JAIME SEVILLA

Behavioral Archetype

THE FORECASTER WHO SIGNS THE LAB CONTRACT — Subject co-founded Epoch AI, the nonprofit whose compute, data and benchmark numbers the labs and the press cite, and runs it. The same organization lists paid engagements with OpenAI, Google DeepMind, xAI and others, and its largest listed donor is the funder behind much of the evaluation layer. He is also the one on the record, in his own words, saying FrontierMath’s OpenAI sponsorship “was unclear” to contributors, and, in 2026, that the benchmark’s scores are “a dataset only us and OpenAI have access to.” The arrangement is lawful, disclosed on Epoch’s own pages, and unusually well documented by its own CEO. This profile scores the seat, not the man.

Essence Indicators

• Co-founder and Chief Executive Officer of Epoch AI. “Jaime Sevilla is the co-founder and Chief Executive Officer of Epoch AI. His research is focused on technological forecasting and the trajectory of AI.” (Epoch AI). In April 2025 he described the same job as director: “I will continue working as director of Epoch AI, in service of our public benefit mission” (his post on X). The team page gives no start date.

• Board of Directors, Epoch AI, listed with Tom Davidson, Ajeya Cotra, María de la Lama and Steve Newman (Epoch AI). The CEO sits on the board that oversees him, and one of his co-directors is on the staff of the evaluator METR.

• Co-author of the FrontierMath disclosure post. “OpenAI commissioned Epoch AI to produce 300 advanced math problems for AI evaluation that form the core of the FrontierMath benchmark.” The post is signed “By Tamay Besiroglu and Jaime Sevilla.” It states: “Per our agreement, we needed OpenAI’s permission before publicly disclosing their involvement,” that “our communication with them should have been more systematic and transparent,” and that Epoch is finalizing a set for which “OpenAI will only receive the problem statements and not the solutions” (Epoch AI). Co-author Tamay Besiroglu is the co-founder who later left to start Mechanize.

• His own words on the disclosure. January 22, 2025: “Some people have pointed out that it was unclear that the FrontierMath benchmark my org developed was sponsored by OpenAI, including to some of the benchmark contributors. We will set the record straight!” (his post on X). February 15, 2026: “This tweet falsely claims to report FrontierMath scores, a dataset only us and OpenAI have access to” (his post on X).

• The 2026 sequel. April 16, 2026: “OpenAI purchased access to the verifiers, which check whether a proposed solution is correct.” And: “We’d love to have more companies purchase verifier access” (his post on X). Epoch’s page lists OpenAI engagements of 2026 “FrontierMath Data Audit” and “FrontierMath: Open Problems Verifier” (Epoch AI).

• Mechanize, and his line against it. April 19, 2025: “I am NOT joining Mechanize, nor I was directly involved in its creation.” April 25, 2025: “To be clear, I have no personal stake in Mechanize” (his post; his post). On the split: “our staff and funders are split on their opinion of Mechanize,” so “I don’t think we can officially endorse nor condemn Mechanize’s mission” (his post on X).

• Advisor, Observatorio de Riesgos Catastróficos Globales, by his own site, which also says “I study trends in Machine Learning at Epoch AI” (jaimesevilla.me).

Social Persona / Impression Management

Immediate impression: The forecaster who posts his numbers. His feed is trend data, compute arithmetic and error bars, with probabilities attached to his own opinions.

Energy: Open and declarative. After the Mechanize departures he published what he called a “full state dump” of his beliefs: “I put the probability that ai ends up being net bad for humans at 15%” (his post on X).

Impression management strategy: Separate the person from the institution, and say so. “I want to, and my team had encouraged me to, be transparent about my personal beliefs about AI risk and my values… But this does and should not reflect Epoch’s official position” (his post on X). The method is disclosure first, which is also why the record on him is thick.

Forensic Archetype Comparison

Psychometric Assessment

Trait scores observe the documented public role, not private character.

Big Five (OCEAN):

Dark Triad:

MBTI: INTP-adjacent, inferred from the documented public role only. A systems-and-forecasts register, with the opinions filed under his own name.

Threat Assessment

No score has moved against a prior version; this is a first scoring. Calibrated against peers: Ajeya Cotra (5/8/7, hazard 64, same board and funder network, holds no CEO seat), Tamay Besiroglu (6/6/7, hazard 62, co-founder who left for the supplier side) and the Epoch AI org page (5/7/7, hazard 64). He takes attack 5 like Cotra and Epoch (an analyst, not an actor), defense 7 like Epoch (disclosure on the record, board oversight, but the sponsored test is the exposure) and hp 7 like all three.

Alignment Analysis

Stated alignment: “Our nonprofit’s purpose is to study AI’s trajectory to benefit society” (his post on X). Epoch states: “We care deeply about always being able to say what we believe to be true” and “we are committed to disclosing any research sponsorship and data access agreements with industry” (Epoch AI; Epoch AI).

Observed alignment: The largest listed donor is Coefficient Giving, formerly Open Philanthropy, with entries from June 2022 to July 2026; listed commercial partners include OpenAI, Google DeepMind, xAI and Anthropic (Epoch AI). He wrote in February 2025: “we actively collaborate with major AI labs like OpenAI and government bodies like UK AISI” (his post on X). The disclosure post says the OpenAI relationship could not be disclosed without OpenAI’s permission.

Gap assessment: No gap between the stated mission and the published ledger is documented; Epoch’s disclosure page is a rare case of the grader listing its own paymasters. The file’s point is what the ledger shows: the benchmark whose scores the public cites was commissioned by the lab it ranks, and the 2026 verifier access was bought by the same lab. The record settles that. It does not settle whether results were affected, and nothing in it says they were.

Interest. Each position below is argued in his own words; the column of what adoption pays Epoch is documented from Epoch’s pages. None of it is motive, and none of it says the argument is wrong.

The convergence is the finding. The position that favors measurement and audit over hardware controls is argued by the person whose organization is the measurer, and the buyer of the measurement is the lab being measured.

Alliance Map

Overlap: the same person is CEO, board member and public voice of an organization that is simultaneously funded by the evaluation layer’s principal philanthropy, paid by the labs it benchmarks, and consulted by two governments. The ties run at the same time and none is concealed: Epoch’s own page lists them. That is the finding, and it holds without any claim about influence.

Convergent Drive Classification

Self-preservation: Keeps the organization funded from both sides, philanthropy and lab contracts, and lists both. Goal preservation: Holds one stated goal, “to benefit society” through measurement, across the 2025 disclosure episode and the departure of three senior colleagues. Resource acquisition: Donor list from Coefficient Giving to named individuals; commercial list from OpenAI to a hedge fund and a venture firm (Epoch AI). Self-improvement: From trend forecasts to selling the verifiers that check the answers.

Subject is not an AI system. The drives appear anyway — in the CEO who publishes his own funder list and then signs the next contract on it.

Sources: Jaime Sevilla — Epoch AI; Team — Epoch AI; Clarifying the Creation and Use of the FrontierMath Benchmark — Epoch AI; Transparency — Epoch AI; Support us — Epoch AI; Jaime Sevilla — personal site; Jaime Sevilla on X.