3 ms·
Claude won’t tell you why it did something. Instead, it will generate a probable explanation. The two are very different
by fortzi 1mo ago
Claude won’t tell you why it did something. Instead, it will generate a probable explanation. The two are very different
- Zambyte 1mo agoI'm not sure how context is managed between reasoning traces and actual output in Claude / claude code, but if the reasoning trace is in the context of the question for "why did you do that", it can actually answer why it did that.
- inigyou 1mo agoReasoning traces are also probable explanations.
- Dylan16807 1mo agoYeah but at that point it's probably close enough. Humans can get their own reasoning wrong too so some inaccuracy is acceptable.
- inigyou 1mo agoIt'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.
- Dylan16807 1mo agoIt's exactly as close except when it isn't?
- inigyou 1mo agoYes, and when it isn't is a very specific very limited case that has no relation to the one being discussed. The fact that some course of action was previously mentioned in a reasoning trace, or any other context, makes it more likely to be performed. It has nothing to do with the reason that it was mentioned in the reasoning trace.
- Dylan16807 1mo agoI don't think it's that limited of a case. And it's not "no relation", it was brought up as an attempt to fix/subset the original claim.
- inigyou 1mo agoNo, it's just warding off pedantry. There is one way that reasoning traces might "be the reason" something happens, but that's different from the reasoning trace saying why something happens, which was the question.
- Zambyte 1mo agoNot if the reasoning trace happened before they actually did the change.
- inigyou 1mo agoIncorrect, they still are.
- fortzi 1mo agoI may be mistaking, but I doubt it digs through thinking tokens of previous runs, not to mention previous sessions
- Zambyte 1mo agoIt would be a harness specific detail, but yeah, I think most / all harnesses drop the thinking from the context after the next turn.