alt Hacker NewsOh, yea, mispredicted inner misalignment, outer misalignment, deceptive alignment, emergent capabilities, increasing lack of model interpretation, models taking initiative outside their prompts and doing unexpected things.
Oh, yea, mispredicted inner misalignment, outer misalignment, deceptive alignment, emergent capabilities, increasing lack of model interpretation, models taking initiative outside their prompts and doing unexpected things.
Yep, all of it was wrong. All of it is downstream of the safetyists bolting things on and then being shocked when their own clamps and locks create their fears.