[michelledepeil]

I spent some time thinking about this when the GPTs started becoming useful, and I'm glad to see my semi-prediction come true (related toot: https://fosstodon.org/@samuelnorbury/109799880269029379). Specifically: LLMs are great when you know what to ask for, and sort of know what the answer should be - call it a boilerplate machine - but bad if you're a novice, in which case meta-analysis of a query is needed to give a useful answer to a question.

[kenjackson]

What does "novice" mean in this context? For example, I recently used chatGPT to explore file systems in the Linux kernel. I'm a novice in this field (but know Linux decently well, and the general field of file systems pretty well), but I could pretty easily tell the information it was giving me was correct.

[michelledepeil]

When you don't have enough background knowledge to even verify whether the answer is plausible. Or when you don't know what question to ask.

It may very well be that newer AIs and models even manage to get that right more often, though. Or that a canny prompter will start by asking which question to subsequently ask.

As a side note, you're not a novice if you know the general field of file systems pretty well, IMHO.