AI Weirdness

Janelle Shane has been training neural networks to do odd, deliberately impractical things since well before most people had heard of GPT, naming paint colors, writing recipes, generating pickup lines, and then writing up the often hilarious ways the results go wrong.

It's the rare AI newsletter that treats failure as the interesting part rather than something to minimize. Shane has a background in physics and electrical engineering, and that shows in how clearly she explains what's actually happening inside the model, even while the subject matter stays delightfully absurd.