Talk to it
loading the weights…
The map on the left is the model's own vocabulary: two words are tied together
when each is among the other's most similar vectors, measured across all 64 dimensions.
Nothing labelled those groups — the model only ever read sentences and was only ever asked
for the next word. Type on the right and the map turns into an explanation: the words of your
prompt light up, dashed lines show which of them the last token actually attended to, and
solid rays fan out to the words the model is considering next, thick in proportion to their
probability. The word it picks is the one the ray leads to.