is less a rule of physics and more a temporary workaround for inefficient architecture. As investors calculate the return on ever-larger clusters, they are ignoring a more fundamental lesson: what is built today is rarely what defines tomorrow.

The shift from brute force to elegance is not just likely; it is a mathematical necessity. Much of modern AI is built on transformer architectures that exhibit quadratic complexity. Double the input, and the requirements for compute and memory grow fourfold. Quadratic consumption of any limited resource will eventually consume everything. Efficiency is not an optional optimization; it is a condition for survival.

Breaking the exponent

Computer science has reached similar limits before. In the early days of computing, simple tasks like sorting data were hamstrung by quadratic-time algorithms. As datasets grew, the math became an anchor. The breakthrough did not come from faster machines, but from adopting n \log n methods, such as Quicksort. This redefined the relationship between data and time, making modern software possible.

A similar leap occurred in linear programming. By refining the simplex method and interior-point algorithms, researchers reduced the computational burden of optimization by orders of magnitude. These shifts unlocked the modern world, from global logistics to high-frequency trading. They proved that changing the exponent is always more powerful than adding hardware.

AI is now awaiting its own “Quicksort moment”. This transition will be driven by five structural unlocks. The signals of this shift are already occurring in the market. If the marginal cost of a token collapses, demand will go up, but liabilities do not flex.

The architectural pivots

The first pivot addresses the foundational bottleneck of memory and attention. Most transformers struggle with long documents because their memory requirements explode as context expands. Nvidia recently addressed this issue in their real-time video World Model by leveraging Hybrid Linear Diffusion Transformers. Subquadratic, a startup founded by industry veterans, is attacking this at the architectural level. While their results are still being validated, initial independent performance is promising. By breaking the scaling law bottleneck directly, they enable efficient, large-context inference that avoids the ruinous costs of conventional transformers.

The second shift moves from modelling language to “world models” via Joint Embedding Predictive Architectures (JEPA). Current large language models are limited by generative training, which wastes compute predicting every irrelevant pixel or word. Yann LeCun’s AMI Labs is instead pursuing architectures that focus on abstract representations and physical dynamics. This allows systems to reason and plan in complex environments like manufacturing or aerospace.

The third unlock concerns the static nature of intelligence. Most models today are “frozen” after training, meaning they cannot learn from new experiences without expensive retraining. Letta is solving this through “continuous learning” in token space. Their agents maintain and restructure persistent memory across sessions. This transforms intelligence from a one-off pretraining event into a dynamic runtime property that improves with every interaction.

The fourth pivot is the “compilation” of skills to prevent redundant effort. Currently, AI agents often approach every task de novo, rediscovering workflows each time they are deployed. Glean is codifying the agent lifecycle by treating execution as something inspectable and reusable. By tracing successful tool calls and decisions, firms can “compile” these paths into repeatable patterns. This turns bespoke agent actions into scalable, industrial processes.

The final pivot separates reasoning from data. The common fix for domain-specific knowledge, known as RAG, acts as a clumsy external appendage that increases latency. The “Memory Decoder“ research suggests a more elegant alternative: a plug-and-play memory module. This allows a general model to inhabit new knowledge without modifying its core parameters. It avoids the “catastrophic forgetting” of fine-tuning while keeping the model’s reasoning sharp.

The new resource mix

These shifts will do more than lower costs; they will reorder the balance of power in the tech stack. This does not mean the end of the massive GPU cluster, but it does mean that algorithmic efficiency, rather than just raw silicon, will increasingly solve resource shortages. As architectural pivots reduce the dependence on brute-force scaling, the “mix” of hardware required by a data centre five years from now will look nothing like the procurement lists of today.

The risk for investors is to over-index on “selling the shortage” based on current constraints. Subsidizing the construction of yesterday’s architecture is a recipe for stranded assets. In the history of technology, the greatest returns have rarely accrued to those who simply bought the most hardware, but to those who understood how the math was changing. Ideas travel much faster than silicon.

This introduces a fresh risk: the translation of technology volatility into financial instability. AI research progress is naturally non-linear, a “good volatility” that drives innovation. Yet when this is amplified by investors using leverage and debt-backed valuations to chase linear extrapolations, it risks “bad volatility” for the wider economy. When the “Quicksort moment” arrives and pivots the industry toward a different resource mix, the capital tethered to yesterday’s brute-force assumptions could face a painful de-leveraging. Underwriting a revolution with a ruler is a dangerous way to run a market.

Scaling laws aren’t meant to be the end of the story. They are in fact the prologue.

Author’s note: Over the last few months I’ve spent close to 500 hours with coding agents, reviewing code bases and speaking with professionals to get an executive perspective on the technology. The conclusions are both exhilarating and concerning. This article examines the drivers of this technical debt, the illusions of “vibe coding”, and a reformist path toward using slop as a tool rather than a trap.

Artificial intelligence has made writing software unnervingly easy. Describe a feature and it appears; ask for a dashboard and one materializes, complete with authentication and database hooks. In demos, the process feels like magic. In production, it increasingly feels like something else.

Call it “coding slop”: plausible, functional, often impressive software that meets the prompt but not the professional standard. It looks like a duck and quacks like a duck. Only later does the owner discover it is made of wood.

The danger lies in the architecture. Modern commerce is erecting a vast digital skyline on faulty foundations. As firms add floors—scaling users or layering features—they are ignoring the structural rot at the ground level. Eventually, these buildings will fall.

The mechanics of the mandate

Software engineering is, by nature, a multi-objective optimization problem. There is a constant, grinding tension between scope, time-to-market, and technical debt. Coding agents thrive in this friction, but only by ignoring the long term. They are extraordinarily effective at pursuing local goals while failing to preserve global coherence.

This is the limitation of next-token prediction. Optimization systems simply exploit whatever objective is specified, usually “make this work now.” Since architecture and systemic resilience are rarely part of the prompt, they are rarely part of the output. The agent satisfies the user’s immediate whim while quietly poisoning the codebase.

The problem is not hypothetical. The maintainability of AI code has become a primary area of research. Many experienced developers are surfacing issues. Sam Hogan recently observed on X “All the best programmers I know are starting to write code by hand again”. An AWS engineer did a deep dive and concluded they’d return to writing code by hand, warning that AI velocity “makes you think you’re winning right up until the moment everything collapses simultaneously.” Great work is possible, but still hard. This is not a Luddite retreat; it is a diagnosis of misaligned incentives.

The high cost of “vibe coding”

The scale of this phenomenon is visible in the data. GitHub repositories are proliferating at record pace, pull requests are multiplying, and App Store submissions have surged 84%. But this explosion of artifacts masks a hollowing out of quality. Software output is being mistaken for software progress.

At a recent technology dinner, a private-equity portfolio manager described replacing SaaS products across his firms with tools “vibe coded” by recent graduates using AI. On a spreadsheet, the savings look compelling. In reality, this is the digital equivalent of subprime lending: immediate gratification exchanged for systemic risk.

The implications materially degrade return on investment. Maintenance typically consumes 40 to 80% of a software lifecycle’s cost. Coding agents make the initial build “free” while steepening the cost of future security patches. Security debt is particularly pernicious; models routinely reproduce vulnerabilities that are invisible to non-experts but fatal under adversarial conditions.

The casino of shame

Why, then, are we not hearing more about this? There is a significant information asymmetry at play. It takes expert code review to determine that a functional “duck” is actually a wooden decoy. In today’s ecosystem, where “success stories” are a steady drumbeat, few executives have the time or the vantage point to look under the hood.

This creates a pervasive sense of shame. If a developer or a firm is not achieving “magic” results, they assume the fault lies in their own prompting skills rather than the tool’s inherent limitations. When employment is judged by “tokenmaxxing,” speed becomes a status signal that hides structural failure. Like a gambler at a slot machine, the user believes one more prompt will surely deliver the jackpot.

Binding to the mast

The answer is not to ban AI, but to treat it with transparency. Used deliberately, “slop” is a powerful tool for rapid prototyping, akin to using low-poly placeholders in game development. It accelerates feedback loops and validates core primitives. The catastrophe occurs only when these placeholders are shifted to general availability without the intended, rigorous rewrite.

To survive the siren song, firms must ground themselves in reality. Those that implement architectural guardrails and mandatory rewrites for core systems will move faster over time. Those that reward solely for shipping velocity will continue to build castles of sand.

I remain bullish on the technology and optimistic for the future. The productivity gains are real, but only for those who recognize the structural trade-offs. In Homer’s tale, sailors survive by acknowledging the danger and binding themselves to the mast. Software leaders must now do the same, or prepare for their buildings to fall.

The code may be free, but the bill is coming.

The current artificial intelligence boom appears to be a study in radical contradiction. On one side, a new technology is performing feats, from folding proteins to generating code, that suggest a historic leap in productivity is underway. On the other, the landscape is increasingly cluttered with “AI slop”: a deluge of low-quality content, hallucinations, and redundant bots. The prevailing assumption is that this waste is a sign of a bubble waiting to burst.

That diagnosis is likely backwards. The surge of unproductive output is not an anomaly; it is a predictable feature of an infrastructure build-out in its most fertile phase. In the history of transformative technologies, the “slop” is often the very soil from which durable markets grow. The paradox of the AI era—unleashed productivity alongside massive, visible waste—is not a contradiction. Rather, it is the normal pattern of major infrastructure build-outs across modern history.

The jetpack and the iceberg

AI has given individuals something like a jetpack. A curious person can now prototype in days what once took a team months. Marketing copy, code scaffolding, legal drafts and design mocks can be generated at will. The constraint is no longer the cost of trying, but the imagination to prompt.

The result is a Cambrian explosion of features, products and experiments. Barriers to entry have collapsed. The same tool that helps a scientist analyze protein structures also enables thousands of derivative chatbots and thin wrappers around foundation models. Speed has improved dramatically; quality, more unevenly.

Beneath this surface productivity lies a growing mass of technical debt. Many AI-powered products work...until they do not. Systems are built quickly, stitched together from APIs and prompts, without the slow, invisible labor that makes software scalable, secure and maintainable (see The software engineering factory). From the outside, the features are intact. Inside, the structure resembles driftwood lashed into a boat.

At the infrastructure layer, the pattern repeats financially at larger scale. Billions are being poured into data centers, specialized chips and energy capacity. Just last week Allbirds, the perennial shoe company, announced its transformation into Newbird AI. Not all of this capital will earn an adequate return. Some facilities will sit underutilized; some model providers will vanish; some applications will never justify the compute they consume. The waste is visible and politically salient.

But that does not make it aberrant.

The long history of overbuilding

Consider railroads. In the 19th century they were the essential infrastructure of industrial expansion, binding together continents and creating national markets. They also destroyed fortunes. By some estimates, 20 to 25% of American railroad mileage ended up in receivership. Lines were overbuilt, misrouted or rendered redundant by competition.

Or take the Florida land boom of the 1920s. Speculative development ran far ahead of demand. Projects proliferated; many collapsed. The bust was severe. Yet the episode left behind roads, utilities and urban layouts that later generations would use. Waste and foundation were laid down together.

The dot-com bubble followed a similar script. Vast sums were invested in companies with flimsy business models. Many failed spectacularly. Yet the fiber-optic cables buried during that era, much of it initially “dark”, later powered the modern internet. Out of the wreckage emerged a handful of giants that reshaped global commerce.

These episodes share a statistical signature: power laws. Most projects fail or muddle along; a small minority generate extraordinary returns. The aggregate gains justify the aggregate excess. It is not that markets allocate capital perfectly during booms. It is that they allocate it experimentally, at scale.

AI appears to be conforming to the same distribution. A great many applications will prove trivial, redundant or unsustainable. A few will become foundational. The difficulty, as ever, is that the identity of the winners is unknowable ex ante.

Lottery tickets

If AI resembles past booms, it also differs in one important respect: participation. Building a railroad required vast, concentrated capital. Launching a dot-com start-up demanded meaningful funding. Experimenting with AI requires little more than a subscription and an idea.

That low barrier to entry has democratized experimentation. It has also multiplied noise. When anyone can build, everyone does. Development is 10x easier but distribution is 10x harder. The volume of half-formed products, brittle systems and speculative ventures rises accordingly. Waste is no longer confined to balance sheets of industrialists; it also diffuses across millions of small bets.

There is a subtler shift, too. Individuals now contribute directly to the infrastructure build-out not only as investors, via pension funds exposed to technology stocks, but as consumers. Monthly payments for AI tools help finance data centers and model training. Unlike a share in a railroad, these expenditures are consumptive. The capital is spent in pursuit of personal projects that may never endure.

This consumption-driven experimentation adds a personal dimension to the boom. Success or failure is often borne privately: a founder’s runway exhausted, a side project abandoned, a workflow rebuilt yet again. The aggregate effect is still infrastructural expansion. But the microeconomics feel more like subscription churn than asset accumulation.

Critics look at this landscape and see a bubble inflated by hype and easy money. They may yet be proved right about particular firms or facilities. Some valuations will not survive scrutiny. Some capacity will be written down. That, too, would be normal.

What would be abnormal is a transformative general-purpose technology arriving without overbuilding.

Out of the ashes, a new era

The mere presence of waste is not evidence of folly. It is evidence of search and creative destruction. After all, virtually all species that have ever existed have gone extinct.

If nothing changes, the AI boom will likely proceed as others have: a surge of capital, a shake-out, consolidation and, eventually, a quieter phase in which the surviving infrastructure underpins more prosaic gains. The headlines will move on. The servers will hum.

The paradox of AI—unleashed productivity alongside torrents of the unproductive—will resolve itself the way it always has: through failure, selection and time. Innovation at scale is messy. The alternative is tidier, and far poorer.

Those hunting for a boom without waste are asking for progress without risk. History suggests they will wait a long time.

Part of a series: Excellence in the Age of AI

In a recent essay, I argued that artificial intelligence is not merely improving firms but outgrowing the corporate structures built to contain it. The bottleneck is no longer the technology. It is the organization. Nowhere is that tension more visible than in software engineering, the function most exposed to AI’s rapid advance.

The headlines, however, remain confusing. Models hallucinate, introduce subtle bugs and fail unpredictably in live systems. At the same time, they are solving frontier problems and posing systemic security risks warranting attention by the Federal Reserve. AI appears both unpredictably stupid and incomprehensibly smart – “jagged intelligence“ in the words of pioneer Andrej Karpathy.

Both claims are true, and together they point to a deeper shift. The capabilities are real enough to rewire how software is produced, a shift akin to the industrialization of farming. What is emerging is not a better toolchain but a new production system. At the frontier, it resembles a “dark factory”: an engineering organization where most of the journey from idea to delivery happens without direct human touch.

Dark factories, dimly lit

To see the shape of this system, it helps to think in levels of autonomy rather than individual tools. The industry’s fixation on copilots captures only the lowest rung. AI Autocomplete leaves the human firmly in charge, accelerating execution but leaving the surrounding system intact.

Above that sits AI Pair Programming, where agents collaborate with developers in real time. Then come AI Delegates, where engineers specify intent and receive multi-step outputs such as features, tests and documentation. Further along is AI Coworkers, where agents triage, decompose and execute tasks with limited oversight directly from backlogs.

At the frontier lies something more consequential: Dark Factories with persistent, autonomous workflows. They are colloquially called “dark” in that they would need no lights to run, as no humans are present. Here, agents plan, execute, validate and iterate across the lifecycle.

Early versions are visible inside firms such as Stripe and Block, where internal systems handle ticket flows, generate code and run tests in continuous loops. These are not yet true dark factories. They are partial systems, missing key components and operating within tight constraints. But they demonstrate that the ceiling is far higher than faster coding. The direction of travel is toward end-to-end automation.

Shedding light on the machinery

The obstacle is not ambition but complexity. Software engineering is a system of interdependent activities: specification, planning, implementation, validation, documentation, deployment and maintenance. Each layer carries requirements for security, performance, cost control and audit-ability. Weakness in any one stage constrains the whole.

Today’s AI systems are strongest in execution. They write code with impressive fluency. But the surrounding machinery remains underdeveloped. Without it, autonomy fragments.

Validation is a case in point. In an agent-driven pipeline, testing must be continuous, automated and adversarial. Code must be assessed not just for correctness, but for security flaws, performance regressions and cost implications in real time. The components for this exist, but they are not yet integrated into a reliable system.

Context is another constraint. Human engineers rely on tacit knowledge: architectural intent, historical decisions and unwritten norms. For agents to operate effectively, that knowledge must be made explicit and machine-readable. Emerging “context graphs” attempt to do this by mapping codebases as living systems, but they remain incomplete.

Coordination poses a further challenge. A functioning factory requires multiple agents working in concert, some decomposing tasks, others implementing and others validating or documenting. This demands orchestration layers that can route work, enforce guardrails and track provenance. In most organizations, such control planes are still rudimentary.

The result is a familiar pattern. Intelligence has outpaced the systems designed to contain it. The technology is ready to scale, but the tooling and the engineering organizations are not.

Progress over perfection

It is tempting to respond to this gap with extremes. Some leaders dismiss the vision, citing failure modes as evidence that autonomy is illusory. Others attempt to leap directly to full automation, replacing human workflows wholesale.

Neither approach is likely to succeed. Industrial systems are not built in a single step. They evolve by accumulating reliable components, each extending the range of automation. The most effective organizations are not waiting for perfect systems. They are building toward them in stages.

The first is Agent-Assisted Execution, where coding agents are widely deployed within clear guardrails. This phase is less about autonomy than about learning, generating data on where systems succeed and fail.

The second is a Division of Human vs AI Labor. Agents take on decomposition, routine implementation and continuous validation. Human engineers shift toward architecture, oversight and exception handling. Incentives play a decisive role, as organizations that encourage experimentation discover capabilities faster.

The third is Supervised Autonomy. Orchestration layers begin to connect the pipeline, enabling multi-agent workflows that operate with limited intervention. Validation becomes embedded, and context systems evolve into shared memory. Humans move to governing the system rather than executing within it.

If nothing changes, most firms will remain stuck in the first stage. They will produce code faster but remain constrained by human-speed processes and legacy coordination. A minority will build the missing machinery and move toward genuine autonomy. Their advantage will be structural, not incremental.

The dark factory is not yet fully built. But its outlines are increasingly hard to ignore. If the corporation must be rethought for the age of AI, engineering is where that rethink will either begin in earnest or quietly fail.

By now it is conventional wisdom that artificial intelligence will make firms more productive: fewer workers, faster output, tidier margins. The previous article argued this is the least interesting outcome. The real shift is organizational.

The modern corporation, for all its dashboards and matrix structures, still defines excellence as disciplined execution within a familiar form. If intelligence can be replicated at near-zero marginal cost, that form begins to look less like an achievement than a constraint. This is not a hypothetical future; it is one a handful of large firms are already pursuing.

The assumption underlying today’s org chart is ancient. Intelligence is scarce and unevenly distributed. Information must be routed through people. Coordination requires layers. From the Roman legion to the railroad and the post-war conglomerate, the span of control has been bounded by a simple constraint: a human can effectively manage only a handful of other humans. Excellence meant optimizing within that limit.

That constraint is breaking. This is not because people have changed, but because the coordination layer has. As artificial intelligence moves from tool to agent – planning, monitoring, debugging and triggering work autonomously – the hierarchy’s primary function of information routing can be meaningfully absorbed by software. What follows is not flatter organizations, but different ones.

Spans without layers

Consider Block, the payments firm led by Jack Dorsey. Rather than sprinkling “AI copilots” across existing teams, it is attempting something more radical (see “From Hierarchy to Intelligence”: to build a “company world model,” a continuously updated representation of its operations, resources and priorities. In a remote-first environment where decisions, code and discussions are machine-readable, AI systems can maintain context that once sat in middle managers’ heads.

The ambition is to widen span of control without thickening the hierarchy. If a system can track what is built, what is blocked and how customer behavior is shifting, senior leaders need fewer human relays. Directly Responsible Individuals can pull resources across functions without negotiating through layers. The intelligence sits in the system. Humans operate at the edge, applying judgment where models should not.

A similar pattern is emerging at Harvey, a legal-technology firm (see “Legal is Next”). Its internal agent, Spectre, monitors bug reports, customer feedback and internal messages. It triggers engineering work without waiting for a human prompt. Engineers have become so productive that implementation is no longer the bottleneck; coordination is. Spectre acts as the beginning of a coordination substrate, an internal nervous system.

These are early experiments. Yet they point to a broader redefinition of excellence: not how well a company manages layers, but how effectively it substitutes for them.

Shifting bottlenecks

Such substitution is not automatic. It requires re-engineering the firm’s technical plumbing. Jeff Dean of Google has advocated that agent-based systems generate dozens of trajectories at machine speed, only to collide with tools designed for humans (see “Advancing AI’s Next Frontier”). A compiler’s start-up time, trivial for a programmer, becomes an Amdahl’s-law bottleneck when an agent runs 50 times faster. Reducing the model’s speed alone yields only marginal gains if the surrounding system continues to lag.

Excellence, in other words, now includes eliminating the latency tax embedded in legacy processes. Digitization cannot be partial. Every artifact – decision logs, contracts, code, and customer interactions – must be accessible to machines. Workflows built for email and meetings must be redesigned for autonomous iteration. Firms that merely bolt agents onto human-speed infrastructure will discover that their smartest systems spend most of their time waiting.

At the extreme we find one- and two-person companies. Medvi, a some what controversial telehealth provider of weight-loss drugs, reportedly reached over $1 billion in run-rate sales with two employees and a web of AI tools (see the New York Times article). Marketing copy, customer service and performance analytics were automated, while doctors and compliance were outsourced. Revenue approached that of listed rivals with thousands of staff.

This is not a parable about one founder’s hustle; it illustrates a structural shift. When coordination, coding and customer interaction can be automated, the minimum efficient scale of a firm collapses. Total addressable markets expand because the cost of entering them falls. More complex systems combining payments, lending, logistics and compliance become feasible for teams that would once have been laughably small.

Compounding competitive gaps

If top performers can oversee broader domains with system-level intelligence, the gap between fast and slow firms will widen exponentially. Market leaders will compound daily through rich learning and data loops, increasing their breadth of winning. The laggards will face a choice: specialize narrowly, outsource to faster platforms or fade.

What changes is the definition of excellence. It becomes an order of magnitude test. How broad is a leader’s effective span? How fully is the firm’s knowledge digitized? How quickly can it recompose capabilities into new offerings? How large a market can it plausibly serve without losing coherence? Speed, judgment, scale and ambition converge.

If nothing changes, many incumbents will pursue AI as a cost-saving device by trimming headcount, improving margins and declaring victory. They will build faster “trains”. Meanwhile, competitors organized as intelligences will pursue “planes”. The performance gap will at first look like luck, then like strategy, and finally like inevitability.

The task for executives is prosaic but demanding. The time to rethink your company is now. Eliminate bottlenecks. Invest in infrastructure that allows agents to operate at machine speed. Redefine managerial roles around coaching and ethical oversight rather than status reporting. Above all, resist the comfort of incrementalism.

For two millennia, excellence meant mastering the hierarchy. In the next decade, it will mean outgrowing it.

On May 26th 1934 a silver train slipped out of Denver at dawn and reached Chicago just after dusk. The Burlington Zephyr covered more than 1,000 miles in just over 13 hours, shattering records and, for a moment, the public’s sense of what rail could be. Sleek, diesel-powered and built of stainless steel, it seemed less a train than a promise: faster, cleaner, unmistakably modern.

Yet even as crowds gathered to celebrate, the future had already taken flight. The Wright brothers’ first powered flight lay three decades in the past. By the early 1930s commercial aviation, though nascent, was advancing steadily. The writing was there for those willing to read it. But railways, flush with incremental triumphs, looked along their tracks rather than up at the sky. Within a generation, aircraft would render the Zephyr’s achievement quaint. What appeared to be a culmination was merely a waypoint.

History suggests that what limits progress is not machinery but imagination. Before flight, even optimists conceived of faster trains, not airborne networks. Before the internet, digitization meant more efficient filing cabinets. Humans extrapolate linearly from the present because institutions, incentives and habits anchor them there.

The same trap shapes how artificial intelligence is discussed today. The prevailing fear is that such systems will overtake knowledge work: that lawyers, analysts and programmers will be displaced as surely as steam engines displaced stable boys. But this frames the question too narrowly. The greater disruption may lie in organizational form. If intelligence can be replicated cheaply and scaled elastically, the boundaries of the firm begin to blur. Hierarchies built to coordinate scarce expertise may come to look as dated as rail barons in a jet age.

The incremental trap

It is tempting to treat AI as another step in a familiar progression: software that writes faster code, drafts sharper emails and answers more questions. That is the equivalent of building a better train. It may be profitable. It is unlikely to be transformative.

Technology’s deeper pattern points elsewhere. Progress has rarely come from doing the same things more efficiently, but from expanding what is conceivable. History should instill confidence. Each wave of computing has widened the frontier. Punch cards automated arithmetic. Assembly language disciplined logic. Compilers abstracted complexity. High-level languages democratized programming. Cloud computing dissolved infrastructure into a utility. At every stage, both the scope of what could be built, and who could build it, expanded.

Artificial intelligence accelerates this shift. Autonomous coding agents generate and refine software with minimal oversight. Research assistants synthesize entire fields in minutes. Early “zero-person companies” are beginning to stitch together functions once thought inseparable. The risk is not that firms fail to adopt these tools, but that they adopt them too narrowly—compressing existing tasks rather than questioning why those tasks exist at all.

Growing software

The most consequential applications of AI may not resemble today’s software. In The MANIAC, Benjamín Labatut recounts John von Neumann’s suggestion that a truly advanced machine “would have to grow, not be built”. The distinction is instructive. Traditional software executes instructions; emerging systems explore possibilities.

Modern AI models are trained on vast data, refined through feedback and deployed in environments where they adapt through use. Extend this logic and one arrives at a different kind of product: not a static tool, but an evolving organism. Such systems can simulate countless potential outcomes before a user articulates a need, selecting responses that appear uncannily well judged. The effect is less like using software than interacting with something that has already rehearsed the future.

This shift is difficult to grasp because it departs from familiar metaphors. Firms are used to specifying workflows, not setting objectives for systems that determine their own paths. Managers allocate tasks; they do not typically oversee agents that generate and test alternatives autonomously. What changes is not just capability, but posture: from directing work to shaping the conditions under which it unfolds.

Aiming for the exponential

The immediate instinct for firms is to deploy AI where returns are easiest to measure: fewer support staff, faster development cycles, marginal cost savings. These are sensible steps. They are also the least interesting ones.

A more ambitious approach begins by discarding the assumption that existing workflows are fixed. If software can be generated on demand, product development need not proceed in stages. If knowledge can be synthesized instantly, strategy need not be bounded by precedent. The constraint shifts from execution to intent—what is worth building, and why.

For individuals, the adjustment is similar. The point is not to compete with machines at tasks they perform cheaply, but to expand the range of problems one attempts. Delegating routine work is a starting point. The larger opportunity lies in pursuing questions that were previously impractical: those that require synthesis, judgement and a tolerance for uncertainty.

The advantage will accrue to those who treat AI not as a tool for optimization, but as a prompt to raise their ambition. In the words of Robert Browning: “Ah, but a man’s reach should exceed his grasp, Or what’s a heaven for?”

For most of recorded history, there was, in essence, a single economy. People combined labor with capital to produce goods and services. Farmers planted, factories assembled, merchants traded. Finance existed, but largely as a servant to production—a way to smooth consumption, mobilize savings and allocate capital to tangible enterprise.

Speculative manias flared from time to time. Tulips bloomed and crashed; railways boomed and busted. Yet these episodes were sideshows. The dominant force was what might be called the “real” economy: the aggregate output of human effort applied to the physical world.

That settled picture changed in the 20th century. Finance did not merely grow; it detached. What began as credit expansion in the roaring 1920s and the widening ownership of equities evolved into a sprawling ecosystem of derivatives, securitization and leveraged intermediation. By the eve of the global financial crisis, the stock of financial assets in America had swollen to many multiples of GDP—around 6 times net assets, outpacing GDP by 275% since the 1950s. The tail was no longer content to follow the dog.

Finance developed its own language, measurements and cycles. During the global financial crisis of 2007-09, balance-sheet fragilities in mortgage-backed securities and repo markets convulsed the broader “real” economy. A decade later, meme-stock frenzies and crypto booms surged even as underlying productivity growth remained pedestrian. Finance could roar while Main Street limped—or collapse and drag it down. It had become distinct, requiring its own measures of performance, risk, innovation and regulation.

Today a third limb is emerging. Artificial intelligence is coalescing into an economic domain that can expand, contract and compete for resources with relative independence from human labor. It is the first genuinely new “part” of the economy in a century.

When output no longer belongs to workers

What qualifies something as a separate economic sphere is not novelty. It is the ability to decouple. Finance earned that status when its booms and busts took on a life of their own. AI is approaching a similar threshold. It can prosper even when large swathes of the labor market struggle.

Consider a marketing agency. In the old model, growth meant hiring: more copywriters, more designers, more account managers. Output rose more or less with headcount. Technology helped, but mostly at the margin.

Now picture the agency with pervasive AI. A small cadre of senior staff directs swarms of software agents that draft copy, generate video, test variants and fine-tune media buying continuously. Work that once took teams several days is completed before lunch. The firm handles four times as many accounts with half as many people.

That creates an awkward statistical problem. Output per worker may soar eightfold. But when most of the incremental work is being done by machines, it is not obvious that “labor productivity” remains the right lens. No one judges modern farming by asking how many seeds a laborer can scatter by hand. Yet much of economic measurement still assumes that human effort is the central input.

Productivity growth averaged 2 to 3% per year for decades – AI will rebalance the equation

As digital workers multiply, they will come to outnumber human ones in many firms. Key Performance Indicators will then begin to shift: from worker efficiency to compute efficiency, and from human management to the AI equivalent. As with finance before it, a new economic sphere will demand its own measures of performance, risk, innovation and regulation.

Competition for resources—and votes

Separation does not mean isolation. Finance competes with the real economy for capital and political favor. The pull from the AI economy will be more amplified.

The most obvious arena is wages. If AI systems can perform cognitive tasks at decreasing marginal cost, the demand for certain categories of human labor will weaken. But competition extends further: to electricity grids strained by data centers; to advanced chips whose supply constraints have already reshaped electronics markets; to commercial real estate repurposed for server farms rather than lakefront property.

In such a world it is conceivable to see rising investment, booming valuations for AI firms and surging demand for compute even as employment or wage growth stagnates. The AI economy could be in expansion while the labor economy feels recessionary pressure.

Unlike capital, however, AI agents do not vote. When tensions sharpen, political systems will mediate the clash. Three fault lines are already visible.

The first is taxation. If machines substitute for workers, should tax systems shift from labor income towards capital—or even towards AI systems themselves? Proposals for “robot taxes” are clumsy, but the underlying concern is fiscal erosion as payrolls shrink.

The second is income support. Wage subsidies, expanded earned-income credits or even universal basic income may move from fringe to mainstream if displacement is rapid. Such policies aim not to halt automation but to socialize some of its gains.

The third is ownership. If AI systems become the dominant productive asset, who owns them matters enormously. Concentrated corporate control could amplify inequality and political backlash. Broader ownership—through pension funds, public stakes or profit-sharing—would align the fortunes of citizens with those of the AI economy, much as mass equity ownership partially aligned workers with finance.

Each of these responses, taken to extremes, could strain liberal capitalism. Heavy-handed automation bans would stifle innovation. Punitive taxation could drive investment offshore. Sweeping nationalization would sap competitive dynamism. Yet ignoring the tensions would be equally hazardous.

Updating the architecture

The emergence of a third economic sphere need not herald collapse. Finance, for all its excesses, has lubricated growth when properly regulated. AI promises similar gains: faster drug discovery, personalized education, more efficient logistics and cleaner energy systems. The potential uplift to global living standards is substantial.

But potential is not destiny. If current trends persist, the AI economy will grow rapidly, capital will chase its returns and labor markets will adjust unevenly. Periods of exuberance and correction are likely. So, too, are political reactions when benefits appear narrowly distributed.

The prudent course is neither panic nor complacency. Governments could modernize tax systems to remain neutral between labor and capital while preserving incentives to innovate. Competition policy can prevent AI markets from ossifying into unassailable monopolies. And ownership structures could evolve so that citizens have a tangible stake in the systems reshaping their work.

A century ago, policymakers learned—sometimes painfully—that a vast financial superstructure required new rules and measurements. They may be about to relearn that lesson. The economy is no longer just what humans make of it. It is increasingly what their machines make alongside them.