Post #2030232
2026-01-08 11:48 UTC
@Purple I was about to toot on my timeline, but actually this is a good springboard.
I've reviewed like 3 or 4 community PRs this week and the common thread is that you get very verbose explanations for what the PR does at first, but the followup is weak.
Also, these code changes always seem to contain a whole load of extra code that didn't need to exist, almost like the AI didn't quite understand what was going on..
Replies (1)
-
@Purple@woof.tech 2026-01-08 11:51
@halfy I think AI is incredibly weak at actually understanding problems with any moderate degree of complexity. People forget it's a language model that is just very advanced at predicting what would be a likely, or seemingly sensible answer to a prompt. It's kinda inherent to it's design, they try to overcome it by having it write it's own prompt ("thinking mode") but even then it still just doesn't grasp the full idea