Dario
Amodei

The first thing to understand about Dario Amodei is that he never wanted to run a company. He wanted to understand how the brain works. Then his father died.

Riccardo Amodei - a leather craftsman from Tuscany who had emigrated to San Francisco - died in 2006 of a rare illness while Dario was a graduate student at Princeton. Four years later, a medical breakthrough made that same illness 95% curable. Dario has said the timing of that fact changed everything for him. Not the grief. The arithmetic. The cost of delay.

That calculus - what does it cost the world when progress arrives one year late? - runs through everything Amodei has built since. It is why he pivoted from theoretical physics to biophysics at Princeton. It is why he pivoted from biophysics to AI at Baidu. And it is why, in February 2021, he walked out of OpenAI with his sister Daniela and six colleagues and incorporated Anthropic, with the stated goal of making sure the most consequential technology in human history does not arrive badly.

Fortune called him "the accidental CEO." He would probably agree. But there is nothing accidental about what he is building.

Amodei grew up in San Francisco - the son of an Italian craftsman and a Jewish-American mother who managed library renovation projects around the Bay Area. His mother, Elena, instilled what he calls "a strong sense of right and wrong." His childhood pastime, according to his sister Daniela: declaring "counting days" and seeing how high he could get.

In 2000, he made the US Physics Olympiad team. At Caltech during the Iraq War, he wrote opinion pieces criticizing student inaction. During the dot-com boom, when his classmates were dreaming of IPOs, he was uninterested. "Writing some website actually had no interest to me whatsoever," he has said. "I was interested in discovering fundamental scientific truth."

At Princeton, he published a dissertation on how large populations of neurons operate as collective systems - work that translated almost directly to what artificial neural networks now do. His advisor, Michael Berry, called him the most talented graduate student he had worked with. He won the Hertz Thesis Prize in 2012. Then he went to Stanford to study cancer detection, then to Baidu to work under Andrew Ng, then to Google Brain, then to OpenAI.

At OpenAI from 2016 to 2021, Amodei rose to VP of Research. He led GPT-2 and GPT-3. He co-authored what became a foundational paper in AI safety - "Concrete Problems in AI Safety" - in 2016, alongside Chris Olah and others. He co-invented Reinforcement Learning from Human Feedback (RLHF), the alignment technique that is now embedded in virtually every major language model on earth. He helped discover the scaling laws that predicted, mathematically, how much better models would get as you threw more compute at them.

He also watched OpenAI take a $1 billion investment from Microsoft and worry that the company's mission was drifting. In December 2020, he and his sister began planning an exit. In February 2021, they left.

The story of how Anthropic was founded is a story about disagreement - but not the kind that makes for good drama. No screaming match. No single moment. Just a gradual divergence about what kind of company you become when you take a billion dollars from a technology giant whose primary interest is not AI safety.

Amodei has been careful not to attack OpenAI publicly. He tends to frame the departure as a philosophical difference rather than a personal one. But the structural concern is clear: he believed that safety research needed to be the organizing principle of an AI lab, not an important team within a lab organized around deployment and revenue.

He brought with him Daniela (now President of Anthropic), Jared Kaplan (co-author of the scaling laws paper, now Chief Science Officer), Chris Olah (now leading interpretability research), Tom Brown, Sam McCandlish, Jack Clark, and Benjamin Mann. Eight co-founders. All from OpenAI. All aligned on the idea that the most dangerous thing in the world right now is building powerful AI without understanding what it is doing.

Anthropic's first major technical contribution was Constitutional AI - a method for training models to be harmless by having AI systems critique their own outputs against a written set of principles, rather than relying solely on human labelers. It was Amodei's answer to the problem he had spent years thinking about: how do you instill values into a system that learns from data?

The company he built has been funded with more than $20 billion from Amazon, Google, and institutional investors. As of February 2026, Anthropic is valued at $380 billion. Its annualized revenue is estimated at around $30 billion. Claude - Anthropic's model family - competes directly with GPT-4o and Gemini. Claude Code has seen 5x usage growth since mid-2025. None of this was the plan. All of it was necessary.

Amodei's position in the industry is genuinely paradoxical. He is one of the most vocal critics of concentrated AI power. He is also one of the people most responsible for concentrating it. He does not hide from this. "I'm deeply uncomfortable," he told 60 Minutes in November 2025, "with these decisions being made by a few companies, by a few people. And yet here I am."

Amodei's position on AI safety is not the doomsday position it is sometimes caricatured as. He does not think AI will necessarily destroy humanity. He thinks it might, if the people building it are not careful, and that being careful is a full-time job requiring serious engineering, not just good intentions.

His contributions to safety are technical, not rhetorical. He co-invented RLHF. He developed Constitutional AI. He created Anthropic's AI Safety Level (ASL) framework - a tiered system modeled on biosafety levels, where each level corresponds to a threshold of capability and a corresponding set of safety requirements. Current Anthropic models are at ASL-2. ASL-3 triggers additional precautions around weapons and cyberattacks. ASL-4 is the level where, by Anthropic's own policy, they would pause deployment.

He is also a strong proponent of chip export controls to China, describing the rolling back of such controls as potentially "the single most disastrous national security decision made in this term." He co-wrote an op-ed on the subject with a former US deputy national security adviser.

On AGI timelines, he is specific: "If you just kind of eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there by 2026 or 2027." He calls himself 90% confident AGI arrives by 2035. He is not hiding from these numbers. He built his company around them.

In February 2026, when Pete Hegseth's Department of Defense asked Anthropic to remove contractual bans on using Claude for mass domestic surveillance or fully autonomous weapons systems, Amodei refused. The refusal was not publicized as a PR move. It was reported as a business dispute.

He gets irritated at being called a doomer. "I get really angry when someone's like, 'This guy's a doomer,'" he has said. He considers himself an optimist - an optimist who takes the math seriously.

Most tech CEOs have a research background the way most people have a college degree - useful for getting in the door, mostly irrelevant after that. Amodei's research background is structural to everything he does. His h-index is 34. He has 173,000 Google Scholar citations. For someone primarily known as a company founder, this is extraordinary.

His 2016 paper "Concrete Problems in AI Safety" - co-authored with Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mane - defined five categories of AI accident risk and has accumulated more than 2,300 citations. It remains one of the foundational documents of the AI safety field. He wrote it while working at OpenAI, before Constitutional AI, before the Claude models, before any of the current alignment infrastructure existed.

The scaling laws paper he co-authored with Jared Kaplan in 2020 gave the industry a precise mathematical framework for predicting model capability from compute and data. It was not just academically significant. It was operationally significant - it told AI labs exactly how to allocate resources to get the most capability per dollar. Every major lab's training decisions since 2020 have been shaped by this work.