Global Distillation

Inspiration

On 27 January 2025, DeepSeek-R1 wiped $589B off Nvidia's market capitalization in a single session — the largest one-day loss in stock-market history. Two days later, OpenAI accused DeepSeek of distilling its models. Within weeks, distillation became an export-control question, a Senate-hearing question, and an EU AI Act question.

Distillation is the quiet engine behind most models people actually pay for — GPT-4o-mini, Gemini Flash, Claude Haiku, Llama 3.2, the entire R1-Distill family.

What it does

Global Distillation is a public compendium of AI model distillation, examined from eight perspectives on one site:

Academic — 46 papers from Bucilă 2006 to the 2026 reasoning-distillation wave, with 24 teacher/student benchmark pairs showing exactly how much capability survives compression. Financial — 63 models' list prices, 13 documented training runs (Sky-T1 at $450, s1 at ~$50 of compute), and a TCO model showing where self-hosting actually overtakes an API. Political — 25 policies across jurisdictions and 14 named disputes, from the OpenAI/DeepSeek accusation to the EU AI Act's GPAI obligations. Company — 18 labs scored on a single question: do they use distillation, sell it, ban it, or accuse others of it? Most do several at once. Developer — 27 tools and 8 costed, step-by-step recipes. Customer — 74 models compared on price, quality, latency, context and licence, plus a decision guide for 10 concrete use cases. Library — 27 distillation methods explained properly: what it is, how it works, the loss function in LaTeX, when to use it, what it costs, what it needs from the teacher. Timeline — 365 dated events, 2006 to today. Every figure carries its unit, its provenance and a numbered citation — 599 primary sources across 44 tables and 47 charts. Nothing is asserted without a link.

Four things make it more than a static report:

It updates itself daily. A GitHub Action pulls live signals at 06:17 UTC — arXiv (5,372 papers matching "knowledge distillation" and climbing), Hugging Face (16,331 distilled models), 15 GitHub repos, and Hacker News — regenerates the whole static layer, and redeploys.

A Three.js hero that argues rather than decorates. A 900-point teacher network transfers soft targets along Bézier paths into a 260-point student that is smaller and three times denser. It is the thesis of the site in one image, and it costs 0.2 ms per frame.

How we built it

A static site with no build step and no runtime dependencies — ECharts, KaTeX, and Three.js are vendored, the data is plain JSON, and the whole thing deploys as files. About 18,600 lines across the renderer, the stylesheet, the hero, and eight generator scripts.

Adversarial verification put a second agent on every file whose only job was to refute it — spot-check 20 claims, fetch 8 URLs, and prove the numbers wrong. A third agent applied the corrections. Roughly 18 corrections per file survived that gauntlet.

Design ran as a blind tournament: three designers built competing prototypes (editorial, premium-SaaS, scientific-instrument) which three judges scored without knowing who made what. The winner became DESIGN.md, with the runners-up's best ideas grafted in.

Quality was enforced by a harsh critic loop. Judges screenshotted our site alongside Epoch AI, Our World in Data and Artificial Analysis, blind, and ranked all four.