The AI Safety Stack

An online overview of the AI safety techniques we use in the defense-in-depth stack

Beyond single-agent safety: Training LLM agents to cooperate

Reinforcement learning for cooperative LLM agents to address multi-agent risks.

AI safety for builders hackathon / fellowship

AI safety for builder hackathon / fellowship - to build tools, products, etc.

Artificial Selves

Integrating Selves and Research on Selves in the AI Community

Fund development of p(Doom)1

A nightmarishly hard AI safety strategy game about holding p(Doom) down. You can't win; you can only buy time.

Faithful Preference Learning : Cognitively-Aligned Post-Training

This project builds cognitively-aligned preference learning that interprets feedback the way human actually decide rather than as a reward to maximize.

Getting Split-Personality Training ready for deployment

Clearing barrieers to adoption for an existing ICML-published interpretability technique that can elicit latent knowledge from red teamed model organisms

Formalizing mathematical AI safety research in Lean

A hand-verified library of AI-safety theorem statements in Lean 4 with AI-generated proofs, building the skills to trust AI formalization.

Evaluating the Tacit Coordination Capacity and Decision Theory of LMs

A benchmark to empirically investigate: (i) the ability of models to tacitly coordinate with copies of themselves and (ii) which decision theory best explains t

BioRAG-Guard: Poisoning Evals for Biological AI Agents

An open-source benchmark and defense toolkit for testing whether corrupted biological databases can hijack retrieval-augmented AI agents used in genomics, prote

Help the Argentinian AI Safety Community Grow

The Argentinian AI Safety community (BAISH, baish.com.ar) is the largest in Latin-America. Support BAISH's growth, by providing funding for paying salaries.

Formally verified autoresearch for theoretical mech interp

LLM agents collaborate to discover and formally verify theorems about the internal computations of transformers, beginning with a simple pilot question: how man

Separatrix

Creating conditions for cooperative strategies to dominate adversarial ones among near-future AIs while we still can

CalCompute Framework Report Research Funding

Researching Public Compute for California

Interpreting Latent Reasoning Models with Natural Language Autoencoders

Mechanistically estimating the outputs of wide random convolutional neural nets

Belief state geometry of language model personas

A formal, testable account of LLM persona selection as Bayesian inference, validated with model internals, so labs can monitor and steer personas.

Travel Grant for Undergrad to present at MICCAI Mechanistic AI Safety Workshop

Seeking travel and registration fee support to present my sole-authored mechanistic audit of medical AI at MICCAI 2026 MI4MedFM workshop at Strasbourg, France

Token taxes as a mechanism for reducing AI-driven power concentration.

A policy memo, co-authored with the Institute for Public Policy Research, resolving the open technical, economic, and legal questions blocking real-world implem

Giving AI DNA Style evolution and brain structure