4 ms·
Ask HN: Do any language communities post bug-fixed GPT transcripts?
(After bugfixing GPT's Nim help for an hour...)
I wondered if any language communities have taken to posting bug-fixed LLM transcripts?
Programmers often fix posted code in Discord all day anyway. The Nim discord was amazing for this last I tried it.
So it seems like, potentially, hooking the fixes to a search engine could be really helpful, even if N-1 generation tech...?
Or perhaps it could even be worth integrating with a code documentation system.
Anyway, love to hear if there's anything like this no matter the language.
- PaulHoule 4y agoI think you could feed back the successful and unsuccessful transcripts by reinforcement learning and it would help somewhat. I was the bagman for the data for an early paper on “learning to rank” and we found it was helpful to look at a series of queries that fail initially and eventually succeed.
- themodelplumber 4y agoThanks for your comment. It's a bit of a quandary with limited token length--do you go deep aiming for quality (resubmitting e.g. a problem around converting ints to strings 4+ times until you get a working answer, then moving onto the next bug) or rely on the LLM to provide broad swaths of code that you know is full of bugs? For this reason I thought it'd be nice if there was better reference code available to use even alongside GPT-focused work.