XVIS / AI systems

Inside models

Attention, embeddings and an agent's trace — the three views that explain what a model did.

Input
A sentence, 600 vectors, an agent run
Output
Attention arcs + matrix, 2-D embedding map, span timeline
Demos
3 · all run in the browser
Loading engine
Static
  • Attention weight

Built with Self-attention

Toy transformer head over your sentence: causal mask, temperature, arcs and matrix.

  • Canvas 2DRasterized plots and overlays
  • D3 7Scales, layouts, force graphs