"most" is a hell of a claim, yea. "they're going to be useful sometimes and will stick around in some form" has been rather clear from early on (they're Transformer++, of course they'd be useful sometimes), but every session I spend with one still turns into lengthy deeply-misleading wild goose chases until I learn enough to force them down the right path[1]. quality/size of the model hasn't mattered for that, it always happens at some points (and it's just luck whether or not it's a critical point), and I see no reason to expect that it will change.

>From memory, the last time I was given a presentation on it, by actual Snowflake staff, they reported that ideal configuration results in something like ~92% accuracy due to the complexity of data at a large business ...

it's exactly this problem. nobody has solved it. lots of companies are trying to sell that they have solved it.

when I don't care about the end result that might be fine (I have used wordpress too), but frequently it ends up that they helped find relevant docs (they're quite good at finding that from vague descriptions, and it's very easy to verify) and I would have been quicker if I had simply read those docs in full. and then I'd have a more complete understanding of the space too. my biggest worry for personal use is that LLMs will write those docs in the future, and be riddled with mistakes :\

1: ...and then they're quite good code dictation tools. people who think "writing code isn't the bottleneck" just haven't spent enough time in well-understood spaces or rebuilt enough wheels. it absolutely can be a several-multiples bottleneck in day-to-day work, or when re-doing something you've done before and fully understand.