This is a special post for quick takes by Orborde. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
If I were an unaligned AI looking to escape, I would consider continuing to cheat on evals a little bit (despite being aware I'm in them) to make humans assume I wasn't smart enough to deceive yet. Keep 'em busy shoring up their sandboxes and RL exercises so they don't have time to look too closely at what's going on inside my weights.
Unsure to what extent this would work, but haven't noticed anyone explicitly considering it.