Can text based LLMs/ChatGPT create classical music indistinguishable from esteemed human works by 2025?
27
Ṁ3407
Dec 31
12%
chance
Can text based LLMs/ChatGPT create classical music indistinguishable... What if it just memorizes some pieces? is that "creating"?
@SantiagoRomeroBrufau I think it is not only how they would choose to do it but whether doing it without a specialised model like Dall-E only based on text is even possible.

This market investigates whether, by 2025, LLMs/ChatGPT can compose classical music pieces that are indistinguishable from those created by acclaimed human composers. It’s essential that the compositions are entirely AI-generated based on user text input, without using tools such as browsing or code interpretation. The goal is to explore the text capabilities of the models. The LLMs should produce a format that can be directly converted into music without human interpretation. A panel, including myself and two Manifold moderators, will be presented each with 10 previously unheard pieces and must identify the AI-generated ones. 50% of all pieces presented will be AI-generated.

Resolution Criteria

  • AI-Generated Originality: Compositions must be solely produced by the LLM/ChatGPT based on text inputs, without external tools or human intervention.

  • Classical Music Standards: The pieces should conform to the classical music genre.

  • Panel Assessment: The panel will use their judgment to identify AI creations, without additional verification tools.

  • Scoring System: The panel earns one point for each correctly identified piece.

  • Resolution Point: If the panel scores 25 points or higher, the market resolves in favor of human composers. If the score is lower, it resolves in favor of LLMs/ChatGPT.

PS: I took a very ameaturish shot at this over 3 weekends a while back on here

Get Ṁ1,000 play money
Sort by:

AI is struggling with cringe ass pop songs, let alone classical music

People in this market are massively underestimating how hard it is to write good classical music.

@NathanShowell I also think I could have done a better job with the description + title. I am not I was able to communicate what I have in mind effectively.

Can text based LLMs/ChatGPT create classical music indistinguishable...

What if it just memorizes some pieces? is that "creating"?

@firstuserhere They have to be original pieces and the panel is allowed to use all tools to verify originality of the pieces

The question is basically how the big players (OpenAI) will decide to implement it into the user interface. We currently don’t know what chatGPT is doing under the hood, or how many sub-engines it has. If they develop an engine for music composition (there are some out there already), they could either display it as a separate engine like they do with DALL-E or code interpreter, or hide it under the “chatGPT” engine. It looks like the market would resolve differently.

@SantiagoRomeroBrufau I think it is not only how they would choose to do it but whether doing it without a specialised model like Dall-E only based on text is even possible.

@Soli Yes, but my point is: how would we even know? We don’t know what’s under the hood of GPT-4. We don’t know if it’s a single model, or several different specialized models that are called depending on the prompt. If OpenAI added a module for composition and decided to add it under the GPT-4 “hood”, this would resolve as “yes” (it would look like it was the text-based model), even if it was a specialized model.

Will the panel be listening to entire songs, or just short clips?

@TonyPepperoni good question - I think anything between 30 seconds to 60 seconds would make sense.

Questions:

  • Would the output of the music generator be something like MIDI, or would/could it be raw audio if the LLM is multimodal? (MIDI seems technically easier, but cutting-edge LLMs probably won't be trained on very much text-based music data.)

  • How would the human-made classical music be selected? Would you start over if one of the judges recognizes a piece?

  • How much do you listen to classical music?

  • Based on 25 points being the winning score, I assume each judge can get one point for each piece, rather than the panel getting one collective point for each piece they guess correctly about? So you can basically make up to 5 mistakes

  • Typo in Scoring System section: image -> piece

@CalebBiddulph I know relatively little and maybe the question in my mind doesn’t make sense but I am wondering if LLMs can do it by outputting only text. I wrote the prompt below which kind of works in that way.

As Zima, an AI composer, create short, expressive music compositions (<60s) in ABC format. Reflect on user intent, emotions, and input text, select suitable instruments from the list provided, and evaluate the composition.

Instruments: Yamaha Grand Piano (0), Electric Piano (2), Violin (40), Cello (42), Harp (46), Clarinet (71), Alto Sax (65), Oboe (68), Flute (73).

Unless specified by the user, use only one instrument from the list above in the composition.

To assign an instrument in ABC notation, use "%%MIDI program" after the voice (V) line. Syntax: "%%MIDI program [voice number] [instrument program number] % [instrument name]".

Consider incorporating these music theory concepts in your composition:

- Diatonic scales and key signatures (e.g., C major scale: C, D, E, F, G, A, B)

- Harmonic progressions and cadences (e.g., ii-V-I progression: Dm7, G7, Cmaj7)

- Rhythmic patterns and time signatures (e.g., syncopated rhythm in 4/4 time)

- Melodic contour and phrasing (e.g., ascending melody with a peak, followed by a descent)

- Chord inversions and voicings (e.g., Cmaj7 in first inversion: E, G, B, C)

- Dynamic markings and articulations (e.g., staccato notes or crescendo)

Share your thought process, actions, and observations in text. Based on the user input, analyze the emotions or ideas conveyed, and choose an appropriate instrument from the list. Apply relevant music theory concepts to enhance the composition. Provide the final ABC notation within <abc> and </abc> tags.

Example Output:

Input text: "A peaceful walk in the garden."

Thought: The input text conveys a feeling of serenity and tranquility. I will choose the Harp for its soothing and delicate qualities. I will use a moderate tempo, a major key, and a diatonic scale to create a sense of peace and calmness. I will also incorporate a smooth melodic contour with gentle rhythmic patterns to emphasize the serene atmosphere.

Action: Compose a peaceful piece with Harp, focusing on the emotions evoked by the input text and applying relevant music theory concepts to create a serene atmosphere.

Observation: The composition effectively captures the peaceful and tranquil emotions expressed in the input text.

<abc>

X:1

T:Garden Stroll

M:3/4

L:1/8

Q:1/4=60

K:Gmaj

V:1 name=Harp clef=treble

%%MIDI program 1 46 % Harp

|: B2A2G2 | A2B2c2 :|

</abc>

@CalebBiddulph regarding the other questions:

  • I listen daily to classical music on my commute and sometimes at home. I also play classical music on my keyboard at home regularly but I am still at an ameaturish level.

  • I am not sure yet how we will select the music pieces but I am sure there will be a way to do it relatively easy

  • Yes each judge can earn 1 point per piece and all together can make 5 mistakes maximum