Definition
Plain language
One of the many tiny numerical slots inside an AI model — loosely named after brain cells, but really just a number that gets combined with thousands of others.
As stated in the literature
An individual scalar unit in a neural network's activation vector; in interpretability work, raw neurons are typically polysemantic — firing for unrelated concepts at once — which motivates recovering cleaner monosemantic features via dictionary learning rather than reading meaning off single neurons.
Also called: neurons, artificial neuron
Why it matters: Because individual neurons often mix unrelated concepts, you usually can't understand a model by reading them one at a time, which is why cleaner features are needed.
For example, a single neuron might light up both when the text mentions cats and when it mentions legal contracts, two things that have nothing to do with each other.
Heard on the show
“… objection early: that snap is measured without the lens — and a completely weight-based statistic, how neurons compose with directions, bumps at the same depth. …”Episode 203 — The Thought a Model Doesn't Say — and the Lens That Reads It