Yes it’s using “an AI”. But that doesn’t mean anything. You can’t just use any AI and have the same result. Just cause AI got a global hype doesn’t mean this is new either. Neural networks have existed for many decades, which is likely what they’re using. The hard part is to get the training data. That is where the value (or usefulnes) comes from. And that source is all their users, listening to all the music, importantly including newly released music, all the time. It’s the basic idea of “people who liked X also liked Y”. What songs people combine together in a playlist. That sort of thing.
We don’t have that data to train “an AI” so we have a local version of this. They have it for millions of users. That’s why their AI is incredibly good at this task. Sure, they also let labels pay them to rank things higher so they get more listens, and that is anything but transparent when and how that happens. But over all, you can’t just magically do what they are doing locally.
Ok, so it doesn’t exist because of the training data, couldn’t that be achieved by the self hosted users sharing anonymous training data? Or maybe sharing curated playlists to help the suggestions for others? I totally understand some may see it as a privacy concern, but since the ultimate privacy is becoming disconnected to everyone and everything, there might be a level of sharing that can be achieved where users could feel comfortable.
Note, I just woke up and this makes total sense to me right now, so don’t hate me please.
Isn’t Spotify just using an AI? Couldn’t one self host an AI that plugs into the music library and makes recommendations?
Yes it’s using “an AI”. But that doesn’t mean anything. You can’t just use any AI and have the same result. Just cause AI got a global hype doesn’t mean this is new either. Neural networks have existed for many decades, which is likely what they’re using. The hard part is to get the training data. That is where the value (or usefulnes) comes from. And that source is all their users, listening to all the music, importantly including newly released music, all the time. It’s the basic idea of “people who liked X also liked Y”. What songs people combine together in a playlist. That sort of thing.
We don’t have that data to train “an AI” so we have a local version of this. They have it for millions of users. That’s why their AI is incredibly good at this task. Sure, they also let labels pay them to rank things higher so they get more listens, and that is anything but transparent when and how that happens. But over all, you can’t just magically do what they are doing locally.
Ok, so it doesn’t exist because of the training data, couldn’t that be achieved by the self hosted users sharing anonymous training data? Or maybe sharing curated playlists to help the suggestions for others? I totally understand some may see it as a privacy concern, but since the ultimate privacy is becoming disconnected to everyone and everything, there might be a level of sharing that can be achieved where users could feel comfortable.
Note, I just woke up and this makes total sense to me right now, so don’t hate me please.