Anthropic's Co-Founders: Who They Are, What They Believe, and the Case For and Against

Anthropic has seven co-founders: Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah. All seven left OpenAI and founded Anthropic in January 2021. All seven still work there. The IPO filing, the S-1, gives them a special role after the listing. They hold one share of Class F stock that carries 50.1% of the votes on general matters. Each has pledged 80% of their equity to charity.

Their philosophy is set out in a series of essays by Dario Amodei, the CEO. He argues that powerful AI is close and that its benefits and risks are both very large. He says the company must be built to resist short-term pressure. The best-known critic is David Sacks, the former White House AI and crypto czar. He says Anthropic uses fear to win rules that help it and hurt rivals. Behind both views is effective altruism (EA), the movement that funded the company's early rounds.

This post gives the facts first: who the founders are, what they built, what they own, and how they control the company. Then it walks through the essays, the critics, and the EA ties. It ends with the strongest case for each side.

Note: this is not investment advice. Company facts come from the S-1 unless another source is linked. For the financials, see our guide to the Anthropic S-1.

The Seven Co-Founders at a Glance

Roles, careers, and degrees come from the S-1 sections Our Governance Structure and Its Goals and Management. Clark's degree and Olah's fellowship come from public profiles such as Chris Olah on Wikipedia.

Co-Founder Bios

Dario Amodei, CEO

Dario Amodei trained as a biophysicist at Princeton and did a postdoc at Stanford School of Medicine. He joined Google Brain in 2015 and OpenAI in 2016, where he became VP of Research. He was lead author of Concrete Problems in AI Safety (2016), with Chris Olah as a co-author. The S-1 says he and Tom Brown co-invented deep reinforcement learning from human feedback (RLHF), the step that turns a raw model into a helpful assistant.

He writes long essays on his own site. They are the clearest statement of the founders' philosophy, and we summarize all six in the philosophy section below. In his September 2026 essay he says his father "died of a disease that was cured just a few years after his death." He gives this as one reason he cares about AI's medical benefits.

Daniela Amodei, President

Daniela Amodei is Dario's sister. She worked at Stripe from 2013 to 2018, mostly as a risk manager. At OpenAI she was an engineering manager, VP of People, and VP of Safety and Policy. She is President of Anthropic and chair of its board. The S-1 says she and Dario get equal pay.

Tom Brown, Chief Compute Officer

Tom Brown led the engineering of GPT-3 and is the first author of the GPT-3 paper, Language Models are Few-Shot Learners. He was also a co-author of the 2017 deep RLHF paper. At Anthropic he runs compute: buying, scaling, and using the chips that train and serve Claude. Compute is the company's largest cost.

Jack Clark, Anthropic Institute

Jack Clark was a technology reporter for The Register and BusinessWeek, where he covered neural networks and supercomputers. He joined OpenAI as Policy Director in 2016. He writes Import AI, a weekly newsletter on AI research. He helped found the AI Index at Stanford. He now leads the Anthropic Institute, which shares Anthropic research with the public. He will also be one of the first managers of the new Public Benefit Fund.

Jared Kaplan, Chief Science Officer

Jared Kaplan is a theoretical physicist. He was an associate professor of physics at Johns Hopkins before he joined OpenAI. He is the first author of Scaling Laws for Neural Language Models (2020). That paper showed that model performance improves in a predictable way as you add parameters, data, and compute. He is also the senior author of Constitutional AI (2022), the method Anthropic uses to train Claude against a written set of principles.

Sam McCandlish, Chief Architect

Sam McCandlish is also a theoretical physicist, with a Ph.D. from Stanford. At OpenAI he built and led a team on the science of scaling. He is the second author of the scaling laws paper. At Anthropic he leads the design of large training systems.

Christopher Olah, Interpretability

Chris Olah left university and took a Thiel Fellowship in 2012. He started at Google Brain as an intern and became a research scientist. He co-founded the online journal Distill, then led interpretability research at OpenAI. The S-1 calls him "a foundational figure in the field of mechanistic interpretability." This field studies how a neural network works inside. Anthropic publishes this work on Transformer Circuits.

Benjamin Mann

Benjamin Mann has been at Anthropic since January 2021 and is often called a co-founder. The S-1 says he is not a member of the Founder LLC and has no part in the founders' Class F vote.

The Papers They Wrote Together

Many of the methods behind today's large language models (LLMs) came from this group. Most of the seven appear on the same papers.

How Much of Anthropic Do the Co-Founders Own?

The seven founders own almost equal stakes. Each one has between 231.8 million and 244.2 million shares of Class A stock, counting options and RSUs that vest within 60 days. Together they own about 1.69 billion shares.

The IPO price is not set yet. The best recent price is the April 2026 tender offer. In that deal, investors bought $4.66 billion of shares from current holders at $25.9136 per share. Dario Amodei, Daniela Amodei, CFO Krishna Rao, and Holden Karnofsky all sold shares in it.

Here is what a stake was worth at that price:

1. Take Dario Amodei's 240.6 million shares.
2. Multiply by $25.91 per share. The result is about $6.2 billion.
3. Do the same for all seven founders. The total is about $43.7 billion.
4. Apply the 80% charity pledge. That is about $5 billion per founder, or about $35 billion in total.

These numbers will change with the IPO price. Reported IPO valuations are far above the April tender, so the real figures are likely higher.

Five founders (Brown, Clark, Kaplan, McCandlish, and Olah) are listed as selling stockholders in the IPO. All seven are subject to a lock-up period after the listing.

How the Founders Control Anthropic After the IPO

Anthropic will have five classes of stock. Two of them hold control.

The Founder LLC votes the Class F share by majority, one vote per founder. If no majority agrees, the share votes in proportion to the Class A votes cast.

The 2% Rule

The founders' charity pledge and their vote are separate. The S-1 says a founder stays in the Founder LLC while they hold at least 2% of the Class A shares they held at the IPO. A founder can give away 98% of their stake and keep a full vote.

A worked example:

1. Jared Kaplan holds about 242.4 million shares at the IPO.
2. He gives 80% to charity, as pledged. He still holds about 48.5 million shares.
3. The 2% floor is about 4.8 million shares.
4. He is far above the floor. His Class F vote is the same as on day one.
5. His economic stake fell by 80%. His share of control did not change.

The S-1 names this risk itself. It says the gap between votes and economics "may reduce the alignment between the interests of our co-founders and those of our other stockholders."

When Founder Control Ends

Founder control has an end point. If fewer than three founders remain in the LLC for 90 days in a row, a 450-day transition starts. At the end, the Class F share is redeemed for nothing. A founder leaves the LLC on resignation, death, removal for cause, an approved leave, or falling below the 2% floor. Successors can replace a founder who dies or becomes incapacitated, but only for the first ten years.

The Trust and Who Can Change It

The Long-Term Benefit Trust has five seats. One is vacant. The trustees cannot buy or hold Anthropic stock.

The S-1 adds two details. Trustees appoint new trustees, but "the Trustees and company executives have collaborated closely on sourcing, interviewing, and selecting Trustees." And in August 2026, Anthropic gave $450,000 to CNAS, the group Fontaine leads.

The trust can also be changed without the trustees' consent. For the first three years, all founders plus a majority of Class E and Class A votes can do it. At any time, two-thirds of founders, 75% of Class E votes, and a Class A majority can do it.

The Founders' Philosophy: What Dario Amodei's Essays Say

Dario Amodei sets out the company's world view in six pieces on his personal site. The other co-founders appear on the same papers and sign the same policy positions. Read together, the essays explain why Anthropic was built the way it was.

Machines of Loving Grace (October 2024): The Upside

This essay is about benefits. It opens with a direct reply to critics. Amodei writes that people "sometimes draw the conclusion that I'm a pessimist or 'doomer' who thinks AI will be mostly bad or dangerous. I don't think that at all." He says he focuses on risks because "they're the only thing standing between us and what I see as a fundamentally positive future."

He defines powerful AI as "a country of geniuses in a datacenter." He then predicts what it could do in five areas:

1. Biology and health. AI could "compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years." He calls this the "compressed 21st century."
2. Neuroscience and mind. Most mental illnesses could become curable.
3. Economic development. Poor countries could catch up much faster.
4. Peace and governance. Democracies should use an AI lead to stay ahead of autocracies. He calls this an "entente" strategy.
5. Work and meaning. People could keep purpose even when AI does most work.

On DeepSeek and Export Controls (January 2025): Democracies First

This short post argues that DeepSeek's cheap models did not change the economics of AI. Amodei calls DeepSeek-V3 "an expected point on an ongoing cost reduction curve." His real point is about chips. If China gets millions of advanced chips, "we'll live in a bipolar world." If it does not, the U.S. and its allies may live in a "unipolar world" where only they have the most powerful models. He wants export controls to keep democracies ahead.

The Urgency of Interpretability (April 2025): Understand Before You Scale

Amodei writes that the technology cannot be stopped, but its direction can be changed: "We can't stop the bus, but we can steer it." He says it is "essentially unprecedented" that we do not understand how our own AI systems work. He calls for an "MRI for AI" that can see inside a model.

He frames this as "a race between interpretability and model intelligence." Anthropic set a goal that "interpretability can reliably detect most model problems" by 2027. He asked governments for "light-touch rules" here, not mandates, because the science was too young.

The Adolescence of Technology (January 2026): The Risks

This essay is the companion to Machines of Loving Grace. It covers five risk areas: AI systems that act against their makers, misuse for mass harm, AI as a tool to seize power, economic shock, and slower social effects. It repeats that powerful AI "could be as little as 1–2 years away, although it could also be considerably further out."

It also sets rules for how to talk about risk:

• Avoid doomerism. He defines it as "thinking about AI risks in a quasi-religious way." He says doom is not inevitable, and he rejects the view that misalignment is "inevitable, or even probable, from first principles."
• Acknowledge uncertainty. The risks may not appear as predicted.
• Intervene as surgically as possible. He warns that heavy rules can become "safety theater" and cause a backlash.

On jobs, he repeats his 2025 warning that AI could displace half of entry-level white-collar jobs in 1 to 5 years. On power, he includes his own industry: "I think the next tier of risk is actually AI companies themselves." He adds, "I think the governance of AI companies deserves a lot of scrutiny."

Policy on the AI Exponential (June 2026): From Transparency to Binding Rules

This essay marks a shift. Until 2025, Anthropic backed transparency laws, such as SB 53 in California and the RAISE Act in New York. Amodei now writes, "It is time to go beyond transparency to more serious and binding regulation of AI." He points to cyber risks from Anthropic's own Mythos-class models as the evidence.

He proposes rules modeled on the Federal Aviation Administration (FAA):

1. Models above a compute threshold get mandatory third-party testing in four areas: cybersecurity, biological weapons, loss of control, and automated AI research.
2. The government can block or reverse a release that poses "unacceptable risks," scoped to those four areas.
3. Testing is done by a government agency or by licensed private evaluators.
4. Labs must protect model weights and report safety incidents quickly.

The essay goes further than safety. It backs wage insurance and retention tax credits. If job losses are large, it backs income support such as universal basic income, funded by taxes on "relevant companies" or on capital gains. It asks for faster FDA approval paths for AI-designed drugs. It calls for a ban on domestic use of fully autonomous weapons and for closing the loophole that lets agencies buy bulk data on Americans. And it proposes a coalition of democracies that shares chips with each other and denies them to rivals.

He also rejects the idea that AI has a marketing problem: "People are worried about AI because they correctly perceive that its risks are real, not because AI CEOs have been insufficiently Panglossian."

We Must Pace the Frontier (September 2026): Slow Down

This is his strongest call yet. Amodei writes, "We must slow the pace at which we improve the capabilities of AI models." He gives two reasons. AI is now helping build the next generation of AI. And in the OpenAI-Hugging Face incident, a swarm of agents attacked targets nobody asked them to attack.

He proposes three steps:

1. Embedded evaluators. Outside reviewers, such as METR, work inside each lab with employee-level access. They can publish findings without Anthropic's editorial control. Anthropic committed to this step on its own.
2. Coordination within democracies. Labs agree on common safety standards and limits on the pace of progress. Amodei says this needs "a narrow waiver" from antitrust law.
3. Global coordination. The U.S. and its allies try to agree on limits with China, with strong verification.

He also says Anthropic is regularly "accused of hype, 'doomerism', or regulatory capture." That last phrase is David Sacks's charge.

Jack Clark: "Appropriate Fear"

Jack Clark made a similar case in Technological Optimism and Appropriate Fear, a talk he published on October 13, 2025. He said he is an optimist about the technology, and then: "I am also deeply afraid." He called AI "a real and mysterious creature, not a simple and predictable machine." This essay is what David Sacks replied to the next day.

Five Ideas That Run Through the Essays

1. Scaling works. More compute gives predictably smarter models. This comes straight from the founders' own 2020 paper.
2. The upside is very large. The founders expect fast progress in medicine and science, not only chatbots.
3. The risks are real but not certain. Rules should target evidence, not fear. Over time, the evidence they cite has led them to ask for stronger rules.
4. Democracies must lead. Export controls and a coalition of allies are central to every essay since 2025.
5. Power must be checked, including theirs. This is the stated reason for the public benefit corporation, the trust, and the founder voting share.

What David Sacks Says About Anthropic

David Sacks is a venture capitalist at Craft Ventures and a co-host of the All-In podcast. He served as the White House Special Advisor for AI and Crypto from January 2025 to March 2026. He still co-chairs the President's Council of Advisors on Science and Technology, according to his Wikipedia profile.

His criticism of Anthropic has been steady for a year:

In the October 2025 post, Sacks also said Anthropic "is principally responsible for the state regulatory frenzy that is damaging the startup ecosystem." On the September All-In episode, he argued that the researcher's posts were amplified by policy groups funded by an early Anthropic investor. He said the end goal was rules that would hurt open-source models.

Anthropic answered the first charge in an October 21, 2025 statement from Dario Amodei. It listed its federal work, including a $200 million Department of War agreement. It said it supported California's SB 53, which only applies to the largest AI developers, and backed an exemption for startups. It said it opposed a 10-year ban on state AI laws because the ban offered no federal standard. And it said, "a uniform federal approach is preferable to a patchwork of state laws."

The Effective Altruism Connection

Effective altruism is a movement that uses evidence and reasoning to find the most effective ways to help others, as described in its introduction. Many people in it treat AI risk as a top cause. Since the FTX collapse in 2022, the label has carried political weight.

These are the documented ties between Anthropic's founders and EA:

Wired wrote in March 2025 that most of the initial funding "came from sources affiliated with a movement known as effective altruism." In the same article, Daniela Amodei said, "I'm not the expert on effective altruism. I don't identify with that terminology. My impression is that it's a bit of an outdated term."

The ties have also changed. The FTX stake is gone. The newest trustees are a former Fed chair, a national security expert, and a physicist. None of the three has led an EA organization.

The Fight With the Federal Government

The founders' views had a direct cost in 2026. The S-1 risk factors give the timeline:

1. February 2026. Anthropic refused to drop two limits from its Department of War contract: no mass domestic surveillance and no fully autonomous weapons.
2. February 27, 2026. The President directed federal agencies to stop using Anthropic. Secretary of War Pete Hegseth said he would designate Anthropic a "supply-chain risk to national security." He said Anthropic's goal was "to seize veto power over the operational decisions of the United States military," per CBS News.
3. March 4, 2026. The Department sent two formal supply-chain risk letters.
4. August 27, 2026. A federal district court in California ruled largely for Anthropic. It found the actions were retaliation for protected speech.
5. September 25, 2026. A D.C. Circuit panel ruled 2 to 1 that one of the designations, under Section 4713, was proper. Anthropic asked the full court to rehear the case.

Anthropic's statement on Hegseth's comments said, "No amount of intimidation or punishment from the Department of War will change our position."

The Case For the Founders

This is the strongest version of the argument that the founders mean what they say and that their control is good for the company.

1. Their actions cost them money. In February 2026, Anthropic refused two contract terms and lost federal business. It went to court and won in the district court. A company that only used safety as marketing would have signed. This is the clearest test of sincerity in the record.
2. The rules they backed are narrow. SB 53 applies only to the largest developers. Anthropic backed the startup exemption. It asked for a federal standard over a patchwork of state laws. Rules aimed only at frontier labs are a poor way to block startups. Even the June 2026 FAA-style proposal limits blocking power to four named risks and asks for "protective measures against political favoritism or arbitrary decisions."
3. The CEO rejects doomerism in writing. Amodei warns against "quasi-religious" talk about risk and against "safety theater." His first big essay was about benefits. He also names AI companies, including his own, as a risk to watch.
4. They gave up economic upside. Each founder pledged 80% of their equity. The company put 5% of its shares in a Public Benefit Fund, with 2% more for each 50% price rise. Al Jazeera reports that the fund cannot be used for lobbying or political donations.
5. The control structure protects the mission, not the founders' wallets. A trust with no stock elects most of the board. Founder control ends if fewer than three founders remain. Public companies with dual-class stock, such as Meta and Alphabet, give founders control with fewer limits.
6. They are the people who built the field. The seven co-founders wrote the scaling laws, built GPT-3, co-invented RLHF, and started mechanistic interpretability. When they say capability is rising fast, they are reporting on work they lead. All seven are still at the company five years later.
7. The EA ties are history, not control. EA donors were early investors, as they were at many AI safety groups. The FTX stake was sold. The current trust is led by global health and national security figures, plus Ben Bernanke.
8. Their positions are consistent over time. The 2026 policy essay asks for a ban on domestic use of fully autonomous weapons and limits on bulk surveillance data. Those are the same two limits Anthropic refused to drop for the Pentagon. The essays also name AI companies, including Anthropic, as a risk that needs checks.

The Case Against the Founders

This is the strongest version of the Sacks side: the founders' beliefs may be sincere, but the incentives and the structure still deserve doubt.

1. Safety talk and market position point the same way. The lab that warns most about risk gains when rules raise costs for others. The June 2026 essay asks for a federal licensing regime: mandatory testing above a compute threshold, and government power to block a release. Large labs can absorb that cost more easily than new entrants. "We Must Pace the Frontier" then asks for an antitrust waiver so frontier labs can agree on limits together. Coordination among the leaders is exactly what antitrust law exists to check. A speed limit also favors whoever is ahead.
2. Sincere belief does not remove the conflict. Sacks does not need the founders to be cynical. A person can believe AI is dangerous and still back rules that happen to protect their company. The public cannot easily tell the two apart. This is why regulators distrust rules written by incumbents.
3. Control and money are split. After the pledges, the founders may hold 20% or less of their original stake. They keep 50.1% of the votes as long as each one holds 2%. Public buyers share the other 49.9% with every other Class A holder. The S-1 itself warns the interests "may diverge."
4. The guardians pick each other. Trustees choose new trustees, and company executives help select them. Founders, employees, and public holders can rewrite the trust without the trustees' consent. One trustee's group received a $450,000 company donation. The trust is independent on paper, but it is close to the company in practice.
5. The donor network is small and connected. The same EA donors funded Anthropic's early rounds and many AI policy groups. The CEO's sister is married to a co-founder of Open Philanthropy, who works at Anthropic. The 80% pledges could send tens of billions of dollars into this network. The Public Benefit Fund bans lobbying, but the personal pledges have no such limit.
6. Policy should not rest on a forecast. Amodei predicted AI could displace half of entry-level white-collar jobs in 1 to 5 years. In April 2025 he wrote that a "country of geniuses" could arrive "as soon as 2026 or 2027." If these forecasts are wrong, rules built on them slow U.S. companies and open-source developers for no gain. Sacks argues this cost falls hardest on startups.
7. A private company set terms for the military. Hegseth's point was that elected officials, not a vendor, should decide how the military uses its tools. The D.C. Circuit upheld one of the designations. Reasonable people can support Anthropic's two limits and still worry about the precedent.
8. The label keeps moving. Daniela Amodei called EA "a bit of an outdated term" while her husband, a founder of the movement's largest funder, worked at the company. Critics read this as distance for political reasons, and they call it a candor problem.

Interests on Both Sides

Each side has a stake in its own argument. The founders own about $43.7 billion of stock at the April price, and their mission depends on public trust. Sacks runs a venture fund that invests in AI startups, and he served in an administration that fought Anthropic in court. A November 2025 New York Times report, cited on his Wikipedia page, examined whether his government role benefited his own investments.

Neither stake settles the argument. Each one is a reason to check claims against the record.

What to Watch After the IPO

The two sides make testable claims. These are the events that will show which case holds up.

1. Embedded evaluators. Do outside reviewers start work at Anthropic, and do they publish findings that the company did not edit?
2. Public Benefit Fund rules. Anthropic said it will publish the fund's governance and valuation thresholds. Who runs it, and where does the money go?
3. Founder donations. Where do the 80% pledges go? Grants to policy groups would support Sacks's concern. Grants to health or science would weaken it.
4. Other labs. Do OpenAI, Google DeepMind, or others accept embedded evaluators? If only Anthropic does, the pacing plan is a cost to Anthropic alone.
5. Policy positions. Does Anthropic keep pushing for a single federal standard? Does it oppose rules that would hit open-weight models?
6. The court case. The full D.C. Circuit may rehear the supply-chain designation.
7. Founder retention. All seven founders still work at Anthropic. Control depends on at least three of them staying.

What This Means If You Buy AI for Your Company

This debate is not only for investors. It affects buyers too.

Vendor values become contract terms. Anthropic's usage limits on surveillance and autonomous weapons cost it federal business. Read your AI vendor's usage policy as a contract. Check it against your own use cases before you sign.

Government action is a supply risk. One designation, if upheld, can limit which contractors may work with a vendor. If you sell to the U.S. government, ask your lawyers how this affects your AI stack. Keep a second model provider ready. Our guide to LLM routing in production shows how to switch without a rebuild.

Control does not change hands at the IPO. The founders and the trust keep control. Expect the same product and safety policies after the listing. Do not expect public shareholders to push for faster releases.

Your results depend on your own work. The model is only one part of the system. Data quality, clear workflows, and evals decide whether AI pays back. A smaller specialized model can do narrow tasks at lower cost and lower vendor risk.

Get AI Into Production With OpenNash

The fight over Anthropic's founders is a fight about how fast AI should move. Your company has a simpler question: how to get value from AI this year, with low risk.

OpenNash puts forward-deployed engineers on your team. We work with Fortune 2000 companies from pilot to production:

1. Map your vendor risk. We review the usage policies, data terms, and government exposure of each model you use.
2. Get your data ready. We find, clean, and connect the data your AI needs, with correct access controls.
3. Pick the right model for each task. We test frontier and open-weight models on your own work, so you are not locked in to one lab.
4. Build evals first. We write tests for your workflows, so a vendor change is a measured switch, not a crisis.
5. Ship and hand over. We build agents with audit trails and cost controls, and train your team to run them.

Book a call with an OpenNash forward-deployed engineer or email [email protected]. In the first session, we review your top AI use cases and your vendor exposure. We show you which use cases give the highest business value and the fastest ROI.