Maybe that’s the next step with LLMs/transformers? Have them write logic/tools instead of using them directly?

Also see if you can catch all the pumpkin references. :) (The pie at the end is one I made using my recipe!)

Example screenshots from Facebook posts I saw.

I’m wondering if that bit about matching the vibe was the source of a pattern I’ve been seeing - have you observed this too? bsky.app/profile/theh...

It makes me wonder if the “memory” feature has picked up on biases and is adjusting to them, or if the people specifically asked ChatGPT to adopt a certain perspective for these sessions in earlier prompts.

These could be faked, sure, but they didn’t seem like that level of propaganda. And the answers shown were very different from what I saw with similar prompts from a new, logged-out session. And again, some of these examples were from weeks or months ago.

For example, someone showed asking ChatGPT about how to break down “traditional gender norms” and it gave an answer aligned with the requestor’s views on feminism. Then even a flat-earther shared screenshots where the answers reinforced some of their suspicions about science.