PREVIEW — simulated numbers. No recorded run is loaded and no GPU is connected, so this page is animating the run's real structure with mock values. The live verified backend exists (gated by scripts/helix_serve_gate.sh); open ?source=sse when it's running.

Watch a transformer think — on a stack you can rebuild from 299 bytes.

Every box below explains itself — hover for a hint, click for the full story. The model continues your text token by token; the spine shows each step of that work as it happens, and where it is powered by Helix.

Key: anything marked with this symbol is powered by Helix — compiled from Helix source by kovc, the compiler that rebuilds from 299 hand-typed bytes.

Conversation greedy · fp32

Run settings

Level of effort ?

Model ?

Try:

?Pick a seed — or type anything — and press Continue. The spine lights up step by step as the model works.

everything on this spine traces back to 299 hand-typed bytes — click to see how

Honest note: these are verified transformers — the spine shows their real work, never an invented number. Recorded-run GPU pacing was ≈ 10 s per token by design (the recorded GPT-2-XL replay: 195.5 s / 20 tokens). Residuals: fp32 · verified to PTX, not SASS · single GPU (sm_86) · oracle shares the spec. This page never invents a number. full residuals & gate records → · expert playground → · models & paths →