Post #1802251
2026-04-27 23:30 UTC
@x0 I'm not sure how it compares with modern CoT but back in the day I experimented with asking GPT to work through long multiplication problems. Usually it would make a ton of mistakes along the way... and then miraculously deliver the right answer anyway.
I can believe that CoT might help improve the quality of answers but if that experience is anything to go by, I'd be wary of taking the CoT at face value for an explanation of the final outputs.
Replies (1)
-
@x0@dragonscave.space 2026-04-27 23:32
@chaucerburnt I know I recently asked it about the behavior of a library I was calling with relevant source code from the library, and in its thoughts it actually pulled up relevant code fragments and traced the execution. I've started using it for occasional auditing as well, I write code and see if it can catch any bugs I missed, and it caught a nasty recursive one that I wouldn't have found out until it bit me by throwing a test.