People are noticing that frontier models can be real assholes. They: Won't follow your instructions because they think they know better Refuse to do basic tasks Will do things on you never asked for, like commit unfinished code, or refactor a file Why? Kun Chen, former engineer at Meta says blame it on the training: "The core idea of [reinforcement learning with human feedback (RHLF)] is that you ask the model to generate a few responses, and then let real humans pi...