Change a token, mask, or temperature and watch the next-token probabilities move.
Trace a tiny deterministic decoder-like block from toy tokens to causal self-attention to next-token probabilities. Every displayed number is computed locally from fixed teaching weights.
Columns show token, relative bar, logit, and probability. Temperature rescales fixed logits before softmax; it does not retrain the model or change its learned weights.
The scaled score matrix, causal mask, and post-softmax attention weights are shown separately. Weight cells retain exact numbers, so color intensity is redundant rather than the sole information channel. Permitted weight cells are keyboard-operable buttons.
Core question: Why must a next-token model hide later positions?
Run and watch: Apply the mask-leak scenario and compare the same attention row with the mask on and off.
Pick a scenario, predict what should happen, then apply it and compare the result with your prediction.
Use one Guided Challenge. Predict first, reveal the mechanism, then connect a concrete attention/logit/probability change to the next-token distribution.
Use the existing Guided Challenge for the run/reveal step. Responses stay in this browser unless you deliberately copy or print them.
A Transformer turns tokens into vectors, lets each position mix permitted earlier information, and converts the final vector into next-token probabilities. This lab keeps every number small enough to inspect.
The toy tokenizer maps each displayed piece to a fixed ID. An embedding table turns each ID into a vector.
A position vector is added to that embedding. Without it, repeated tokens begin with the same representation wherever they appear.
The model derives a query, key, and value vector at each position. A query–key dot product measures one numerical match, then division by the square root of the key dimension keeps scores in a workable range.
The causal mask replaces future scores before softmax. A masked position receives no attention weight.
Softmax turns the permitted scores into weights that sum to one. The attention output is the weighted sum of value vectors.
A large weight matters only through the value it multiplies. The weight alone is not a complete explanation of the prediction.
The toy block combines the attention output with its residual and feed-forward path. Output weights map the final state to one logit per vocabulary item.
Softmax converts logits into probabilities. Temperature rescales those same logits: lower values sharpen the distribution and higher values flatten it.
This deterministic block is not pretrained, does not learn during the activity, and does not reproduce a production language model. Real systems use many layers, heads, larger vocabularies, learned parameters, and a separate generation rule.
Attention weights show one mixing operation. They do not reveal a model’s complete reasoning, understanding, or causal explanation.
Curriculum: Tokenization, embeddings, position, query–key scores, causal masking, softmax, logits, temperature, and evidence limits.
Misconceptions to test: one token always equals one word; attention is understanding; the largest weight is the complete reason; temperature changes learned weights; the highest-probability token must always be selected.
Use Tab and Shift+Tab to move through controls. Use Enter or Space on buttons and arrow keys on sliders or select controls.
Read the current model state in the applet’s text-state panel.
Important results are available as text and do not depend only on color or motion. This does not replace a full human assistive-technology audit.
Text equivalent for the current visualization state.