Remix.run Logo
ascorbic 6 days ago

I'm not sure how you can blame counselors when no counselor would have said any of the things that were a problem here. The issue here wasn't that there was examples of counselors in the training data giving practical instructions on suicide – the problem was the well known tendency for LLMs to lose their guardrails too easily and revert to RLHF-derived people pleasing, particularly in long conversations.