Position: Linear projection of activations into 3D subspace*.

Color: Activation strength of the studied bilinear form.

Label: Shows the current token and the predicted token by the model.

*There may be illusory gaps near the origin due to thresholding of low activations.