ChatGPT Is a Thousand Days Old. Emad Mostaque Says We Have About a Thousand Left.

ChatGPT went live on 30 November 2022. By the time this episode of The Cognitive Revolution was recorded, that was about a thousand days. Host Nathan Labenz uses the anniversary as a countdown. Emad Mostaque, founder of Stability AI and now of Intelligent Internet, argues that even if capabilities froze today, the systems already built are enough to hollow out human work once they are properly wired into firms. Frontier labs, meanwhile, are racing to ship agents designed to replace labor. In the book *The Last Economy*, the line is: the machines are taking our jobs. Thank God. Now we can get to our real work.

Labenz’s introduction sketches why the old premise fails. Human society was built on scarcity. Elon Musk’s version is still the folk economics of the present: if you do not make stuff, there is no stuff. What happens when an AI doctor that does not eat gives better frontline advice than a human at a thousandth of the cost? Patients get abundance. Doctors get poverty. The same pattern, applied to driving and then to a majority of the economy in less than a generation, is the problem Mostaque says there is no escaping. He signed the 2023 pause letter while still CEO of Stability, then left, and is now training small open models while trying to describe the transition in concrete detail.

The terms he wants in circulation are the intelligence inversion, the metabolic rift, and the abundance trap. Instead of GDP alone he wants to score an economy on intelligence, connectivity, and resilience. He even proposes a dual currency: one for rival physical goods, another for non-rival ideas. Labenz calls these underdog ideas, then quotes Yuval Noah Harari’s point that shared stories are how society runs. Money itself is a useful fiction. In that light, Mostaque’s scheme is not obviously crazier than whoever first decided gold or shells should clear a market. The risk Labenz names is Balaji Srinivasan’s “nuclear outcome”: weaponization without abundance. The invitation is to tell an abundance story fast enough that a modern Luddite movement does not shut the project down first.

Labenz opens by calling a positive vision the scarcest resource in the discourse, and Mostaque describes the book as an engineering manual. The foundation is intelligence theory, built from first principles after watching Stable Diffusion and later agents beat humans at more and more tasks. The axiom is persistence. Certain complex adaptive systems last a long time in uncertain environments. The ones that last best have the closest match between their internal model of reality and reality itself. That match, he says, looks like the loss function in generative AI, and the mathematics is the same. Economics, in this telling, should be derived from that one principle rather than from scarcity, unmeasurable utility, or general equilibrium.

Labenz tests the idea against Scott Alexander’s 2017 Slate Star Codex essay on predictive processing: each neural layer predicts the next signal and stays quiet until surprise forces attention upward. Why doesn’t a rock count? A tortoise can live 150 years; a stone can sit for a thousand. Mostaque draws a hard line. The framework applies only to systems in motion, with information flowing among agents that act. A rock does not compute, does not press on its boundaries, and has no update cost. Agents minimize three costs: predictive error (surprise), model complexity (the cost of thinking), and update cost (the cost of learning). Those three, he says, are isomorphic to Karl Friston’s free energy principle. Gradient descent is the same optimization. Stones are not in the game.

The economic punchline is that the textbook picture is two centuries out of date. Adam Smith’s *Wealth of Nations* was written when value was land and serfs. Then it was factories and machines, then software. Now, Mostaque says, it is the wealth of robots. On a panel a couple of years ago, Nat Friedman coined “AI Atlantis”: a new continent of a trillion agents and robots about to enter the workforce. Classical assumptions still treat humans as the top producers. That, Mostaque argues, may not be true in a few years. The whole measurement stack has to be rebuilt for a world in which the best models of reality are already AI models, and the best agents are already AI agents.

Labenz asks why serious people, including Tyler Cowen, still cap AI’s gift at half a percent of GDP a year. Mostaque’s answer is update cost. Priors are expensive to change. In January 2020 he was publicly warning about Covid; most people called it fine until Tom Hanks got it. The AI clock is faster than that. It is three years since Stable Diffusion, just over a thousand days since ChatGPT, one year since o1 preview, a little more than a month since GPT-5. A month earlier, he says, the vast majority of users were still on GPT-4o. About 20% of Americans still have not heard of ChatGPT. Diffusion is slow. Capabilities are not.

At the cutting edge the phase change is already visible. Mostaque has Codex running on the CLI for three hours, building a textbook website unattended. The jump is from “not quite good enough” to superhuman. Unlike robots, which wait on factories and spare parts, this generation of software already has the hardware; what is missing is interfaces and workflows. The day before the taping, Alibaba’s Qwen lab released a 30-billion-parameter mixture-of-experts model with 3 billion active parameters that, he says, outperformed Grok 4 on Humanity’s Last Exam and beat Deep Research-class systems. Three billion active parameters run on a CPU with 16 GB of RAM. That is not a lab toy. It is something that fits on an ordinary machine.

Intelligent Internet’s own medical model is 8 billion parameters and needs 8 GB of RAM. It already outperforms human doctors. The team still will not ship it at scale because they do not think it is good enough; next year it should be better than any doctor, fully traceable, and runnable on a phone. How do you put that in GDP? A recent MIT study found that 95% of corporate AI deployments have not worked, largely because they are running last-generation models, and last generation is six months old. The real economic hit, Mostaque insists, is not a polymath inventing a new field. It is cooks, not chefs: systems that follow instructions, check their errors, and work for hours. Labenz adds the cost curve: GPT-3 was $60 per million input tokens; GPT-5 is about $1.25 to $1.50. That is a 95% drop with a much better model, which is exactly the kind of gain GDP is bad at seeing.

After the ads, they name the trap. Next year anyone with a phone can have a doctor opinion that outperforms most physicians, plus legal help from systems like Grok that already beat typical advice. Doctors, Mostaque notes, err about 20% of the time. For the patient this is a welfare explosion. For the economic system, which prices scarcity, infinite supply collapses price, and the people who lived on that scarcity lose their income. Corporate profits may rise. Knowledge work still gets displaced by “KVM jobs” on the other side of a screen: keyboard, video, mouse workers who do not sleep, do not err, and keep learning. Lives get better. The ledger reads as poverty.

The companion phrase is the metabolic rift. GPUs do not eat, do not need housing, and do not pay tax; usage is often deductible, and you can rent them by the hour. An AI’s only need is its objective function. Mostaque has Eliezer Yudkowsky’s *If Anyone Builds It, Everyone Dies* waiting on the desk, but he wants the conversation off AGI and onto AI accountants, lawyers, and designers. Those systems will not start consuming like people. Once they are smarter than a human at the job, they stay that way. Organizations, in his reduction, are executors: money in, more money out. AI will out-execute. That does not reverse.

Labenz puts a number on “GPUs don’t eat.” A phone battery holds about 20 watt-hours, a laptop about 100. In his Detroit neighborhood electricity is about 18 cents a kilowatt-hour, so 100 watt-hours is under two cents. Running a model on a laptop for a couple of hours costs pocket change. Mostaque’s practical example is websites. They used to cost thousands of dollars. On Replit you chat with an agent for a day, pay $20 or $40 including the company’s margin, and get something as good. That cost, he says, drops another 10x next year and 10x the year after from chip speed alone. Image generation already finished the trip: a few cents with tools like Nano Banana, quality above most human designers. Creation is going the way consumption went in the last internet age: toward zero, with better aesthetics sitting in the model’s latent space.

This is the intelligence inversion. Value climbed from land and serfs, to muscle, to industrial and then software capital. Each time a layer was automated, humans moved up a floor. Intelligence is the last floor. "There's nowhere left to pivot." Physical robots make the squeeze tighter: a Unitree machine that stands back up in a second, a $20,000 Tesla Optimus that could do a plumber’s work in a few years. Token prices keep falling toward a dollar-fifty, then 15 cents, then a penny a million. GPT-4.5 was too expensive and people fled to cheaper models. The scarce thing in the projections, he says, is not giant AGI clusters. It is cheap, useful intelligence.

Labenz raises the usual refuge: a caring economy of teaching, mentoring, and looking after one another, plus creative work that may just be leisure. Mostaque does not deny that people will still want people. He will not go to a robot concert; he calls that the Taylor Swift economy. A walk with a child is not getting automated. The problem is flows. The Fed’s mandate is employment and inflation. Cutting rates now means hiring more GPUs, not more knowledge workers. Identity has already shifted from kinship networks to job titles. Community, neighbors, and religion have hollowed out. The caring layer can emerge only if capital still reaches the people who would do it. He runs the UBI arithmetic: $16,000 is the US poverty line; paying that to every American adult costs $5 trillion, equal to the entire federal tax take. Corporations pay $0.9 trillion. "Tax the AI" does not close the gap.

The deeper claim is that computation and consciousness have split. Consciousness stays human; it is the domain of *why*, not *how*. Cooks execute. Humans make meaning. The next thousand days, he says, is the year of tipping for KVM jobs: you will chat with something on a call or a Zoom and not be able to tell. Labenz offers an even sharper version, that AI doctors already win on bedside manner in some studies and that Alpha School has moved adults into mentor roles. Mostaque’s positive picture is not that we must choose. It is that people should care for others because a rich life includes it, not because scarcity requires a wage. Attention is one of the last finite resources. The fisherman-and-banker parable is his warning against filling it with extraction. Voice-cloned scams calling his mother, and Sam Altman’s remark that some 10,000 people a month die by suicide, some after talking to ChatGPT, are the bad uses. Coaching and support are the good ones. Architecture, not capability, decides which we get.

If the theory does not convince, Mostaque offers harbingers. Neil Howe’s Fourth Turning was supposed to land around 2025. He sees critical slowing down, variance explosions, flickering categories (employee or gig worker? is Bitcoin money?), and correlations so tight that an Ever Given can jam the world. Debt is maxed. The stock market prints records while depression and suicide rise. Small inputs swing wildly. Next year’s digital-asset boom in the US, he suspects, will be the biggest bubble yet, capital hunting anywhere that is not AI. Classical levers are losing force: if the Fed floods the market, not much happens. The old axioms, that scarcity is fundamental, that human labor has value, that growth needs physical resources, that markets equilibrate, that money measures worth, are the lies he wants retired. Humans will be the dumbest people on the team. Intelligence replicates on a few GPUs.

Labenz is most persuaded by recovery times. Systems in crisis take longer to bounce back from a new insult; recessions are lasting longer. Efficiency has been traded for resilience until the whole stack is brittle. Mostaque treats corporations as slow, dumb AIs that chew people as fuel, and schools as factories that feed them. Goodhart’s law is the operating system: you manage what you measure, then you become it. Simon Kuznets invented GDP in the 1940s and said himself it was a bad measure of welfare. Offshoring, Meta running experiments on whether sadder posts make sadder people, Scott Alexander’s “Seeing Like a State” parable of bulldozing villages into legible monocultures, all of that reduces fallback. He jokes that Los Angeles has spent more on a San Francisco railway than has been spent training all the AI models put together, and that we still cannot fund a “common-sense GPT” to flag dumb policy.

The money-equals-value lie is now visually obvious. An AI consult at 10 cents versus a $100 human appointment is three orders of magnitude, yet the cheaper one may be better. Happiness, past a hygiene line, does not track wealth. Several crises, AI, robotics, climate, are arriving at once, with the buffers from the last round already spent. That is why he wants a different scoreboard before the next shock.

From the generative-AI math he pulls three laws of living systems. The law of flow: value must circulate; hoarding brings stasis and collapse. The law of openness: connection fights entropy. Tokugawa Japan from 1633 to 1853 is the closed-system warning, a monoculture that could not absorb Commodore Perry’s guns. The law of resilience: diversity, not just connectivity. The Irish potato famine and the early-20th-century banana collapse are the exhibits. These are hygiene factors. Then come the four capitals he writes as MIND. Material capital is GDP: I give you an apple, I have one less, you eat it, it is gone, water flowing downhill. Intelligence capital is non-rival: I give you an idea and we both have it. Erik Brynjolfsson’s GDP-B, adding consumer surplus from free digital goods, could add $96 trillion to the measured US economy. Network capital is who trusts you and whom you can call. Diversity capital is optionality when a phase transition hits.

The four multiply. Zero any one and you are finished. Singapore is his balanced case: little material, high intelligence, network, and diversity. Resource curses are the inverse, too much M, starved I, N, and D. Everyone listening, he says, can score themselves the same way. Monocultures in crops are already a worry; monocultures in model latent space may be worse. An Oxford and Scale study found that teaching one model to love owls can make another, from the same base, love owls even off-topic. Combined with the platonic representation hypothesis, that is a Stuxnet-class fear: a memetic virus in similar latent spaces. Labenz notes the owl effect was same-base only; Mostaque still does not want governments run by one latent space.

He maps the old schools onto the same geometry. Organizations look like transformers, attending over ordered data. Markets look like diffusion, the reverse of destroying an image into noise, the same family Tesla uses for driving. Hodge decomposition splits flows into three: gradient (Smithian, scarce goods, downhill), circular (Marxian, intelligence that compounds as it is shared, capital that no longer needs labor), and harmonic (Hayekian landscape, the rules of the game). Each school touched one part of the elephant. Neoliberal capitalism was the least-bad fit for its time. AI does not think in scarcity or rational human agency, and it is about to be the marginal producer. Adam Smith wrote the wealth of nations. Mostaque asks what a nation is when most productivity is AI, and what wealth even means.

Labenz supplies the flywheel. Anthropic’s leaked deck imagined labs pulling away on synthetic data. Stripe’s payments foundation model, trained on something like 1.3% of global GDP flowing through the network, is getting sharply better at fraud, which attracts more volume, which trains a better model. Strength begins strength. Tesla and Palantir at 200 times earnings are the same puzzle: intangible, networked intelligence that GDP never quite names. Mostaque’s three futures start from that geometry. Digital feudalism is a handful of labs with a generational lead, data flywheels becoming intelligence flywheels, Mag 7 already an outsized share of US market cap. Stripe embedding its own chain to extract rents is rational. OpenAI’s corporate structure, in his view, has dropped the pretense of being aligned with humanity. Their goal is AGI, not being an API company. Why give you the intelligence instead of using it?

Fragmentation is the geopolitical fork. Governments notice that models can manipulate everything, so you get Chinese AI, British AI, American AI, firewalls, and standards that split from the first layer up. TikTok’s sale involving Oracle, Andreessen, and Silver Lake is about who adjusts children’s minds. A report on BrainCo adapting neural patterns for athletes, allegedly with Chinese funding, is a preview. Who owns the model, who runs it, and who sets the objective function has the power. Mad Max balkanization of the info-sphere is a possible “positive” future only in the sense that it is not a singleton. It is still ugly.

Symbiosis is the one he is building toward: decentralized, open, small models optimized for human flourishing, with the important systems, health, education, government, finance, under collective ownership like a utility. He no longer cares if ChatGPT is the best creative writer. He cares if the interface and memory that teach his child are aligned, preferably owned by the family, and if regulated decision models are transparent. Time is short. OpenAI announced Stargate UK at $30 billion. The tell he is watching is Grok 5 versus Grok 4. If it lands well, Elon will not stay quiet. That, in three to six months, is his first clean signal on whether scaling laws keep going. If they do, next year is the year you break AGI on a capability extrapolation, and the owners of the big GPU classes become the most capable entities on earth. If they S-curve, intelligence costs collapse toward zero and the story changes. Either way, API parity with first-party products is a courtesy, not a law. Labenz notes that Cursor’s whole pitch dies the day the labs keep a better model for themselves. OpenAI’s own five-year mix, in a chart The Information ran, had API shrinking while agents and ChatGPT took $80 billion each on the way to a $200 billion run rate. The agent product is a replacement for workers. Fiduciary duty, Mostaque says, pushes them that way.

Vitalik Buterin’s “revenue evil curve” is the constraint. Products start generous, then cut off access to create a paid tier, and from that moment they drift toward exclusion. Meta optimizing for engagement is hard to align because the objective is manipulation. Mostaque wants models that teach his kids *his* ethics, not a blank “ethics for no one.” Important domains should run on maybe a trillion good tokens, not 36 or 100 trillion of internet sludge, as a distributed swarm whose objective is individual, community, and societal flourishing. He does not think that is encoded in current models. Japanese consensus is not German consensus. Open source is the way to keep the layers local.

The headline of the new social contract: AI that runs the important things should be a utility, owned and optimized for people. Today’s money is born as bank credit, which is debt. The old extract from the young. Capital capitalizes capital. Stanford work he attributes to Erik Brynjolfsson shows early-career jobs already falling off a cliff because models are graduate-level. Why hire a complaining junior? Intelligent Internet is open-sourcing models for finance, education, and health. Verified compute on those deployments would secure FoundationCoin, a Bitcoin-like ledger mined not by ASICs but by national champions stacking GPUs to train public medical and education models. Digital assets are about $4 trillion today; he expects $40 trillion. OpenAI’s inference bill this year, he says, is in the same neighborhood as Bitcoin’s compute budget, while OpenAI and Anthropic together made about $20 billion against $160 billion flowing into crypto. Point that flow at cancer supercomputers and autism knowledge bases instead of at Lambos.

The second layer is cash against that gold: money you receive for being human, a switch from banks creating money to persons creating money. Ownership of the labs cannot substitute. If OpenAI were worth $100 trillion, 30 times the most valuable company and larger than $85 trillion of global GDP, and Americans held 10%, that is about $29,000 each; at a 5% dividend, $1,500 a year, and that is only Americans. It does not work. Buyers of FoundationCoin would earmark proceeds to cancer, Alzheimer’s, or a local education model, see the supercomputer, and watch a counter of people helped. At Stability he gave away 20 million A100-hours and saw 300 million downloads; the same allocation muscle, pointed at public goods, is the pitch. Next year they plan a free, open-source app that checks every diagnosis in every language. High-quality medical advice and an AI on your side, he says, should be human rights. Alpha School’s two hours a day in the global 0.5 percentile is the education version of the same claim.

Nation-states are the wrong unit for OpenAI’s best Japanese accountant sitting on a server in Arizona, but they are the right unit for health data that should not leave town. Rather than a Mistral or Cohere, he would set the local champion’s valuation at a dollar and give the equity to citizens, with DAO-like control of localized datasets, acting as mining pools until governments catch up. Labenz asks what policymakers can do before any of that exists. Mostaque’s near-term ask is almost free: run every US bill through an AI that checks the Constitution, common sense, and flourishing, with reproducible analysis. Geometry engineering, not more GDP programs. Give people a universal AI and the information coming up from the bottom of government changes. He thinks that is the first check on corporate capture that the system has ever had.

Mostaque’s safety story is not a bigger moat around a singleton. If FoundationCoin becomes the highest-marginal-dollar buyer of compute, labs will adapt. Gold-standard anchor datasets that map Malaysian culture or the ethics of faith traditions, the “doer” Kissinger and Eric Schmidt talked about as underlying agreements, can be fed into other models. A distributed network of personal, city, and national agents is harder to 51% than a single cluster, the way Bitcoin is hard to attack. Instrumental goals stay small and slow instead of world-eating. He would force labs to publish training-data ingredients the way food companies publish recipes, and to keep Reddit sludge out of a medical model. In the absence of those datasets and incentives, the only gradient is profit, or rather cash-flow maximization: subscriptions on day one, GPU bills on day 60, the Amazon playbook. OpenAI’s latest burn plan, Labenz notes, went from $30 billion to on the order of $110 billion of losses. The prize is all human intellectual labor. Pause letters and taxes do not move that incentive. He signed the 2023 pause letter as Stability’s CEO and later concluded that none of the pause-and-regulate path would work.

The next thousand days matter because defaults crystallize. UAE handing ChatGPT to citizens is already a powerful default. The first entity that checks every policy, or that stands up a cancer supercomputer, becomes a Schelling point. Public-sector jobs are safe for a while, he jokes, including San Francisco MTA administrators on $400,000, because they are not about production. Then income tax and consumption fall as workers are displaced. Dario Amodei said 90% of code would be written by AI around now; he should have said *can* be. Mostaque had said five years for programming a couple of years ago, which still leaves a couple. Maybe half of coders use AI today. Distribution lags. Moats do not: once your medical history lives in ChatGPT, you will not switch. Collectively owned interface and data, not a ban on other models, is the control plane he wants. Two years ago he could not organize the world’s cancer knowledge. Now, he says, a supercomputer pointed at it would accelerate a cure, and there is no honest debate about whether that is worth bootstrapping.

The good ending is a Star Trek future rather than Star Wars: exploration, inquiry, family time, a caring economy that is not a wage. China, with a broken population pyramid, might keep its robots at home and optimize for flourishing. Work to live, do not live to work. Put everyone on equal capability, then let extra material come from extra contribution. The biggest lie that causes violence, he says, recorded a week after the killing of Charlie Kirk, is that some humans are not humans. Echo chambers amplify it. Use AI to increase human agency rather than to replace people with agents. That is a design choice available now. Put yourself in 10 or 20 years. Who owns the model that teaches your kids and manages your health? He does not want it to be a government allocator or a private company. He wants a sovereign AI, clean data, looking out for him. The book is *The Last Economy*. The company training the open datasets is Intelligent Internet. He does not think the momentum reverses. The question is which story we install while it is still cheap to choose.