OpenAI’s GPT4 now has a bio tool that can store and remember information across chat sessions. Here I’m using it to store and later execute Python in a different session.

The bio tool doesn’t respond with the info when you store it -it instead updates the “model set context” which you can query to see what they’ve decided to remember about you for future conversations.

I was also playing around with getting it to dump its full prompt ala @simon https://simonwillison.net/2023/Oct/26/add-a-walrus/#the-leaked-dall-e-prompt

Here it was giving a little more insight into how it first search()’es then uses mclick() on the returned results to lookup several reputable results in parallel.

@singe an opportunity to poison the state of the model?

@buherator I’m not sure - you can influence your own model, but that’s more a feature than a bug.

@singe I have no clue how this works internally, but attacking ones personal storage to make them get nasty answers (for psyops?) seems like an interesting venue to me

@buherator for sure, if you could break into someone’s session and modify their prompt without it showing up in the “custom prompt” setting that’s an avenue. I’m surprised by what gets stored in bio - for eg I was asking some questions about Plato in light of the scroll they deciphered - and it decided to permanently remember the new date of Plato’s enslavement.

Ah, and now they announce it.