The Amodeis
A physicist and an English major grew up in the same house, took completely different paths, and co-founded the most consequential safety-focused AI lab in the world.
Daniela Amodei studied English literature at UC Santa Cruz. She played classical flute on a partial scholarship. She ran a congressional campaign in Pennsylvania, making eleven thousand voter calls to flip a House seat. She had never published a research paper, never trained a neural network, never written a line of code that shipped to production. And in 2021, she co-founded what is now the most consequential safety-focused research lab in the world.
That fact alone tells you something important about what Anthropic actually is. It is not just a research lab. It is an argument about how you build an organization around dangerous technology. And to understand that argument, you have to understand two people who grew up in the same house in San Francisco, went in completely different directions, and ended up in the same room making the same bet.
Dario Amodei was born in 1983. His father, Riccardo, was an Italian leather craftsman from a small town in Tuscany called Massa Marittima. His mother, Elena, was a Jewish American from Chicago who worked as a project manager for libraries. The family had a streak of moral seriousness. Daniela’s told this story about their maternal grandmother chaining herself to the Italian consulate in Chicago in the 1930s to protest Mussolini’s invasion of Ethiopia. Whether that story is perfectly accurate or a little embellished by family memory, it says something about the household these two grew up in. You were expected to care about the world and act on it.
Dario was the science kid. By his own telling, the dot-com boom barely registered. He said, and I’m quoting him here: “Writing some website actually had no interest to me whatsoever. I was interested in discovering fundamental scientific truth.” He made the USA Physics Olympiad team in 2000. He started at Caltech, transferred to Stanford, graduated with a degree in physics in 2006.
That same year, his father died after a long illness. Dario was twenty-three. And something shifted. He moved away from theoretical physics toward biophysics, toward biology, toward the parts of science that might address human suffering directly. His PhD at Princeton studied the collective behavior of neural circuits, how hundreds of cells in retinal tissue fire together. He won the Hertz Thesis Prize for it. And somewhere in that work, studying how biological networks process information, the bridge to artificial networks started forming.
Daniela, four years younger, took a completely different path. She graduated summa cum laude with her English degree, won a concerto competition as a flute soloist, and then went into the world. International development. Then politics. In 2012, she was a field director for Matt Cartwright’s congressional campaign in Pennsylvania’s 17th district. She recruited more than eighty volunteers. Cartwright won. She went to D.C. as his scheduler and legislative correspondent. Then she left politics for Stripe, the payments company, where she was employee number forty-five.
At Stripe, Daniela did two things that would matter later. First, she grew the company from forty-five people to three hundred, personally recruiting ninety-two engineers with a close rate above seventy-five percent. Second, she moved into risk management, leading three teams of twenty-six people and driving the company’s loss rate down seventy-two percent from its peak to the lowest in Stripe’s history. She was building organizations and managing risk. Not writing papers. Not training models. Building the structures that let technical work happen at scale.
Meanwhile, Dario’s path through research was accelerating. After his postdoc at Stanford, he joined Baidu’s Silicon Valley AI Lab. He was first author on Deep Speech 2, a paper that showed you could build a speech recognition system by throwing more data and compute at a deep neural network instead of hand-engineering every component. MIT Technology Review named it one of the ten breakthrough technologies of 2016. But what mattered to Dario was not the speech recognition itself. It was the curves. The scaling curves. Performance improved smoothly and predictably as you added more compute. He said it had a big impact on him.
He went to Google Brain. Co-authored a paper called “Concrete Problems in AI Safety” with Chris Olah and others. That paper helped establish safety as a legitimate technical research area, not a side conversation for philosophers. Then OpenAI recruited him. He rose to VP of Research. He co-authored the foundational paper on reinforcement learning from human feedback, the technique that would later make language models conversational. He led the development of GPT-2 and GPT-3. OpenAI entrusted him with over half the company’s compute budget for GPT-3. By any measure, he was at the center of the most powerful AI research effort on the planet.
And he walked out.
December 2020. Dario has been clear about what happened, and what did not happen. People said the departure was about the Microsoft deal. He says that is false. The real issue was a disagreement about how to build. Dario and Daniela, who had joined OpenAI as VP of Safety and Policy, believed safety had to be baked into the training process from the beginning, not bolted on after you had already built a powerful system. The leadership at OpenAI had a different vision. Dario’s words: “It is incredibly unproductive to try and argue with someone else’s vision.”
So they did not argue. They left. And they took people with them. Tom Brown, the lead author on the GPT-3 paper. Chris Olah, a self-taught researcher with no formal degree who had pioneered a field called mechanistic interpretability, which is essentially the science of opening up a neural network and understanding what it is doing inside. Jack Clark, a former Bloomberg journalist who had been OpenAI’s policy director. Sam McCandlish. Jared Kaplan, a Johns Hopkins professor who had co-authored the scaling laws paper with Dario. Seven co-founders in total. All from OpenAI. All willing to leave the most resourced lab in the industry because they believed the approach was wrong.
They incorporated Anthropic as a Public Benefit Corporation. Their founding mission statement reads: “the responsible development and maintenance of advanced AI for the cultural, social and technological improvement of humanity.” Their early meetings happened in Precita Park in San Francisco because it was 2021 and everything was still outside. Jaan Tallinn, the co-founder of Skype, led the Series A at a hundred and twenty-four million dollars.
The founding thesis was simple and radical. Safety and capability are not in tension. Getting safety right is what enables the beneficial outcomes. This was not the consensus view. Most of the industry treated safety as a brake, something that slowed you down, constrained what you could ship, cost you market share. Anthropic’s bet was that safety was the engine.
The clearest expression of that thesis is Constitutional AI. Here is what it is in plain terms. The standard way to make a language model behave well is called reinforcement learning from human feedback. You hire people to look at the model’s outputs and rate them. Good answer, bad answer. The model learns from those ratings. The problem is that human raters are inconsistent. They bring their own biases, their own bad days, their own cultural assumptions. And none of those values are written down anywhere. They are implicit, invisible, and impossible to audit.
Constitutional AI takes a different approach. You write down the principles you want the model to follow. Literally write them down. A constitution. Then you train the model to critique its own outputs against those principles and revise them. In the second phase, an AI evaluator judges which responses better follow the constitution, and you use that feedback to improve the model further. The constitution Anthropic published in January 2026 is about twenty-three thousand words long. It draws from the UN Universal Declaration of Human Rights, Apple’s terms of service, research from DeepMind, and non-Western cultural perspectives. It has a priority hierarchy: safety and human oversight first, then ethical behavior, then Anthropic’s own guidelines, then helpfulness.
Now let me talk about Daniela, because she does not get her full weight in most of these stories, and that is a mistake.
Anthropic exists as an organization because Daniela built it. The research is Dario’s domain. The company is hers. She has been president since day one. She built the hiring systems, the operational infrastructure, the fundraising engine, the policy relationships, the culture. When Anthropic grew from seven people to thousands, that did not happen by accident. It happened because someone who had scaled Stripe from forty-five to three hundred people, who had managed congressional campaigns, who had run risk teams, applied all of that experience to building a structure around the science.
She has said, and this is revealing: “I have probably been the leader who’s been the most skeptical and scared of the rate at which we’re growing.” That is not a weakness. That is operational wisdom. Growth kills companies. Especially companies doing dangerous work. The person holding the growth rate accountable is as important as the person pushing the research forward.
When someone says you need a PhD to lead in this space, Daniela is the counterargument. An English literature degree, a flute scholarship, a political campaign, a risk management career, and now the presidency of one of the most valuable private companies on earth.
The tools they have shipped reflect the philosophy. Claude, named after Claude Shannon, the father of information theory, is trained on Constitutional AI principles. The thing that distinguishes Claude from other models is not raw capability. It is honesty about uncertainty. Ask Claude something it does not know, and it will tell you it does not know. Ask it something nuanced, and it will give you the nuance instead of collapsing it into a confident wrong answer. That is not a product decision. That is a training philosophy showing up in every interaction.
But here is where I have to be honest, because this story is not hagiography.
Anthropic has taken billions from Google and Amazon. Google owns fourteen percent of the company. Amazon has invested eight billion dollars. The question that every serious observer asks is whether a company can self-regulate when the financial incentives are this large.
The Responsible Scaling Policy was supposed to be the answer. Version 1.0, published in September 2023, was the first framework of its kind: safety levels modeled after biosafety levels, with clear thresholds that would trigger additional safeguards as models became more capable. It was genuinely novel. Then version 3.0 arrived in February 2026, and it dropped the binary safety thresholds. Jared Kaplan, who by then was Anthropic’s Responsible Scaling Officer, said the quiet part out loud: “We didn’t really feel it made sense to make unilateral commitments if competitors are blazing ahead.”
That is a reasonable business argument. It is also a retreat from the original promise.
So what do the Amodeis represent? Not sainthood. Not villainy. A bet. The bet that you can build the most powerful technology in human history and do it in a way that does not destroy the thing you are trying to improve. That you can write down your values and train a system to follow them. That the person who builds the organization is as important as the person who builds the model. That a family from San Francisco, one kid who studied physics and one kid who studied literature, could look at the same problem from different directions and build something that matters.
Whether that bet pays off is the story of the next decade. But the people who made the bet, and the tools they shipped along the way, are worth understanding. Because if they are right, the approach they are proving out changes everything. And if they are wrong, we will need to understand exactly where the logic broke down.