What the model is thinking about, token by token

Each word below was generated by the model while a lens pack read its hidden states. Shade the text by one concept, or colour each token by the area of knowledge it reads as. Click any token to see what fired there and at which depth of the model.

Prompt

Response

Fields across the response

Each column is one generated token. Darker cells are further above that Field's background; blank cells are below the 75th percentile. Click a Field to shade the text by it.

Field profile