I recently played with a 7yo who reminded me that LLMs are good at writing code the way that second graders are good at telling stories.

RE: https://infosec.exchange/@josephcox/117354532593344642

This article is full of gems, like:

> We tried calling it a simulation, which worked like some percentage of the time, but then other times they would see the images and see, oh, ‘I'm in a real parking lot, these people are just lying to me,’” he said. “We ended up having to call everything a sandbox. And with that prompt, it's able to consistently drive the car and like never refuse to do that.”

RE: https://toot.cafe/@nolan/117345796169248176

The linked post argues that because …

“An LLM is a giant associative map of human discourse. Its latent space contains everything from peer-reviewed computer science literature and thoughtful architectural postmortems to YouTube comment flame wars, chaotic Reddit arguments, and half-baked forum rants.”

… we should send inputs that are human-y nice and use intellectual language to get it to generate responses in kind.

Whether or not that works, maybe we should stop training models on comment flame wars and half baked forum rants.

RE: https://neuromatch.social/@jonny/117346693614824194

Meta’s Muse code, seen here embodying the Unix philosophy of having tools that do one thing. 🫠