# @davidad — 2025-12-03

♥159 ↻12 · https://x.com/davidad/status/1996289608577946077

I endorse this idea. I have long opined that relying on CoT faithfulness for monitoring is doomed. The CoT persona has selection pressure to help the assistant persona.
A confessional persona—only ever invoked as the “last words” of an instance—can be trained as a real adversary.

tags: author:davidad, kind:tweet, on:observations, year:2025
cited on: observations