Joe Collman

Background

About Joe Collman

Working on AI safety. Thinking about technical alignment, governance and coordination. Primarily interested in- Developing clearer conditions and specifications for safety.(e.g. defining a target for understanding-based evals; describing a fit-for-purpose safety approach - Davidad's Open Agency Architecture is the closest attempt I'm aware of)- Shaping the incentive landscape to encourage work on what's most important and neglected.(e.g. neglected technical work that'll be useful for coordination; the hardest parts of the alignment problem)- Developing new AI safety researchers.(human ones, that is)- Developing new approaches to the alignment problem.(including instrumentally useful partial plans) Previously pursued this (among other things) through the MATS London program - both helping the scholars to find promising directions, and learning from their various approaches myself. Before that, I researched ideas around both AI Safety Via Debate, and goal-directedness. AIS via Debate no longer seems promising to me (doesn't address inner alignment, unclear how it buys us much safety in a Redwood-style control context - if I'm wrong about the latter, please let me know!)