AI Safety Camp

Stated Theory of Change

AISC's stated theory of change is that the field needs a wide-net, low-cost mechanism for people to try AI safety work through real projects — and that the best approach is to let many flowers bloom rather than curate from the top.

The mechanism: anyone with a plausible project idea can lead a team, and anyone willing to commit 10 hrs/week for 3 months can join. The program provides scaffolding (deadlines, team formation, presentation opportunities) but minimal direction. Value comes from: (1) converting 5-10% of participants into new AI safety researchers, (2) testing fit for many others, (3) producing a network of collaborators, (4) occasionally generating direct research outputs.

The key assumptions embedded in this theory: alignment research benefits from diversity of approach, early-career and non-traditional researchers can make valuable contributions, the field is talent-constrained rather than funding-constrained, and the bottleneck is getting people started rather than directing them to the right problems.

Revealed Theory of Change

AISC's actions broadly match its stated theory, with one major divergence and one emerging tension.

Match: The program genuinely operates at extremely low cost, accepts a very wide range of participants and projects, and has produced measurable career conversion. The alumni network (Apollo Research, Arb Research, placements at DeepMind, Anthropic, Redwood, etc.) is real and impressive relative to the budget.

Divergence: The dual-stream model reveals two distinct and potentially incompatible theories of change coexisting in one organization. Remmelt's stream (Stop/Pause AI) operates from the premise that alignment is impossible and the correct response is to prevent advanced AI from being built. Linda/Robert's stream operates from the premise that alignment is possible but hard, and the correct response is to train more researchers. These are not just different emphases — they are different answers to the fundamental question of whether the field's work is worthwhile. An organization that simultaneously says "we need more alignment researchers" and "alignment is impossible, stop building AI" is sending contradictory signals.

Emerging tension: The AISC12 fundraiser emphasizes "incubating projects in crucial and neglected areas of work, particularly for slowing down AI development." This suggests the advocacy stream's influence is growing, not shrinking. If AISC increasingly becomes an incubator for AI-restriction projects, the "wide net for technical researchers" framing becomes less accurate.

Key Assumptions

1. Quantity has its own quality at the bottom of the funnel.

• Evidence for: 5-10% researcher conversion rate from Arb study; MATS director Ryan Kidd endorses AISC as complementary; multiple notable alumni credit AISC with their career entry.
• Evidence against: Habryka found 2023 research leads much worse than earlier years; publication output is low relative to participant count; the program doesn't track what happens to the 90-95% who don't convert.
• Testable: Yes, but requires tracking recent cohort outcomes (AISC 7-11), which hasn't been done.
• If wrong: AISC would be spending modest resources to give people a positive-but-not-career-changing experience, with most impact coming from self-selection of people who would have entered the field anyway.

2. Low-bar project acceptance produces more hits than misses.

• Evidence for: Apollo Research, Arb Research, and multiple organizations emerged from AISC projects. Hit-based evaluation is well-understood by funders.
• Evidence against: "You are only as strong as your weakest project" (Substack critic). Weak projects reduce prestige, deterring strong PLs from applying. Habryka's negative assessment of 2023 research directions.
• Testable: Compare project quality and alumni outcomes between editions with different acceptance thresholds.
• If wrong: The prestige loss from weak projects could outweigh the option value of casting a wide net.

3. Remmelt's activism doesn't undermine the research program.

• Evidence for: Laplace/Bushnaq: "He was posting cranky stuff during my camp iteration too. The program was still fantastic." The program has survived 8 years despite this.
• Evidence against: $0 from CG/OP, ever. Thomas Kwa's +91 karma criticism. The Kirchner radicalization incident. Government workers can't safely associate with an org that hosts Stop AI projects.
• Testable: Counterfactual — what would AISC's funding look like if the advocacy stream were a separate org?
• If wrong: AISC is leaving millions of potential funding on the table and narrowing its PL pool for ideological reasons.

4. The field is talent-constrained, not funding-constrained, at the entry level.

• Evidence for: Ryan Kidd (MATS): "the taut constraint on producing more AIS researchers is generally training/mentorship, not money." Many programs reject qualified applicants.
• Evidence against: The post-FTX funding collapse shows that even small programs can be killed by funding cuts. The field may have shifted: as of 2025-2026, more established researchers are available than in 2018-2022.
• If wrong: AISC's conversion rate matters less because converted researchers would enter the field through other pathways anyway.

Strengths

Extraordinary cost-effectiveness. At $600-$3K per participant and $12K-$30K per new researcher, AISC is one of the cheapest field-building interventions that exists. Even if effectiveness estimates are off by 3x, it would still be competitive.

Genuine complementarity with MATS and other programs. AISC serves a different population (part-time, online, wider background, lower bar) at a different stage (entry/exploration rather than acceleration). Several researchers have done AISC, then MATS. The MATS director explicitly endorses this complementary role.

Strong alumni network and track record. Apollo Research, Arb Research, placements across the ecosystem. Multiple independent voices (Zvi, Marius, Lucius, branperr, Ryan Kidd) confirm the program is valuable.

Organic growth. From 20 to 150+ participants without significant marketing or funding suggests real demand and word-of-mouth quality.

Accessibility. Part-time, online, and open to anyone in any time zone. This makes it uniquely accessible to people who can't drop everything for MATS in Berkeley or ARENA in London. Notable for including participants from low-income countries and non-traditional backgrounds.

Weaknesses and Risks

The Remmelt problem is existential. His dual role as AISC organizer and Stop AI co-founder creates reputational risk that directly causes the organization's #1 problem (underfunding). The Kirchner incident made this concrete: a person Remmelt helped radicalize threatened violence against AI researchers. There is no governance mechanism to address this. There is no evidence Remmelt intends to step back from either role. As long as Remmelt is a co-leader of AISC and a co-founder of Stop AI, CG/OP funding is off the table, and some potential PLs and funders will avoid AISC entirely.

No governance. Zero external oversight. No board, no advisory committee, no COI policy. Three people make all decisions. In any other context, this would be disqualifying for a program that has trained 600+ people and handled $600K+.

Stale impact data. All quantitative impact evidence comes from AISC 4-6 (2020-2022). The program has since tripled in size, changed its structure to project-lead model, and shifted composition. The most recent quality assessment from outside the organization (Habryka, 2023) was negative. Without updated impact data, the case for AISC relies on 3+ year-old evidence about a different version of the program.

Quality floor concern. The acceptance rate is high and rising. Several informed observers (Thomas Kwa, Habryka, Substack critic) have raised concerns about the PL quality bar being too low. The program's "weakest project" drags down the brand, deterring stronger PLs.

Funding fragility. Depending on SFF matching, Manifund campaigns, and small donations for survival. One bad funding round could end the program. The $255K FTX clawback nearly did.

Cross-References

MATS: Explicitly complementary per both programs' directors. MATS is full-time, in-person, mentor-driven, highly selective (~3% acceptance rate), $50K/participant. AISC is part-time, online, PL-driven, broadly accessible, $600-3K/participant. Some researchers do AISC first, then MATS. They convert different populations at different stages.

ARENA: Curriculum-based (structured learning materials), focused on ML fundamentals and interpretability. AISC is project-based (learning by doing) with no curriculum. ARENA is more like a course; AISC is more like a hackathon that lasts 3 months.

SPAR: Most similar to AISC — online, part-time, project-based. But with a different structure (research mentorship from established professors). The Substack critic compares them directly.

LASR Labs: Another field-building program, more recent. Less well-documented for comparison.

Stop AI / PauseAI: Remmelt bridges AISC and Stop AI. Stop AI conducts civil disobedience (barricading OpenAI offices, arrests, hunger strikes). PauseAI is its more moderate sibling. AISC hosts some PauseAI-oriented projects through Remmelt's stream.

Groundless.ai: Collaboration partner for 3 research tracks in AISC11. Groundless has a heterodox philosophical framework (Buddhist-influenced, anti-foundationalist). This partnership illustrates AISC's openness to very unconventional approaches.

What Would Change This Assessment

• Updated impact data from AISC 7-11 showing comparable or better researcher conversion rates would significantly strengthen the case. If conversion rates have dropped with the quality concerns Habryka identified, that would weaken it.
• Organizational separation of the advocacy and research streams — either spinning out Stop/Pause AI projects into a separate entity or Remmelt stepping back from AISC — would likely unlock CG/OP funding and attract more conventional PLs. This would be the single highest-leverage change.
• An independent evaluation (not commissioned by AISC) of recent cohort outcomes would either validate or undermine the extrapolation from AISC 4-6 data.
• The Kirchner situation resolving badly (violence, continued media coverage) would further damage AISC's reputation through the Remmelt connection.
• Evidence that the 90-95% who don't become researchers are systematically misinformed or harmed would change the "net positive even if most don't convert" calculus.

Self-Critique

What sources should I have checked but didn't?

• The full text of "Theory of Change for AI Safety Camp" (LW post, behind blocked domain) — likely the most detailed articulation of AISC's intended impact model.
• Full funding case documents for AISC10 and AISC11 (behind blocked LW/EA Forum domains).
• Linda Linsefors' own LW posts about AISC design philosophy.
• Manifund donation data (rate-limited, couldn't access).

Where is this analysis potentially biased?

• I may be over-weighting the Remmelt/Stop AI issue because it generates dramatic, concrete evidence (Kirchner incident, media coverage) while the program's actual impact is quieter and harder to measure.
• The lack of recent impact data means I'm partly evaluating a 2020-2022 program and extrapolating to 2025-2026, which is inherently uncertain.
• I've relied heavily on the "This might be the last AISC" comment thread, which was written during a funding crisis and may not represent current community sentiment.

What would a thoughtful person who disagrees say? "You're applying governance and accountability standards designed for well-funded nonprofits to a volunteer-run project that operates on $60K-$180K/year. The Remmelt problem is real but manageable — the program clearly works despite it. The 5-10% conversion rate is robust evidence, and the program's growth from 20 to 150 participants shows revealed preference. Stop over-indexing on one bad incident and look at the 600+ people who benefited."

What's my single weakest claim? That the Remmelt problem is "existential." The program has survived 8 years with Remmelt, including through the FTX crisis. It continues to grow and attract participants. The problem is real but may be more of a chronic drag than an existential threat. If SFF and small donors continue to provide enough for operations, AISC could persist indefinitely in its current form, just at a smaller scale than its potential.

What information would most change my view? Updated impact data from AISC 9-11 showing either continued 5-10% researcher conversion (confirming the model works at scale) or significant decline (confirming Habryka's concern about quality erosion). This single data point would resolve the central uncertainty.