Epoch AI

Stated Theory of Change

Epoch's stated theory of change is that better information about AI trends leads to better decisions by all stakeholders. They aim to be the "William Nordhaus of AI" -- creating the empirical foundation that informs policy, investment, and research regardless of one's priors about whether AI is beneficial or dangerous.

The specific mechanism: collect and curate data on AI model training compute, hardware trends, algorithmic efficiency, and data availability. Analyze these trends to produce forecasts and identify bottlenecks. Publish findings openly. Engage with policymakers, labs, and the public to ensure decisions are "informed by the best possible evidence."

This is a deliberately neutral theory of change. Epoch does not advocate for specific policies or take a position on AI progress. Their bet is that the world is better off with high-quality public information about AI, regardless of which direction decisions ultimately go.

Revealed Theory of Change

Epoch's actions mostly align with their stated mission, but with notable tensions:

Alignment. They genuinely produce the best public data on AI compute trends. Their databases are cited by governments, labs, and researchers across ideological lines. They publish findings that sometimes contradict their clients (the o3-mini FrontierMath discrepancy). Their "Three Challenges" paper documented limitations of compute-based regulation that Epoch itself helped establish.

Divergences.

1. The FrontierMath episode revealed a willingness to subordinate transparency to client relationships. Epoch agreed to NDAs that prevented disclosure to their own contributors. The "corrective actions" came only after public scandal. This suggests the neutrality commitment can be overridden by financial considerations.

2. The pattern of departures suggests Epoch functions partly as a capabilities pipeline. Three co-founders left: one to a capabilities startup (Mechanize), one to AI safety evaluations (Apollo), one to compute governance policy (RAND). The Mechanize departure is the most revealing -- Besiroglu used Epoch's research on automation economics to build a company literally aimed at "full automation of all work."

3. Revenue diversification toward AI labs creates subtle dependencies. Paid clients include OpenAI, Google DeepMind, xAI, Anthropic. While Epoch aims to charge "industry consultant rates," the growing commercial relationships create incentives to produce work that labs find useful rather than work that might embarrass them.

4. Investing reserves in semiconductor/AI stocks creates a financial interest in AI progress for an organization that claims to be neutral about whether AI progress is beneficial.

Key Assumptions

Assumption 1: Better public information about AI leads to better decisions.

• Evidence for: Epoch data informed compute thresholds in EU AI Act and Biden EO. Policymakers across ideological lines cite their work. CG uses their estimates to calibrate timelines and funding allocation.
• Evidence against: Information is necessary but not sufficient. Having good data about AI progress doesn't prevent misuse if institutions are captured or incentives are misaligned. Also, good information about AI capabilities can help capabilities developers as much as safety researchers.
• Testable: Yes -- do policies informed by Epoch data turn out to be well-calibrated?
• If wrong: Epoch becomes a useful public good but doesn't meaningfully reduce AI risk.

Assumption 2: Organizational neutrality is possible and desirable.

• Evidence for: Epoch is genuinely trusted by multiple camps. Policy credibility comes partly from not having an agenda.
• Evidence against: Neutrality on existential risk is itself a position. All funding comes from AI safety philanthropy that explicitly frames Epoch as risk reduction. Being neutral while funded by safety-oriented donors could mean producing work that happens to be useful for safety without ever taking hard positions that might be needed.
• Testable: Would Epoch publish findings strongly suggesting AI progress should be halted or dramatically slowed? We don't know.
• If wrong: Epoch provides a veneer of scientific objectivity to a debate where the outcome depends on advocacy, not just facts.

Assumption 3: Compute trends are the right thing to track for understanding AI trajectory.

• Evidence for: Compute has been the best predictor of AI capabilities. Epoch's own research shows training compute growing ~4-5x/year with strong predictive power.
• Evidence against: Their own "Three Challenges" paper argues compute as a proxy is breaking down. Reasoning training, distillation, and deployment scaffolding increasingly drive capabilities beyond what compute alone predicts. Steven Byrnes's critique suggests algorithmic progress may be much less than Epoch's headline estimates.
• Testable: Yes -- compare Epoch forecasts against reality over time.
• If wrong: Epoch's core methodology becomes obsolete, though they're already pivoting toward benchmarking and capability indices.

Assumption 4: Independence is maintained despite extreme funder concentration.

• Evidence for: Published FrontierMath results contradicting OpenAI's claims. Published "Three Challenges" paper questioning their own compute governance framework. Internal diversity of views (Erdil's bearish timelines vs. Barnett's bullish ones).
• Evidence against: Board is 50% funder-controlled. No independent directors. No published conflict of interest procedures. FrontierMath NDAs show willingness to compromise transparency for client relationships.
• Testable: Would Epoch publish findings that directly threatened CG/OP funding or lab client relationships?
• If wrong: Epoch's research subtly self-censors on questions that might displease funders.

Strengths

Unique positioning. No other organization occupies Epoch's niche. There is simply no comparable public database of AI compute trends, and no other nonprofit doing this kind of quantitative AI forecasting with this level of rigor. This makes them irreplaceable.

Credibility across divides. Being trusted by governments, AI labs, safety researchers, and policy institutes simultaneously is extremely rare. This credibility is their most valuable asset.

Research quality and output. 21 people producing 38 data insights, 14 papers, 40 newsletters per year is remarkable productivity. The research is technically sound, openly published, and widely cited.

Intellectual honesty. Sevilla's candid admissions of uncertainty ("I'm just very confused" about x-risk), willingness to correct past estimates (data limits paper revision), and publication of findings contradicting clients (o3-mini FrontierMath scores) all suggest genuine intellectual integrity at the leadership level.

Policy impact. Epoch data has materially shaped the most significant AI regulations to date (EU AI Act, Biden EO). This is concrete, measurable impact.

Weaknesses and Risks

Funder concentration. 91% of known funding from a single source (CG/OP) is extreme dependency. Two of four board members from that funder. The organization's survival depends on maintaining this relationship.

Governance gaps. No independent board members, no published conflict of interest procedures, dual-hatting of board + executive roles (de la Lama). These are standard nonprofit governance failures that become more concerning given the funder concentration and lab client relationships.

Capabilities pipeline risk. The Mechanize departure is not an isolated incident but part of a pattern. Four of seven co-founders left, including three to capabilities-adjacent work. Epoch's "neutral" research environment may be training and socializing researchers who subsequently apply their understanding of AI scaling to accelerate capabilities.

FrontierMath as cautionary tale. The disclosure failure revealed systemic issues: willingness to accept NDAs that prevented transparency, a "verbal agreement" as the only safeguard against training data contamination, structural ownership giving one company exclusive access. These are governance failures, not mere communication mistakes.

Methodological vulnerability. The core algorithmic progress claim (~3x/year) faces substantive challenge from Byrnes and others. If this number is significantly wrong, it undermines Epoch's most-cited research. The compute database, while unique, involves significant estimation and has never been externally audited.

Semiconductor stock investments. Creating a financial interest in AI sector growth while claiming research neutrality on AI progress is a structural conflict. Even if it doesn't actually influence research, the perception alone damages credibility.

Cross-References

Complementary to: METR (model evaluations -- Epoch is developing a benchmark with METR), GovAI/Centre for the Governance of AI (compute governance policy that builds on Epoch data), AI Safety Institute/UK DSIT (government capacity that uses Epoch analysis).

Revenue relationship with: OpenAI, Google DeepMind, xAI, Anthropic (paid clients). These relationships provide useful information flow but create subtle dependency.

Talent pipeline to: Mechanize (Besiroglu, Erdil, Barnett), Apollo Research (Hobbhahn), RAND/GovAI (Heim), Principia Labs (Glazer). Epoch has been remarkably productive at spawning new organizations, though the direction is mixed from a safety perspective.

Funded by same source as: Many AI safety organizations (CG/OP funds a large fraction of the field). This creates a correlation risk -- if CG/OP changes strategy, many orgs are affected simultaneously.

What Would Change This Assessment

Positive updates:

• Independent board members appointed with no CG/OP or lab affiliation
• External audit of compute database methodology published by credible third party
• Published forecasting track record showing calibration quality
• Revenue diversification reducing CG/OP to <50% of total
• Published conflict of interest and recusal procedures
• Evidence that Epoch published findings clearly uncomfortable for funders or clients

Negative updates:

• Evidence that research conclusions were influenced by funder or client preferences
• Further departures to capabilities-focused organizations
• Failure to implement stated corrective actions on benchmark transparency
• Continued expansion of lab client relationships without proportional governance strengthening
• Suppression or delay of findings embarrassing to OpenAI or other clients

Self-Critique

Weakest claim: My assessment of the semiconductor stock investment conflict may overstate the risk. The amounts involved are likely small relative to total assets, and the framing as a "hedge" is logically coherent even if optically concerning. I may be pattern-matching to financial conflicts of interest in ways that don't map well to a small nonprofit.

Potential biases: I may be too sympathetic to Epoch's mission because their work is genuinely useful and their intellectual honesty is refreshing. The FrontierMath episode could be more damning than I'm framing it -- it might reveal a systematic willingness to prioritize revenue over transparency that hasn't been tested in other domains yet.

Missing perspective: I haven't engaged deeply with the view that Epoch's work is net negative for safety because it helps capabilities developers more than safety researchers. Habryka's critique is the closest to this, but a more developed version would argue that quantifying AI progress precisely makes it easier for labs to allocate resources efficiently to capability gains.

What a thoughtful disagreer would say: "Epoch is the ideal laundering mechanism for AI labs. Labs fund neutral research through philanthropy, the research is genuinely high-quality and trusted, and then the same labs use that research to build more powerful AI. The whole theory of change -- that better information leads to better decisions -- ignores the fact that labs have better information regardless and that public information mostly helps them recruit, fundraise, and shape policy in their favor. Epoch's neutrality is not a feature, it's a bug."

Information that would most change my view: A 990 filing showing that commercial revenue from AI labs is a much larger share of total revenue than I've estimated, or evidence that a research finding was delayed or softened due to client or funder pressure.