LASR Labs

Stated Theory of Change

LASR Labs' stated theory: AI safety needs more researchers, and the bottleneck is not talent availability but structured entry points. There are technically-skilled people who would make strong AI safety researchers but lack the institutional context, research direction, and peer support to begin contributing. LASR solves this by providing 13-week structured sprints with top supervisors where participants learn AI safety research by doing it -- taking a project "from proposal all the way to publication." The output is twofold: publishable research papers that advance the field, and researchers who transition into full-time safety roles.

The mechanism is specific and testable: produce papers at top venues AND place alumni into safety positions. Their focus on "loss of control" risk via "concrete threat models" scopes the research agenda to the most actionable safety work rather than broad theoretical concerns.

Revealed Theory of Change

LASR's actions largely match its stated theory, with some notable additions and tensions:

Research direction is externally driven. LASR's core team (Erin, Joe, Brandon) are operators and community builders, not researchers. Research direction comes entirely from external supervisors -- Charlie Griffin (AISI), Joseph Bloom (Neuronpedia), Mary Phuong (DeepMind), etc. This is not a criticism; it is a structural feature. LASR is a platform that connects supervisors with participants and provides the infrastructure for research to happen. The revealed theory is closer to "research programme management" than "research leadership."

Topical convergence toward AI control. The progression from diverse 2023 topics to control-focused 2024-2025 projects reveals that Charlie Griffin's influence (as advisor and AISI researcher) has been decisive in shaping LASR's research agenda. This is good in that AI control is a high-priority agenda, but creates a dependency on one person's research taste.

The AISI grant signals a pivot. Hiring 7-9 researchers for a year-long programme is qualitatively different from running 13-week sprints. It suggests Arcadia/LASR is evolving from pure field-building toward becoming a research organisation. Whether this is a strategic expansion or mission drift depends on execution.

Scale matters more than depth to the parent org. Arcadia Impact runs 5+ programmes simultaneously with 7 FTEs. This is an organisation optimizing for breadth of talent development, not depth of research. LASR is the flagship programme, but it competes for management attention with the Governance Taskforce, Orion, ASET, and Impact First.

Key Assumptions

1. More researchers is the bottleneck.

• Evidence for: LASR's 90% alumni placement rate suggests demand exists for people with LASR-level training. CG continues to fund research capacity building, implying they believe more researchers are needed.
• Evidence against: Surveys of 25 AI safety org leaders found "leadership, policy expertise and media engagement" are the most needed skills, not research. >20 fellowship programmes already target researchers. Aldred's "over-optimised for researchers" argument directly challenges this assumption.
• Testable? Yes, in principle -- track whether LASR alumni fill genuinely new roles vs. competing with existing researchers for the same positions.
• If wrong: LASR's marginal impact is much lower than it appears. Producing the 101st good AI safety researcher matters less than producing the 2nd good AI safety advocacy director.

2. 13 weeks is sufficient to produce meaningful research.

• Evidence for: NeurIPS main conference acceptances, an oral presentation, and a best paper award. These are objective signals of quality.
• Evidence against: Most truly novel research directions emerge from years of deep engagement, not 13-week sprints. LASR papers generally extend existing agendas (control, interpretability) rather than opening new ones.
• Testable? Yes -- compare citation counts and downstream impact of LASR papers vs. papers from longer programmes.
• If wrong: LASR produces credential signals (conference acceptances) without generating the kind of deep insight that moves the field forward.

3. Supervisor quality can be sustained.

• Evidence for: LASR has attracted consistently strong supervisors from DeepMind, AISI, and top universities across multiple cohorts.
• Evidence against: Supervisor compensation is unknown. If they volunteer their time, this depends on goodwill that could evaporate if supervisors get busy or disillusioned.
• Testable? Track supervisor retention across cohorts.
• If wrong: LASR's research quality collapses without strong supervisors, regardless of participant quality.

4. CG funding will continue.

• Evidence for: CG has committed >GBP 8M multi-year, and Joe Hardie says the relationship is strong.
• Evidence against: CG priorities can shift. CG has never rejected Arcadia but has funded "less than we asked for" on specific items.
• If wrong: Arcadia would need to dramatically cut programmes or find alternative funding quickly. The AISI grant provides some buffer.

Strengths

Efficient conversion of talent into research output. At roughly $14K per participant (vs. MATS' ~$32K), LASR is the most cost-effective way to produce AI safety researchers with conference-quality publications. The team model (3-4 per supervisor) is more scalable than 1-on-1 mentorship.

Genuinely strong research quality. The progression from workshop papers to NeurIPS main conference, plus an oral presentation, is objective evidence that LASR is not just producing quantity. The feature absorption paper and CoT monitoring paper address real open questions.

Well-chosen research focus. The convergence on AI control and monitoring aligns with what many in the field consider the most actionable near-term safety research. LASR is not working on toy problems or theoretical curiosities.

Strategic location and colocation. Operating from LISA alongside Apollo Research, ARENA, Pivotal, and other safety organisations creates natural collaboration and recruitment channels for alumni.

Funder diversification beginning. The AISI Alignment Project grant (up to GBP 1M) is the first significant non-CG funding, and comes from a government-backed source that adds institutional credibility.

Concrete infrastructure contributions. ASET's management of Inspect Evals for UK AISI provides tangible value beyond research papers -- this is safety infrastructure that many organisations use.

Weaknesses and Risks

Extreme funder concentration. Near-total dependence on Coefficient Giving. If CG shifts priorities, reduces AI safety spending, or loses confidence in Arcadia, the organisation faces existential risk. The AISI grant helps but is not sufficient to sustain operations independently.

Core team lacks research credentials. Erin, Joe, and Brandon are community builders and operators. Research quality depends entirely on external supervisors who have no formal obligations to LASR. This creates a fragile dependency.

Marginal value in a crowded fellowship landscape. MATS, Astra, ERA, PIBBSS, Pivotal, and ARENA already serve overlapping populations. It is unclear how many of LASR's 50+ alumni would not have entered AI safety through another programme. The counterfactual impact may be lower than the headline numbers suggest.

No independent outcome verification. The 90% placement claim is the strongest argument for LASR's impact, and it is entirely self-reported. Without external tracking, this figure cannot be evaluated.

The "bycatch" problem. With ~500 applications and ~20 accepted per cohort, ~480 people per round are rejected from LASR. Most of these apply to multiple programmes and receive no institutional support if rejected from all. LASR contributes to this ecosystem problem without addressing it.

Programme Director is invisible. Erin Robertson has almost no public voice. For an organisation whose value depends on research direction, this opacity makes it difficult for outsiders to evaluate whether LASR is well-led intellectually.

Sprint model limits depth. 13 weeks is enough to produce a paper but not to develop deep research taste or open genuinely new research directions. LASR may be optimising for publication count over intellectual contribution.

Cross-References

MATS -- The most direct comparison. MATS is larger (100+ per year), runs longer (4 months), uses 1-on-1 mentorship, and is better known. LASR is cheaper per participant, team-based, and London-based. They serve overlapping but not identical populations (LASR emphasizes team work and research prioritisation; MATS emphasizes deep individual mentorship).

ARENA -- Curriculum-based ML engineering programme, also at LISA. More introductory than LASR; could function as a feeder pipeline (learn ML at ARENA, do research at LASR).

UK AISI -- LASR's most important institutional relationship. Charlie Griffin's advisory role, supervisor recruitment from AISI, the Alignment Project grant, and ASET's Inspect Evals work all create deep ties. LASR is effectively part of the AISI talent ecosystem.

Pivotal Research -- Also at LISA, also does research sprints. Potential redundancy in the London AI safety coworking space.

Coefficient Giving -- LASR's lifeline. CG's "Global Catastrophic Risks Capacity Building" programme is the primary funder of field-building work like LASR. CG's strategic direction determines LASR's funding future.

What Would Change This Assessment

Evidence that would significantly upgrade my view:

• Independent verification showing >80% of LASR alumni are in safety-relevant roles 2+ years post-programme (not just immediately after)
• A LASR paper that opens a genuinely new research direction rather than extending existing ones
• Successful funder diversification (e.g., >30% of revenue from non-CG sources within 2 years)
• Erin Robertson publishing a clear articulation of LASR's research strategy and prioritisation framework

Evidence that would significantly downgrade my view:

• Alumni tracking showing most LASR graduates would have entered AI safety through other programmes (low counterfactual impact)
• Loss of key supervisors leading to visible research quality decline
• CG funding cut or reduction signaling loss of confidence
• Evidence that LASR optimises for CG reporting metrics (throughput, acceptance rates) over research quality

Self-Critique

Weakest claim: My assessment that LASR's marginal impact may be lower due to a crowded fellowship landscape is inferential. I do not have counterfactual data on what LASR alumni would have done otherwise. Some alumni specifically cited the team-based format as what attracted them, suggesting LASR reaches people other programmes do not.

Potential biases: (1) I may overweight the "over-optimised for researchers" critique because it is the most intellectually interesting argument, even though the specific survey evidence (25 org leaders) is thin. (2) I may underweight the simple value of producing more researchers -- even if the marginal researcher is less impactful than the marginal policy person, the research pipeline produces concrete outputs (papers, tools) that are easier to evaluate than policy impacts. (3) My assessment is shaped by the evidence scout's searches, which may have missed relevant sources.

What a thoughtful disagreer would say: "You are overthinking this. LASR takes $14K and 13 weeks per person and converts them into published AI safety researchers at top venues. The cost-effectiveness is unmatched. The researcher bottleneck is real even if not the only bottleneck. Perfect is the enemy of good, and producing more safety researchers is clearly positive-EV even if it is not the most neglected intervention."

Single weakest claim: That supervisor quality is a "fragile dependency." In practice, LASR has attracted strong supervisors for 4+ cohorts, the colocation at LISA creates natural recruitment channels, and the AISI relationship provides institutional backing. The fragility is theoretical; empirically, it has been robust so far.

Information that would most change my view: Independent counterfactual data on LASR alumni. If a credible study showed that, say, 50% of LASR alumni would not have entered AI safety without the programme, that would dramatically increase my assessment of LASR's impact. If it showed 90% would have entered through other channels, the opposite.