Post

Simon Willison
@simonwillison.net

What are some of the things you've learned about how LLMs (and LLM-powered systems like ChatGPT) work that were non-obvious but most helped you build a more effective mental model of how to use them?

Jan Ehrhardt
@jehrhardt.dev

LLMs are actually stateless. You put in the historic context as input. Chat is only a UI constraint to build this context iteratively.

4 reposts59 likes

asura
@asura.dev

This is a big one. There was a whole thread on bsky the other day where people thought you could "teach" chatGPT. A lot of people need to hear this one. You can't "teach" it (outside of providing context as mentioned).

asura
@asura.dev

I'll try to find it when I'm on a PC. A quick check back at my mentions, etc, and there's nothing there. It had to have been deleted and/or I was blocked. :P Saying ChatGPT doesn't eat cats and isn't secretly withdrawing $30/month from your wallet is a Speedrun Any% Block here. :P

Joe Fabisevich
@mergesort.me

There is one caveat with that, which is ChatGPT memory. ChatGPT never remembers a thing I say, so I was shocked to find out that it remembers and can recall months old conversations for @joeyabanks.bsky.social. He’ll ask some esoteric thing me mentioned in June, and it’ll pick up right from there.

asura
@asura.dev

In Gemini, I have the option to continue chats or start new ones. With those I can have it "know" stuff from months ago. You can also use this feature with their Gems.