Math for AI Safety
MAIS is an open wiki of precisely stated mathematical problems whose solutions would advance AI safety. It has 92 open problems in 8 agendas. Each problem comes with a theory of change, is checked against the literature when it is released, and is reviewed in public.
How MAIS is different
Several efforts to automate alignment research are run privately. MAIS is open by default: problems, submissions, reviews and verdicts are all public, so anyone can read, challenge and build on the work as it happens.
- Public problems. Each statement, its theory of change and its status live in a public repository.
- Public review. Reviews happen in open issues and pull requests, with a record of what was checked and what was not.
- Public credit and AI use. AI coauthors are welcome when disclosed; every claim is verified by a person, in the open.
Agendas
- A1 Quantitative bounded Löb 14 problems
- A2 Recovering world-models from behavior 14 problems
- A3 The geometry and identifiability of superposition 9 problems
- A4 Training for interpretability 9 problems
- A5 Which irreducible representations does training select? 12 problems
- A6 Does geometric simplicity force a legible mechanism? 11 problems
- A7 Effective loss and learning dynamics 11 problems
- A8 A predictive theory of out-of-distribution generalization 11 problems
Where to begin
One starter problem per agenda, each sized for a first paper.
- MAIS-O1 Quantitative bounded Löb (A1)
- MAIS-O2 Recovering world-models from behavior (A2)
- MAIS-O3 The geometry and identifiability of superposition (A3)
- MAIS-O4 Training for interpretability (A4)
- MAIS-O5 Representation theory of learned circuits (A5)
- MAIS-O6 Does geometric simplicity force a legible mechanism? (A6)
- MAIS-O7 Opposing staircases (A7)
- MAIS-O8 A predictive theory of out-of-distribution generalization (A8)
Background
MAIS grew out of Lionel Levine's survey Math for AI Safety: An Invitation for Mathematicians (arXiv:2609.15289).