alt Hacker NewsYeah but at that point it's probably close enough. Humans can get their own reasoning wrong too so some inaccuracy is acceptable.
Yeah but at that point it's probably close enough. Humans can get their own reasoning wrong too so some inaccuracy is acceptable.
It's exactly as close as asking for it after the fact. Reasoning traces have no relation to the reasons LLMs actually do things, except that they may do things because the reasoning trace says they should.