The People Building This Are Also Scared of It

Dario Amodei left OpenAI because he didn't trust them. Now he's running the most powerful AI lab in the world — and he's not sure anyone should trust him either.

Dario Amodei grew up in San Francisco while the first internet revolution was happening around him. He had no interest in it. He was interested in math, in understanding the universe, in science fiction. His sister Daniela was into reading and arts. They were complements — and they always wanted to do something big together.

He studied neuroscience, moved into AI, ended up at OpenAI, and helped develop the scaling laws that powered GPT. Then he left. When asked why, he's direct: it wasn't just disagreement on safety. It was that he stopped trusting the people. When you don't trust someone and don't share their vision, the answer is simple — you go do your own thing.

So he did. He and Daniela founded Anthropic in a park during the pandemic, seven co-founders gathering on the grass for lunch, talking about what they were building. They're the only AI lab of their scale that still has all its original co-founders. Dario says that almost never happens.

The Business Model Is the Values

Early on, while others chased consumer apps, Anthropic bet on enterprise and coding. Dario's reasoning is simple: if your business model conflicts with your values, you either betray the values or become irrelevant. Consumer social media optimizes for engagement — which means addiction, which means slop. Enterprise means working with biotech to discover drugs, with researchers to accelerate science, with companies to make energy more efficient. The incentives align differently.

Claude Code, which automated large chunks of software engineering, wiped $285 billion in market value from software stocks overnight. Traders called it the SaaS apocalypse. Dario's response: the pie is getting bigger. Some incumbents will shrink. Some will disappear. The overall software industry gets larger, not smaller — but only for those who see what's coming.

The Uncomfortable Honesty

He's said AI could eliminate half of all entry-level white collar jobs in the next one to five years. He hasn't walked it back. His framing: you automate 90% of a job, people are ten times more productive in the remaining 10%. Eventually it gets close to 100%. Then you have to find something else for people to do — and that transition will be painful.

When asked if that sits comfortably with him, he says no. It's very uncomfortable. He chose Anthropic because if this technology is coming regardless, he'd rather be inside trying to make it go better. That's not reassurance. It's just the honest version of why someone builds what they're afraid of.

Mythos

The lab built a model so capable at finding cybersecurity vulnerabilities that early testers told them it was a superweapon and begged them not to release it. They didn't. The decision cost them enormously commercially. They called it Project Glasswing and gave restricted access to select organizations. The Pentagon wanted in. Then the Pentagon demanded access without guardrails. Anthropic refused. They got blacklisted. Trump's Defense Secretary called Dario an ideological lunatic.

His response: he's been called worse.

The Question He Can't Answer

He puts the probability of civilizational collapse from AI at somewhere between 10 and 25%. When pushed — if there's a 25% chance the plane crashes, you don't get on — he agrees. 25% is too high. The goal is to make that number much lower. He can't guarantee it will be zero. No one can. But the alternative — stepping back and letting someone with less concern for the outcome build it — doesn't lower the probability either.

His role model isn't Oppenheimer. It's Leo Szilard, who first conceived of nuclear chain reactions and then spent the rest of his life trying to prevent nuclear war. Oppenheimer, in his view, is the failure case. Too centered on himself, not enough on the system.

When he needs to find calm, he goes to Italy. Sits next to a horse named Calypso. She doesn't know about any of this. She's just a happy horse.

He finds that helpful.