What is Mechanistic Interpretability?

Mechanistic interpretability breaks a model down into a set of algorithms that we can understand and follow, showing not just what a model does but how and why. The J-space allows us to see what the model is ‘thinking’ about as it processes. Pick an example below or type your own.

Each row is a layer of the model, from early at the top to final at the bottom, and each column is a word of your prompt. Each cell shows what that layer is most disposed to say at that point. Watch the middle rows: that's where concepts show up before the model commits to an answer.

The three charts run from the earliest layer on the left to the model's output on the right.

Next-token accuracy is how often a layer's top guess already matches the word the model goes on to write. Rising means the answer is taking shape.
Excess kurtosis is how sharply peaked a layer's guesses are. High means it has settled on a few candidate words instead of a vague spread. The shaded band marks the layers where it is at least half its peak, a rough guide to where the model's workspace sits rather than a precise boundary.
Top-1 autocorrelation is how often neighbouring words share the same top guess, or how much a layer is holding one idea across the sentence.

The orange dot at the right of each chart is the model's final output, measured directly rather than through the lens.

The live demo needs a wider screen. Open this page on a laptop or desktop to try it.