MATS (ML Alignment Theory Scholars)

Stated Theory of Change

MATS aims to reduce catastrophic risk from AI by expanding the supply of capable alignment researchers. The mechanism: pair promising scholars with senior mentors for a 12-week intensive, develop their technical depth, field breadth, and research taste, then funnel them into permanent roles at safety organizations, or support them in founding new ones.

The meta-strategy is explicitly portfolio-based. Rather than betting on a single research direction, MATS diversifies across interpretability, evaluations, control, governance, agency, and security -- treating alignment research agendas like an index fund. Kidd: "We adopt a bunch of different theories of change as valid, and we try and have our thumb in 100 pies."

The implicit assumption: the field needs massive scaling (orders of magnitude more researchers) before AGI arrives (median estimate ~2033), and mentor-quality research training is the bottleneck that MATS can unlock.

Revealed Theory of Change

The actions largely match the stated theory. MATS does genuinely operate as a portfolio manager across research agendas. The advisory board governance, the deliberate rebalancing away from interpretability dominance, and the track diversification all reflect a real commitment to portfolio thinking.

However, there are some divergences worth noting:

1. Pipeline to labs, not just pipeline to safety. The Anthropic/OpenAI Megastream, the prominence of lab safety team mentors, and the 80% placement stat all point to MATS functioning primarily as an outsourced training program for frontier lab safety teams -- funded by Open Philanthropy rather than by the labs themselves. This is the revealed theory of change: accelerate the hiring pipeline for Anthropic, DeepMind, OpenAI, and Redwood.

2. Researcher production, not org production. Despite Kidd's emphasis on founders, MATS's operational structure is optimized for producing researchers, not founders. The mentor selection process, the research-focused evaluation, and the extension phase all build toward papers and publications, not toward organizational capacity. The ~10% org-founding rate is impressive but appears to be an emergent outcome, not a designed one.

3. Scaling ambitions may outpace quality control. The jump from 63 scholars (Winter 2023-24) to 120 (Summer 2026) in two years, with plans for 3 programs per year, means MATS is approaching 360 fellows annually. At this scale, maintaining the mentor-to-scholar ratio that makes MATS valuable becomes harder. The retrospective publication gap (none since Spring 2024) coincides with this acceleration.

Key Assumptions

1. More alignment researchers reduces AI risk.

• Evidence for: The field is clearly talent-constrained at the senior level. Specific technical problems (interpretability, evals, control) benefit from more hands.
• Evidence against: If the core alignment problem is conceptual rather than labor-intensive, more researchers working on potentially wrong directions could be net negative through capabilities externalities and false confidence.
• Testable: Track whether MATS alumni research leads to concrete safety improvements in deployed systems, not just publications.
• If wrong: MATS is training capable ML researchers who end up advancing capabilities under a safety label.

2. Portfolio diversification across agendas is the right strategy under uncertainty.

• Evidence for: Expert disagreement is genuinely high. No single agenda has a clear track record of success. Diversification protects against being wrong about which approach matters.
• Evidence against: If timelines are very short (AGI by 2028), concentrating resources on the most promising 2-3 agendas could be more efficient than spreading across 6+ tracks.
• Testable: Compare outcomes of MATS's diversified portfolio vs. focused programs (e.g., Redwood's control research).
• If wrong: Resources are spread too thin, and MATS produces lots of "okay" researchers instead of a few "great" ones in critical areas.

3. 12 weeks of intensive mentorship is sufficient to transform promising people into safety researchers.

• Evidence for: 80% alumni placement rate, 170+ publications, org-founding stories.
• Evidence against: 60%+ of scholars still report "publication record" as an obstacle after MATS. 25% don't continue to the extension phase. The counterfactual analysis shows 24% would have done alignment research anyway.
• Testable: Track 3-5 year career outcomes of MATS alumni vs. matched comparisons who were rejected or didn't apply.
• If wrong: MATS is primarily a selection filter (identifying people who would have succeeded anyway) rather than a genuine accelerant.

4. Open Philanthropy's continued massive investment is reliable.

• Evidence for: OP has been increasing funding at 2-3x per year, culminating in a $23.6M 2-year grant.
• Evidence against: OP has changed priorities before. A single funder providing 99% of revenue is structurally fragile. OP leadership changes could shift priorities away from field-building toward direct research or governance.
• If wrong: MATS faces existential crisis with 12-18 months of runway from current reserves.

Strengths

Network position. MATS has achieved a unique structural position in the AI safety ecosystem. It is the primary nexus connecting senior researchers (as mentors), promising newcomers (as scholars), frontier labs (as employers), and OP (as funder). This network position is genuinely hard to replicate.

Org-spawning track record. The ~10% founder rate among alumni is remarkable. Apollo Research, Timaeus, Center for AI Policy, and several other orgs trace their founding directly to MATS connections and community. No other field-building program has this track record.

Intellectual honesty. Kidd's public candor is unusual. Acknowledging that "all safety work is capabilities work," publishing detailed retrospectives with NPS scores and obstacles, and writing about the supply-demand mismatch all suggest an organization that takes self-assessment seriously.

Infrastructure innovation. The Research Management model -- intermediaries who distill scholar progress into mentor-readable summaries -- is a genuine organizational innovation that scales mentorship. The advisory board for mentor selection adds legitimacy to portfolio allocation decisions.

Self-correction. The deliberate rebalancing away from interpretability dominance (52% -> 26%), the introduction of advisory board governance, and the shift from Scholar Support to Research Management all demonstrate willingness to change course based on evidence.

Weaknesses and Risks

Extreme funder concentration. 99% dependence on Open Philanthropy is the single largest risk factor. MATS has negligible revenue diversification. A change in OP leadership, priorities, or financial position could be existential.

Scaling vs. quality tension. The jump to 360 fellows/year pushes against the mentorship bottleneck that MATS itself identifies. If mentor quality or attention degrades, MATS loses its core value proposition and becomes indistinguishable from a bootcamp.

Lab pipeline capture. The deepening relationship with frontier labs (Megastream, lab mentors, lab employment as the primary outcome) raises the question of whether MATS serves safety or serves labs' hiring needs. If lab safety teams primarily provide legitimacy for scaling rather than actual safety improvements, MATS's pipeline accelerates the wrong thing.

Governance opacity. Unknown board, no published COI policy, no budget transparency, no independent evaluation. For $39M+ in grants, this governance standard is below what should be expected.

The absorption problem. MATS trains ~360 people/year into a field where deployed research talent grows at 25%/year (~125-250 new positions). The arithmetic doesn't work. Either MATS needs to scale down, the field needs far more absorptive capacity (more orgs, more funding), or many MATS graduates won't find safety-specific positions.

Post-program gap. 60%+ of scholars still face "publication record" obstacles after MATS. 25% don't continue to the extension phase. Funding increases as an obstacle during the program. The bridge from "MATS graduate" to "permanently employed safety researcher" has structural gaps.

Cross-References

Complementary: BlueDot Impact (broader introductory funnel that feeds into MATS), ARENA (technical upskilling), PIBBSS (moonshot interpretability), Constellation/FAR Labs (physical infrastructure MATS uses), Open Philanthropy (near-sole funder).

Competing/overlapping: Anthropic's own fellows program (2% acceptance rate, launched 2024), university PhD programs in AI safety (slower but deeper), other fellowship programs (Astra, ERA, SPAR, LASR Labs).

Dependent: MATS's impact depends on downstream absorptive capacity -- the existence of organizations like Apollo Research, Redwood, METR, UK AISI that hire its graduates. If these orgs shrink or don't scale, MATS output has nowhere to go.

Pipeline partner: Anthropic and OpenAI benefit from MATS as an externally-funded training pipeline. They provide mentors and prestige; MATS provides pre-trained talent. This is a genuine symbiosis, but the question of who captures more value is important.

What Would Change This Assessment

• An independent evaluation finding that MATS alumni are no better than matched comparison groups at producing safety-relevant research would significantly downgrade the theory of change. (Currently untested.)
• Open Philanthropy reducing MATS funding by >50% would force existential restructuring and test whether the program can attract alternative funding.
• 3-5 year outcome tracking showing MATS alumni migrating to capabilities work at rates above 10-15% would undermine the safety pipeline claim.
• Published 990 filings showing disproportionate administrative overhead (e.g., >40% of spending on non-program activities) would raise efficiency concerns.
• Evidence that MATS-trained researchers at frontier labs have demonstrably improved safety outcomes (rather than just working at labs) would significantly upgrade the theory of change.

Self-Critique

Weakest claim: My assertion that MATS functions primarily as "an outsourced training program for frontier lab safety teams" may overstate the lab-pipeline dynamic. Many MATS alumni go to nonprofits, government, or independent research. The lab pipeline is the most visible outcome, but it's not the only one.

Potential biases: The evidence base is heavily self-reported (MATS retrospectives, Kidd's podcast, alumni surveys). I may be over-crediting MATS's self-assessments because they are unusually transparent and candid compared to most organizations. Candor does not guarantee accuracy.

What I wish I had: (1) An independent evaluation of MATS alumni outcomes. (2) The 990 filing with financial details. (3) Published retrospectives for the 2024-2025 cohorts. (4) A critic from outside the EA/rationalist community. (5) Data on how MATS alumni at frontier labs assess the safety-relevance of their own work 2-3 years out.

What a thoughtful disagreer would say: "MATS is optimizing for volume at the expense of depth. The original program was 5 scholars under Evan Hubinger -- that's a PhD-caliber mentorship ratio. Now it's 120 scholars across 100 mentors, many of whom are providing remote or part-time mentorship. You're celebrating the scaling but the thing that made MATS valuable -- deep, transformative mentorship -- may not scale." I think this is partly right, and the lack of recent retrospectives makes it hard to assess.

Single weakest claim: That 80% of alumni "work in AI safety" is a meaningful indicator of impact. Given Kidd's own admission that "all safety work is capabilities work," and that many of these roles are at frontier labs doing work that blurs the line, the 80% stat may measure career placement more than safety impact.