Watch the six films and explore the labs
Six original narrated videos centered on empowerment papers by Benjamin Eysenbach and collaborators: 6:18–6:58 each, about 40 minutes total, 36 chapters. Visual edition 2 replaces the original system voice and text-heavy layout with neural narration, persistent visual experiments, and short equations introduced after the intuition. The companion retains all 108 derivation steps, interactive labs, transcripts, and source notes.
| Film | Primary paper | What the lesson develops |
|---|---|---|
| 1 · DIAYN | Diversity Is All You Need, Eysenbach et al., ICLR 2019 | Skills as messages, mutual information, variational decoding, intrinsic reward |
| 2 · Geometry | The Geometry of Empowerment, Ji et al., October 2026 preprint | Discounted occupancy, the probability simplex, information radius, channel capacity |
| 3 · Scaling | Scaling Empowered Agents, Ji et al., October 2026 preprint | Inner skills, outer preparation, decoder and value approximations, measurement bottlenecks |
| 4 · Representations | Learning to Perceive the World Through Control, Bastankhah et al., ICML 2026 | Forward and backward equivalence, independent nuisance cancellation, minimality assumptions |
| 5 · Human assistance | Learning to Assist Humans without Inferring Rewards, Myers et al., NeurIPS 2024 | Human influence, two contrastive predictors, density-ratio cancellation, takeover |
| 6 · Language agents | Training LLM Agents to Empower Humans, Ellis et al., ICML 2026 | Entropy versus surprisal, likelihood proxies, stopping thresholds, selective fine-tuning |
The teaching sequence combines intuition, derivation, worked examples, and geometric visuals, inspired by the requested Jia-Bin Huang and 3Blue1Brown presentation references. The scripts, diagrams, and animations are original. Narration uses the original af_heart synthetic voice from Kokoro 82M, rendered locally through kokoro-onnx. It is not a clone of any referenced educator. Complete utterances determine animation timings; the finished speech is never time-stretched. This independent collection is not produced or endorsed by those creators or the paper authors.
What can an agent learn with the skills it has? is a searchable, annotated guide to 19 primary papers on dynamics identification, information-seeking exploration, skill composition and acquisition, curriculum selection, and transfer limits. It includes eight Eysenbach-coauthored papers, a suggested reading order, a proposed research structure, a rope-manipulation experiment sketch, and downloadable BibTeX/JSON. The literature guide is a written companion; it is separate from the six video topics. Rebuild it with python3 src/literature.py.
Open index.html locally, or use the hosted collection. Keep the assets, media, previews, captions, transcripts, and code directories with the HTML. The complete player, lab code, and lesson data are embedded; playback and experiments require no API key or network access. Captions are optional, so narration does not cover the diagrams. If a browser restricts caption tracks under file://, use the local server below or the MP4 subtitle stream in a desktop player.
For a local server with video seeking support:
python3 serve.py --port 8772Open http://127.0.0.1:8772. Films are H.264/AAC, 1600 × 900, 24 fps. Each MP4 includes six chapter markers and a selectable subtitle stream; the browser CC control enables captions. The player occupies the full lesson width and the chapter equation panel is collapsed by default. Each lesson has a direct HTML entry point, MP4 download, VTT/SRT captions, and plain-text transcript.
The labs are exact finite-channel teaching experiments, not reproductions of the papers’ trained agents or benchmarks. Information quantities use natural logarithms and nats. The browser code and standard-library Python implementation compute:
- Mutual information, skill posteriors, and a variational decoder bound.
- Finite-channel capacity with Blahut–Arimoto and an explicit upper/lower gap.
- A two-context example separating potential, learned skills, decoding, and estimator bias.
- Independent nuisance cancellation and the distinction between state entropy and control.
- Human action priors, path clearance, takeover, and a hypothetical task outcome.
- Cumulative surprisal, a strict stopping threshold, and the corresponding sequence-probability floor.
The full Python source is visible inside every lab and downloadable as code/empowerment_lab.py.
python3 code/empowerment_lab.py
python3 -m unittest discover -s code -p 'test_*.py'The scripts preserve important limits: mutual information is a weighted average KL for a fixed prior; channel capacity optimizes that prior; population variational bounds do not certify arbitrary neural predictions; representation results require their stated optimality and minimality conditions; the LLM method uses a proxy and target selection rather than exact empowerment policy-gradient RL.
Use Python 3.12+, FFmpeg, and ffprobe. Narration runs locally with the full-precision Kokoro v1.0 model (Apache-2.0) and ONNX Runtime. The setup script downloads about 337 MiB from the upstream release and verifies pinned SHA-256 checksums; model weights stay outside Git. No API key is required. Audio is cached by utterance, voice, and speed. Caption and visual cue boundaries come from measured audio durations. Each cue is a complete utterance, never a fragment split by character count.
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python src/download_voice.py
.venv/bin/python src/neural_audio.py
.venv/bin/python src/prepare_math.py
.venv/bin/python src/render_all.py
.venv/bin/python src/render.py --previews
.venv/bin/python src/build_site.pysrc/film_script.py contains the revised spoken-first scripts. src/content.py retains the detailed equation and source material. content.json adds measured narration, chapter, beat, and utterance timestamps. src/film_scenes.py draws the revised animations. src/prepare_audio.py and src/scenes.py retain the original rendering implementation for reference; the edition 2 build uses neither the old voice nor its slide layout. src/labs.js computes the interactive examples. No third-party browser scripts, analytics, or CDN assets are needed.
src/verify_media.py checks the 5–7 minute duration constraint, streams, chapter counts, ordered beat/caption timing, audio levels, and full video/audio decoding. qa/visual-verification.json records a layout audit of all 216 narrated poses: text stays on screen without text overlap, and the diagrams average roughly 10–15 visible words per pose. Storyboards were also inspected visually. With the local server running, qa/verify_site.py checks all six players, chapter navigation, slider boundaries, desktop/mobile layout, offline labs, and Python/JavaScript numerical agreement. Browser checks use an installed Google Chrome through Playwright; change its executable path for another platform.
Primary paper URLs, consultation dates, and PDF checksums are recorded in sources/manifest.json. PDFs were read for research and are linked rather than redistributed. The transcripts and notes identify original examples and distinguish them from paper claims. The October 2026 work is labeled as preprints rather than peer-reviewed results.
Related collections: Reset-free RL · TAMP.