MATS RESEARCH

Institutional Archetype

THE FELLOWSHIP THAT STAFFS THE FIELD — MATS describes itself as “an independent research fellowship that connects talented researchers with top mentors in the fields of machine alignment, transparency, and security” (MATS). Applicants are admitted to a ten to twelve week program in Berkeley and London, matched to a mentor who works at a lab, an evaluator, a think tank or a university, and the best are offered a funded extension of six to twelve months. The mentors and the employers of the alumni are the same organizations: the labs whose models are tested, and the evaluators who test them.

The other evaluator-layer profiles on this site test models, fund testers or write the rules for testers. MATS supplies the people. Apart Research runs a weekend-hackathon funnel into the same employers; Halcyon Futures seeds the evaluators as organizations and also leads a MATS stream; FAR.AI and Redwood Research are both among the organizations whose staff mentor MATS fellows. This profile maps who funds the pipeline, who teaches in it and where the graduates go, from MATS’s own pages. It scores the position in the apparatus, not the people in it.

Mandate & Origin

MATS’s team page states its goal: “The main goal of MATS is to grow the AI safety & security research fields.” Its About page adds: “MATS exists to help grow the next generation of researchers who can address the technical and strategic challenges posed by increasingly capable AI systems. We aim to identify promising individuals, support their development through hands-on research experience, and connect them with the broader ecosystem of AI alignment organizations, labs, and mentors.” The homepage’s headline claim: “Since late 2021, 631 researchers have participated in MATS, producing more than 220 research papers, joining leading AI labs, and founding new organizations driving progress in AI alignment, transparency, and security” (MATS, Team; MATS, About; MATS).

The funder’s account of the origin is on the record. Coefficient Giving’s 2025 Letter from the CEO says: “We made our first $250,000 grant to the program that would eventually become ML Alignment & Theory Scholars (MATS) in 2019, when it was a side project by some students affiliated with the Stanford Existential Risks Initiative,” and that the first cohort “had 5 fellows and no permanent full-time staff” (Coefficient Giving, 2025 Letter from the CEO). The funder’s grant feed carries the program under its earlier name, “SERI MATS”, with the Berkeley Existential Risk Initiative and AI Safety Support as recipients of record, before any grant to MATS Research itself (Coefficient Giving, grants search). This site’s research page The EA to AI-Safety Funding Pipeline sets the same account beside the funder’s other capacity-building grants.

The leadership, from MATS’s team page: Ryan Kidd, “CEO & Co-Founder,” who is “a Co-Founder and Board Member of the London Initiative for Safe AI (LISA), a Manifund Regrantor, and advisor to Halcyon Futures, Catalyze Impact, Pivotal Research, AI Safety ANZ, and BASE,” with a physics PhD from the University of Queensland and earlier independent alignment research “for the Stanford Existential Risks Initiative”; Christian Smith, “COO & Co-Founder,” also a co-founder of LISA. The UK branch is led by Matthew Wearden, Managing Director, and Maria Campbell, Deputy Managing Director, formerly Chief Operating Officer and Chief People Officer at Monzo and Snyk. The research team is led by Laura Vaughan, Research Director, with 28 staff whose titles end in Research Manager, among them a former MATS scholar (Rich Barton-Cooper) and Jason Green-Lowe, who before joining MATS “led the Center for AI Policy (CAIP), which lobbied Congress to pass strong AI safety legislation” (MATS, Team).

The board, on the same page: Josh Jacobson, Chief Operating Officer at the Future of Life Foundation, with earlier work “for METR, FAR AI, Anthropic, on a research grant, for the Future for Humanity Institute (EpiFor), the Berkeley Existential Risk Initiative”; Abi Olvera, Research Director at the Golden Gate Institute for AI and a former U.S. diplomat; and Michael Aird, AI Program Director at Longview Philanthropy and “a board member at the MATS Program.” ProPublica’s record of the FY2024 return lists Jacobson as President, Olvera as Treasurer and Aird as Secretary, each at $0, and Kidd and Smith as Co-executive Directors at $123,879 and $112,551 (MATS, Team; ProPublica Nonprofit Explorer).

The legal entity. MATS’s transparency page: “MATS Research is an independent 501(c)(3) public charity (EIN: 99-0648563),” and “Prior to 2024, MATS Research Inc. was fiscally sponsored by BERI.” It also reports “a separate entity in the UK, which is a private company limited by guarantee, where MATS Research Inc is its sole member and guarantor,” MATS Research UK Ltd, registered in England and Wales (no. 17234700). It posts its bylaws, conflict-of-interest and whistleblower policies, Form 1023, determination letter and the 2023 Form 990-N. ProPublica lists Mats Research Inc, Berkeley, California, “Tax exemption issued: April 2024,” with a Form 990 for the year ending December 2024 filed 14 November 2025 (MATS, Transparency; ProPublica Nonprofit Explorer).

Funding & Backers

MATS publishes its donors by name, date and amount from the date it incorporated independently; the page says it “exclusively reflects donations and documentation from the period following our independent incorporation” and points to BERI’s own transparency page for the earlier years (MATS, Transparency). The record, as printed:

• 2025 cash gifts. Coefficient Giving via National Philanthropy Trust, $11,823,980 (23 July 2025); Coefficient Giving via Silicon Valley Community Foundation, $3,629,280 (12 September 2025); Good Ventures Foundation, $1,520,150 (6 May 2025) and $3,038,300 (7 May 2025); Berkeley Existential Risk Initiative, $450,000 (5 March 2025); the Survival and Flourishing Fund via Silicon Valley Community Foundation, $289,000 (7 July 2025); Cooperative AI Foundation, $68,294; Schwab Charitable, $50,000; Manifund, $10,000 and $400; one individual, $12,000. The lines sum to $20,891,404, of which Coefficient Giving and Good Ventures account for $20,011,710, or 95.8 percent.
• 2024 cash gifts. Coefficient Giving via Silicon Valley Community Foundation, $3,000,000 (11 June), $382,029 (15 November) and $660,000 (30 December); Founders Pledge, $584,000; the Foresight Institute, $70,000; Vanguard Charitable, $44,000; Manifund, $8,365; and BERI in four gifts of $1,000, $499,000, $500,000 and $500,000. The lines sum to $6,248,394, of which Coefficient Giving’s three gifts are $4,042,029 and BERI’s four are $1,500,000. The FY2024 return reports revenue of $6,308,393, all contributions, expenses of $2,926,442 and net assets of $3,381,951.
• Compute credits from labs. Anthropic API credits, estimated value $35,000 (late 2024); Google Gemini Academic Program award, estimated value $100,000 (6 November 2025); Thinking Machines research grants (10 November 2025), no value given; OpenAI ChatGPT Pro and Codex access, estimated value $120,000 (23 March 2026).
• The funder’s own postings. Coefficient Giving’s grant feed, under its earlier name Open Philanthropy, lists a $2,047,268 grant to BERI for the SERI-MATS collaboration (March 2023), two grants totalling $2,641,368 to BERI and $732,631 to AI Safety Support for the winter 2023-24 cohort (March 2024), two grants totalling $3,382,029 to MATS Research for “MATS Program 6.0 + 7.0” (February 2025) and a $660,000 grant to MATS Research for winter 2024-25 research expenses (May 2025), alongside $2,321,550 and $709,567 to AI Safety Support in 2025 (Coefficient Giving, grants search; Coefficient Giving, grants search, page 2; Coefficient Giving, grants search, page 3). See Open Philanthropy.
• The Survival and Flourishing Fund. SFF-2024: MATS Research, $639,000 in the Main track, recipient MATS Research, Inc., “General support,” funder Jaan Tallinn; SFF-2025: ML Alignment & Theory Scholars Research, Main $289,000, recipient MATS Research, Inc., “General support,” funder Jaan Tallinn. The 2026 table does not list MATS (SFF 2024; SFF 2025; SFF 2026). See Survival and Flourishing Fund and Jaan Tallinn.
• Longview Philanthropy. MATS’s March 2026 report on hiring needs thanks “Coefficient Giving, the Survival and Flourishing Fund, and Longview Philanthropy, without whose donations we would be unable to run upcoming programs or retain team members essential to this report.” Longview does not appear on the transparency page’s donor lines; a MATS board member is Longview’s AI Program Director (MATS, AI Safety Talent Needs in 2026; MATS, Team).
• The funder’s own account. Coefficient Giving’s CEO letter: MATS now runs “a year of around 100 scholars with an admission rate of 4-7%,” and “over 80% of their alumni are now working full-time in AI safety and security (accounting for a meaningful portion of safety staff at some of the biggest companies and government institutes).” A footnote discloses that MATS’s “few dozen full time staff” include the CEO’s wife, “who joined in December as a research manager after a decade as a data scientist at Uber” (Coefficient Giving, 2025 Letter from the CEO).

Institutional Voice & Intent

MATS’s own account of what it sells is addressed to three audiences. To applicants: mentorship, “Compute, stipend, housing, and travel support to minimize barriers to participation,” a “Research Symposium,” and the extension; its homepage prices the fellowship at “a $1250 stipend per week,” housing and “$2k per week of compute resources.” To employers: its homepage pairs each alumni testimonial with an employer name and says “joining leading AI labs.” To funders: “Donate” is a header link on every page, with donation routes for donor-advised funds, wires and stock (MATS, About; MATS, Donate).

Stated intent: Grow the number of researchers working on AI alignment, transparency and security, and place them with the organizations that do the work.

Observed intent: Run an admissions-and-matching funnel in which the mentors are staff of the employers, select fellows through “a multi-stage admissions process” that can include interviews, work tests, coding assessments and writing samples, and report alumni outcomes (75 percent working in the field, 10 percent founding organizations) as the program’s output. Its March 2026 report states the hiring logic in its own words: fellowship programs “play a crucial role as hiring pipelines, by providing extended observation periods and opportunities for candidates to grow research taste, execution speed, communication quality, and independence” (MATS, Autumn 2026; MATS, AI Safety Talent Needs in 2026; MATS).

Gap: MATS discloses its donors by line, its board, its officers’ pay and its alumni destinations. What its pages do not publish is an admissions list by selecting organization, a count of alumni by employer, the share of mentors who are also hiring managers at the employers its alumni join, or any measure of the field’s output absent the program. The homepage lists more than forty alumni-founded organizations and a set of employers, with no figures per employer. The structure the page describes is a pipeline whose teachers, funders and employers overlap; whether it creates people the field would not otherwise have, or sorts and credentials people the field would have found, is a question the pages do not settle and do not claim to.

Interest: MATS argues that AI safety is held back by a shortage of trained people and that fellowships are the remedy. What adoption would pay it is set out from its own pages and its funders’ postings.

• Fund fellowship pipelines for the field. Position: “The main goal of MATS is to grow the AI safety & security research fields.” What it pays: every gift on its transparency page is a payment for that goal, $20,891,404 in 2025 alone, and the funders post the same program under their capacity-building focus area (MATS, Team; MATS, Transparency; Coefficient Giving, grants search). Best case, its own: hiring managers it interviewed said the field is “capacity constrained by a lack of senior researchers who can mentor and supervise junior talent,” and a fellowship gives the extended observation that interviews and credentials cannot. Convergence: the labs and evaluators that hire from the pipeline do not pay for the observation period; philanthropic donors do, while 21 OpenAI staff, 18 Anthropic staff, 12 Redwood Research staff and seven Google DeepMind staff appear on the mentor list (MATS, AI Safety Talent Needs in 2026; MATS, Mentors).
• Make mentor testimonials and policy placements the field’s hiring currency. Position, from the report’s recommendations: “Standardized Mentor Testimonials,” so that “If fellows left with a structured evaluation from their mentor that covers research taste, execution speed, communication quality, independence, and other key dimensions, it could dramatically reduce hiring uncertainty,” and “Direct Policy Embedding,” in “partnerships with congressional offices, think tanks, government agencies, or initiatives like the Frontier Model Forum.” What it pays: a mentor’s written evaluation issued by MATS becomes the credential employers read, and the placements it proposes run through organizations already on its mentor and board lists. Best case, its own: “Generic academic recommendations provide little value,” while calibrated evaluations “carry substantial weight.” Convergence: the Frontier Model Forum is named as an embedding venue, and the evaluators and labs that would read the testimonials are the same organizations whose staff write them (MATS, AI Safety Talent Needs in 2026).

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

• The program. The Autumn 2026 cohort runs ten weeks in Berkeley and London from 28 September to 4 December; the page says “Fellows will receive mentorship from world-class researchers and at organizations like Anthropic, Google DeepMind, OpenAI, Redwood Research, and ARC,” and that “For the first time, we are running Founding & Field-Building and Biosecurity tracks.” Winter 2027 applications are closed. Of fellows who applied for the extension phase in recent cohorts, “roughly 80% of fellows who applied to the extension phase were accepted” (MATS, Autumn 2026; MATS).
• The mentors. The Mentors page lists 201 mentors across its cohorts, on twenty-one result pages. By listed organization: OpenAI 21 (among them Tomek Korbak and Micah Carroll), Anthropic 18 (among them Ethan Perez, Samuel Marks, Fabien Roger, Nicholas Carlini, Sam Bowman, Jack Lindsey and Kyle Fish), Redwood Research 12 (among them its CEO Buck Shlegeris), the UK AI Security Institute 9, Coefficient Giving 7, Google DeepMind 7 (including Neel Nanda), SecureBio 7, RAND 4, the Alignment Research Center 5 (four listed as ARC and one as the Alignment Research Center), Apollo Research 3 (its CEO Marius Hobbhahn among them), Transluce 3 (including Jacob Steinhardt and Sarah Schwettmann), AVERI 2 (Miles Brundage and Patricia Paskov), Halcyon Futures 2, METR 2 (Neev Parikh and Megan Kinniment), Thinking Machines 2, and one each at Eleos, BlueDot Impact and Sentinel Bio. Sunishchal Dev is listed under “Center for AI Standards and Innovation; RAND,” and Paul Christiano’s bio reads “the executive director of the Alignment Research Center and a technical advisor at the Center for AI Standards and Innovation within NIST.” Sydney Von Arx is listed under Nightingale with the line “works on adversarial stress-testing at METR.” Esben Kran is listed under Juniper Ventures and his bio says he “founded Apart Research, an AI safety research organization where he now serves as board chair” (MATS, Mentors).
• The evaluators’ own streams. The Autumn 2026 Anthropic stream lists “Improving the robustness of LLMs to red teaming,” “Control,” “Scalable oversight” and “cybersecurity evals”; the Apollo Research stream says “We plan to evaluate scheming behaviors directly relevant to internal deployments” and says the work requires “building a science of scheming” (MATS, Anthropic stream; MATS, Apollo stream).
• The Halcyon stream. Autumn 2026, led by Halcyon Futures’ CEO Mike McCormick and its investor and grantmaker Ross Matican: “Backing projects focused on product development and organization building in the areas of AI safety and alignment, biosecurity, and critical cybersecurity.” The stream’s field-building project: “we are prioritizing a talent matching project. This includes processing thousands of individuals in our CRM, and finding how they may pair with our portfolio companies and other areas of high priority in our network.” Ryan Kidd is listed on MATS’s team page as an advisor to Halcyon Futures. Halcyon’s portfolio includes evaluators on this site (MATS, Halcyon stream; MATS, Team).
• Where alumni go, in MATS’s own words. “75% of alumni are now working in AI alignment, transparency, and security”; “10% of our alumni have co-founded AI safety organizations and research teams during or after their time at MATS.” “MATS alumni have been hired by leading organizations like Anthropic, Google DeepMind, OpenAI, Meta AI, UK AISI, Redwood Research, METR, RAND CAST, Coefficient Giving, ARC, FAR.AI, Apollo Research, Truthful AI, Goodfire, LawZero, MIRI” and others. Alumni-founded organizations it lists include Apollo Research, ARENA, Center for AI Policy, Leap Labs, LISA, PRISM Eval, Timaeus, Poseidon Research, General Analysis and WeaveMind. Alumni profiles name current employers: Anthropic (Nina Panickssery, Johannes Treutlein, Dan Valentine), Redwood Research (Cody Rushing), the UK AI Security Institute (Joseph Bloom, “Head of Model Transparency”), and a contractor stint at METR evaluating GPT-4 before release (Quentin Feuillade-Montixi, who then co-founded PRISM Eval). Marius Hobbhahn’s testimonial: “Apollo almost certainly would not have happened without MATS” (MATS; MATS, Alumni). See Apollo Research and METR.
• The funder’s staff. Coefficient Giving’s own team pages list the program in two staff biographies: Rohan Subramani, who joined in May 2026, “worked on technical AI safety research at the University of Toronto, the MATS program, and UC Berkeley’s Center for Human-Compatible AI (CHAI)”; Steven Veld, who joined in September 2026, “worked on AI scenario modeling and AI policy through the MATS Program and the Institute for AI Policy and Strategy (IAPS) Fellowship.” Seven Coefficient Giving staff (among them Adin Richards, Aman Patel and Damon Binder) are also listed as MATS mentors (Coefficient Giving, grants search; MATS, Mentors).
• Research output. The homepage counts more than 220 publications, “over 19,000 collective citations and an organizational h-index of 59,” and lists research agendas its fellows helped start, among them sparse autoencoders, activation engineering, emergent misalignment, evaluating situational awareness and externalized reasoning oversight (MATS).

Actions & Leadership Choices

CONDUCT: WATCH — ONE PIPELINE, THE SAME FUNDERS AND THE SAME EMPLOYERS ACROSS THE LABS AND THE EVALUATOR LAYER. No wrongdoing is alleged. The structural point is overlap: one philanthropic network funds the training that places people in the labs and the evaluators; the labs’ and evaluators’ staff mentor in it; the lab donates compute to it; its board and leadership sit on the boards and advisory lists of the organizations that fund, seed and hire from it. All of it is disclosed on its own pages, which is how the overlap can be read at all.

• Late 2021. The count of researchers who have participated in MATS begins (MATS).
• 2022 to 2024. Open Philanthropy grants for the program go to BERI and AI Safety Support (Coefficient Giving, grants search).
• 2024. MATS Research Inc incorporates independently; the IRS exemption is issued in April; fiscal sponsorship by BERI ends (MATS, Transparency; ProPublica Nonprofit Explorer).
• February and May 2025. Coefficient Giving posts $3,382,029 and $660,000 in grants to MATS Research (Coefficient Giving, grants search).
• May to September 2025. Gifts of $15,453,260 from Coefficient Giving and $4,558,450 from Good Ventures Foundation are recorded on the transparency page (MATS, Transparency).
• 14 November 2025. FY2024 Form 990 filed (ProPublica Nonprofit Explorer).
• 24 March 2026. Publishes its report of 23 interviews with hiring managers, research leads and funders (MATS, AI Safety Talent Needs in 2026).
• Autumn 2026. Founding & Field-Building and Biosecurity tracks open for the first time, including the Halcyon Futures stream (MATS, Autumn 2026; MATS, Halcyon stream).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. MATS’s cash comes from the network that also funds Apart Research, Redwood Research, AVERI and Palisade Research: Coefficient Giving appears in each funder record, and SFF in several. Its mentors include staff of Redwood Research, Apollo Research, Transluce, AVERI and METR, and the founders of Apollo Research and PRISM Eval are alumni. Its CEO advises Halcyon Futures, which leads a MATS stream, seeded Transluce and is a listed funder of AVERI; its board seats an employee of a thanked funder (Longview) and a former METR, FAR.AI and Anthropic staffer. Its board also includes the Research Director of the Golden Gate Institute for AI; Fathom’s partner list names a “Golden Gate Institute,” and neither page says the two names are one organization. The documents show disclosed arrangements, not coordination.

Reach Assessment

Institutional: Talent supply at scale to the labs and the evaluator layer, and a hiring credential those employers read. Memetic: The research agendas it credits to fellows (activation engineering, emergent misalignment, sparse autoencoders) and the alumni-founded organizations are the work most likely to travel outside the field.

Civilizational: If one fellowship becomes the default way a researcher is observed, scored and introduced to the field, and the mentors scoring the fellows are staff of the employers, the field’s norms of what counts as a good researcher and a good research question propagate through its alumni. MATS’s pages are open about the structure: donors by line, a board with named affiliations, the employers of its alumni, and a report that says fellowships function as hiring pipelines. What a reader cannot check is how much of the field’s growth the program caused rather than sorted.

Score basis. Atk 6 is reach: MATS counts 631 researchers since late 2021, says 75 percent of alumni are working in the field, names Anthropic, Google DeepMind, OpenAI, the UK AISI, Redwood Research and METR among employers, credits its alumni with founding Apollo Research, and lists 201 mentors, 21 of them at OpenAI and 18 at Anthropic (MATS; MATS, Mentors). The points under a higher number are that no statute or release gate makes any lab ask for its output and that it publishes no evaluation of a named lab model. Peers on atk 6: Apart Research holds 5, a weekend funnel with 8,700 Sprint participants and a policy channel but fewer named lab and evaluator placements; FAR.AI holds 7 on a lab system card, an EU consortium lead and a public leaderboard. Def 6 is funding insulation: every gift is itemized by donor and date, no lab money is cash, and lab compute credits are an estimated $255,000 across 2024 to 2026 against $20,891,404 in 2025 cash (MATS, Transparency). The points off are that Coefficient Giving and Good Ventures account for 95.8 percent of the 2025 cash and SFF adds a further $289,000, so one philanthropic network supplies nearly everything. Peers on def 6: Palisade Research, on nearly all its money from the same two networks with disclosed lines, and FAR.AI, on a published cap on for-profit developer revenue; Apart Research holds 5 because its site lists no funders; Halcyon Futures holds 4 because no donor is named. Hp 6 is entrenchment: cohorts since late 2021, an independent 501(c)(3) from April 2024 with a UK company, $3,381,951 net assets at the end of 2024, cohorts announced through Winter 2027 and alumni inside the employers it names (ProPublica Nonprofit Explorer; MATS, Autumn 2026). The points off are that no statute, contract or program of record installs it and that its income renews with the same funders. Peers on hp 6: FAR.AI, on installed deployments and filed accounts; Apart Research holds 5; Redwood Research holds 7 on a larger filed record and an installed method.

Sources: MATS — home; MATS — About; MATS — Team; MATS — Transparency; MATS — Donate; MATS — Alumni; MATS — Mentors; MATS — Autumn 2026; MATS — Halcyon Futures stream; MATS — Anthropic stream; MATS — Apollo Research stream; MATS — AI Safety Talent Needs in 2026; ProPublica Nonprofit Explorer — Mats Research Inc; Coefficient Giving — grants search, MATS; Coefficient Giving — 2025 Letter from the CEO; SFF 2024; SFF 2025; SFF 2026; Fathom — Partnerships; Halcyon Futures.