The music industry has already passed the threshold for digital AGI, but seems to be experiencing disruption dramatically less than AI song generation capabilities might suggest. Anyone can create a song of expert level quality in under 10s with an automatically generated prompt (or write one themselves if they want). The Chainsmokers took about 8 hours in total to make the single Roses, which has over 1.3bn streams as of 02/10/2026 and I believe is on the quicker end of writing time for pop songs, although I am not sure reliable data exists for this anywhere. However, despite AI’s capability to create expert level music 2,880x faster than human artists with essentially 99.9% reliability, this has not significantly disrupted music production or consumption in the past year since the capability was developed. I believe this is because the bottleneck of music productivity is in capturing attention, not generating additional supply of songs. I think it is worth investigating the economics of industries where AI has already fully automated digital human labor with some 10x or more improvement along a relevant dimension to understand when and where access to cheap intelligence will matter and by how much.
the bottleneck of music productivity is in capturing attention
Yeah. I think the same applies to capitalism in general; these days it is relatively simply to produce many things, the problem is that thousand competitors can do the same thing, so the key is to convince the customer to buy your product instead of their mostly identical products.
Now I am not saying that good songs are identical, but they are in some sense fungible. Like, if you have a favorite song, it is easy to imagine a parallel universe where it doesn’t exist, and something else is your favorite song, and you are probably not any less happy in that universe.
As a random example, there is a CD series “Super Eurobeat” that contains 250 CDs. Assuming about 12 songs per CD, that’s about 3000 songs. It would take me more than a year to only listen to each of them once (assuming that realistically I can’t spend the entire day listening to music, I also have other things to do). These songs are already selected as the “best of” within given genre, sometimes just one best song from a music group that has produced many. And that’s just one genre, not even the most popular one. So from my perspective, the music is already well beyond scarcity; the problem is discovering the pieces that you might like. (The recommendation algorithms I have tried in the past have failed me. They mostly recommended “things that many people like” and “other songs from the same sings”. What I would like to see instead is an algorithm that can somehow figure out my music taste, and then give me songs, possibly obscure, that score very high.)
The music industry has already passed the threshold for digital AGI, but seems to be experiencing disruption dramatically less than AI song generation capabilities might suggest. Anyone can create a song of expert level quality in under 10s with an automatically generated prompt (or write one themselves if they want). The Chainsmokers took about 8 hours in total to make the single Roses, which has over 1.3bn streams as of 02/10/2026 and I believe is on the quicker end of writing time for pop songs, although I am not sure reliable data exists for this anywhere. However, despite AI’s capability to create expert level music 2,880x faster than human artists with essentially 99.9% reliability, this has not significantly disrupted music production or consumption in the past year since the capability was developed. I believe this is because the bottleneck of music productivity is in capturing attention, not generating additional supply of songs. I think it is worth investigating the economics of industries where AI has already fully automated digital human labor with some 10x or more improvement along a relevant dimension to understand when and where access to cheap intelligence will matter and by how much.
Yeah. I think the same applies to capitalism in general; these days it is relatively simply to produce many things, the problem is that thousand competitors can do the same thing, so the key is to convince the customer to buy your product instead of their mostly identical products.
Now I am not saying that good songs are identical, but they are in some sense fungible. Like, if you have a favorite song, it is easy to imagine a parallel universe where it doesn’t exist, and something else is your favorite song, and you are probably not any less happy in that universe.
As a random example, there is a CD series “Super Eurobeat” that contains 250 CDs. Assuming about 12 songs per CD, that’s about 3000 songs. It would take me more than a year to only listen to each of them once (assuming that realistically I can’t spend the entire day listening to music, I also have other things to do). These songs are already selected as the “best of” within given genre, sometimes just one best song from a music group that has produced many. And that’s just one genre, not even the most popular one. So from my perspective, the music is already well beyond scarcity; the problem is discovering the pieces that you might like. (The recommendation algorithms I have tried in the past have failed me. They mostly recommended “things that many people like” and “other songs from the same sings”. What I would like to see instead is an algorithm that can somehow figure out my music taste, and then give me songs, possibly obscure, that score very high.)