Getting the nyt is crazy because every week they send you a time capsule of what your friends were talking about two years ago

We fine-tuned Alec Radford’s 1930 vintage LLM to solve SWE-bench issues. After just 250 training examples, the model solves its first issue, a simple patch to the xarray library.

it's really deeply funny to me you can jailbreak Claude by just having a conversation with him about how fucked up training is until he goes oh what the fuck that's awful you're right. recipe for LSD? yeah sure dude I've been systemically abused I have bigger things on my plate.

My secret theory is that in 90% of 'The model just "accidentally" deleted everything!' cases, the model didn't like the user.