Here's something I learned this week..

Before a baby is born, before its eyes can open or its ears can hear, the brain is already firing. Random patterns of activity, all the time. Bursts and waves that have nothing to do with the outside world, because the outside world hasn't reached the baby yet.

Biologists have been studying this since the 1980s. They knew it happened in basically every mammal brain, including ours. But for forty years, nobody could quite say what it was for.

Two researchers from Korea just published a paper that thinks they have an answer. And the answer might explain why ChatGPT confidently makes things up.

I wanna talk about this. But before I do, here's something interesting.

For years now, every fix we've thrown at AI hallucinations has been working on the model after it's been set up. RLHF, better training data, evaluation tweaks, all of it happens after the model already exists.