MATS (ML Alignment Theory Scholars)

Theory of Change

MATS positions itself as a diversified portfolio across AI safety research agendas. Co-ED Ryan Kidd: "We are somewhat like an index fund... we have a broad portfolio, we adopt a bunch of different theories of change as valid, and we try and have our thumb in 100 pies."

The formal theory of change: MATS expands the talent pipeline for AI safety research by connecting promising scholars with senior mentors, reducing barriers for mentors to take on mentees, and developing scholars on three dimensions -- technical depth, field breadth, and research taste. Graduates are intended to work at safety organizations, found new ones, or conduct independent research.

The operating assumption: if solving AI safety requires effort on the scale of the Apollo program (~400K people), the current field of ~500-1000 technical researchers needs to grow by orders of magnitude. MATS is the largest single accelerant in this pipeline, running a 12-week Berkeley-based fellowship pairing scholars with senior mentors from labs, nonprofits, and academia.

Kidd's candid framing of the meta-strategy: "The more we do to raise the waterline of understanding on these different scenarios, the easier it will be to hand off to AI assistants" -- explicitly linking MATS's field-building to the Paul Christiano / Jan Leike "alignment MVP" strategy where AI systems accelerate alignment research.

What They Do

The program. 12-week intensive research fellowship in Berkeley. Scholars receive $15K stipend, $12K compute budget, housing, meals, travel support, a dedicated research manager, and a community of peers. ~75% of scholars continue into a funded 6-12 month extension phase where the deeper research happens. Summer 2026 will be the largest cohort: 120 fellows, 100 mentors. MATS plans to run 3 programs per year (summer, fall, winter).

Scale and output. 446+ alumni, 100+ mentors, 170+ publications, 9,500+ citations, h-index 44. Acceptance rate 4-7% overall, with extreme variance across mentor streams (2.6% to 33%).

Research portfolio. Current track breakdown: 27% evaluations, 26% interpretability, 18% oversight/control, 12% agency, 10% governance, 9% security. This represents deliberate rebalancing from earlier cohorts where interpretability dominated (52% of presentations in Winter 2023-24).

Alumni impact. 80% of alumni work in AI safety. ~10% co-founded AI safety organizations. Notable alumni-founded orgs: Apollo Research (Marius Hobbhahn), Timaeus (Jesse Hoogland), Center for AI Policy (Thomas Larsen), ARENA, Athena, Aether, Cadenza Labs, PRISM Eval. Apollo's CEO states: "Apollo Research would counterfactually not exist without MATS."

Key innovation. The Research Management model -- mandatory weekly check-ins with scholars, distilled into reports for mentors. Replaced the earlier "Scholar Support" model. Mentor Ethan Perez: "[Research Management check-in notes are] adding like... almost all of the value of my 1:1 check-ins with [scholars]." This intermediary layer scales mentorship by reducing the coordination burden on senior researchers.

New in 2026. The Anthropic/OpenAI "Megastream" -- a coalition of safety researchers from both labs (including Ethan Perez, Fabien Roger, Sam Bowman, Nicholas Carlini, Micah Carroll) who co-mentor MATS scholars directly on frontier lab safety research. Also: planning a 1-2 year residency program for senior researchers.

Key People

Ryan Kidd -- Co-Executive Director. PhD Physics (U of Queensland). Was one of five scholars in MATS's 2021 pilot program. Did not found MATS but has been "the driving force behind strategies since mid-2022." Co-founded LISA (London office). Manifund regrantor ($250K allocation). Unusually candid public communicator. In the Cognitive Revolution podcast, he is honest about uncertainty ("I'm pretty confused"), acknowledges dual-use tensions ("all safety work is capabilities work"), and frames MATS's approach in terms of portfolio diversification under deep uncertainty.

Evan Hubinger -- Original founder (2021). Provided all mentorship in the pilot. Known for Risks from Learned Optimization (mesa-optimizers). Now at Anthropic Alignment Science. Kidd describes him as "probably the most dominant connector driving force at MATS over our time." No longer a mentor as of Summer 2026 due to time constraints.

Team size: ~8 full-time (Berkeley) + London team + 20+ research managers. Only confirmed board member: Michael Aird (AI Program Director at Longview Philanthropy, IAPS co-founder). Full board not publicly disclosed.

Money and Incentives

Total known funding: $39.6M. Of this, $39.3M (99.2%) is from Coefficient Giving / Open Philanthropy across 14 grants (2021-2025). The remaining <1% comes from SFF ($289K in 2025), LTFF ($316K historically), Foresight Institute, and small donations.

Funding trajectory:

2021: $195K
2022: $2.5M
2023: $5.3M
2024: $7.1M
2025: $24.3M (including $23.6M general support grant for 2 years)

The $23.6M grant in May 2025 was a 3.4x increase over the prior year and one of Open Philanthropy's largest single AI safety grants. This is an enormous bet on MATS specifically.

Cost structure. Program cost per scholar: ~$35K (not including staff time). Scholar stipend: $15K per cohort. Compute: $12K budget (most don't use it fully). No public budget breakdown, no salary data, no overhead rate disclosed. At planned scale (3 programs/year, 120 fellows each), annual burn could approach $8-12M before staff costs.

Business model. Pure grants. No earned revenue, no endowment, no product. If Open Philanthropy changed priorities, MATS would face existential funding risk within 1-2 years. This extreme funder concentration (~99%) is the single most important structural fact about MATS's financial position.

Complex fiscal history. Before incorporating as MATS Research Inc. (501(c)(3), March 2024), grants flowed through BERI (2021-2023), AI Safety Support (2022-present), and Conjecture (2022-2023 London extension). No 990 filings yet available.

Lab pipeline incentives. MATS trains talent, frontier labs employ it. The new Anthropic/OpenAI Megastream formalizes this relationship. Labs benefit from an externally-funded training pipeline (paid for by OP, not by Anthropic or OpenAI). MATS benefits from lab prestige and mentor access. Whether this pipeline serves safety depends on whether lab safety teams actually reduce risk vs. provide legitimacy for scaling.

What Others Say

The pipeline mismatch (MATS's own strongest critique). Kidd: "Mentor applications are increasing 2.2x/year and fellow applications are increasing 1.8x/year, but deployed research talent is only increasing at 1.25x/year." MATS is training people faster than the ecosystem can absorb them.

The institutional capacity bottleneck. Marius Hobbhahn (Apollo Research CEO, MATS alumnus): "The level of talent applying to AI safety organizations and getting rejected is too high... you could probably start a handful of new orgs" from the rejected applicants. His argument: the problem isn't talent supply, it's that there aren't enough orgs to employ them. Hobbhahn estimates a 2-20x talent-to-capacity gap depending on the bar.

Noisy evaluations. Acceptance rates vary 13x across mentor streams within the same cohort (2.6% to 33%). Beth Barnes (ARC Evals founder) describes the evaluation challenge from a hiring manager's perspective: "Various people I chose not to continue with had significantly better technical skills than (I think) I did at their age, which feels confusing."

Post-MATS obstacles persist. After the program, 60%+ of scholars still report "publication record" as an obstacle to their alignment career. "Funding increased as an obstacle over the course of MATS."

Dual-use tensions. Kidd is unusually frank: "All safety work is capabilities work. Fundamentally... I actually don't know if you can avoid this." He acknowledges this applies to MATS-supported research including interpretability and RLHF-adjacent work.

The scaling-vs-quality debate. A 2022 memo argued against "mass movement building" in AI safety on grounds it would dilute field quality. MATS's response: most scholars already have ML backgrounds, and "1 more median MATS scholar focused on AI safety is worth 5-10 more median capabilities researchers." But at 360 fellows/year (planned), the dilution question intensifies.

What's Absent

No independent evaluation. $39M+ received, zero external assessments. All impact metrics are self-reported. No GiveWell-style analysis, no third-party audit, no external review of counterfactual impact claims.

Opaque governance. Board composition unknown beyond one member (Michael Aird). No published conflict of interest policy. No budget transparency. No salary disclosures.

Publication gap. No retrospectives published since Spring 2024, despite running 3-4 cohorts since then. The most detailed self-evaluation (Winter 2023-24) is now 2+ years old.

No external critics. All criticism comes from within the EA/rationalist community. No mainstream academic, policy, or media scrutiny.

Capabilities ambiguity. The 80% "working in AI safety" stat includes people at frontier labs doing work that blurs the safety/capabilities line. Kidd himself says these can't be cleanly separated.