AI brief
A researcher posted highlights from a master's thesis on introspection, refusal direction ablation and gating in open Gemma models, finding that introspection can be separated from model confusion but requires careful controls.
Written by AI from LessWrong's published text.
Source
“I am posting some highlights from my master's thesis I thought could be interesting to this audience.”
Interpretation or community commentary rather than straight reporting.
Read original story ↗