In December 2020, Dario Amodei told an all-hands video call at OpenAI that he and six colleagues were leaving to start a company that would put AI safety at its center. The chat filled with surprise and goodbyes. Jack Clark typed: “I love you all.” Then the seven screens went dark. They had been removed from the meeting, their work email and Slack access revoked, their laptops locked remotely.
A few hours earlier they had been senior researchers and managers at one of the leading AI labs. Now they had no jobs, no office systems and no clear next step. Where would the GPUs come from? The money? What would the company be called? Even buying everyone new laptops was an open question. Almost six years later, the company is called Anthropic, its product Claude competes head-on with OpenAI, and the founders are preparing to take it public.
That departure scene comes from an excerpt of Kevin Roose’s new book, published in The Atlantic. It is a better starting point than “seven executives start an AI company”: these were people who had already helped build frontier models, and they were trying to write their disagreements about risk, speed and organizational responsibility into the rules of a new company. Source: The Atlantic
Why they left
The departure is often boiled down to “OpenAI didn’t take safety seriously enough.” That gets the direction right and flattens the rest. Safety was the core disagreement, but it was tied up with who could decide when a product shipped, how the company was run, and whether a safety team could change anything once AI research had become a commercial race.
Dario was OpenAI’s vice president of research. He put conditions to Sam Altman for staying, including giving the safety team the power to block product launches. While those talks were still going, the seven set up a private Slack group called BATNA, a negotiation term for “best alternative to a negotiated agreement.” They were bargaining and preparing an exit at the same time. The talks did not produce a deal.
Daniela negotiated the exit on the group’s behalf. Both sides brought in lawyers and settled on a window during which the departing team could recruit colleagues, plus a one-year mutual no-poach agreement. Dario announced the decision at the all-hands, and Jack left his goodbye in the chat. The details matter: Anthropic was not a group of star scientists who barely knew each other. They had worked inside the same organization and went through a hard collective exit together.
When TIME interviewed him later, Dario was restrained about why they left, putting the weight on mutual trust, shared values and a common mission. He has also argued that AI risk is an empirical question: you have to build and study sufficiently powerful systems to learn what the risks actually are. That is the tension built into Anthropic. It is wary of powerful AI, and it believes it has to push powerful AI forward itself in order to study and manage it. Source: TIME on Anthropic
None of this made the conflict between safety and business go away. It moved the conflict inside Anthropic. Research needs scale and compute. Safety needs time, evaluation and restraint. Commercial growth needs products that keep improving and customers who keep adopting them. What the founders chose together was not to leave the race, but to build a lab that says it will run the race differently.
Six founders’ skills link into one research loop; Daniela’s job of running the company is the outer ring holding it together. Drawn from public roles; not an actual reporting line.
Seven jobs, one company
Dario Amodei: turning scientific judgment into company direction. Dario trained in physics and biophysics and worked at Google Brain and then OpenAI, where he was vice president of research, helped develop large language models, and worked with colleagues on reinforcement learning from human feedback. As Anthropic’s CEO he sets research direction and strategy. His main influence is less any single technique than a question he keeps asking at the level of the organization: if model capabilities grow as expected, what safety machinery does the company need to have in place first? His public writing stresses risk and also describes large possible benefits, for example in biomedicine. He holds both at once, which is part of why Anthropic keeps building while studying how to constrain what it builds. Sources: Dario Amodei’s website, Anthropic leadership
Daniela Amodei: organizing research into a company that ships. Daniela is president and chair of the board, and Dario’s sister. Before Anthropic she worked at Stripe and at OpenAI, in roles spanning safety and policy, people, and engineering management. Her remit covers research, engineering, product, governance and commercial execution, and she manages the executive team. If Dario is usually seen as the company’s technical vision, Daniela’s job is to keep the research organization, product teams, customer demands and company rules all running at once. Her role in negotiating the exit shows she was drawing the company’s organizational boundaries from day one, not holding a symbolic management title. Sources: S-1 management biographies, Anthropic leadership
Tom Brown: making large-model training work as engineering. Tom led engineering on GPT-3 and is now Anthropic’s chief compute officer, responsible for acquiring, scaling and using compute. He is also first author of the GPT-3 paper, “Language Models are Few-Shot Learners.” The 175-billion-parameter GPT-3 showed that scaled-up models could handle many tasks from a few examples. Behind the paper sat hardware scheduling, distributed training, reliability and cost control. Tom’s career connects the scientific ideas behind models with the enormous machinery needed to run them. Compute today is not a back-office budget line; it sets how fast a lab can experiment and where it stands against competitors. Sources: S-1 founder biographies, GPT-3 paper
Jared Kaplan: finding the regularities between scale and capability. Jared is a theoretical physicist and Anthropic’s chief science officer. “Scaling Laws for Neural Language Models,” which he wrote with Dario, Tom, Sam and others, studied how parameters, data and training compute affect a model’s loss. Its empirical regularities help researchers estimate how to split extra compute between model size and data. It did not prove that any capability will appear if you just make models bigger, and it does not replace experiments. What it offers is a testable framework, so frontier experiments are not pure guesswork. Jared now leads Anthropic’s technical research, linking basic science, model development and AI safety. Sources: Johns Hopkins University, Scaling laws paper
Sam McCandlish: making the research loop faster and more reliable. Sam is a physicist who built and led OpenAI’s team studying the science of machine learning and scaling, and he co-authored the scaling laws paper. Anthropic lists him as chief architect, focused on pretraining, research productivity and reinforcement learning infrastructure. Training a large model is not one press of a button. Teams run experiments over and over, track down failures, compare results and reuse systems. Sam’s specialty is the infrastructure for that loop. Tom focuses on compute resources and how they are organized; Sam sits closer to the training pipeline and research tooling. Each depends on the other to turn resources into model progress. Sources: S-1 founder biographies, Anthropic leadership
Chris Olah: trying to open the black box. Chris Olah is an early driver of mechanistic interpretability. He worked at Google Brain, led interpretability research at OpenAI, and co-founded the journal Distill. He describes his goal as reverse-engineering neural networks into algorithms humans can understand. Anthropic’s research on Claude has identified millions of concept features. In one well-known demo, researchers amplified a “Golden Gate Bridge” feature, and the model began describing itself as the Golden Gate Bridge in answer to all kinds of questions. The demo made abstract research concrete, and it also showed that when internal representations are pushed artificially, a model’s behavior can turn strange. Olah’s field has not made models fully transparent, but it has produced new ways to inspect how they work inside. Sources: Chris Olah’s website, Anthropic research
Jack Clark: translating lab language into public questions. Jack was a technology journalist covering supercomputers, neural networks and advanced technology before joining OpenAI to run policy and communications. He is now Anthropic’s head of public benefit and leads the Anthropic Institute, which studies the social effects of frontier AI. His background differs from the technical founders’: while engineers talk about evaluations and training scale, Jack has to connect those ideas to policy, jobs, national security and public understanding. His reporting experience is a bridge between the company and the outside world, not a substitute for technical work. A company studying the public interest also faces a credibility test: who sets the research agenda, and can the conclusions stand apart from the company’s commercial interests? Sources: S-1 founder biographies, Anthropic leadership
Together the seven cover research direction, operations, compute, basic science, training systems, model internals and public policy. That does not mean each guards a separate patch. The research agenda sets compute needs; infrastructure sets experiment speed; explaining model behavior feeds safety evaluation; policy and governance decide how the company answers to the outside world. Anthropic’s early advantage may well come from having all these functions founded by people who had already worked together, rather than bolted on after the company grew.
From papers written at OpenAI in 2020 to the amended S-1 in November 2026. The seven announced their departure in December 2020; Anthropic was founded in 2021.
Writing safety commitments into the rules
The founders left OpenAI with a research plan and also with the intention to change the company’s legal and governance structure. Anthropic is a public benefit corporation, with “the responsible development and maintenance of advanced AI for the long-term benefit of humanity” written into its corporate purpose. It also created a Long-Term Benefit Trust, whose trustees have no financial stake in the company and which shapes the board through a special class of stock.
The amended IPO filing also has Founder LLC hold a single share of Class F stock, giving the seven founders 50.1% of the vote on matters put to shareholders generally. That is a significant founder-control arrangement. The governance picture is not “the founders decide everything,” though. The 50.1% covers general shareholder votes, including the election of directors other than Class T directors. The Long-Term Benefit Trust, through Class T stock, generally has the power to elect and remove the Class T directors, and those directors make up a majority of the board. Economic ownership, shareholder votes and board seats are allocated by different mechanisms. That separation is the system of checks Anthropic is trying to build. Source: amended S-1, November 10
Governance design can change the bar for a decision. It cannot guarantee a good outcome. TIME reported that under its Responsible Scaling Policy, Anthropic committed to holding back releases when a model crosses certain capability thresholds before matching safety measures are ready. The company argues that studying strong models is itself necessary for understanding risk and writing rules. Critics ask how firm these self-imposed limits will stay as revenue, investors and competition grow. Source: TIME on Anthropic’s structure
Left, control: the founders hold the shareholder vote, and the Trust elects most of the board. Right, money: each founder plans to sell about $250 million, roughly 1.4% of Bloomberg’s estimate of their net worth. Figures use the midpoint of the filing’s price range.
This is not an abstract governance debate. The more Claude succeeds, the more compute, talent and customers Anthropic needs. The more of those it has, the more safety research it can afford, and the more it runs into conflicts between growth speed, product launches and risk management. The problem the company was founded to solve turns into a harder one once it is operating at scale.
Wealth from the IPO, and a test
The amended filing of November 10, 2026 shows each of the seven founders planning to sell about 3.38 million shares in the IPO. At $74, the midpoint of the $70–$78 range, that comes to about $250 million each and about $1.75 billion for all seven. Bloomberg, citing the Bloomberg Billionaires Index, estimates each founder’s net worth at about $18 billion. The planned sale is a share count multiplied by an assumed price; net worth is an estimate of everything a person owns. Neither is cash already received, and the seven do not hold identical assets. Sources: amended S-1, Bloomberg
Alongside the wealth is a pledge. When they founded the company, the co-founders pledged to give 80% of their Anthropic equity to charity. The amended filing shows the company has already contributed 5% of its shares to the Anthropic Public Benefit Fund, with Jack Clark and Buddy Shah as its initial managers. In an October announcement, the company estimated that these arrangements, together with an employee equity donation program, would eventually put about 20% of its shares toward the public benefit. The filing also states that the founders’ 80% pledge is voluntary and revocable and does not create legally binding obligations. How fast the money is given and where it goes will depend on future disclosures. Sources: Anthropic Public Benefit Fund announcement, The Chronicle of Philanthropy
The scope of “seven co-founders” is also easy to misread. Anthropic’s public list and Founder LLC include Dario, Daniela, Tom, Jack, Jared, Sam and Chris. Ben Mann played an important role early on and is often called a co-founder, but the filing states that he is not a member of Founder LLC. The “seven” in the news refers to the people in this founder voting structure, not a complete ranking of early contributors. Source: S-1 founder disclosure
Told only through IPO wealth, this story quickly becomes seven new billionaires splitting the take. The better question is whether the organizational problem they left a leading lab to solve has actually been solved. When AI capability and company revenue speed up together, who has the authority to say “wait”? Can internal safety evaluations hold against a market timetable? How will the charitable equity turn into results the public can see?
The seven gave Anthropic an unusual mix: people who had trained large models by hand, people who understood the cost of compute, people who studied the regularities and internals of models, and people who ran the organization and talked to the public. The company they built has shown that a safety story can run alongside commercial competition. The next problem is showing that this is more than an early-stage ideal, and that it still holds once scale, profit and power have all grown fast.