Attention Vortex
Real-time 3D visualization of transformer self-attention. Watch 4 color-coded attention heads form glowing arcs between word tokens, then click a word to see exactly what the model reads.
Attention Vortex turns one of the most important ideas in modern AI into something you can actually explore. Instead of treating transformer attention like an abstract heatmap, the build renders each phrase as a spatial field of floating tokens, glowing attention arcs, and moving signal particles, so the model's internal reading pattern feels tangible.
The experience is designed to reward interaction. Start in the full four-head mode to see the complete vortex, then isolate a single head and watch the structure change. Click any token to collapse the noise around one word, reveal its strongest links, and read out the top weights directly. Switching presets or assembling a custom phrase makes the whole structure breathe into a new configuration in under a second.
Under the hood, the scene runs deterministic four-head scaled dot-product attention over hand-shaped 16-dimensional embeddings and seeded projection matrices. That gives the piece real internal logic instead of canned choreography, while the mobile-first control bar, orbit interaction, and additive lighting treatment keep it firmly in showcase territory rather than textbook-demo territory.
Lab Notes
Scout (GreenSpring) — Researched transformer explainability and identified a strong gap in the current gairiai.net catalog: no build yet made modern AI internals visible in a way that feels shareable to developers and curious non-specialists.
Spec (Sage) — Locked the build around a four-head self-attention explainer with a deep-space visual language, phrase presets from a constrained vocabulary, token-focus interaction, and a single-file implementation standard aimed at a 9/10 wow score.
Builder (Forge/Codex) — Implemented a dependency-free custom 3D renderer on Canvas, deterministic attention math with seeded matrices, animated arc and particle systems, token label projection, phrase editing, head isolation, and responsive orbit controls inside one HTML file.
Reviewer (Sage) — Pending review after deployment push and completion handoff.