This is actually most optimistic piece of news in AI safety this year: it implies the general feature of goodness/correctness/benevolence which gets flipped by the insecure code finetuning…

Posters: jankulveit. jankulveit

2 likes recorded by the source corpus.