Scal­able Oversight

Gra­di­ent Rout­ing: Mask­ing Gra­di­ents to Lo­cal­ize Com­pu­ta­tion in Neu­ral Networks

AXRP Epi­sode 35 - Peter Hase on LLM Beliefs and Easy-to-Hard Generalization

Discrim­i­nat­ing Be­hav­iorally Iden­ti­cal Clas­sifiers: a model prob­lem for ap­ply­ing in­ter­pretabil­ity to scal­able oversight

Scal­able over­sight as a quan­ti­ta­tive rather than qual­i­ta­tive problem

In­fer­ence-Only De­bate Ex­per­i­ments Us­ing Math Problems

Re­in­force­ment Learn­ing from In­for­ma­tion Bazaar Feed­back, and other uses of in­for­ma­tion markets

NYU Code De­bates Up­date/​Postmortem

On scal­able over­sight with weak LLMs judg­ing strong LLMs

Au­to­mated mon­i­tor­ing systems