Limits

The failures aren’t random. They’re the mechanism, visible.

You can now predict what a language model gets wrong, which is a rare superpower. Here are three classic faceplants.