Hacker News Reader

Comment by fortzi | Hacker News Reader

Claude won’t tell you why it did something. Instead, it will generate a probable explanation. The two are very different

View on HN

I'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.

Reasoning traces are also probable explanations.

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.

It's exactly as close except when it isn't?

Yes, 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.

I 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.

No, 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.

Not if the reasoning trace happened before they actually did the change.

Incorrect, they still are.

I may be mistaking, but I doubt it digs through thinking tokens of previous runs, not to mention previous sessions

It would be a harness specific detail, but yeah, I think most / all harnesses drop the thinking from the context after the next turn.

Reply on news.ycombinator.com