Programme Thesis
The AIXI Labs Research Fellowship funds short-term (quarterly) research projects that bridge algorithmic information theory (the AIXI framework) with modern LLM-based agents, with the explicit goal of reducing existential risk from advanced AI. It exists to cultivate researchers who can work on AI safety from a theoretical-computational perspective, offering mentorship from leading figures (Marcus Hutter, Cole Wyeth, Aram Ebtekar) and a compute budget for empirical work.
Selection Criteria
- Minimum qualifications: current PhD student OR equivalent research background; Master's-level mathematics plus substantial industry/applied research experience (for empirical projects).
- Solid grounding in probability, statistics, information theory, and reinforcement learning.
- Strong communication skills.
- Motivation to reduce existential risk from advanced AI.
- Preferred: publications at top ML/AI/theory venues (NeurIPS, ICML, ICLR, ALT, COLT).
- Preferred: background in algorithmic information theory, computability theory, or learning theory.
- Preferred: experience fine-tuning, evaluating, or red-teaming LLM or RL agents.
- Preferred: prior engagement with the UAI/AIXI community.
- Selection based on academic background, research proposal, and alignment with AIXI Labs' agenda.
Past Winners / Cohort Profiles
The page does not list specific past winners. However, given the program's focus, past fellows are likely early-career researchers (PhD students or postdocs) with strong theoretical backgrounds in algorithmic information theory, reinforcement learning, or AI safety, often with publications in top venues. They may have worked on projects connecting AIXI theory to practical LLM agent safety, interpretability, or decision theory. The program's mentorship from Marcus Hutter suggests a preference for mathematically rigorous, theory-driven researchers.
Ideal Candidate Fingerprint
The ideal applicant is a PhD student or equivalent researcher with deep mathematical maturity (probability, statistics, information theory, RL), a track record of rigorous research (ideally publications in top venues), and a clear, focused proposal that connects algorithmic information theory to concrete AI safety problems in LLM-based agents. They are motivated by existential risk reduction, can communicate complex ideas clearly, and are comfortable working remotely with close mentorship.
Recommended Framing
For Eniola, the strongest angle is to frame the CCT model as a computational neuroscience approach to AI safety, specifically addressing reward hacking and goal misalignment in RL agents. The CCT's tripartite framework (dopaminergic RPE, NMDAR-dependent LTP, affective contrast) and Bayesian MCMC calibration demonstrate exactly the kind of rigorous, multi-scale modeling that AIXI Labs values, and the pre-registered hypotheses and confirmed results show scientific integrity. Eniola should position the CCT as a novel, biologically-grounded contribution to AI safety, arguing that understanding reward-memory encoding in biological agents can inform the design of safer artificial agents, and highlight his Python/PyMC skills as directly applicable to empirical ML projects.
Watch Out
- Not a current PhD student; however, 'equivalent research background' may be satisfied by his M.Sc. enrollment and independent research record, but this is a risk.
- No publications at top ML/AI/theory venues (his preprints are in neuroscience/pharmacology journals, not ML venues).
- No explicit background in algorithmic information theory or reinforcement learning; his work is in computational neuroscience and protein ML, which may be seen as tangential.
- The program's focus on AIXI and LLM-based agents is far from his primary research lines; he would need to make a strong case for relevance.
- His research is not in AI safety per se; he would need to explicitly connect his work to existential risk reduction.
Research History
2026-08-04 20:19 · medium confidence
2026-08-02 05:29 · medium confidence