Animated, interactive explainers for Technical AI Safety research, made mostly by Claude. It's partly an experiment in how well an agentic system handles a teaching task, and partly because I (Agustin) wanted good visualizations of these topics to exist. Browse them at https://agustinbrusco.github.io/tais-explainers/.
Heads-up: unless a piece is marked as checked by me, I haven't verified its facts myself. The reviews described below are done by other Claude models, not by people. If you find a mistake, please open an issue.
The aim is a visceral understanding: you drag a lens across the layers of a model and watch what it decodes, and you watch a chain-of-thought monitor lose its grip as reasoning moves into latent space. That understanding also has to be correct, so every claim traces back to a specific place in a source.
See projects/README.md.
- Brief: what you'll be able to see, predict or explain afterwards, and which misconceptions it defuses.
- Research: the primary sources, actually read, in references/.
- Claims ledger: every statement in the piece → a source, a location, and a status (
sourced,simplified,speculative). - Script → the hardest visual first → build, as a web explorable (D3 / Three.js, no build step), a Manim video with local TTS narration, or both.
- Review: rendered frames checked against a visual checklist, an adversarial fact-check, and a simulated learner read-through, all before a human sees it.
Every figure carries a badge: schematic, real (with the model and layer), or speculative.
| path | what |
|---|---|
| projects/ | one folder per explainer: brief, claims ledger, script, narration, web/, manim/ |
| references/ | per-topic dossiers and source lists (PDFs are fetched locally, not committed) |
| kit/ | the shared visual language: semantic color tokens, Manim helpers, web CSS/JS, starters |
| scripts/ | TTS, paper fetching, contact sheets, headless screenshots, scaffolding |
| learner/ | the learner profile, concept map and journal the explainers are tailored to |
| .claude/ | Claude Code skills (explainer, research, review) and review agents |
Linux or macOS, with uv, Node ≥ 20, and TeX Live (for Manim math).
# Manim needs Cairo/Pango headers (Debian/Ubuntu):
sudo apt install libcairo2-dev libpango1.0-dev pkg-config
./scripts/setup.sh # Python env, Kokoro TTS model (~340 MB), headless Chromium
./scripts/setup.sh --interp # optional: CPU torch + TransformerLens, for visuals built from real activations
./scripts/doctor.sh # checks every tool and prints a fix for anything missingNo system ffmpeg is needed (a static build ships with imageio-ffmpeg), and no GPU is needed.
TBD.