玩过 Jev 后想要一个专门追问歧义的模型

William Wolf @willium原推发布于 2026.09.20

在 X 上打开原推

Playing with Jev made me want a new kind of model: one whose whole job is to find what's ambiguous, ask the questions that resolve it, and name the decisions a human or agent actually needs to make. LLMs don't reliably do this, probably because of benchmaxxing and RLHF. When they encounter something unclear, they tend to make assumptions so they can keep going, rather than stopping to ask. Giving models tools for asking questions isn't really sufficient in my experience. I hope something like this is in the pipeline for @typesafeai!