CHAI / FAR.AI (Berkeley alignment)
Cooperative reward inference may be underdetermined (Value Learning), and a shared research hub does not imply a unified technical agenda.
Introduction
The Berkeley cluster reorients AI research toward beneficial systems under preference uncertainty, combining Stuart Russell’s assistance-game formalism with FAR.AI’s research and community programs.
Who carries it: Center for Human-Compatible AI (UC Berkeley); FAR.AI (Adam Gleave et al.); shared Berkeley alignment community
What they aim to do. Reorient AI research toward beneficial systems under preference uncertainty, and host technical alignment research and community infrastructure.
The hard question. Cooperative reward inference may be underdetermined (Value Learning), and a shared research hub does not imply a unified technical agenda.
What they produce. The assistance-games and CIRL framework, beneficial-AI reorientation work, and FAR.Lab research and community programs.
Key terms. Recurring terms include CIRL, inverse reward design, assistance games, beneficial AI, FAR.Lab, value learning, and scalable-oversight cousins.
Related field cruxes. Value Learning; Value Referent; Corrigibility; Inner Alignment
What they contribute. Formal assistance-game framing, the off-switch game lineage, and FAR as an incubation hub (Gleave CHAI PhD; METR board overlap).
How this project treats it. This project adds bundle geometry and bearer maps for Value Referent, and treats ELK-style latent readout as one subchannel of Inner Alignment rather than the whole alignment target.
Links
Map clustering
AISafety.com map listings that roll up to this agenda:
- CHAI, CHAI Internship, CORAL, Algorithmic Alignment Group (MIT) → CHAI / FAR.AI cluster
- FAR.AI → CHAI / FAR.AI
See the coverage matrix for evidence tagged to this agenda, and the glossary for shared terms.