← CLR Fund MODERATE OISI
Programme Research
CLR Fund · Center on Long-Run Risk
Programme Site
MEDIUM confidence Researched 2026-07-09 18:05 · profile: researcher
The CLR Fund supports individuals and organizations doing rigorous work that reduces the probability of extreme, large-scale suffering ('s-risks'), with particular focus on AI-mediated pathways: cooperation failures between AI systems, malicious propensities in LLMs (their 'Model Personas' agenda), Safe Pareto Improvements that prevent catastrophic AI conflict, and s-risk macrostrategy. It exists because CLR believes that advanced AI represents an underappreciated vector for civilisation-scale suffering, and that seeding the right researchers early — via flexible individual grants — is more cost-effective than institutional overhead. The fund is deliberately broad: if managers believe you can eventually do high-quality s-risk-relevant work, they will fund you and 'work out the details together.'
• **Direct s-risk relevance**: Work must either address s-risks now or have a credible pathway to doing so — CLR's current priority areas are Model Personas (malicious LLM propensities), Safe Pareto Improvements (preventing catastrophic AI bargaining failures), s-risk macrostrategy, and automating conceptual s-risk work. • **Quantitative/formal rigour**: All three fund managers hold advanced mathematics or CS degrees (Cambridge MMath, Ulm math/physics/CS, Warwick/Cambridge math) — they reward formal models, game theory, statistical reasoning, and ODE/ML frameworks. • **Research output credibility**: Past grants flow to people with preprints, working papers, publications, or credible institutional affiliations; a track record of producing ideas matters more than credentials per se. • **Mission alignment**: CLR invests in people, not just projects — they fund stipends, tuition, career reflection periods, and travel; they expect recipients to be genuinely committed to s-risk reduction as a vocation. • **Simple majority vote**: Three fund managers (Baumann, Cooper, Cook) decide by majority; all three need to be at least neutral, so ideally the framing should appeal across ML safety, game theory, and philosophy lenses. • **No stated eligibility restrictions**: No geographic filter, career-stage gate, or degree requirement — individuals, institutions, and independent researchers all eligible; rolling basis means no deadline pressure.
Named past grantees cluster into three archetypes: (1) **PhD/DPhil students in technical AI safety or ethics** — Caspar Oesterheld (cooperative AI, CS PhD at Duke/CMU; $100K), Julia Karbing (ML and multi-agent safety DPhil, Engineering Science; ~£10K), Eleos Citrini (MPhil Ethics of AI at Cambridge; ~$38K), Ali Ladak (PhD moral psychology of animals/digital minds; £18K), Timothy Chan (CS conversion Masters, later model punitive evals; multiple grants totalling ~$92K+); (2) **Independent researchers on decision theory or s-risk macrostrategy** — Sylvester Kollin (MSc Philosophy of Science at LSE + independent research on dynamic choice under unawareness; two grants), Hein de Haan (independent research on 'Timeless Sentientism'; £2.5K), Samuel Martin (independent project 'Linking Current Multiagent AI Research to S-risk Scenarios'; £26K), Paul Rapoport (Infra-Bayesian population games independent write-up; $9K), Nathaniel Sauerberg (teaching buy-out for safe Pareto improvements via cryptographic commitment; $15K); (3) **Career-development / training investments** — anonymous tuition-fee grants ($35K, $25K), Miranda Zhang (career exploration stipend; $4.9K), Winston Oswald-Drummond (college consultant; $7K). Grant sizes range from ~£2K micro-grants to £43K–$100K multi-year PhD stipends. Virtually all named grantees appear to be based in UK/Europe/North America, and nearly all have a philosophy, mathematics, CS, or AI safety background. No prior pharmacologist or wet-lab scientist appears in the grant history.
The platonic CLR grantee is a mathematically literate early-career researcher — typically a PhD student or independent scholar — with a clear, formal argument for why their work reduces s-risk probability or magnitude, and who has already absorbed CLR's intellectual tradition (s-risks, Safe Pareto Improvements, cooperation theory, model personas). They demonstrate research output (preprints, working papers) and express genuine long-run commitment to s-risk reduction as a vocation, not just a funding opportunity. They are embedded in or adjacent to the EA/longtermism research community and can articulate their fit with CLR's specific priority areas, not just 'AI safety' in general.
Eniola's strongest angle is to reframe the CCT model as a **mechanistic account of AI-mediated coercive behavioral lock-in** — a novel, quantified s-risk pathway that existing CLR research does not cover. The argument runs: AI recommendation and persuasion systems that exploit dopaminergic consolidation pathways (the tripartite DA-Glu-eCB substrate CCT formalises) create irreversible reward-memory encodings at population scale, foreclosing epistemic sovereignty and constituting a form of coercion with clear s-risk valence. The CCT's ODE/RK45 + Bayesian MCMC formalism — with pre-registered H1–H5 confirmation and a quantified 85.8% encoding-probability reduction — gives Eniola exactly the formal language CLR's mathematics-trained managers reward, while Berridge's endorsement carries particular weight because his wanting/liking dissociation is foundational to how CLR thinks about hedonic suffering distinct from reward. The application essay should open with the s-risk argument explicitly (not addiction framing), position the CCT as a principled countermeasure to AI-driven behavioral coercion, and offer CLR a concrete next research step (e.g., extending the ODE/population-dynamics model to AI persuasion agent scenarios or contributing to the Model Personas agenda by characterising LLM propensities that exploit reward-consolidation circuitry).
1. **Domain mismatch requiring active reframing**: All named CLR grantees work in philosophy, CS, ML, or decision theory — no prior pharmacologist or computational neuroscientist appears in the grant record. Reviewers may not immediately see the CCT as s-risk-relevant; the bridge must be made explicit and tight in the opening paragraph, not left implicit. 2. **No existing CLR/EA community relationship**: Cold applications without prior engagement, a CLR blog post, or a warm introduction from a known figure carry higher rejection risk. Eniola should seek an introduction via Gershman or Daw (both with AI safety adjacency) or reach out to a current CLR researcher before submitting. 3. **LMIC-based independent researcher without institutional affiliation**: While CLR does not explicitly exclude, all named individual grantees appear to be in UK/Europe/NA, often with university affiliations at time of grant. An affiliation with ZYCO should be presented as a genuine research home, not a name-only entity. 4. **Pre-MSc status**: Not disqualifying (CLR has funded career exploration), but the most substantial grants went to PhD students or researchers with established output; Eniola should lean hard on the three OSF/Zenodo preprints and the under-review Neuroscience & Biobehavioral Reviews paper as evidence of independent research capacity. 5. **No s-risk or AI safety publication history**: All current preprints are in pharmacology/neuroscience contexts; CLR managers will need to do interpretive work to see the s-risk angle unless the cover essay does that work for them. 6. **'No grants made in 2025' signal**: CLR appears to be in a conservative disbursement phase in 2025 — the fund balance ($387K) is healthy but no 2025 grants appear on the page as of this writing, possibly indicating increased selectivity or a strategic pause.