Epoch AI

Theory of Change

Epoch AI positions itself as the empirical foundation for AI policy and strategy decisions. Jaime Sevilla draws an explicit parallel to Nobel laureate William Nordhaus's climate science work: "We want to do something similar for artificial intelligence to what William Nordhaus did for climate change. He set the basis for rigorous study and thoughtful action guided by evidence."

The causal chain: collect data on AI trends -> analyze and publish findings -> inform policymakers, researchers, and the public -> better decisions about AI. The org is deliberately neutral on whether AI progress is good or bad. Sevilla (June 2025): "Our staff, and the AI community more broadly, is split on whether advancing AI will ultimately benefit society. As an organization, we are decidedly neutral on this question."

The founding mission (2022) was more directly safety-oriented: "clarify when and how transformative AI capabilities will be developed," working in close collaboration with Open Philanthropy and Rethink Priorities. The current framing emphasizes broad societal benefit rather than existential risk reduction specifically. All funding from CG/Open Phil is classified under "potential risks from advanced artificial intelligence."

What They Do

Core data products. The compute trends database tracks training compute for 120+ significant ML models since 1950 -- widely acknowledged as the most comprehensive such dataset. The founding paper "Compute Trends Across Three Eras of Machine Learning" (Feb 2022) corrected an earlier OpenAI estimate, finding compute doubles every ~6 months rather than 3.4 months. Additional databases track ML hardware, GPU clusters, frontier data centers, and AI companies.

Key research. "Will We Run Out of Data?" (2022, updated 2024) estimated text data exhaustion between 2026-2032, identifying a bottleneck years before it became mainstream. "Can AI Scaling Continue Through 2030?" (commissioned by Google DeepMind) concluded bottlenecks are surmountable. "Three Challenges Facing Compute-Based AI Policies" (co-authored with Google DeepMind) argued that compute as a regulatory proxy is breaking down. The Epoch Capabilities Index (ECI) identified an acceleration in AI capabilities around April 2024. The GATE macroeconomic model analyzes conditions for explosive economic growth from AI automation.

Benchmarking. FrontierMath: 300 research-level math problems evaluating frontier AI reasoning. Currently the most prominent private math benchmark, though structurally compromised by OpenAI's exclusive ownership and access (see Money section). Tier 4 completed in 2025. Also developing a software engineering benchmark with METR and a benchmark on open math problems with Schmidt Sciences.

Policy engagement. Briefings to U.S. Congress, UK AI directorate, Capitol Hill. Consultations with EU AI Office, UK DSIT. Epoch data informed compute thresholds in the EU AI Act and Biden Executive Order.

2025 output. 38 data insights, 14 reports, 40 newsletter issues, 5 podcasts; 987K website users; $5M spent, 21 staff.

Key People

Jaime Sevilla, Director. Spanish (~30), math + CS background, paused Aberdeen PhD to found Epoch. Candid about uncertainty, intellectually honest. His quip after co-founder launched a capabilities startup: "Yay just what I wanted for my bday: a comms crisis."

Notable departures form a striking pattern. Tamay Besiroglu (co-founder, associate director) left April 2025 to found Mechanize, a startup aimed at "full automation of the economy," backed by Nat Friedman, Patrick Collison, Jeff Dean. Took researchers Ege Erdil and Matthew Barnett. Marius Hobbhahn (co-founder) left April 2023 to found Apollo Research (AI safety evaluations). Lennart Heim (co-founder) left for RAND/GovAI (compute governance). Elliot Glazer (lead mathematician) left to join Principia Labs. Four of seven co-founders have departed, three to capabilities-adjacent or commercial work.

Team. 21 full-time staff as of 2025, remote-first, globally distributed.

Money and Incentives

Total known funding: ~$23.8M (2022-2025). Roughly 91% from a single source:

Source	Amount	% of Known Total
Coefficient Giving / Open Phil	$21,773,611	~91%
Jaan Tallinn	$600,000	~2.5%
Likith Govindaiah	$400,000	~1.7%
Leopold Aschenbrenner (via Manifund)	$200,000	~0.8%
Sentinel Bio	$165,000	~0.7%
Carl Shulman	$100,000	~0.4%
Paid engagements (est.)	Unknown	Unknown

Annual spend. ~$5M in 2025 for 21 staff. Fundraising target: $3M more, could deploy $10M.

Business model. Primarily philanthropic grants (~80%+ from CG/OP), supplemented by paid consultations with AI labs, governments, and investors. Clients include OpenAI, Google DeepMind, xAI, METR, UK DSIT, Sequoia Capital, Bridgewater, EU AI Office, Anthropic (small pilot).

FrontierMath financial structure. OpenAI commissioned and funded 300 benchmark problems. OpenAI retains ownership and has access to all problem statements and solutions (except 50-problem holdout where only statements are shared). Epoch cannot share problems with other AI companies without OpenAI's written permission. Only a "verbal agreement" that OpenAI won't train on the data.

Semiconductor stock investments. Epoch invests reserves in semiconductor/AI stocks. They frame this as a hedge -- "any gains from such investments increase our capacity to advance our mission in scenarios when our work matters most." This creates a financial interest in AI sector growth for an organization studying AI trajectory.

Board-funder overlap. Two of four board members (Tom Davidson, Ajeya Cotra) are from Coefficient Giving/Open Philanthropy, the source of 91% of funding. No independent board members exist -- the other two are the director (Sevilla) and the CFO/Secretary/Head of Ops (de la Lama, who holds board + executive roles simultaneously).

Key incentive questions. (1) Does extreme funder concentration constrain what Epoch can research or publish? (2) Does revenue from AI labs create pressure to avoid conclusions that displease lab clients? (3) Do semiconductor stock investments create pressure toward bullish AI progress findings? (4) Could Epoch survive if CG/OP withdrew support? There is evidence of editorial independence: Epoch published FrontierMath evaluation results contradicting OpenAI's claimed scores (11% vs. 32% for o3-mini). But structural incentives are structural regardless of current behavior.

What Others Say

The Mechanize critique. Oliver Habryka (LessWrong founder): "This seems like approximate confirmation that Epoch research was directly feeding into frontier capability work." Anthony Aguirre (FLI): "Huge respect for the founders' work at Epoch, but sad to see this." Holly Elmore (PauseAI): "This is the opposite of keeping the world safe from powerful AI! You are a traitor."

Paul Romer on FrontierMath (Nobel laureate, Jan 2025): Argues Elliot Glazer "does not understand the world in which he now works" -- that the holdout set is "an obvious sham" because OpenAI owns all problem statements, and that "monopolistic impunity is corroding the social norms of integrity." A structural argument about how tech industry incentives differ from academic norms.

Methodological critique. Steven Byrnes (LW, 114 karma): argues Epoch's headline ~3x/year algorithmic progress number is misleading, conflating distinct types of improvements. Real "stereotypical learning algorithm improvements" may be <10x over the entire 2018-2025 period (~30%/year), far less than the exponential suggested. Three other independent estimates arrive at similar ~3x numbers, but Byrnes argues they're measuring different things.

Policy credibility. "Epoch is highly trusted by all camps. At least in DC federal policy circles, I've heard people say 'I like what Epoch wrote'... You're not seen as having 'motivated reasoning.'" -- Golden Gate Institute for AI. "Epoch is probably the single organization I cite the most in my writing." -- Institute for Progress. "A key crux for CG's AI grantmaking strategy is AI timelines. Epoch reports have been a frequent source of parameter value estimates." -- Coefficient Giving managing director. "We used the Epoch Capabilities Index to determine what the recent rate of software progress has been." -- AI Futures Project.

What's Absent

No 990 financial filings yet (too new as independent nonprofit; first filing expected late 2025/early 2026). No independent board members -- all four have financial relationships with Epoch or its primary funder. No published conflict of interest or recusal procedures. No external audit of the core compute database methodology. No published forecasting track record assessing accuracy of past predictions. No breakdown of revenue from commercial vs. philanthropic sources. No documented policy for managing the tension between semiconductor stock investments and research neutrality.