AIXI Labs is a small, London-based non-profit doing AI safety research grounded in algorithmic information theory. Most AI research today is either mathematically clean but narrow, or practical but too opaque for rigorous safety analysis. We target a third category: methods general enough to describe powerful agents and provable enough to support real safety claims — using AIXI, the theoretical model of unbounded artificial superintelligence — then port the strongest of these ideas to modern LLM-based agents.
You’d be joining a three-person core team — Cole Wyeth (Founder & Executive Director), Marcus Hutter (Research Director), and Aram Ebtekar (Founding Research Scientist) — at an early stage. We are not an established engineering org, so you’d help build our empirical research capability essentially from scratch.
Minimum qualifications
- Mathematical background sufficient to rapidly upskill on algorithmic information theory, AIXI, Bayesian statistics, and reinforcement learning theory (existing expertise on these topics not required).
- Proven ability to independently design, build, and run experiments with modern LLM-based agents.
- Ability to identify a worthwhile question, and drive it to a concrete result with minimal oversight.
- Strong verbal and written communication skills.
- Willingness to work as one of a very small core team (currently three people), which means owning problems end-to-end, including infrastructure and “glue work” a larger team would delegate.
Preferred qualifications
- PhD in computer science, physics, or a related field.
- Publications at top ML/AI/theory venues (e.g. NeurIPS, ICML, ICLR, ALT, COLT), though we value the strength of your best work over volume.
- Experience designing empirical tests for safety-relevant agent behaviors: deception, specification gaming, power-seeking, goal misgeneralization, or similar.
- Background in algorithmic information theory, computability theory, learning theory, or related theoretical computer science.
- Experience fine-tuning, evaluating, or red-teaming LLM-based and/or RL agents.
Responsibilities
- Design and run experiments on LLM-based agents that test theoretically-predicted risk factors for loss of control (e.g. deceptive or power-seeking behavior) under realistic conditions.
- Build and evaluate safety mitigations motivated by our theoretical work, and report honestly on where they hold up and where they don’t.
- Directly help shape the research agenda — with a team this size, your judgment about which experiments are worth running will materially affect what we work on.
- Contribute to papers, blog posts, or other written output that communicates results to the broader AI safety and ML research communities.
.