Conversation
Turbopack hands a worker its chunk list in the script URL's fragment (`/_next/static/chunks/turbopack-worker-….js#params=…`). A fragment is not part of a Request, so answering one from the service worker's asset cache loses it and the worker dies on the first line with "Missing worker bootstrap config". In a production build with the service worker in control, that killed every worker on the site: the diffusion model, the head camera's training, the path tracer's scene builder. Worker scripts now go straight to the network. Found while adding article 014, whose figures are three workers; `e2e/music-ai.spec.ts` covers it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…tes them The model is lib/ml's Transformer again (the one № 004 builds), 64,704 parameters: 56 words, a 16-step window, two layers. A chorale becomes one token per voice per eighth note — a pitch, HOLD or REST — so writing music is writing the next word. Beside it: what a piece is measured by (in key, voice crossings, parallel fifths, held notes, chord variety), a copy detector against the training set, the five-line rules composer, and DPO for the preference figure. Data: Craig Sapp's edition of Bach's 370 four-part chorales, CC BY-NC-SA 4.0, moved to C major or A minor, 333 to train on and 37 kept back. scripts/music rebuilds both .bin files byte for byte from the corpus clone and a checkpoint; the corpus and the runs live outside git (docs/research/music-ai). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every sample is computed here, with none of the browser's own instruments or effects: three soft voices (a warm pad, an electric piano by FM, a flute with breath), a damped stereo reverb laid out like Freeverb, and a low-pass over the mix. The first organ was sharp — 6 to 16 dB more in the 1–4 kHz band the ear is most sensitive to (docs/research/music-ai, "Timbre"). The player owns the page's single AudioContext and the synthesiser worker: playing something stops whatever was playing, changing the instrument keeps the position and never leaves a gap, a piece counts as finished when the music ends rather than when its reverb tail does, and the last figure to leave stops the sound and closes the context. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
One worker file serves all of it. Training keeps a worker to itself for the three minutes it holds the thread; figures 02 and 03 share a second one, so the trained model and the chorales are fetched and parsed once instead of twice; the player has a third for synthesis, so a note never waits behind a training step. Every message carries which figure asked, because a shared worker's replies reach both and loading one figure used to reset the other. Workers are made in one place (worker-factory.ts): three `new Worker(new URL(…))` call sites for the same file left the production bundle unable to boot any of them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… wrote it Fig. 01 trains a composer in about three minutes and improvises at six moments on the way, so the reader hears it stop noodling; the snapshot they picked is never taken from under them by the next one, and the numbers underneath say what changed. Fig. 02 is the reader as the judge: two pieces, one pick, one DPO step against a frozen copy of the model. About twenty picks are audible; at the page's learning rate that moves held notes from 48 % to 64 % without the model forgetting Bach (docs/research/music-ai). Fig. 03 is the blind test: Bach, the model and five lines of maths, shuffled, with the piano rolls and the answers held back until the reader has guessed — a chorale and a block-chord machine look nothing alike. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…self A small Transformer learns Bach's chorales in the reader's browser in three minutes, and they hear it learn. Then the honest half: the copy detector shows it composing rather than remembering (5 to 10 eighths at the end, once 32), the judge figure shows it gaming the score — asked for fewer parallel fifths it learns to stop moving — and five lines of maths get a blind test of their own. Numbers are the page's own recipe, measured twice in node and averaged (docs/research/music-ai). Sound is checked by hand: CI cannot hear, so e2e/music-ai.spec.ts covers training, a pick, the blind test's secrecy, and that nothing heavy is fetched before it is asked for. The smoke test's single-article tag moved to `multi-task`: `llm` now has two. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
RESULTS.md carries every number the article quotes and the ones it does not: the licence trail for the data, both training recipes, the preference-tuning tables at the research push and at the page's gentler one (where the reward-hacking contrast does not reproduce, which the article now says), the timbre work, and two ways of measuring harshness that were wrong. DESIGN.md §4 gains "Instruments that make sound": one player per page, sound that stops when the figures go, a piece that ends with its music and not its reverb, and the fact that CI cannot hear. HANDOFF.md gets article 014, and COMMITS.md two scopes that were already in the log. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
№ 014 — 三分鐘訓練你的 AI 作曲家
A 64,704-parameter Transformer (lib/ml's own, the one № 004 builds) learns Bach's four-part chorales in the reader's browser in about three minutes, and a synthesiser written sample by sample plays what it writes.
Three figures
Honest by construction: every figure's numbers are proxies for "like a chorale", never for "good", and the article says so; the copy detector is there to catch the model remembering rather than composing (5 to 10 eighths at the end of training, once 32); the reward-hacking contrast that only reproduces under a harder push is reported as exactly that.
Data: Craig Sapp's edition of Bach's 370 chorales, CC BY-NC-SA 4.0. The tokens and the trained model in
public/posts/music-ai/carry the same licence, with attribution in the README, the article and the handoff.scripts/music/rebuilds both files byte for byte.A fix that matters beyond this article
fix(pwa): Turbopack hands a worker its chunk list in the script URL's fragment, which a Request does not carry, so the service worker's asset cache was answering worker scripts without it. In a production build with the service worker in control, every worker on the site died on its first line — the diffusion model, the head camera's training, the path tracer's scene builder. Worker scripts now bypass the cache.Checked
docs/research/music-ai/RESULTS.md).🤖 Generated with Claude Code