Watch a language model think.
Advanced view →
Give it something to complete
The capital of France is
Model:
Show me its thinking
Takes ~15-30s on the demo CPU. The advanced view runs a deeper analysis and takes longer - this view keeps it snappy.
You asked
The model replied
How it came up with
Where it was looking
The bigger a word here, the more the model looked at it while choosing the next word. Small gray = ignored.
Want to poke under the hood?
Open the advanced view
to see every layer, head, and per-head ablation.