Type any sentence and see what GPT-2 predicts comes next at each token position.
Click a token in the attention strip to see what meaning it contributes to the prediction.
The model runs entirely in your browser — no data leaves your machine.
Loading GPT-2 (124M parameters)
Initializing…
First visit downloads ~210 MB (model + weights). Cached for subsequent visits.
Try:
Tokens — click to set prediction point
Running inference…
Next-token predictions
Attention from prediction point — click a token to inspect
Attention by layer \u25B6
Causal contribution (ablation)
How does this token change the prediction? Green = predictions
boosted by this token’s presence. Red = predictions
suppressed. Computed by re-running the model with this token masked out
and comparing the two prediction distributions.