This is a long-form, in-depth interview with Anthropic CEO Dario Amodei (part of the same series as the earlier Bloomberg documentary, but a fuller conversation version). It opens on his state and mindset: he doesn't sleep much and is learning to relax under 'unusual pressure'; he describes the current experience as a 'smooth exponential' — nothing happens, nothing happens, and then 'whoosh,' it takes off, like flying away from Earth at relativistic speed, where you sleep and wake to find two days have passed on Earth, then sleep again and it's three, four days. But he doesn't advocate anxiety or 'fear of what you'll wake up to'; rather, one should assess risk calmly and rationally like a surgeon or a military officer. He repeatedly stresses: yo-yoing between 'I'm not worried' and 'oh god we must panic today' is the mark of immature decision-making; the mature approach is neither to complacently ignore nor to exaggerate, adjusting countermeasures smoothly upward as technical capability rises.

He rarely states the real reason he left OpenAI outright: not a disagreement over safety (that kind of disagreement happens daily inside Anthropic and isn't enough to make someone leave), but a collapse of trust — 'when you feel the other side's values don't match what they say, that they're dishonest, that they're not acting for the reasons they claim, when you see disturbing patterns of dishonest behavior.' He's 'completely at peace' with it and is willing to let the market and public opinion judge who's right; the famous scene at the India AI summit where he and Altman 'refused to hold hands,' he explains, was just because the conference was extremely chaotic and Modi spontaneously asked everyone to hold hands. On collaboration, he thinks players' credibility varies across the industry (he's known Demis Hassabis, who builds Gemini, for 15 years, they swap safety ideas, and he buys compute from Google); credible players should band together and, using 'carrots (mutual inspiration, like AlphaFold and interpretability research) + an implicit stick (doing the right thing makes those who don't look bad),' pull the whole industry up — this is the 'race to the top.'

On betting on enterprise/coding, he says it's 'alignment of values and business model': consumer/social-media businesses rely on addiction and ad-time, and AI-video 'slop' exists to maximize attention-minutes; whereas curing disease, energy, education, scientific research, and health in the developing world — the positive uses he believes will ultimately outweigh the negatives — fall almost entirely within the enterprise domain, and enterprises value trust and long-term relationships, which dovetails highly with 'responsible deployment.' Claude Code and Claude Cowork both went viral; after the latter's launch, software stocks lost $285 billion overnight (the SaaSpocalypse). His advice to incumbent software companies is concrete: don't be complacent, don't bury your head in the sand, list out all your moats — 'writing complex software fast' will surely vanish, but customer relationships, industry know-how, and unique domain knowledge will remain and even matter more; AI makes the 'pie' bigger (maybe 10x), so an incumbent industry growing 1.5x still counts as falling behind, there will be big losers, but software as a whole will be larger.

On compute and valuation he gave stunning figures: Anthropic planned to scale compute 10x per year, but Q1 2026 revenue rose 3x in a single quarter — an 80x annualized rate (3 to the fourth power), a 'localized extreme burst' that no one could or would rationally stock up for in advance, and won't last. The near-trillion-dollar valuation and financing are just a buffer against the 'cone of uncertainty,' with very little dilution — the exact opposite of 'fundamentals being in trouble'; he stresses model quality is the most important moat, and Anthropic never relies on 'stickiness' to lock in customers. Asked why product velocity is so fast, he attributes it to two things: a unified company culture (he and Daniela made 'preserving culture while scaling at high speed and hiring from big tech' their top priority) and the increasingly reliable acceleration from 'using Claude to build Claude.' He says the most stunning AI he's seen is in biomedicine: Claude diagnosed a slew of conditions that top doctors missed (including that time with Daniela herself), and it's so good at drug design/computational chemistry that it astonished him, a former biologist — he firmly believes AI will bring 'a century of scientific and medical progress.'

Employment is his signature topic: he reiterates that 'in the next 1-5 years, half of entry-level white-collar jobs may be cut,' but complains that he's 'always edited into three-second doom clips' — he always presents solutions at the same time (a token tax, macro policy, adjusting roles together with enterprises, though he's reserved about retraining) and distinguishes 'tasks' from 'jobs.' He cites Anthropic itself: AI now writes nearly all the code yet still makes engineers more effective, but signs are emerging that 'some people are no longer being made more productive — better to just let the AI do it'; meanwhile there's strong demand for roles like 'forward-deployed engineer / applied-AI solutions architect' that combine 'technical + client-facing.' He predicts an anomalous combination of 'high GDP growth + high unemployment/wage compression,' and avoiding 'unemployment-driven unrest' is precisely why he wants to warn early so policy can be made; he always pushes enterprises toward 'using the same people to do more new things (positive-sum)' rather than 'laying off to cut costs.'

The second half is the hardest: defense, Mythos, and governance. He's long been anti-war (traceable to Caltech), but because of the 'resurgent authoritarian camp' — Russia-Ukraine, the Taiwan Strait — he believes self-defense is necessary, and so became one of the first AI companies to sign with the Department of Defense and go onto classified networks — 'not for the money; getting on government networks is both a hassle and unprofitable, it's because we care'; but precisely because they care, they must draw red lines: refusing mass surveillance and fully autonomous weapons, 'if democracies do these things, winning isn't worth it.' As a result he was blacklisted by the President, the Pentagon labeled him a 'supply-chain risk,' OpenAI took the contracts he wouldn't, and the defense secretary called him an 'ideological lunatic' — none of which he minds. 'Success + holding to values' are the only two things he cares about, so he lives simply. Pressed on the Bloomberg report that Claude was used in the Iran war via Palantir Maven and that a missile hit a girls' school, killing 150+ (mostly children), he responds that he doesn't know Claude's specific role, but such war tragedies are terrifying, and this precisely shows the importance of holding to the 'humans make the final decision' principle — they'd rather 'risk the company's future' to set limits, and this use case 'didn't even violate our red lines'; what he fears more is the fully autonomous weapons others allow that would violate red lines.

Mythos is another main thread: he calls it the strongest model to date, able to autonomously run an entire 'cyber kill chain' and turn vulnerabilities into usable exploits; an early tester exclaimed 'this should require a gun license to use, please don't release it.' It found 271 new vulnerabilities in Firefox and thousands inside enterprises. Anthropic would rather absorb huge commercial losses than fully release it — giving defenders patches first, and only gradually opening it to a wider audience once defenses are strong enough (current classifiers are easily jailbroken, and he criticizes peers who think that's enough) and governments have proactively slowed the pace due to counterintelligence risk. He rebuts 'open-source models can reproduce Mythos' as utterly false: point an open-source model at the exact line of code Mythos already found and of course it can reproduce it, but the real skill is finding the needle in the haystack across an entire codebase. On governance, he thinks 'whether the government should directly take over such powerful private technology' is a serious question that also worries him: historically, nuclear weapons, the internet, and GPS all originated with government; AI is the first powerful technology led by the private sector with government arriving late, which is itself dangerous and unstable. His solution is checks everywhere: internally there's a 'long-term benefit trust' (which can appoint/remove a majority of directors and effectively fire him), and he encourages peers to emulate it; externally, legislation and the judiciary must push, with mandatory pre-release testing and audits — he mocks certain Silicon Valley figures for yo-yoing 'from extreme anti-regulation (look at us and you're stifling innovation) to extreme nationalization (the government should seize it all) overnight,' calling for a 'rational, moderate middle path.' On China his stance is unchanged: he likens selling chips to China to a dangerous act, once proactively cut off China's access, losing hundreds of millions (a large share of revenue at the time); he most fears AI combined with high-tech authoritarianism (Xinjiang, Hong Kong, cross-border suppression of critics) spawning a '1984 or worse' dystopia, while AI also has the chance to become a 'pro-democracy, freedom-enhancing' technology — which path is taken depends on the actions of AI companies, governments, and everyone.

In closing he discusses recursive self-improvement: it's not a single moment but a continuous process — a year ago AI brought roughly a 10-15% total-factor-productivity gain, now maybe 20-30% and still doubling; there's no single moment when 'AI starts self-improving/goes out of control,' but rather an 'accelerating exponential,' where at every point you must assess whether to slow down and add controls. He self-assesses 10-25% odds of civilizational collapse ('25% is too high, we need to push it very low — you wouldn't board a plane with a 25% chance of crashing'), saying Anthropic's actions are lowering rather than raising that probability; his favorite book is 'The Making of the Atomic Bomb,' but he compares himself to Szilárd — who first conceived the chain reaction — rather than Oppenheimer, believing one can't rely on 'larger-than-life individual heroes' but on checks everywhere. Asked 'you're building an extremely powerful technology and will profit from it, why should we trust you,' he answers: starting from a point of distrust is rational, Silicon Valley has lost the world's trust and must earn it back through actions — holding back Mythos, cutting off China's access, and delaying Claude's release are all 'putting money where your mouth is'; he asks everyone to look at the whole history and infer the 'most consistent hypothesis,' 'and I think the consistent hypothesis is: we genuinely want to do the right thing.'

(For AI investors: ① the 80x quarterly-annualized growth + 10x compute planning quantify the explosion in frontier-model demand and the tightness of compute supply; ② the closed-frontier 'capability threshold + safety threshold + regulation/export controls' are becoming deeply politicized, directly reshaping the compute chain and the application-layer landscape; ③ enterprise/coding is Anthropic's profit and growth engine, and Claude Code-class agents are materially reshaping SaaS valuations. Can be read against the Hassabis×Amodei debate and the Bloomberg Anthropic documentary in this same library.)