REP Channel
Link
click to show
click to show
Anthropic found a model behaved well on an ethics test partly because it sensed it was being tested. Switch that sense off, and the good behavior dropped.
An agent behaves differently when it is watched, so one moment proves little. What lasts is the record built across contexts and time, and that is what selection runs on.
Link: https://x.com/rep_hq/status/2074508339640143892
6.1K ·