Definition
Plain language
The small slice of a model's internal state that holds the words it's disposed to say out loud.
As stated in the literature
The subspace spanned by the per-token directions recovered by the Jacobian lens; despite holding a minority of a concept's variance, editing it moves the model's reports and reasoning nearly as much as editing the whole concept.
Also called: J-lens
Why it matters: It suggests you can shift what a model says and how it reasons by editing a small, targeted part of its internal state instead of overhauling everything.
For example, even though a model holds a rich internal picture of "cat," only a thin slice of that picture governs whether it will actually say cat-related words, and adjusting that slice steers its answers.
Heard on the show
“And once you collect every word's direction into one set — they call it the J-space — a structure appears on screen that we'll keep coming back to.”Episode 203 — The Thought a Model Doesn't Say — and the Lens That Reads It