ABAB Pedia

Tom Brown

Tom Brown is a co-founder and the Chief Compute Officer of Anthropic. He previously led research engineering for GPT-3 at OpenAIOpenAIOpenAI is an artificial intelligence research and product organization founded in 2015 that develops the GPT model series, ChatGPT, Codex and a developer API, and conducts research on image, speech, video and intelligent agents. After its 2025 restructuring, the nonprofit OpenAI Foundation continues to control the for-profit OpenAI Group PBC. Its growth has involved large-scale fundraising and compute partnerships, and has also been accompanied by controversies over governance, copyright, privacy and AI safety.Open the full entry and has worked on reinforcement learning from human feedback and large language model research.

Contents25 sections
Key facts

Studies at the Massachusetts Institute of Technology

MIT lists him as a member of the Class of 2010.⁠[2]

Starting Out in Startup Engineering

In the summer of 2009, while still a student at MIT, Brown became the first employee of LinkedLanguage; after returning to school, he joined the mobile advertising company MoPub.⁠[3]

From February to November 2011, he was a founding engineer at MoPub.⁠[4]

Founding SolidStage and Grouper

He then ran Grouper together with Michael Waxman, turning online matchmaking into social outings in which three friends met another group of three friends in person.⁠[6]

Turning to Artificial Intelligence Research

In 2015, Brown decided to move from internet startups to AI and joined South Park Commons. In his own recollection, the community’s machine learning discussions and peer support helped him overcome his worries about starting over; after about six months of intensive study, he joined OpenAI.⁠[7]

He took on short-term contract work for Twitch to cover his living expenses during his subsequent self-study.⁠[3]

First Stint at OpenAI

In May 2016, Brown joined OpenAI, where he worked until May 2017.⁠[4]

Human Preferences and Reinforcement Learning

Brown co-authored “Deep reinforcement learning from human preferences” with Paul Christiano, Jan Leike, Dario AmodeiDario AmodeiDario Amodei is an artificial intelligence researcher and the co-founder and chief executive officer of Anthropic. Trained in physics and biophysics, he conducted research at Baidu, Google Brain, and OpenAI, where he led research including work on GPT-2 and GPT-3. His career spans large-model development, AI safety, interpretability, and public policy.Open the full entry, and others. The research had humans compare short clips, then learned a reward function and trained agents, reducing the need for direct human feedback in games and simulated locomotion tasks; the paper also discussed limitations such as reward exploitation and mismatches in training feedback.⁠[8]

Google Brain and Adversarial Examples

“Adversarial Patch,” on which he was first author, demonstrated adversarial patterns that can be printed and placed in real-world scenes: even with changes in position, angle, and scale, the pattern could still induce an image classifier to output an incorrect label chosen by the attacker. The work revealed the vulnerability of vision systems to interference in the physical world.⁠[9]

Return to OpenAI

In December 2018, Brown returned to OpenAI as research engineering lead for GPT-3, serving until December 2020.⁠[4]

Fine-Tuning GPT-2 from Human Preferences

Brown took part in research on fine-tuning GPT-2 using human preferences. The team applied rewards derived from human comparisons to text style and summarization tasks, and also found that the labeling approach could change model behavior: labels favoring copying from the original text could train a system that is good at extracting passages rather than an ideal summarizer. The research emphasized the importance of feedback quality and ongoing checks.⁠[10]

Scaling Laws for Language Models

Brown is a co-author of “Scaling Laws for Neural Language Models.” Within its experimental range, the study described power-law relationships between model parameters, data size, and compute on the one hand and predicted loss on the other, providing an empirical basis for allocating training resources; these relationships depend on conditions such as other resources not becoming bottlenecks.⁠[11]

GPT-3 and Few-Shot Learning

Brown is listed first on “Language Models are Few-Shot Learners” and is marked as an equal contributor alongside several other authors. The paper studied GPT-3, with 175 billion parameters, which completes a variety of tasks from task descriptions and examples in the prompt without updating model weights for each task. It also documented limitations including repetition, incoherence, factual errors, bias, and inference cost.⁠[12]

Co-Founding Anthropic

In May, the company announced that it had raised $124 million, setting reliable, interpretable, and steerable general AI systems as its research direction.⁠[13]

Brown later became Chief Compute Officer, leading the organization that acquires, scales, and efficiently uses compute to support frontier model development, deployment, experimentation, and safety research.⁠[1]

Research on a Helpful and Harmless Assistant

Research that Brown took part in applied reinforcement learning from human feedback to improve an assistant’s helpfulness and harmlessness, combining preference models, red teaming, and iterated training. The work showed that behavioral alignment and some capability gains can occur together, but that trade-offs remain between helpfulness and refusing harmful requests.⁠[14]

Constitutional AI

Brown is a co-author of “Constitutional AI: Harmlessness from AI Feedback.” The method has a model critique and revise its responses according to a set of principles, then trains the model on AI-generated preference feedback, reducing reliance on item-by-item human harmlessness labels; humans remain involved in choosing the principles.⁠[15]

Partnership with AWS on Trainium

As Anthropic’s compute needs grew, the company made AWSAmazon Web ServicesAmazon Web Services (AWS) is the cloud computing business of Amazon. Launched in July 2002, it made its public debut with the Amazon S3 storage service on March 14, 2006, and has since added hundreds of services, including Amazon EC2; its segment net sales were about $46.0 billion in the third quarter of 2026.Open the full entry its primary cloud and training partner, and the two collaborated on optimizing Trainium chips and software, including low-level kernels and the Neuron compiler tooling. This became part of the compute system for which Brown is responsible.⁠[16]

Compute Supply and Multiple Hardware Paths

In a Y Combinator interview, Brown said that different workloads and supply conditions suit different chips; using GPUs, TPUs, and Trainium together expands available compute but also increases engineering complexity. He also described power as a constraint on expansion and expressed support for nuclear and renewable energy.⁠[3]

Compute Agreement with SpaceX

AMD Partnership and Future Deployment Plans

AMDAMDAMD (Advanced Micro Devices, Inc.) is a publicly traded U.S. semiconductor company founded in 1969 and headquartered in Santa Clara, California; its stock trades on Nasdaq under the ticker AMD. The company develops CPUs, GPUs, FPGAs, adaptive systems-on-chip and related software, mainly for the data center, personal computer, gaming and embedded markets; its product brands include Ryzen, EPYC, Radeon and Instinct.Open the full entry and Anthropic announced a partnership of up to 2 gigawatts, with the first 1-gigawatt MI450/Helios deployment planned to begin in the first half of 2027. In the announcement, Brown emphasized expanding compute supply and choosing hardware by workload; the two companies also plan to build on their existing experience with MI355X in further collaboration.⁠[18]

July–September 2026: Internal Records in Lyrics Copyright Litigation

An amended complaint filed against Anthropic by music publishers on July 22 states that on February 5, 2023, Brown asked a not-yet-released Claude to output Bob Dylan lyrics, and uses this to support claims of infringement and internal knowledge. The filing names Anthropic as the defendant, and these statements are allegations by the plaintiffs.⁠[19]

Anthropic maintains that model training is fair use.⁠[20]

As of September 16, the earlier lyrics case remained at the stage of the parties’ cross-motions for summary judgment.⁠[21]

G20 Innovation Ministerial

On September 2, the White House published the outcomes of the G20 Innovation Ministerial, listing Brown and U.S. Secretary of Commerce Howard LutnickHoward LutnickHoward William Lutnick is an American financial executive who was sworn in as the 41st U.S. secretary of commerce in February 2025. He spent more than four decades at Cantor Fitzgerald and led its rebuilding after the September 11 attacks.Open the full entry as participants in a fireside chat. The meeting focused on AI, innovation, and cross-border technology cooperation.⁠[22]

Dispute over OpenAI’s Book Training Data

In the copyright litigation against OpenAI, a statement of facts filed by the author plaintiffs on September 17 cites internal records to assert that Brown supported expanding LibGen nonfiction book data and using it for training, and took part in discussions about the questionable nature of that data source. This is the plaintiffs’ account of his conduct while at OpenAI and is not a court finding.⁠[23]

In its filing in the same case, OpenAI argues that pre-training is a transformative use aimed at learning the statistical patterns of language and has caused no cognizable harm to the market for the books concerned.⁠[24]

On September 21, the Authors Guild said that written briefing would continue, with a hearing expected in early 2027.⁠[25]

Views on Trust, Safety, and Innovation

Ownership and governance in the amended IPO filing

In its second amended S-1/A filed with the SECU.S. Securities and Exchange CommissionThe U.S. Securities and Exchange Commission is the federal securities regulator of the United States. Created during the Great Depression by the Securities Exchange Act of 1934 and led by five commissioners, it has a three-part mission of protecting investors, maintaining fair, orderly and efficient markets, and facilitating capital formation.Open the full entry on November 10, 2026, Anthropic listed Tom Brown with beneficial ownership of 244,173,040 Class A shares before the offering, using September 30, 2026 as the ownership reference date. The filing proposed selling 3,378,378 shares in the IPO, leaving 240,794,662 shares after the offering, assuming no exercise of the underwriters’ additional-share option. These were proposed transactions in a registration statement that was not yet effective.⁠[27]

The footnote broke that ownership figure into 241,348,140 shares held of record by him or trusts over which he exercises investment or voting control, 2,185,400 shares underlying options exercisable within 60 days of September 30, 2026, and 639,500 shares issuable upon settlement of RSUs whose service-based vesting condition would be satisfied within that period. The total therefore combined record holdings with equity awards included under the filing’s beneficial-ownership rules.⁠[27]

The filing also named Tom Brown among the seven initial members of Founder LLC. Immediately before completion of the offering, the company planned to issue that collective entity one Class F share without economic rights, carrying 50.1% of voting power on matters submitted to stockholders generally. The Long-Term Benefit Trust would separately retain the right to elect a board majority through Class T shares. Each Continuing Founder would have one vote within Founder LLC; its collective voting power was not an individual founder’s holding.⁠[27]

The prospectus also disclosed each co-founder’s pledge to devote 80% of their personal Anthropic equity to philanthropy. The pledges were voluntary and revocable, with no binding legal obligation or fixed timing, amount, recipients or structure. They concerned personal economic interests and did not change the Founder LLC’s Class F voting arrangement.⁠[27]

Related Organizations, Websites, and Public Accounts

Massachusetts Institute of Technology: Official website: https://www.mit.edu/ (opens in a new window).⁠[28]

South Park Commons: Official website: https://www.southparkcommons.com/ (opens in a new window).⁠[30]

Google Research: Official website: https://research.google/ (opens in a new window).⁠[32]

Sources33 sources

  1. Leadership at Anthropic (opens in a new window)anthropic.com↑
  2. Tom Brown ’10 — MIT News / Forbes press clip (opens in a new window)news.mit.edu↑
  3. Anthropic Co-Founder: Building Claude Code, Lessons From GPT-3 & LLM System Design (opens in a new window)ycombinator.com↑
  4. Tom Brown — LinkedIn (opens in a new window)linkedin.com↑
  5. SolidStage (opens in a new window)ycombinator.com↑
  6. Grouper (opens in a new window)ycombinator.com↑
  7. How I made the switch to AI research (opens in a new window)blog.southparkcommons.com↑
  8. Deep reinforcement learning from human preferences (opens in a new window)arxiv.org↑
  9. Adversarial Patch (opens in a new window)research.google↑
  10. Fine-tuning GPT-2 from human preferences (opens in a new window)openai.com↑
  11. Scaling Laws for Neural Language Models (opens in a new window)arxiv.org↑
  12. Language Models are Few-Shot Learners (opens in a new window)arxiv.org↑
  13. Anthropic raises $124 million to build more reliable, general AI systems (opens in a new window)anthropic.com↑
  14. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback (opens in a new window)anthropic.com↑
  15. Constitutional AI: Harmlessness from AI Feedback (opens in a new window)arxiv.org↑
  16. Anthropic and AWS deepen their work together (opens in a new window)anthropic.com↑
  17. Higher usage limits for Claude and a compute deal with SpaceX (opens in a new window)anthropic.com↑
  18. AMD and Anthropic Announce Strategic Partnership to Deploy up to 2 Gigawatts of AMD Instinct MI450 Series GPUs (opens in a new window)newsroom.amd.com↑
  19. Concord Music Group, Inc. et al. v. Anthropic PBC — Second Amended Complaint (opens in a new window)musicbusinessworldwide.com↑
  20. Music publishers file amended lyrics lawsuit against Anthropic (opens in a new window)musicbusinessworldwide.com↑
  21. Anthropic bought songbooks, scanned them, and destroyed them, music publishers allege. Now they want to know which songs were in them. (opens in a new window)musicbusinessworldwide.com↑
  22. G20 Innovation Ministerial Concludes with Consensus Statement (opens in a new window)whitehouse.gov↑
  23. Class Plaintiffs’ Corrected Rule 56.1 Statement of Undisputed Material Facts (opens in a new window)authorsguild.org↑
  24. OpenAI’s Memorandum of Points and Authorities in Support of Motion for Summary Judgment (opens in a new window)storage.courtlistener.com↑
  25. OpenAI Top Executives Knew Mass Book Piracy Was Illegal, New Filings Show (opens in a new window)authorsguild.org↑
  26. Tom Brown on trust, safety and innovation (opens in a new window)x.com↑
  27. Anthropic, PBC — Amendment No. 2 to Form S-1 (opens in a new window)sec.gov↑
  28. MIT - Massachusetts Institute of Technology (opens in a new window)mit.edu↑
  29. Y Combinator (opens in a new window)ycombinator.com↑
  30. South Park Commons (opens in a new window)southparkcommons.com↑
  31. OpenAI (opens in a new window)openai.com↑
  32. Google Research (opens in a new window)research.google↑
  33. Tom Brown (@NotTomBrown) (opens in a new window)x.com↑