[mattdm]

ChatGPT, for all its amazingness, _never_ follows instructions. It just appears to, as it generates likely text. This is incredibly important to understand — it isn't a generalized AI.

Larger and more sophisticated models will do the party trick more convincingly... but actually following instructions will require a different approach.

[andrewstuart]

Its ridiculously annoying.

The other thing that I would really like to see is at least an option for deterministic responses. I like the creativity of ChatGPT constantly giving different responses but it is hugely problematic when you are trying to iterate and it keeps completely changing the thing you are iterating on.

[kylebenzle]

LLMs are a drunk guy at a party. A context machine, or a really good spell checker, maybe like a really outgoing, kinda drunk guy at a party. He can just walk right up to any group, hear someone say the words, "...thats why we never go on vacations anymore..." and he will just jump right in, "Me and the wife just got back from a great VACATION, we love going to Cabo, its out favorit vacation spot, last year we went..."

The guy has NO IDEA what the group was actually talking about but it doesn't matter anyway. You can give him more context but all he can do id further tailor his story to try and fit in better with the group discussion.

I think people get confused about what LLMs are doing because you CAN ask the drunk guy questions and he CAN kind of answer, or ask him to do a party trick and he can kinda do it, but all he is doing is continuing the context you give him, he IMMEDIATELY forgets anything you told him so he can't do the same exact thing twice.

[ilaksh]

Use the API and specify temperature 0. It will be as close to determinist as the underlying system allows.

[minimaxir]

> ChatGPT, for all its amazingness, _never_ follows instructions.

Yes it can, with robust system prompts.

The webapp is a misleading representation of ChatGPT's capabilities.