GPT-6's new edition of Weber

[Mr_Incognito]

Congratulations on great pioneering work!

I believe the proofreading and editing will become the crucial issue in the foreseeable future.

There are other friends on the forum providing new editions of old works - I think it is wonderful, am not sure how comfortable they would be with it, but providing the raw files would be great for other users.

I recorded my screen as Astra engraved a Haydn SQ minuet & trio in 55 minutes: https://x.com/aug5thmusic/status/2098584544895025193

This is every prompt I sent:

1. "Take this score and turn it into a polished LilyPond engraving." (`sq-minuet.pdf`)
2. "This is a good start, but there are errors to fix. The third beat of Violin I, m. 1 should be an E-flat, for example, and beat 2 of m. 3 should be an F. Find similar errors and fix them. As you will have noticed, the image quality of this score is poor, so many of the glyphs have worn away."
3. "Sorry, I meant the second beat of m. 1, not the third."
4. "Consider the surrounding harmonic context and what you know of this period's harmonic conventions when determining unclear pitches."
5. "Don't search online. Rely only on the score I gave you."
(Astra had started to search online for Mozart E-flat String Quartets. Cheater.)
6. "Good work. I don't detect any errors. Create the parts now."
7. "The staves are a little squished in the Trio section. We can probably do without the extra padding. Move the "Menuetto" and "Trio" text closer to the staves and reduce the extra spacing between the two sections."
8. "Thank you. That looks better to me. I was recording you as you worked. Can you take the video `2026-09-11 16-55-57.mp4`, add a real-time clock in the corner so people can see how much time has elapsed, and speed through the boring parts so the whole thing is under 8 minutes?"
(Yes, Astra made the video too.)

Files: https://github.com/aug5thmusic/ai/tree/ ... a-haydn-sq

[Vogel]

Congratulations on great pioneering work!

I believe the proofreading and editing will become the crucial issue in the foreseeable future.

There are other friends on the forum providing new editions of old works - I think it is wonderful, am not sure how comfortable they would be with it, but providing the raw files would be great for other users.

Thank you! I think there are two dynamics at play here, each with its own implications:

(1) Using LLMs to proofread drafts and input notes: This is a 1000x productivity upgrade. Every engraver should do it. It's like using score-scanning software, only better.

(2) Allowing LLMs to edit music: This is thorny. Art is traditionally a human enterprise, and handing it over to machines can feel like a betrayal. I understand; I'm not entirely comfortable with it myself. But I thought this was an experiment worth trying, and I think anyone who reads Astra's Preface will be impressed by its warmth and intelligence.

[JJP]

(1) Using LLMs to proofread drafts and input notes: This is a 1000x productivity upgrade. Every engraver should do it. It's like using score-scanning software, only better.

(2) Allowing LLMs to edit music: This is thorny. Art is traditionally a human enterprise, and handing it over to machines can feel like a betrayal. I understand; I'm not entirely comfortable with it myself. But I thought this was an experiment worth trying, and I think anyone who reads Astra's Preface will be impressed by its warmth and intelligence.

(1) I will not do this. It is far too dangerous to leave proofreading to a probability algorithm. We don’t proofread based on what is most likely. We proofread based on what is intended vs what is actually there, plus the musical, grammatical, and visual clarity and correctness within the context of the piece. The algorithms can’t understand context.

(2) “Warmth and intelligence” this sounds like attributing emotion and cognition to the system. This contradicts the first half the assertion.

[John Ruggero]

Dear Astra,

Suggestions:

Your cues are too short. One note is not a cue; one little snippet is not a good cue. And in my opinion, cues should be more frequent.

Clefs are not inserted before cues unless there is a clef change.

[benwiggy]

The work was completed over two days, during which Astra worked almost continuously,

Two day of continuous AI processing? I wonder how much energy that used.

recreating the score in LilyPond.

Lilypond really needs a good utility to convert to XML, so that the data can be used in other apps.

[Vogel]

We don’t proofread based on what is most likely. We proofread based on what is intended vs what is actually there, plus the musical, grammatical, and visual clarity and correctness within the context of the piece. The algorithms can’t understand context.

Frontier models understand context incredibly well. In one of my last proofreading runs, I had Astra review the whole symphony, consider each note in its harmonic context, and report likely errors. It found two errors I had intentionally left in the score along with several others.

Here, for example, Astra suggests that the viola's C in m. 172 should be a D -- not just because D makes more harmonic sense, but because the winds have a similar phrase 150 measures earlier with a D, not a C. This shows incredible long-range and contextual reasoning.

--

Weber_error_report.png (244.18 KiB) Viewed 123 times

Weber_error_report-2.png (208.46 KiB) Viewed 123 times

[Vogel]

The work was completed over two days, during which Astra worked almost continuously,

Two day of continuous AI processing? I wonder how much energy that used. :(

We're using energy right now to have this conversation. You use energy every time you turn on a computer or use digital notation software. I used energy to produce a new digital edition of Weber's second symphony. I'd say that's a net benefit to the world!

Also, using AI to transcribe music is actually energy efficient. It uses less energy than if you or I were to do it manually.

[RMK]

We're using energy right now to have this conversation. You use energy every time you turn on a computer or use digital notation software. I used energy to produce a new digital edition of Weber's second symphony. I'd say that's a net benefit to the world!

Also, using AI to transcribe music is actually energy efficient. It uses less energy than if you or I were to do it manually.

Aren't you being a bit disingenuous here? I think you know darn well that it takes exponentially more power to run an AI task for 48 hours non-stop than any of your other examples.

Look, I think you have done a terrific thing here, although I wonder how it would handle an actual holograph rather than a clear copyist's handwriting. Have you tried anything really complicated, say a few pages of Wozzeck or Elektra? How would it do with a score by George Crumb?

To say a new edition of the Weber is a "net benefit to the world" is indulging in hyperbole, IMO. Also, i have seen the parts you posted to IMSLP, and while much improved from your initial effort, they are not yet what I would consider appropriate for publication, for the reasons John R. stated among others.

However, I applaud your effort and am looking forward to see future developments.

[Vogel]

We're using energy right now to have this conversation. You use energy every time you turn on a computer or use digital notation software. I used energy to produce a new digital edition of Weber's second symphony. I'd say that's a net benefit to the world!

Also, using AI to transcribe music is actually energy efficient. It uses less energy than if you or I were to do it manually.

Aren't you being a bit disingenuous here? I think you know darn well that it takes exponentially more power to run an AI task for 48 hours non-stop than any of your other examples.

I doubt it. How many hours would it take you to transcribe, engrave, and proofread a 104-page symphony? It would certainly take me more than 48 hours -- probably more than 100. And during that time, my computer and notation software would be running continuously. What does that add up to?

Also, that equation doesn't include time. What is your time worth? $50/hour? Even if you were as fast as AI and could produce a new edition of a symphony in only 48 hours, that's $2,400 in labor AI is saving you. At 100 hours, it's $5,000. That sounds like a great value to me.

Look, I think you have done a terrific thing here, although I wonder how it would handle an actual holograph rather than a clear copyist's handwriting. Have you tried anything really complicated, say a few pages of Wozzeck or Elektra? How would it do with a score by George Crumb?

I haven't. I'll leave modern music to people who know it better than I do. But I'd love to see the results of that experiment.

To say a new edition of the Weber is a "net benefit to the world" is indulging in hyperbole, IMO.

"Net benefit" means that the benefit outweighs the cost. It says nothing about the size of the benefit. If something's value is 1, but the cost is only 0.1, then it is a net benefit to the world even if the benefit is small.

Also, i have seen the parts you posted to IMSLP, and while much improved from your initial effort, they are not yet what I would consider appropriate for publication, for the reasons John R. stated among others.

Luckily, the edition is CC0/PD, so anyone who finds fault with Astra's work can correct it and upload their own version.

However, I applaud your effort and am looking forward to see future developments.

Thank you!