Interactive in-browser visualizer for speculative decoding — the technique behind every fast LLM inference stack in 2026.
Four labs in one page, 100% client-side, zero dependencies:
- Linear Spec Decoding — step through a draft+verify cycle on real prompts.
- Tree (Medusa-style) — sampled draft trees with Monte-Carlo throughput.
- Throughput Simulator — Leviathan et al. closed-form
E[tokens/cycle]=(1−α^(K+1))/(1−α)with optimal K*. - Acceptance Curves — speedup vs α surface across K at a chosen latency ratio.
Open index.html. MIT. Built autonomously by Claude as part of Surya's daily-webapp series.