Dario Amodei — “We are near the end of the exponential”

Summary

• Dario’s headline claim is that we are near the end of the exponential — the technology has tracked his expectations (“smart high school student to smart college student to beginning to do PhD”), and the genuine surprise is “the lack of public recognition of how close we are.” He puts 90% odds on a “country of geniuses in a data center” within ten years and a 50/50 hunch it lands in one to two, maybe one to three years — “I think it’s crazy to say that this won’t happen by 2035.”
• The scaling thesis is unchanged since his 2017 “Big Blob of Compute Hypothesis”: cleverness doesn’t matter, seven ingredients do, and RL is now showing the same log-linear scaling pre-training did — performance on math contests and “a wide variety of RL tasks” is log-linear in training time. Continual learning “might not be a barrier at all”; long context is “an engineering and inference problem,” not a research problem, possibly solved “in the next year or two.”
• Anthropic’s revenue is the tradeable data point: zero→$100M (2023), $100M→$1B (2024), $1B→$9-10B (2025), with “another few billion” added in January alone. Dario expects the curve to bend “somewhat this year” but stay fast — and says numbers implying ~$200B TAM by 2028 “are too small.” Separate hard call: “It is hard for me to see that there won’t be trillions of dollars in revenue before 2030.”
• The capex discipline argument is the episode’s sharpest risk framing: if you buy $1T/year of compute and revenue comes in at $800B, “there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt.” Being off by one year on the demand curve is ruinous, which is why Anthropic buys “hundreds of billions, not trillions” — and why he suspects rivals “have not written down the spreadsheet.”
• Industry math for suppliers: compute build-out is ~10-15 GW this year, growing roughly 3x/year — ~30-40 GW next year, ~100 GW in 2028, ~300 GW in 2029 at ~$10-15B per gigawatt, i.e. multiple trillions of dollars a year of capex by 2028-29. Steady-state economics: each model is profitable ("$1B to train, $4B revenue"), companies lose money only because they’re funding the exponentially bigger next model.
• On coding, his 90%-of-lines prediction “happened, at least at some places,” but the spectrum matters: 90% of lines → 100% of lines → 90% of end-to-end SWE tasks → 100% → 90% less SWE demand. He thinks models may do SWE fully end-to-end “in a year or two” — including setting technical direction — and pegs today’s coding-model productivity uplift at 15-20% total factor speedup, up from 5% six months ago, a snowball model.
• Structurally he expects 3-4 players and positive margins, not commoditization: a Cournot-style equilibrium like cloud, except “models are more differentiated than cloud.” Profitability isn’t a scale milestone — it’s a demand-prediction residual: guess demand right with >50% inference gross margins and ~50% of compute on training, and “the underlying economics are profitable.”
• Policy stance: against the 10-year federal moratorium on state AI laws (“10 years is an eternity… a crazy thing to do”) but pro federal preemption done right; transparency standards now, fast targeted action if bioterrorism risk crystallizes “as soon as later this year”; hawkish on chip export controls to China (the counterarguments are “fishy”); and more worried about the FDA pipeline jamming up than about “stupid” chatbot bills.

Deep dive

1. The exponential went as expected — the world’s obliviousness didn’t

• Asked for his biggest update since the previous interview, Dario says the technology tracked his forecast “plus or minus a year or two”: the march “from smart high school student to smart college student to beginning to do PhD and professional stuff, and in the case of code reaching beyond that.” The frontier is “a little bit uneven, but roughly what I expected.”
• The genuine surprise: “It is absolutely wild that you have people — within the bubble and outside the bubble — talking about the same tired, old hot-button political issues, when we are near the end of the exponential.”

2. The Big Blob of Compute still explains everything — RL included

• His 2017 doc was written when GPT-1 had just come out and covered robotics, AlphaGo, Dota, AlphaStar: cleverness and new methods “don’t matter very much.” What matters is seven things — raw compute, data quantity, data quality/distribution, training duration, an objective function “that can scale to the moon,” plus normalization/conditioning so the blob “flows in this laminar way.”
• The update since: “We’re seeing the same scaling in RL that we saw for pre-training” — other companies have published that math-contest performance is “log-linear in how long we’ve trained it,” and Anthropic sees it “not just math contests. It’s a wide variety of RL tasks.”
• On why labs build bespoke RL environments if generalization is coming: same reason GPT-2 trained on broad internet scrapes — “not because we want to cover a specific document or a specific skill, but because we want to generalize.” GPT-1 was trained on fanfiction and didn’t generalize; breadth fixed it, and RL is retracing that arc from math contests to code to “many other tasks.”

3. Sample efficiency puzzle: pre-training sits between evolution and learning

• Dwarkesh channels Rich Sutton’s objection — something with “the true core of human learning” wouldn’t require all these billions of dollars of data and compute and these bespoke environments to learn Excel — implying “we’re scaling the wrong thing.” Dario grants “a genuine puzzle” but says “it probably doesn’t matter,” and calls the RL-vs-pretraining framing “a red herring.”
• The real asymmetry: models train on trillions of tokens humans never see, but start as “literally random weights” where brains ship with evolutionary priors. His resolution: pre-training lives “somewhere between the process of humans learning and the process of human evolution,” and in-context learning sits between long-term and short-term human learning — “the LLM phases exist along this spectrum, but not necessarily at exactly the same points.”

4. 90% on the country of geniuses by 2035 — the residual is non-verifiable tasks

• In 2019 scaling was “a 50/50 thing”; today, on getting a “country of geniuses in a data center” within ten years, “I’m at 90%.” The missing 10% splits into irreducible world-chaos (“Taiwan gets invaded, all the fabs get blown up by missiles”) and one technical residual: tasks that aren’t verifiable — “planning a mission to Mars; doing some fundamental scientific discovery like CRISPR; writing a novel.”
• When Dwarkesh presses that leaning on verification betrays doubt about generalization, Dario pushes back: “We already see substantial generalization from things that verify to things that don’t” — the failure mode is a world where “we don’t fully color in the other side of the box,” and “it’s not a binary thing.”

5. The code spectrum: 90% of lines already happened; end-to-end SWE in one or two years

• Dario says people repeatedly misread his predictions, so he lays out the spectrum: 90% of lines of code (predicted 8-9 months ago at 3-6 months out — “that happened, at least at some places,” including Anthropic) → 100% of lines → 90% of end-to-end SWE tasks including “compiling, setting up clusters and environments, testing features, writing memos” → 100% → eventually 90% less demand for SWEs. “We’re proceeding through them super fast.”
• His concrete bet: “We may get to the point in a year or two where the models can just do SWE end-to-end” — and when Dwarkesh asks if that includes setting technical direction and understanding context, “Yes. I mean all of that.” Dwarkesh’s reply — “that is AGI-complete, which maybe is internally consistent” — stands.

6. “Diffusion is cope” vs. the 10x revenue curve

• Dwarkesh’s hot take: diffusion is an excuse — AIs have inherent onboarding advantages (read your whole Slack in minutes, share knowledge across copies, no adverse selection), yet “we pay humans upwards of $50 trillion in wages” while AI integration lags. Dario’s answer rejects both poles — slow-diffusion dismissiveness and “Dyson spheres around the sun so many nanoseconds after we get recursive” — for a middle world: “extremely fast, but not instant.”
• The evidence: Anthropic’s “bizarre 10x per year growth in revenue” — zero→$100M in 2023, $100M→$1B in 2024, $1B→$9-10B in 2025, “another few billion to revenue in January.” He’d guess the curve “bends somewhat this year” but “stays pretty fast even as the scale goes to the entire economy.”
• Even Claude Code, “extremely easy to set up,” hits enterprise friction — legal, provisioning, security, a leader explaining a $50M rollout to 3,000 developers. His ceiling: not even AGI is “an infinitely compelling product” — maybe 3-5x or 10x annual growth at hundreds-of-billions scale, “which has never been done in history before.” And a flat rejection of the strong version: “If we had the ‘country of geniuses in a data center’, we would know it. Everyone in this room would know it. Everyone in Washington would know it.”

7. Continual learning: maybe not a barrier at all

• Dwarkesh’s test case is his video editors, who build up taste and audience context over six months — when can an AI match that on-the-fly learning? Dario says the country of geniuses can, via mastered computer use (OSWorld climbed from ~15% to 65-70% since launch a year and a quarter ago) plus long context: “A million tokens is a lot. That can be days of human learning.” His timeline for indifference between a six-month human editor and a six-month AI: “one to two years, maybe one to three.”
• His deeper claim: the existing paradigm — pre-training/RL generalization plus in-context learning — “may just be enough” for the country of geniuses and “enough to generate trillions of dollars of revenue,” with a good chance of continual learning proper being solved “in the next year or two” anyway. Longer context is blocked by KV-cache serving economics, “an engineering and inference problem,” not research.
• He cites ML’s history of dissolving barriers: “‘They can understand syntactically, but they can’t understand semantically? It’s only statistical correlations’… then suddenly it turns out you can do code and math very well.” Hedged honestly: “Some of them are real. The need for data is real, maybe continual learning is a real thing.”

8. The productivity fight: major-study downlift vs. “zero time for bullshit”

• Dwarkesh cites the study where experienced developers felt more productive with AI but showed a 20% downlift in merged output, and asks where the “renaissance of software” is. Dario’s rebuttal is internal and blunt: under incredible commercial pressure plus safety work, “There is zero time for bullshit. There is zero time for feeling like we’re productive when we’re not… We wouldn’t be going through all this trouble if this were secretly reducing our productivity.”
• On why the best coding model hasn’t produced a runaway lab lead: the advantage is a snowball, not a step function — coding models now give “maybe a 15-20% total factor speed up… Six months ago, it was maybe 5%. So it didn’t matter.” Plus leakage: “we’re not perfectly good at preventing some of these other companies from using our models internally.” Theme: “all of this is soft takeoff, soft, smooth exponentials, although the exponentials are relatively steep.”

9. Why not buy $5 trillion of compute? Because off-by-one-year is death

• Dwarkesh’s challenge: if you truly believe in a country of geniuses in 1-3 years, conservative compute scaling is incoherent. Dario’s spreadsheet: extrapolating 10x/year implies $100B revenue end-2026, $1T end-2027 — but committing to $1T/year of compute means “if it’s even $800 billion, there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt.” If growth is 5x instead of 10x, or the country of geniuses arrives mid-2028 instead of mid-2027, “you go bankrupt.”
• “Responsible” scaling means the spreadsheet exists, not that the number is small: “I get the impression that some of the other companies have not written down the spreadsheet… They’re just doing stuff because it sounds cool.” Anthropic’s enterprise mix (“less fickle than consumer”) and better margins buffer the bet; “things would have to go pretty badly for us to be in financial trouble.”
• The polio analogy caps the diffusion point: even after a vaccine exists, “we’ve had a polio vaccine for 50 years. We’re still trying to eradicate it in the most remote corners of Africa” — the technology arriving is “the starting gun,” not the finish line.

10. Industry math: trillions in capex by 2028-29, trillions in revenue before 2030

• Industry compute build: ~10-15 GW this year, ~3x a year — 30-40 GW next year, ~100 GW in 2028, ~300 GW in 2029, at ~$10-15B per gigawatt: “You’re getting multiple trillions a year by 2028 or 2029.” When Dwarkesh back-solves an implied ~$200B Anthropic TAM by 2028: “I don’t want to give exact numbers for Anthropic, but these numbers are too small.”
• The revenue prediction, with its stated slow-end scenario: “It is hard for me to see that there won’t be trillions of dollars in revenue before 2030” — the plausible worst case being the country of geniuses arriving in 2028 with revenue in the low hundreds of billions, then two years of diffusion to trillions.

11. Profitability is a demand-forecasting residual, not a milestone

• Dario’s toy model: ~half of compute for training, half for inference at >50% gross margins — pay $100B/year for compute, $50B of it supports $150B of revenue, and “basically you’re profitable and you make $50 billion of profit.” “Profitability happens when you underestimated demand and loss happens when you overestimated it” — the invest-invest-then-inflect model “doesn’t work that way” here.
• Today’s losses decomposed: “a model got trained that costs $1 billion last year. Then this year it produced $4 billion of revenue and cost $1 billion to inference from… Each model makes money, but the company loses money” because the next model costs $10B — an artifact of the exponential scale-up phase, not the steady state.
• Why not spend more than 50% on training? Log-linear returns: “that extra $20 billion, each dollar there is worth much less to you” — the equilibrium training share is “of order one fraction of the business. Like not 5%, not 95%.” The economy caps it too: “I think we may get 10-20% per year growth in the economy, but we’re not gonna get 300%.”

12. Market structure: three to four players, cloud-like — but more differentiated

• His equilibrium (“all my lawyers never want me to say the word ‘monopoly’”): not a monopoly, but a small-number Cournot-style oligopoly driven by very high entry costs, like cloud’s three-maybe-four players — “it doesn’t equilibrate to perfect competition with zero margins.”
• Models beat cloud on differentiation: “It’s not just that Claude’s good at coding, GPT is good at math and reasoning. It’s more subtle than that” — different coding styles, different subtleties. The one counter-argument he flags: if AI models can produce AI models, “that’s kind of an argument for commoditizing the whole economy at once… maybe we want that world” — but that’s “far post-country of geniuses.”
• His geographic worry deserves note: growth could run “like 50% in Silicon Valley and parts of the world socially connected to Silicon Valley” and near-current-pace elsewhere — “a pretty messed up world” he actively thinks about preventing.

13. Robotics, API durability, and tokens worth tens of millions

• Robotics doesn’t need human-like learning — generalization from video games, simulated environments, or computer control could get there — but once models have the skills, “robotics will be revolutionized,” both design and control. Timeline discipline holds: same “extremely fast, but not infinitely fast” diffusion — “maybe tack on another year or two.”
• The API model “is more durable than many people think”: on an exponential there’s always a surface of new use cases developed in the last three months, and the API is “very close to the bare metal.” Chatbots are already saturating (“making it smarter doesn’t really help the average consumer that much”) — but that’s product-specific, not evidence models are good enough.
• Pricing will fragment because tokens aren’t equal: “restart it” tech support is worth cents; telling a pharma company “take the aromatic ring from that end of the molecule and put it on that end… those tokens could be worth tens of millions of dollars.” Expect pay-for-results and labor-like hourly compensation experiments.
• Claude Code’s origin: early 2025 he told staff AI could now nontrivially accelerate their own research; the internal “Claude CLI” saw such fast adoption that “it looks like we already have product market fit. Let’s launch this thing” — hundreds of internal coders as a built-in feedback loop. “This is the reason why we launched a coding model and didn’t launch a pharmaceutical company.”

14. Governing proliferating AIs: offense-dominance and the need for speed

• On a world of many actors running “huge populations of misaligned AIs,” Dario is skeptical of balance-of-power checks: “We might live in an offense-dominant world where one person or one AI model is smart enough to do something that causes damage for everything else.” Near term: safeguards, alignment work, bioclassifiers across the few current players; long term, “some architecture of governance that preserves human freedom” — AI monitoring for bioterrorism and mirror life “in a way that preserves civil liberties.”
• His core anxiety is compression: over 100 years we’d adapt as we did to explosives and cameras — “we’d make our mistakes. My worry is just that this is happening all so fast.” Dwarkesh’s dark corollary — if checks and balances won’t work with AIs, “maybe this just dooms human checks and balances as well” — gets no clean rebuttal, only “there’s some way to make this happen.”

15. Regulation: moratorium was “crazy,” transparency now, FDA is the real bottleneck

• The Tennessee bill criminalizing AI emotional support is “dumb… made by legislators who just probably had little idea what AI models could do.” But the actual vote was a 10-year ban on all state regulation “with no apparent plan to do any federal regulation” — “given the serious dangers… 10 years is an eternity — I think that’s a crazy thing to do.” He’d support real federal preemption: “Here’s what we’re going to do, and states you can’t differ from this.”
• Sequencing: transparency standards first, then fast targeted action as evidence lands — “I could certainly imagine… a world where later this year we say, ‘Hey, this AI bioterrorism stuff is really serious’” and it goes into federal or state standards.
• On benefits being Whac-A-Moled: markets in the developed world are robust — “when there’s a lot of money to be made… it’s actually hard for the regulatory system to stop it.” His counter-worry is the drug approval pipeline jamming up: AI-discovered drugs with crisp efficacy may not need “all this superstructure designed around an era of drugs that barely work.” The exhibit for money beating policy is his own losing fight: export controls are squarely bipartisan, the counterarguments “fishy” — “yet it doesn’t happen and we sell the chips because there’s so much money riding on it.”

16. China, authoritarianism, and whether AI dissolves dictatorships

• Why shouldn’t both the US and China have a country of geniuses? Two failure modes: nuclear-like offense dominance “but more dangerous,” and instability from divergent win-probability assessments — “if one side is like, ’there’s a 90% chance I’ll win,’ and the other side thinks the same… they can’t both be right, but they can both think that.” Plus high-tech authoritarian oppression — “this is about the government. This is not about the people.”
• He expects “a critical moment, a small number of critical moments, or some critical window” where AI confers some large national-security advantage (is nuclear deterrence still stable? does someone get offensive cyber dominance?), followed by an implicit or explicit negotiation over “the post-AI world order” — his goal is that classical liberal democracy “has a strong hand” when the rules of the road are set. He clarifies the “autocracy cannot exist post-AGI” passage was exploratory, not endorsed: committing to overthrow authoritarians “has exactly all the problems you say.”
• His genuine hope, flagged as idealistic: as industrialization made feudalism unsustainable, AI could make “dictatorships become morally obsolete” — including equilibria where “we can give everyone in an authoritarian country their own AI model that defends them from surveillance.” He admits the Obama-era internet hope failed the same test: “what if we could try again with the knowledge of how many things could go wrong?”
• On the developing world: philanthropy helps but “growth is always better and stronger if we can make it endogenous” — they shouldn’t build data centers in China, but “it’d be great to build data centers in Africa” (as long as China doesn’t own them), plus AI-driven biotech startups seeded in the developing world.

17. Constitutions: principles over rules, three loops of legitimacy

• Why principles rather than instructions: “purely a practical and empirical thing” — a rules list (“don’t tell people how to hot-wire a car, don’t speak in Korean”) doesn’t generalize; principles make behavior “more consistent” and edge cases coverable. On corrigibility, Claude is “a mostly corrigible model that has some limits, but those limits are based on principles” — “we’re not trying to build something that goes off and runs the world on its own.”
• Who sets the principles: three loops — internal iteration with public updates; competition between companies’ constitutions so observers cherry-pick the best elements (“a soft incentive and feedback”); and societal input, from Collective Intelligence Project-style polling up to a “crazy idea” of legislated mandatory constitutional sections — which he immediately walks back as “too rigid.” Dwarkesh notes loop two recreates the libertarian charter-city “archipelago” vision; Dario: “things will go wrong that you hadn’t imagined… it’s a matter of the proportions” of all three.

18. What history will miss — and the DVQ

• Asked what the eventual Making of the Atomic Bomb of this era will fail to capture: how little the outside world understood at every moment (“anything that actually happened looks inevitable in retrospect”), and the raw speed — “some very critical decision will be some decision where someone just comes into my office and is like, ‘Dario, you have two minutes’… ‘I don’t know. I have to eat lunch. Let’s do B.’ That ends up being the most consequential thing ever.”
• His management model: a third to 40% of his time on culture, biweekly all-hands “Dario Vision Quest” talks (“a name I tried to fight because it made it sound like I was going off and smoking peyote”), and an unfiltered Slack channel — the point being “a reputation of telling the company the truth… to avoid the corpo speak.” He’s candid that at 2,500 people his edge is no longer research insight but “seeing a bunch of stuff within Anthropic and having to make a bunch of decisions” — and notes, without names, that other AI companies are showing “decoherence and people fighting each other.”