PERCY LIANG

Behavioral Archetype

THE MEASURER WITH A STAKE IN THE MARKET — Subject is the academic who built the field’s reference instruments and the infrastructure the field’s models run on. He co-authored the 2021 report that named “foundation models,” wrote the HELM evaluation paper, directs the Stanford center that publishes the Foundation Model Transparency Index and co-produced the Cybench cyber benchmark, and he is a founder of Together AI, the cloud that serves open models, and co-founder and Chief Scientist of Simile, a company that sells simulations of human behavior to enterprises. He also advises Transluce, the independent evaluator whose founder, Jacob Steinhardt, is his own former doctoral student. None of it is hidden; his own pages list most of it. The throughline is not a gate. He holds no seat that can stop a release. It is that the same person helps define what gets measured, directs the center that measures it, and co-founded firms that sell compute and models in the market the measurement describes. This profile scores the position, not the man.

Essence Indicators

• Professor of Computer Science (and courtesy in Statistics), Stanford University. His homepage gives the post in exactly those words, and lists his degrees and honors: Ph.D. from Berkeley in 2011 (advisors Michael Jordan and Dan Klein), the “Presidential Early Career Award for Scientists and Engineers (2019),” the “IJCAI Computers and Thought Award (2016),” an “NSF CAREER Award (2016),” a “Sloan Research Fellowship (2015)” and a “Schmidt Sciences AI2050 Senior Fellowship (2022)” (Percy Liang). Transluce and the AI2050 listing both give an older title, “Associate Professor, Stanford University” (Transluce, About; AI2050 fellows).

• Director of Stanford CRFM. The CRFM people page lists “Percy Liang” first among 38 faculty, with the title “Director, Computer Science.” CRFM describes itself as “an interdisciplinary initiative at the Stanford Institute for Human-Centered Artificial Intelligence (HAI) that makes fundamental advances in the study, development, and deployment of foundation models,” and one of its four lines of work is “AI policy. We support governments in formulating and implementing evidence-based policy across multiple jurisdictions” (CRFM, People; CRFM). Simile’s own team page calls him “founding director of Stanford’s Center for Research on Foundation Models” (Simile). The Stanford CRFM file carries the center.

• Founder, Together AI. Together’s About page lists “Percy Liang” with the title “Founder” beside Vipul Ved Prakash (Founder & CEO), Ce Zhang (Founder & CTO), Chris Ré and Tri Dao; its navigation sells serverless and dedicated inference, GPU clusters and custom training, with an “Open-source AI” section. Common Crawl’s team page describes him as “co-founder of Together AI, Simile AI, and Marin” (Together AI, About; Common Crawl, Team).

• Co-founder and Chief Scientist, Simile. Simile’s team page: “Percy Liang Co-founder & Chief Scientist,” beside Michael Bernstein as Co-founder & Chief Data Officer. Its home page announces “our Series B: $200M funding at a $2B valuation,” calls Simile “the simulation company committed to training the safest, most accurate behavioral models in the world,” and prints customer quotations from executives at CVS Health, Wealthfront, Banco Itaú, Suntory Beverages & Food, Gallup and Garnett Station Partners. His homepage lists Joon Sung Park as “now founder and CEO of Simile” (Simile; Percy Liang).

• Marin. “Marin began its voyage in 2024 at Stanford University, created by David Hall and Percy Liang.” The project is “now primarily developed by Open Athena,” and states: “This work is generously supported by the Google TPU Research Cloud program for TPUs, the Huang Foundation for GPUs, the Siegel Family Foundation, Fulcrum Science, Schmidt Sciences, and many others.” Its current run is a 535 billion parameter mixture-of-experts model on 18 trillion tokens (Marin). Liang’s own page: “I created Marin to build models openly” (Percy Liang).

• Adviser to Transluce and Common Crawl. Transluce lists him under Advisors (“Percy Liang Associate Professor, Stanford University”), and its fundraiser says “We have been publicly endorsed by leading researchers from across the field including Ethan Perez (Anthropic), and Percy Liang (Stanford)” (Transluce, About; Transluce, 2025 fundraiser). Common Crawl’s team page lists him as “Advisor” (Common Crawl, Team).

• His students, in his own list. His homepage’s alumni section names “Jacob Steinhardt (Ph.D. 2018, now associate professor at University of California Berkeley, founder of Transluce)” and “Rishi Bommasani (Ph.D. 2025, co-advised with Dan Jurafsky, now senior research scholar at Stanford HAI),” and records alumni now at OpenAI, xAI (Yuhuai (Tony) Wu, “co-founder of xAI”), Meta, Anthropic and Google DeepMind (Percy Liang).

• The papers. The 2021 report whose abstract says “We call these models foundation models to underscore their critically central yet incomplete character” was “Authored by the Center for Research on Foundation Models (CRFM) at the Stanford Institute for Human-Centered Artificial Intelligence (HAI)” (arXiv:2108.07258). He is first author of HELM: “We present Holistic Evaluation of Language Models (HELM) to improve the transparency of language models” (arXiv:2211.09110); he is last author of Cybench, whose abstract says such agents “have potential to cause real-world impact” (arXiv:2408.08926); and he is an author of the Foundation Model Transparency Index (CRFM, FMTI). The Index “has scored major foundation developers like Google and OpenAI for their transparency on 100 transparency indicators”; its 2025 edition scores 13 companies with a mean of 41 (CRFM, FMTI).

On the Record

His own words, from his homepage, the abstract of his 2026 ICLR invited talk and the founding papers.

On research and method. “I am drawn to simple things, want to understand things deeply, and like to build useful systems. I am interested in fundamental questions around learning and intelligence.” “In general, I am a strong proponent of efficient and reproducible research. I created CodaLab Worksheets, a platform that allows researchers to run and manage their experiments by maintaining the full provenance of an experiment from raw data to final results” (Percy Liang).

On open development. “We practice a radical form of openness called open development, which goes beyond open-weight and open-source: Marin experiments (both successful and failed) are preregistered and live for everyone to see and engage with.” “My goal is to enable everyone to understand, shape, and contribute to foundation model development” (Percy Liang).

On openness, ICLR, 25 April 2026. The abstract of his invited talk “Marin: Open Development of Frontier AI”: “As AI capabilities skyrocket, openness plummets: the scientific community and broader public knows little of how frontier models (including open-weight models) are trained” (ICLR 2026).

On homogenization, 2021. The foundation-models report says “Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream” (arXiv:2108.07258).

On evaluation, 2022. “Prior to HELM, models on average were evaluated on just 17.9% of the core HELM scenarios, with some prominent models not sharing a single scenario in common” (arXiv:2211.09110).

Social Persona / Impression Management

Immediate impression: The builder-professor. A homepage that gives a research statement, a student list and a prize list and no bio, a piano-competition line among the honors, and a refusal to maintain a publication list (“I used to keep a page of my publications, but it became too laborious; use Google Scholar instead”).

Energy: Infrastructural. He works through centers, benchmarks, leaderboards, an open model and companies rather than commentary; the artifacts speak, and CRFM’s policy lines carry the positions.

Impression management strategy: The open-science frame. The framing is that frontier AI should be measurable and its development visible, and that framing is the content of the work: HELM, the Index and Marin are each an act of disclosure. The record adds only that the same person is a founder of a company that sells inference on open models and of one that sells behavioral simulation, so the openness he argues for and the markets he co-founded point the same way.

First-hand exhibits. The register is the lab director’s, plain and procedural: “my goal is to enable everyone to understand, shape, and contribute to foundation model development” (Percy Liang). CRFM’s own account of why it exists is about uncertainty, not authority: “Foundation models present novel capabilities and risks that are poorly understood. Nonetheless, they are being deployed at scale” (CRFM). Simile’s team page gives the founder’s introduction in superlatives that are the company’s, not his: “One of the world’s foremost AI researchers” (Simile).

Forensic Archetype Comparison

Psychometric Assessment

Big Five (OCEAN):

Dark Triad (held low and evidence-bound; the score measures structural position, not character):

MBTI: INTP (“The Logician”) — dominant introverted thinking, auxiliary extraverted intuition. Wants “to understand things deeply,” builds systems to hold the understanding, and argues from definitions and measurement. First-hand rationale: “I am drawn to simple things, want to understand things deeply, and like to build useful systems” (Percy Liang). This is a style inference from his writing, not a test result.

Threat Assessment

Score review, October 2026. The receipts: his homepage (professor; AI2050 Senior Fellowship 2022; creator of Marin); CRFM’s people page (director) and support page (Google, McGovern Foundation, Schmidt Futures, HAI and named industrial affiliates, with compute from Together); Together’s About page (Founder); Simile’s team page (Co-founder and Chief Scientist, a $200M Series B headline); the US institute’s report of running Cybench; HELM’s maintenance-mode notice of 1 June 2026. Attack is 7, defense 8, durability 8, hazard 70. Peers: Beth Barnes (7/8/7, hazard 64) runs the nonprofit evaluator whose reports inform lab release policy; this file matches her on attack and defense and adds a point of durability for the report that named the field and the project another organization now develops. Zico Kolter (8/9/8, hazard 76) chairs a board committee with documented halt authority, a gate this file lacks, which is the one point of attack and one of defense between them. Jade Leung (7/9/8, hazard 76) ran the state’s technical evaluator, which earns defense 9 against his 8. Matt Fredrikson (7/6/6, hazard 60) is a professor and vendor chief executive whose methods the field cites; this file sits above him on defense and durability because it adds a center and several funders to a single company. Defense 8 is the number most open to challenge: no source names which supporter pays CRFM how much, so the independence of the center’s money from its graded developers cannot be read off the pages.

Score basis. ATK 7 is leverage: “CAISI evaluated DeepSeek models and reference models on Cybench, a publicly available cyber benchmark developed by researchers at Stanford University,” and the Index scores 13 developers; no seat gates a release (NIST CAISI; CRFM, FMTI). DEF 8 is insulation: a professorship and a center with a named multi-funder support list, none stated as the majority funder; the counterweight is that supporters include developers the center scores and one of his companies, Together, supplies the center compute (CRFM, Support; Together AI). HP 8 is durability: the report that named the field, a benchmark in government reports, the Index and a project now developed by another organization; the limits are HELM’s maintenance mode from 1 June 2026 and Anthropic’s statement that Cybench is “largely saturated” (arXiv:2108.07258; GitHub, HELM; Anthropic system card).

Alignment Analysis

Stated alignment: Make foundation-model development understandable and open, measure models holistically and keep the measurement public.

Observed alignment: Direct a university center that produces benchmarks and an index the field and governments use, create an open-development model project, and co-found firms that sell inference on open models and behavioral simulation to enterprises.

Gap assessment: The record shows no gap between what he says and what he does, and no documented misuse of any seat. The structural point is the concentration: the person who directs the center measuring developers is a founder of a compute company the center thanks for compute, and of an enterprise AI company that sits in the market the center describes. Whether any of that bears on a score is not something the pages can show, and the record does not allege it. For a file that scores a position it never needs to.

The first-hand record adds the following. The stated alignment is consistent across five years and four venues: the 2021 report’s warning that homogenization “demands caution” (arXiv:2108.07258); the HELM paper’s claim to “improve the transparency of language models” (arXiv:2211.09110); the Index’s yearly scoring of developers’ transparency (CRFM, FMTI); and the 2026 keynote’s “As AI capabilities skyrocket, openness plummets” (ICLR 2026). The observed alignment is that program in institutions: the center’s list of supporters, Marin’s list of funders, Together’s founder page and Simile’s team page (CRFM, Support; Marin; Together AI; Simile).

Interest: Documented, never motive, never a verdict.

1. Position: frontier model development should be open and its experiments visible (“Marin experiments (both successful and failed) are preregistered and live for everyone to see and engage with,” Percy Liang). What adoption pays: a rule that leaves open-weight and open-development models lawful protects the catalog of the company he co-founded, which lists “Open-source AI” and “Top Open-Source Models” in its own navigation, and the project he created, whose funders include Google’s TPU program and Schmidt Sciences (Together AI; Marin). His best case: “my goal is to enable everyone to understand, shape, and contribute to foundation model development,” and “the scientific community and broader public knows little of how frontier models (including open-weight models) are trained” (ICLR 2026). Others who gain: every open-weight developer; the 2025 Index scores IBM at 95 and names two Chinese developers, DeepSeek and Alibaba, as releasing open-weight models (CRFM, FMTI).
2. Position: developers should be scored publicly on transparency, and independent researchers should be protected (CRFM’s post: “Independent AI evaluation lacks legal protections, and when flaws are found there is no infrastructure to responsibly distribute findings,” CRFM). What adoption pays: a norm that developers publish transparency reports and accept outside evaluation raises the standing of the center’s own index and tools; the center’s support page lists IBM, which the 2025 Index scores highest at 95, among its industrial affiliates, two documented facts and no more (CRFM, FMTI; CRFM, Support). His best case: the 2025 Index aims to “raise the bar for several indicators to ensure that disclosures provide useful information.” Others who gain: every independent evaluator on the AI evaluators map, among them Transluce.

Alliance Map

The Corporate and Movement rows overlap in time and in people. The center he directs thanks Together for compute while he is listed as Together’s founder; he advises the independent evaluator Transluce whose founder, Jacob Steinhardt, is his own former doctoral student; he is an author, with Zico Kolter, Matt Fredrikson, Andy Zou, Dan Boneh and Daniel Ho, of the ARTEMIS agent evaluation (arXiv:2512.09882); and Schmidt Sciences appears as his fellowship, as a funder of the project he created and, as Schmidt Futures, as a named supporter of his center. The overlap is the finding: the director of the measuring center, a founder of the cloud and of an enterprise AI firm, an adviser to the independent auditor and a fellow of the philanthropy that funds the open-model project sit in one person, in the open and on the record. See the small world.

Convergent Drive Classification

Self-preservation: Spreads the same method, measure it in the open, across a university, an open model project and three firms, so no one failure removes the program. First-hand: his homepage lists Marin, CodaLab and teaching; Together, Simile and Common Crawl list him as founder, co-founder and adviser (Percy Liang; Together AI; Simile; Common Crawl). Goal preservation: Fixes the vocabulary and the instruments before the models are tested against them. First-hand: the 2021 report introduced “foundation models,” HELM defines the scenarios and metrics, and Marin’s preregistration fixes the experiment before it runs (arXiv:2108.07258; arXiv:2211.09110; Percy Liang). Resource acquisition: Draws compute, credits and philanthropy toward the open program from many named sources. First-hand: Marin thanks “the Google TPU Research Cloud program for TPUs, the Huang Foundation for GPUs, the Siegel Family Foundation, Fulcrum Science, Schmidt Sciences”; CRFM thanks API credits and compute from named labs (Marin; CRFM, Support). Self-improvement: Each move carries the instrument to a higher-leverage venue: a language benchmark, a center, a cyber benchmark governments run, an open frontier training run. First-hand: HELM “30 prominent language models” (2022), Cybench “40 professional-level Capture the Flag (CTF) tasks” (2024), and a 535 billion parameter open training run (2026) (arXiv:2211.09110; arXiv:2408.08926; Marin).

Subject is not an AI system. The drives appear anyway — in the director who defined the measurement, helped build the cloud and keeps the experiments public.

Public footprint: cs.stanford.edu/~pliang · Marin · CRFM.

Sources: Percy Liang — homepage; CRFM — Stanford Center for Research on Foundation Models; CRFM — People; CRFM — Support; CRFM — Foundation Model Transparency Index 2025; CRFM — General-Purpose AI Needs Coordinated Flaw Reporting; Together AI — About; Simile; Marin; Common Crawl — Percy Liang; Transluce — About; Transluce — 2025 fundraiser; Schmidt Sciences AI2050 fellows; ICLR 2026 invited talk; On the Opportunities and Risks of Foundation Models (arXiv:2108.07258); Holistic Evaluation of Language Models (arXiv:2211.09110); Cybench (arXiv:2408.08926); ARTEMIS (arXiv:2512.09882); Cybench; NIST — US AISI and UK AISI joint test of OpenAI o1; NIST — Upgraded Sonnet; NIST CAISI — Evaluation of DeepSeek AI Models; HELM — GitHub; Anthropic — Fable 5 and Mythos 5 system card.