AI · Reading a model’s mind

Why is it so hard to see inside a language model?

Name: Why is it so hard to see inside a language model?
Uploaded: 2026-10-25
Duration: 24 s
Description: A language model is billions of numbers, and we know every one. Yet we can't read why it answers as it does. Each neuron fires for many unrelated ideas at once, cats, cars and poetry, because the model packs more ideas than it has neurons. Researchers now untangle these mixtures into separate features, one thread at a time.

What you’ll see

A language model is billions of numbers, and we know every one. Yet we can't read why it answers as it does. Each neuron fires for many unrelated ideas at once, cats, cars and poetry, because the model packs more ideas than it has neurons. Researchers now untangle these mixtures into separate features, one thread at a time.