ABAB Pedia

Gabe Pereyra

Gabe Pereyra, who publishes research as Gabriel Pereyra, is an AI researcher and the president and co-founder of Harvey. He earned a bachelor’s degree in computer science at the University of Southern CaliforniaUniversity of Southern CaliforniaThe University of Southern California (USC) is a private research university in Los Angeles, California, planned by a group led by Judge Robert Maclay Widney and opened to students in 1880.Open the full entry, pursued doctoral studies in neuroscience at Oxford and worked at DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry and MetaMetaMeta (Meta Platforms, Inc.) is a publicly listed American technology company, renamed from Facebook in 2021, that operates social communication, digital advertising, virtual reality, smart glasses and artificial intelligence businesses.Open the full entry. In 2022, he and Winston WeinbergWinston WeinbergWinston Weinberg is the CEO and co-founder of Harvey, a legal AI company. He earned a Bachelor of Arts from Kenyon College and a JD from the University of Southern California’s Gould School of Law. He practiced securities and antitrust litigation at O’Melveny & Myers before founding Harvey with Gabe Pereyra in 2022 to apply large language models to legal work and other professional services.Open the full entry founded Harvey to apply large language models to legal and professional services. His subsequent research has included legal AI agents and model post-training.

Contents20 sections
Key facts

Batch normalization in recurrent neural networks

On October 5, 2015, César Laurent, Pereyra, Philémon Brakel, Ying Zhang and Yoshua BengioYoshua BengioYoshua Bengio is a professor of computer science at Université de Montréal, founder of Mila and a deep learning pioneer who shared the 2018 Turing Award with Geoffrey Hinton and Yann LeCun; since 2023 he has turned to AI safety research, chairing the International AI Safety Report and founding the non-profit LawZero in 2025. He appeared on the Lex Fridman Podcast in 2018.Open the full entry posted “Batch Normalized Recurrent Neural Networks” on arXiv. They compared normalization at different connections in recurrent networks. Normalizing the input-to-hidden connections accelerated convergence during training, with little improvement in generalization on the language modeling and speech recognition tasks they tested.⁠[1]

Overconfidence in neural networks

A 2017 ICLR workshop listed “Regularizing Neural Networks by Penalizing Confident Output Distributions,” a paper by Pereyra, George Tucker, Jan Chorowski, Łukasz Kaiser and Geoffrey HintonGeoffrey HintonMachine learning researcher.Open the full entry. They compared a penalty for overly concentrated output probabilities with label smoothing on benchmarks for image classification, language modeling, machine translation and speech recognition.⁠[2]

Online knowledge distillation

On April 9, 2018, Pereyra and co-authors Rohan AnilRohan AnilRohan Anil is a machine learning engineer and a co-founder and director of Core Automation. At Google he worked on reducing the cost of training neural networks, helped develop the Shampoo optimizer and held technical leadership roles on PaLM 2 and Gemini. He subsequently joined Anthropic’s pretraining team.Open the full entry, Alexandre Passos, Robert Ormandi, George E. Dahl and Geoffrey E. Hinton posted a paper on online distillation on arXiv. They investigated whether networks trained on separate data subsets could learn from one another’s predictions, use additional parallel computing power, simplify conventional multistage distillation and make predictions more reproducible.⁠[3]

Founding Harvey with Winston Weinberg

Pereyra and Winston Weinberg founded Harvey in August 2022. Pereyra brought experience in AI research and Weinberg in securities and antitrust litigation. They applied large language models to professional tasks including legal research and due diligence. Pereyra became the company’s president.⁠[4][5][6]

Adoption by professional services firms

Harvey’s third anniversary

In an August 4, 2025, anniversary article, the founders said Harvey served more than 500 clients across 54 countries, had exceeded $100 million in annual recurring revenue and employed about 350 people. They also described workflow tools, additions to the company’s data sources and hiring plans in several markets. The figures were the company’s own at the time.⁠[7]

Interview with No Priors

He recalled meeting David Wakeling at Allen & Overy in Harvey's first year. A small pilot expanded firmwide, more quickly than he expected.⁠[8]

Evaluating legal AI at Harvey

In San Francisco, Gabe Pereyra and Harvey’s head of applied research, Niko GrupenNiko GrupenNiko A. Grupen is a machine learning researcher and head of applied research at legal AI company Harvey. He studied multi-agent behavior and reinforcement learning at Cornell University, worked at Apple and Google Brain and helped develop ways to evaluate AI systems on legal work at Harvey.Open the full entry, told Sourcery about their work releasing the Legal Agent Benchmark. They discussed its use of real legal tasks and the effect of model pricing on customers.⁠[9][10]

Harvey Tenet research preview

On August 20, 2026, Pereyra and colleagues published a research preview of Harvey Tenet, a legal model developed with Fireworks through post-training based on Kimi K3. The team adjusted both the model and the framework in which its agents ran. It explored due diligence, structured document review and searches of law firms’ internal knowledge. The paper described research in progress and plans to incorporate selected capabilities into the product.⁠[11]

Education

Pereyra earned a bachelor’s degree in computer science from the University of Southern California and went on to doctoral studies in neuroscience at Oxford. A Stanford Law SchoolStanford Law SchoolStanford Law School is Stanford University’s law school. Its law curriculum began in 1893, and it offers JD and advanced law degrees. Its alumni include technology-media entrepreneur Michael Arrington.Open the full entry event biography identifies DeepMind as his sponsor. Harvey’s biography describes his subsequent move into industry.⁠[6][4]

Research and engineering before Harvey

Before founding Harvey, Pereyra worked as a research scientist at DeepMind and as a machine learning engineer at Meta. He uses Gabriel Pereyra in scientific publications and Gabe Pereyra in professional profiles.⁠[6][12]

Instruction learning with OPT-IML

Pereyra co-authored the OPT-IML paper released on December 22, 2022. The researchers brought together about 2,000 natural language processing tasks to study how task counts, sampling, examples and training objectives affected generalization after instruction tuning. They trained OPT-IML models with 30 billion and 175 billion parameters.⁠[13]

Series A and product direction

Trust in professional work

In the same announcement, the founders identified scalability, transparency and security as foundations for trust. They described layers of review, natural-language feedback and ways to verify AI output. They proposed that engineers and legal professionals develop systems together to support professional judgment and client service.⁠[14]

Series B and custom models

Agents for merger and acquisition due diligence

On September 8, 2026, Pereyra co-authored a Harvey and Baseten research article on a recursive language-model framework. A lead agent divided the document-review work among other agents, which analyzed documents in their own contexts before their findings were assembled into a due diligence memorandum. The team studied post-training with simulated data rooms and held-out tests. Results still fell short of meeting all review criteria, and an extra layer of delegation did not consistently improve them.⁠[16]

New funding for Harvey

Public profiles and publications

Advice Shared on No Priors

He recommended designing around recurring legal work by letting lawyers upload their own documents, act on them and check accurate citations. (36:32)⁠[8]

He advocated designing for secure document sharing and collaboration across the advisers and clients involved in a complex professional project. (28:46)⁠[8]

Sources17 sources

  1. Batch Normalized Recurrent Neural Networks (opens in a new window)arxiv.org↑
  2. Regularizing Neural Networks by Penalizing Confident Output Distributions (opens in a new window)research.google↑
  3. Large scale distributed neural network training through online distillation (opens in a new window)arxiv.org↑
  4. Gabe Pereyra — CodeX FutureLaw 2026 (opens in a new window)conferences.law.stanford.edu↑
  5. Partnering with Harvey: Putting LLMs to Work (opens in a new window)sequoiacap.com↑
  6. Gabe Pereyra — President & Co-Founder at Harvey (opens in a new window)harvey.ai↑
  7. Harvey’s Three Year Anniversary (opens in a new window)harvey.ai↑
  8. No Priors Ep. 142 | With Harvey Co-Founder and President Gabe Pereyra (opens in a new window)youtube.com↑
  9. Harvey Co-Founder Gabe Pereyra on the Token Pricing Reckoning Coming for AI (opens in a new window)youtube.com↑
  10. Harvey Co-Founder Gabe Pereyra on the Token Pricing Reckoning Coming for AI (opens in a new window)podcasters.spotify.com↑
  11. Harvey Tenet Research Preview (opens in a new window)harvey.ai↑
  12. Partnering with Harvey: Putting LLMs to Work — Gabe Pereyra repost (opens in a new window)linkedin.com↑
  13. OPT-IML: Scaling Language Model Instruction Meta Learning through the Lens of Generalization (opens in a new window)arxiv.org↑
  14. Sequoia and OpenAI Back Harvey to Redefine Professional Services, Starting with Legal (opens in a new window)harvey.ai↑
  15. Harvey raises $80M Series B from Elad Gil, Kleiner Perkins, OpenAI and Sequoia (opens in a new window)harvey.ai↑
  16. Post-Training RLM Agents for End-to-End M&A Diligence (opens in a new window)harvey.ai↑
  17. Harvey Raises $550M at a $15.5B Valuation to Help Legal Teams Own Their Intelligence (opens in a new window)harvey.ai↑