1 article found
OpenAI's new confidence-targeted evaluation method reveals we've been rewarding LLMs for confident bullshit instead of honest uncertainty
