An interactive guide to how LLMs actually think — a real language model, dissected live in your browser tab. English and 中文.
You type a sentence; a real (very small) language model splits it into tokens, maps them into meaning-space, reads them with sixteen attention heads at once, and gambles on the next word — every step shown live, on your device. No videos, no mock-ups, no server: the 7.5 MB model runs in plain TypeScript in the page, and nothing you type leaves the tab.
Live at insidethemachine.org — English and 中文.
| # | Essay | One question it answers |
|---|---|---|
| 1 | Inside the Machine (flagship) | What happens between your prompt and the reply? |
| 2 | Why It Lies | Why do confident-sounding answers go wrong? |
| 3 | The Attention-Head Field Guide | What are all those heads actually doing? |
| 4 | Why It Can't Count | What does tokenization break? |
New essays ship monthly after launch — see SERIES.md.
The same widgets, packaged as lesson modules with teacher guides,
printables, slides and an unplugged activity — WCAG-audited, offline-capable,
zero data collection. Start at #/classroom in a running build; the shared
front matter (model card, privacy & safety, tech check, standards crosswalk,
accessibility, letter kit) lives in
src/classroom/about/content/.
npm install
npm run devTests (npm test), DOM-hash baselines (npm run check:hashes,
HASHES.md), accessibility audit (npm run audit:a11y).
- The in-page model is TinyStories-1M run by nano-lm — a zero-dependency TypeScript re-implementation of the GPT-Neo forward pass, verified token-exact against the reference implementation.
- Act 4 optionally wakes a larger model (SmolLM2-135M-Instruct, ~136 MB, explicit click) via transformers.js.
- No analytics, no accounts, no cookies; the GPT-2 tokenizer and the small model's weights are self-hosted next to the site.
If you reference this essay or its widgets in academic work, cite via CITATION.cff (GitHub's "Cite this repository" button):
Shen, Shangyan. Inside the Machine: An Interactive Guide to How LLMs Actually Think. 2026. https://github.com/toolbooth/llm-explainer
An arXiv preprint is planned; once it exists, CITATION.cff's
preferred-citation will point at it — cite the preprint from then on.
Reports of classroom use are warmly welcomed — see the letter kit or open a thread in Discussions.
Code and text: MIT. The TinyStories-1M weights are redistributed unchanged with attribution from an upstream research release and are not under the MIT license — see the model card.