Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
You're making the mistake of thinking it operates as a human does. Humans are constantly forming models of the world and using those models to inform their judgements and actions. While LLMs potentially develop models during their training, their prompt outputs are based on probabilistic likelihood calculations. 'The code being bad' is one likelihood which might emerge for it to disjointedly expand on, but there are many others. It's more like it's exploring probability space while hugging the median than actually contemplating your question; the calculations it runs through are instantaneous.
*A note on its calculations: the probabilities themselves pertain to the text being outputted and not necessarily the underlying concepts, so if it says something about 'the code being bad', that might only indicate calculations pertaining to the very words these ideas are expressed in rather than the underlying ideas themselves. LLM might not have, through its training or anything else, an approximate understanding of what code or 'bad' even are, but instead merely highly elaborate algorithms linking them and other words and word assemblages together.
So since its operating primarily or wholly on a linguistic level, it is impossible to get it to divorce its output from your starting prompt, which sets off the whole probabilistic determinacy cycle.
You are agreeing with ME on the issues with current LLMs.
The underlying models (even current ones, which are fundamentally not ideal for this kind of thing, which is my point) do reliably identify bad code in isolation. But this is irrelevant because any correct analysis they produce is contingent on the framing I provide, which means it's really just my analysis being reflected back, not the "model's" (whatever that means).
More options
Context Copy link
More options
Context Copy link