A few stray thoughts on LLMs.

I love using ChatGPT as a kind of family feud "what will the average (X category of person) think about this phrase".

Kind of an automatic wisdom of the crowds.

If you have one AI, it's got to be boring as hell.

If you have lots of AI agents, they can have very different personalities.

LLMs are a theory-of-mind in a box, magical duct tape you can put onto anything to help it interact with things in the people category.

We tend to assume a human-style intelligence in systems we categorize as a person. Which can cause us to mis-categorize computers.

Someone told the story from 30 years ago of a kid using an 8 bit computer and typing into a file that the OS expected to be a BASIC program the text "What is the capital of Idaho." When it gave an error, the kid said, "See? it's dumb."

LLMs are absurdly good at behaving like they have a coherent human mind, so we figure, implicitly, that they are like humans in a lot of ways, but actually they're way better than humans in some areas (rough vibes based fact recall), and way worse than humans in others (multi-ply theory of mind and rational thought).