In 2024, I started generating music with AI models. I've generally found that the latest models are now capable of outputting any sound that is possible. The limitations of AI music are now generally caused by my selecting the wrong inferences (it usually takes 600-1000 runs plus editing), not understanding song structure, or not understanding what listeners like and dislike.
Every time I publish a new release, which I've been doing about once or twice per month, I will post a poll to Manifold to bring it to users' attention. The polls are bet on with one of the related markets.
Once I feel that I have achieved a song that is satisfactory, I will attempt to get a radio station to play it, even if only as a novelty during a morning show or something like that. "Six Weeks From AGI" (https://soundcloud.com/steve-sokolowski-2/six-weeks-from-agi) may be sufficient or I may wait for another song later this year to try, but I will try before AGI is actually achieved and anyone will be able to generate something better than this with no effort.
If any station plays any portion of any song I have ever created where the entire sound is generated by AI, then this market will resolve to YES. If time runs out, it will resolve to NO.
Whether the song is played once, becomes a Billboard hit, or if a morning show of a small rural station decides to play part of the song to criticize what AI is able to achieve is not relevant to the resolution.
I intend to pay for and use the latest models and software tools as they are released, even if they cost money, although I believe the musical "Turing Test" has already been reached.
RELATED MARKETS:
Update 2025-02-06 (PST) (AI summary of creator comment): AI Created Song Clarification:
If an AGI model outputs MIDI files for instruments, which are then input into music production software to create the arrangement, that song will be considered an AI created song.
Update 2025-06-07 (PST) (AI summary of creator comment): - Paid advertisements featuring the song will not count towards a YES resolution.
Only earned media will count. This means the radio station must choose to play the song as part of its programming (e.g., during a show, as a novelty, for critique), not as a paid advertisement.
1,000People are also trading
I don't know why there are so many traders here, but I might as well make use of the interest.
Should we use Manifold to create a radio hit? The models are clearly capable of it right now, and I know how to use them to get what is needed; the problem is figuring out what is needed. I think we would have a much better chance if we used Manifold polling to select from various options as the song progressed.
Is there anyone who would regularly vote in daily or every other day polls over the course of a month to achieve that?
@SteveSokolowski i would probably not vote every day but i would likely vote up to several times a week
@jim, how do you think we should set this up?
I think we need a few criteria:
A place to post files that can't be censored (ideally, I could host some sort of system myself) and which is easy for non-technical people to download versions of the song and understand how to use it
A canonical forum to discuss; preferably one that is easy for people coming from any platform to join
A structure for the polls to associate a poll question with specific versions - like this poll is voting between version 5 and version 6, which is better and why
We can do it entirely here if the tools exist here. There are people in AI music forums who may also be interested in this and want to join us.
Do you have any suggestions for the technical setup?
-----
Also, for the basic copyright stuff: we should credit any person who contributes substantially, and offer the song for free. If any money is made, it goes to pay back someone who advertises the song first, and the rest goes to charity.
@SteveSokolowski I'm too busy at the moment to be too involved although I wish I could, would be very interested
@TimothyJohnson5c16 No. That wouldn't be in the spirit of the market.
I do intend to possibly use music to advertise Nittany Toastmasters on the radio before the students return in a few months, but such a feature will specifically not count - only earned media does.
@TimothyJohnson5c16 Oh, and to let you know, a 30 second spot during a rush hour daypart runs about $100 near State College.
Wow I like what you've made. It also sounds like a perfect use case for LLMs since vague platitudes is what song lyrics are all about. You could likely do away with lyrics and make it just sound like words, e.g. I think a median language between French and Japanese would sound really nice in a song
It's odd to me that this market is being bid down with time.
It was always obvious that if a song gets on the radio, it's most likely going to be in November or December. Why would I try earlier than necessary? At the end of May, the models passed the knee of the curve and are now improving exponentially.
That's why I continue to buy YES shares; in a rational market where the technology is improving, resolution is going to be determined at the last minute.
Dear traders, and @jacksonpolack , @Ziddletwix , @SemioticRivalry , @Joshua , @Bayesian,
I have placed limit orders for YES at various price points.
Now that I'm waiting for the defendants to reply and for the R&R in the case, I think I'm ready to attempt this. Udio's "styles" feature has progressed the state of the art enough to take a shot.
I just got a hook that can achieve this, but it's going to take a lot of work to get the vocals just perfect. Expect an update in mid-June.
@SteveSokolowski I’m interested in knowing more about the AI model that you’re using to produce MIDI files
@probajoelistic I"m not using one yet. I'm just reserving that possibility for o3 being able to do it.
I think though, that I had a breakthrough that may end up causing that be unnecessary. I'm working on a "reasoning" or "test-time compute" architecuture for music models over the next few days by using o3-mini-high to output Python.
The realization I had is that music models right now are like GPT-4.5 - they just output stuff and have no idea what it sounds like. We can add "reasoning" to them by connecting Gemini Pro 2.0 Experimental 0205, and entering a loop where Gemini improves the prompt until it is satisfied with the results.
In my manual testing by dragging and dropping, this works dramatically well. For example, it can completely eliminate the "AI-sounding vocals" Issue. I'm hoping to have this automated by the end of the week.