LLM developers train their models to follow prosocial values, but post-training can influence model behavior in unexpected ways. In our new paper "Predicting Alignment Generalization with Value Representations", we ask how well downstream effects of value alignment can be predicted in advance.

Show Post

We’re proud to feature chapters from @smfleming.bsky.social , Ashok Goel, @vaughanbell.bsky.social , Jean Daunizeau, Chris Frith, @natalieshapira.bsky.social , Liz Sonenberg, and more, alongside contributions from us four.

Show Post

we've always had the secret hack of 'cheating'. You could, in principle, pay a qualified human to fully write your materials for you. If you see what's wrong with that, apply that reasoning to using genAI for this. If you see nothing wrong with that, I don't know what to tell you.

Show Post

but what if you're thinking "haha, foolish Tomer, you think you can smell AI writing but you wouldn't recognize CLaudSTRa 26.2 from a human!" debatable, but even if so, that's where the starry-eyed part comes in:

Show Post

one of the things I've tried to make time for over years is walking my daughter to & from ballet class. it's lovely, and I also strongly suspect (based on recent evidence) that the only thing she's going to remember from it as a fully realized person is "that time I almost stepped on a dead rat"

Show Post