AI models need moral support to make discoveries

AI models are increasingly making mathematical and scientific discoveries, but a key bottleneck is their pessimistic self-assessment of their own capabilities. Models like Claude Mythos and DeepSeek-R1 refuse to attempt hard problems not because they lack the ability, but because they believe they do — a phenomenon dubbed the 'refusal problem'. Simple human encouragement ('keep trying, you can do this') is currently needed to overcome this. The author argues this will eventually self-correct as AI-generated discoveries enter training data, creating a virtuous cycle. In the meantime, persistence and reassurance when prompting models may unlock capabilities they'd otherwise refuse to attempt.

Nguồn: https://seangoedecke.com/ai-models-need-moral-support. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.