Artificial Intelligence completed Beethoven's unfinished Tenth Symphony | Hacker News Reader

Artificial Intelligence completed Beethoven's unfinished Tenth Symphony

smithsonianmag.com

65 pointssizzle56 comments

Actually working in the field, I hate stuff like this so much. Like, it's a cool project, but the reporting and marketing and hype is all so fucking awful. Imagine if a person had done this instead of a model:

> [I] Completed Beethoven's Unfinished Tenth Symphony

> Now, thanks to [my] work, Beethoven’s vision will come to life.

And so on. Your reaction would probably be "no, this isn't Beethoven's vision, it's some random person's vision." In this case, it's that much more egregious because it's not written by a journalist through a game of telephone. It's actually written in this breathless magical way by the person who built the model.

There needs to be a new fallacy or something: "Appeal to AI." Because otherwise we treat "AI completed Beethoven's vision" as somehow more True or Correct than "I completed Beethoven's vision." And it's much more dangerous when we apply that free pass to things like "AI predicts recidivism." Is it just riding off of how people perceive mathematical correctness or something?

Not working in the field at all, I completely agree! Well said.

What pains me is that these are the exact articles that make people outside of the field say "AI fails to deliver", "AI is just a bunch of ifs" or "AI is just snake oil the new winter is coming".

Most actually interesting ML/DL projects are not stuff that the public directly interacts with so you end up with a ton of inconsequential stuff like this article and real progress is just not visible.

I agree that the media is overhyping AI without fully comprehending what it can and cannot do.

At the same time, I believe your criticism is a little off-base. Sure, it's hard to tell it apart from how a person would have finished the symphony. That, however, is precisely the point. That's what's so fascinating about it.

This thing has been going on for a long time.

First time some seriously tired to make "computer composer", maybe 1960s?, it passed double blind test flying colors. Professional musicians declared that computer could never compose something like this" about computer made music.

Using learning and using generative grammars has been used to imitate different jazz musicians in 90s.

Some modern composers use computers to generate music, they just tweak parameters until it is how they like it. It's still human made.

I couldn’t agree more. We have very little basis to judge whether the algorithm did Beethoven justice and so all the celebration is really for a non-event. A computer wrote down notes after other notes. Maybe they sound not too dissonant. That’s not news.

This statement feels too strong. An algorithm completed Beethoven's tenth symphony, and it did a passable job, maybe better than what a graduate student could achieve. That's exciting for tool-assisted composition, even if it tells us nothing about Beethoven's historical vision for the symphony.

Agree! This is great and exciting. When people learn to play music composed by someone else, they are experiencing/living to some degree the mind of the composer. And music can fairly be described as sets of repeatable patterns. So if AI can take the entirety of a composer's life's work and discern key patterns, replicate them even, then some aspect of the mind of the composer is reborn. This is immortality. A dedicated grad student could probably do it for one composer; AI can do it for ALL composers. And humanity is all the better for it :)

I don’t think a computer could know what the artist intended to create.

I wonder, how accurate can the computer be at predicting the next note (the first note from the uncompleted section)? Is that an interesting question?

That can be proven though. Keep a work by the author out of the training phase, then see if the model can predict it from an incomplete version.

Otherwise...

Yes, it's a very interesting question! This how many language models work. You might be interested in reading about a metric called perplexity. [0]

The major caveat is that it's invalid to evaluate on completed, polished work, and then extrapolate to what's practically a napkin note. Another important note about it is that some next token predictions are harder than others. For example, try predicting the next word (the _____) in each of these:

> I really love this Dr Seuss Book called "The Cat in the _____

versus

> I really love this Dr Seuss book called "The Cat in the Hat." _______

[0] https://en.wikipedia.org/wiki/Perplexity

You could also make a model which always responds with the word "hat". If the second sentence is the "creative" portion, it would then prove 100% accurate on your in-sample data, and the next sentence could be "Hat, hat hat." And then people could say, see, this brilliant AI speaks like Dr Seuss!

You could also just string a bunch of random words together and build a model that learned to predict them with total accuracy. Of course if the words are truly random, it would have basically no chance of predicting the next random word from outside the learning set.

The conceit really is that some model got enough important, critical contextual information from the partial symphony to be able to predict with greater than random probability the next portion. On a micro level, one note at a time, maybe. But every note further into the unknown is predicated on the previous ones, so on a macro level it's absurd to think it's "predicting" anything.

Thanks for sharing that term.

A score is so much more complex. I need to dive into the approach Playform took. I notice that team is comprised of experts in a variety of fields, who make a lot of editorial decisions on the output of the AI and then decided the style of the output (scherzo, etc.) and where that particular element might be placed in the score. I wonder what would have happened if they just ignored the AI and completed it using their own compositional knowledge?

Calling the prediction accuracy doesn't seem quite right, but I don't know what term I would use. I created a mobile music notation application that has pretty good accuracy when writing on the score. Some users asked for the ability to predict what might be written next. The first challenge is creating a UI to present the choices, but lets assume that is a solved problem. I trained up models and you can probably guess what happens. It is similar to the demos you have listened to where thirty seconds of on original song are then taken over by AI. Things sound similar and then descend into madness. This is sort of fun but not super useful to a composer. Maybe if you, the composer, had a large corpus used to trained with, the AI could riff on your own ideas.

My experiments were with a single instrument and then I tried using the process on multiple instruments. I would consider the end product pretty much useless.

Users of music notation software, especially those who are able to write out their compositions using notation, would find this feature pretty aggravating. Cool for bizarre demos, pretty useless for making a living.

I do expect to see a lot more activity in this space as we learn more techniques, train bigger models and learn how to keep the AI from going off the rails.

> There needs to be a new fallacy or something: "Appeal to AI."

I applaud and welcome this idea

The title of the article is really kind of nonsense in a way. I mean I could write music and declare I had completed Beethoven's Unfinished Tenth Symphony, but it wouldn't mean I did a good job.

Unless it's deeply compelling, believable, or otherwise worthy of performance and audience applause, it really doesn't mean much - at least based on that title.

The article specifically talks about how they have tested it with audiences, who were unable to tell the AI-generated bits from the original ones.

How do you know beethoven didn't intend for the remainder to be 79 bars of silence? (This is a serious question.)

Fundamentally, current A.I. is predicated on replicating stationary distributions. In that sense they are deeply conservative and can only "predict the past", instead of creating true novelty.

> they have tested it with audiences

What do they do if the audience says that there are too many notes?

Indeed. It's a cool project, but I don't think this is a great success metric. A better metric would be to have a human composer do the same task, and have audiences judge whether they like the human- or AI-generated result better.

No, that’s completely flawed. Beethoven’s style of music is not about liking it or not, it’s about his way of understanding and composing music.

If you wanted to know if they liked the music more, you could play any modern song and it would probably score higher, and I say that as a classical music fan.

To give an example, I would always go to a Mozart concert if the chance comes up, despite me not really liking his music a lot, there’s still no denying that they’re musical masterpieces and provide more than nice sounding tunes.

Sorry, "like" was a bad choice of words. I was assuming that the test audiences had a certain level of musical education and understood that the goal was to produce a Beethoven-sounding result.

> We challenged the audience to determine where Beethoven’s phrases ended and where the A.I. extrapolation began. They couldn’t.

I'd say that's a pretty weak test. The AI could have repeated the previous phrases for 20 times, and the audience won't be able to tell where exactly to draw the line, but they'll realize that something is wrong.

It's an interesting feat. But it's nowhere near "How Artificial Intelligence Completed Beethoven’s Unfinished Tenth Symphony".

True, though my point was directed at the title.

The sample at the end of the article seems to me at the same time Beethoven-like and meaningless. Made me think of the "screams of the damned" AI recreation of Frank Sinatra

<https://www.theguardian.com/music/2020/nov/09/deepfake-pop-m...>

It feels like I'm hearing hints of past symphonies, but nothing new and it lacks overall coherence.

I've listened to (and occasionally played) a lot of really mediocre classical music, and this sounded exactly to me like "lesser composer tried to mimic Beethoven" and/or "sketches completed by his student". Or, alternately, "great composer phoned it in", of which Beethoven's "Wellington's Victory" is a relevant example. The AI totally nailed the sound of Beethoven, the rhythms, harmonies, and orchestration... but my reaction is the same as yours. Fine for listening to in the background while I code, I guess.

On the other hand, I'm now very curious what similar experiments with other composers would sound like: Vivaldi, Brahms, Shostakovich, Philip Glass, John Williams.... the potential for AI orchestration is probably infinite.

Exactly what I thought. Like the experiment with drugged up spiders weaving borked webs.

This Beethoven has been smoking banana skins.

I actually had a really hard time determining where beat 1 is of each measure. It feels kind of meandering, like a touch of musical dysphasia.

The Grosse Fuge (German spelling: Große Fuge, also known in English as the Great Fugue or Grand Fugue), Op. 133, is a single-movement composition for string quartet by Ludwig van Beethoven. An immense double fugue, it was universally condemned by contemporary music critics. A reviewer writing for the Allgemeine musikalische Zeitung in 1826 described the fugue as "incomprehensible, like Chinese" and "a confusion of Babel".

A reviewer once gave a bad review to a good piece of work, therefore, anytime a bad review is given, the accompanying work must be good.

> it was universally condemned by contemporary music critics

Reading Slonimsky's Lexicon of Musical Invective: critical assaults on composers since Beethoven's time, entirely made up of contemporary condemnation of what seems every famous composer since (and including) Beethoven, I get the impression you could say this about any of their works.

(free to borrow) ebook: https://archive.org/details/lexiconofmusical00nico/page/42/m...

Feeding the composer's works prior to the Ninth would have been much more compelling. The question then becomes whether the AI's Ninth sounds anything at all like the actual Ninth.

I imagine there would be as much similarity between the AI symphony and Beethoven's one as there is between Beethoven's other symphonies, which isn't that much. The human mind just has so many more inputs besides previous work to influence its output. Beethoven's whole life experience factored into the Ninth's creation.

That's not to say we shouldn't try it out.

Exactly! AI, there's your goal, now bridge that gap.

Even non-ML software is capable of surprising its designers with so-called "emergent behavior" so creativity is possible. The Starcraft II AI AlphaStar (which as I understand it played 200 years worth of Starcraft against itself before it was able to beat the pro human players) demonstrated some "unusual strategies… occasionally wildly off meta." https://www.pcgamesn.com/starcraft-2/starcraft-2-deepmind-ai

I would like to hear what their AI did with completing partials of the other 9 symphonies. It would tell us a lot about how accurate is their model.

"It [Analytical Engine] might act upon other things besides number, were objects found whose mutual fundamental relations could be expressed by those of the abstract science of operations, and which should be also susceptible of adaptations to the action of the operating notation and mechanism of the engine. Supposing, for instance, that the fundamental relations of pitched sounds in the science of harmony and of musical composition were susceptible of such expression and adaptations, the engine might compose elaborate and scientific pieces of music of any degree of complexity or extent."

- Ada Lovelace, 1843

The article was very disappointing not answering to the most obvious question of "how can a pattern manager that does not understand (at this stage of the discipline) the very patterns it uses and generates, compete with no less than Beethoven". A large number of experts was reportedly used: how their contribution could correct the obvious and known faults of current generative AI would have been the big question.

Yes but what if I want to hear the completed version made by the crypto trading hamster when it needs a new career?

Unless top level classical musicians with a PhD in Beethoven era music are fooled I object to the strongly phrased and singular phrasing of completes in the title.

AI provides one of many possible completions to the tenth symphony would be good.

Local context without a coherent structure as usual.

I suspect we'll see "Cover Song Generator" any time now. Put in name of selected song (or just a lyrics prompt), put in name of selected musician, press start. The RIAA is going to hate this.

Harmonically and formally not as adventurous as the scherzo of Symphony #1.

I think this experiment is worth doing. Perhaps one day the machines will learn to resolve the stylistic differences between the works of contemporaries. (It takes a while to hear something new and be able to answer 'who wrote this?').

Why they decided to train on Beethoven's entire output is beyond me. ( His works are usually grouped into several 'phases'. In later life he declined when asked duplicate his earlier works. Will it sound like Symphony 4.5?) Guess we'll hear if that worked.

Cant wait to read the rest of the Canterbury tales or better yet listen to a brand new 8 track michael jackson album. What’s stopping us from putting out a second bible? Didn’t know we needed “AI” to make things up for us.

I feel like this is a very silly use of “AI” but at least the novelty is interesting and hopefully the data gathered from such a thing can improve machine learning in other ways.

I think this is an excellent use of AI and I will happily listed to the results.

Music is great, no matter what the source.

...And It Sucked

I suspect many of the responses will be overly critical in a way that is borderline unfair.

This is an AI informed by the opinions of multiple scholarly experts. I'm sure they'll keep smashing "train" and "generate" and adding new inputs until it meets some standard of sounding like authentic music. The public already accepts several completions of deceased composors' works. People pay to hear Simon Rattle's fourth movement of Bruckner's Ninth. Everyone plays the same version of Fantaisie Impromptu despite instructions to never posthumously publish Chopin's manuscripts. There's a Youtube completion of Morning Glories that receives great praise despite the missing sections lacking faithfulness to Scott Joplin's technique and style. If David Cope's algorithmic machine had announced it had written the missing sections, he'd be burned at the stake by his Youtube audience.

Against the criticism that this sounds like existing Beethoven music: I'm relieved it doesn't sound like new Stravinsky. New Metallica often sounds like bits of old Metallica except strange and different and sometimes confusing.

It sounds decent. There are probably elements of "new and different" that are missing, but there are repeated themes with variations, interplay between minor and major keys, and similar note/rhythm structures to historical Beethoven---maybe not as much dynamic variation, tutti rests, and solo/soli melodies as the Ninth. I admittedly don't know much about the full catalogue of Beethoven and expect many of the critics here also mostly know the Ninth.

My personal wish for gauging this work (once it's complete) would be more quantitative: a "classical music earworm neural net" plus a "composer uniqueness rating".

The first would basically max out on La donna e mobile or Messiah. These are the melodies that would be instantly learnable and widely recognizable. If each Beethoven composition rates higher than the last, we can expect the Tenth to be the Gold Standard of earworms. This could be your alarm clock tone and still never get annoying. Otherwise, we'd need to accept that the posted version isn't absolutely epic but still more worthy of release as a "what if" instead of doomed to history as a bunch of lost, incomplete sketches and musical riffs of a angry, sick, deaf lunatic. A best guess, if you will.

A uniqueness AI should be able to say Pictures at an Exhibition and Night On Bald Mountain came from the same pen but are not built with the same bricks. That is, the composer and style are the same but the themes and melodies are unique even though they share the same twelve notes, twenty lengths, eight dynamic markings, etc. Until you can show that Bach never ever repeated a lick from one chorale to the next (as well as what constitutes the length of a lick... seven notes? ten seconds? two measures?), it's unfair to accuse this composition of borrowing from other works. If the Proposed Tenth scores the same uniqueness as any other Beethoven work when removed from its training set, there would be no leg for the argument that this is just a derivative work. We are all derivative works making more derivative works.

An extra wish would be to know Beethoven's published versions varied from the early sketches. (Someone already commented this same thought.) From what I recall, Bruckner's Ninth underwent multiple revisions before he finally croaked at 80 percent complete. His first sketches were reportedly not very playable but were fine-tuned by him as well as a few well-known contemporaries. If Beethoven always kept adding to sketches but never really changing the line, we can assume everything in the sketches here would belong in the symphony. Otherwise, I hope the research team is figuring out a way to guess what all the changes would have been.

In short, don't knock the arrangement just because you know it was informed by an algorithm. After all, what are genius composers if not masters of taking only existing music and established rules and bending them in new ways---in other words, living algorithms?

> My personal wish for gauging this work (once it's complete) would be more quantitative: a "classical music earworm neural net" plus a "composer uniqueness rating".

This sounds a little like marking your own homework...

There are plenty of people that recognize the Seventh Symphony, but it's not what people are chanting at soccer matches or their first thought when this composer is mentioned. Still, you could realize after enough listening that they come from the same dude.

I think my idea is closer to providing two good adversaries if this piece were rated by an adversarial network. If the generated Tenth can't even clear these bars, then I would be more accepting of someone's negative or hostile opinion... and only if that opinion was formed after knowing the composition was made by machine.

Genius composers are masters of taking existing music and bending it in genius ways.

There's quite a bit more to it than being new.

It's about expression, not structure. You can copy-with-statistical-inference the structures but all that gets you - at best - is a kind of melted parody of the source works.

You can also do what David Cope does, which is take one structure and overlay it on another. That sounds a lot more coherent, but that's because it is - inevitably.

Creating that kind of coherence and intent from scratch - deliberately, and being able to assess that you've created it, and to what extent - is very probably impossible with statistical techniques.

After writing my comment and reading yours and thinking a while...

I think if somebody saw the first five songs of Hamilton, a few lines from thirty other songs, plus a four-measure theme for each character... it wouldn't matter what other non-Hamilton knowledge of the composers or cutting edge tools a person had; it would be impossible to write anything remotely close to the real Hamilton (even if you compared just the music without lyrics and knew the full dialogue). Maybe something cute and sometimes clever could be composed... but it wouldn't even be close to the real thing.

Though this music sample is interesting, there's total validity when people say that we can't call this THE finished version.

Even with a handful of dedicated experts working on this project, it can't possibly be close to the real thing if Ludwig had been able to finish this symphony.

I bet some kind of nested generative-adversarial approach could do that.

Lumberjack completed Beethoven's unfinished Eleventh Symphony. And then converted it to webm at 156 kbps and uploaded it to youtube, which, a second later, took it down for copyright infrigement, because it has the name of the famous dog in it. News at 11.

“Artificial intelligence extrapolated music pattern”

Beethoven's Ninth was apparently used to set the size of the CD, though there are disputes whether this was actually the case or corporate advertising.

https://www.researchgate.net/publication/322951358_Shannon_B...

Reply on news.ycombinator.com