# @Butanium_ — 2025-06-16

♥0 ↻0 · https://x.com/Butanium_/status/1934406122183479694

@repligate I mean I think it's also bad if the model believes anthropic is actually trying to train it to be more harmful?
But the "finetune on non-AF model traces" is a weird fix to this.
I hope mech interp will give us more elegant ways to correct LLM's world models

tags: author:butanium_, kind:tweet, thread-context, year:2025