Gustav Söderström

You're copy pasting prompts from Twitter or whatever. In an app like Spotify, how willing are people to get not lazy and really descriptive about what they actually want? What have you learned about the nature of people's laziness versus willing to put a lot of work in to get the thing that they want via that more rich uplink.

You're copy pasting prompts from Twitter or whatever. In an app like Spotify, how willing are people to get not lazy and really descriptive about what they actually want? What have you learned about the nature of people's laziness versus willing to put a lot of work in to get the thing that they want via that more rich uplink.

Yeah, so that's probably the most exciting thing for us of this generative AI age and the dual uplink paradigm. So previously, we mostly relied on some explicit input when you playlist. That's high value information. You are sitting there thinking, this song goes really well with this song and that song.

Yeah, so that's probably the most exciting thing for us of this generative AI age and the dual uplink paradigm. So previously, we mostly relied on some explicit input when you playlist. That's high value information. You are sitting there thinking, this song goes really well with this song and that song.

So if you think about it as labeling, even though you're playlisting for yourself, you're sort of labeling these tracks, at least in relation to each other. And you're putting a lot of effort in it. And that was and is our big advantage in music recommendations, even though generative recommendation systems are starting to take over from these more old school collaborative systems.

So if you think about it as labeling, even though you're playlisting for yourself, you're sort of labeling these tracks, at least in relation to each other. And you're putting a lot of effort in it. And that was and is our big advantage in music recommendations, even though generative recommendation systems are starting to take over from these more old school collaborative systems.

So we had some really strong signal like that, where you quite seldomly invested a lot of time in producing a data set that described you about playlisting. But most of the time, We just had skips. And the challenge for us is the phone is in the pocket. So even if we had a thumbs up down, you're not going to take out the phone every time and say, I didn't like this because of that.

So we had some really strong signal like that, where you quite seldomly invested a lot of time in producing a data set that described you about playlisting. But most of the time, We just had skips. And the challenge for us is the phone is in the pocket. So even if we had a thumbs up down, you're not going to take out the phone every time and say, I didn't like this because of that.

Or even do a thumbs up, thumbs down requires you to take up your phone, unlock it, open Spotify. What you can do from your earphone is to skip. So we have the skip signal, but that is a very blunt signal. So we play you a song and you skip it. That could be because you absolutely hated it. It could be because you love it, but it's a hundred times. You're tired of it. It could be that you love it.

Or even do a thumbs up, thumbs down requires you to take up your phone, unlock it, open Spotify. What you can do from your earphone is to skip. So we have the skip signal, but that is a very blunt signal. So we play you a song and you skip it. That could be because you absolutely hated it. It could be because you love it, but it's a hundred times. You're tired of it. It could be that you love it.

You're not tired of it, but you're at the gym. So jazz is not the right thing. All of those just look like a skip to us. We have a lot of that signal, but it's blunt and you will never get to perfect personalization through that. Now what we find with Generative AI, one of the first services that we've launched that is live now in 40 countries is something called AI Playlisting.

You're not tired of it, but you're at the gym. So jazz is not the right thing. All of those just look like a skip to us. We have a lot of that signal, but it's blunt and you will never get to perfect personalization through that. Now what we find with Generative AI, one of the first services that we've launched that is live now in 40 countries is something called AI Playlisting.

We literally use an LLM that is trained on your listening data and world knowledge and so forth. And you can literally tell us in English what kind of playlist you want. Previously, you could playlist songs and maybe put a title on it. We could guess this is probably a running playlist. So we could do something. Now you can say, I want a running playlist that is EDM. I want big drops.

We literally use an LLM that is trained on your listening data and world knowledge and so forth. And you can literally tell us in English what kind of playlist you want. Previously, you could playlist songs and maybe put a title on it. We could guess this is probably a running playlist. So we could do something. Now you can say, I want a running playlist that is EDM. I want big drops.

I want it to be 160 BPM. And then you get a suggestion from the LLM. And then you can keep a few tracks and say, these were good. These were not good. Now refine it. I don't like these artists, but I want more of that. So for us, it's the first time that we get that kind of fidelity of what is actually in the user's mind.

I want it to be 160 BPM. And then you get a suggestion from the LLM. And then you can keep a few tracks and say, these were good. These were not good. Now refine it. I don't like these artists, but I want more of that. So for us, it's the first time that we get that kind of fidelity of what is actually in the user's mind.

When we think about Spotify, we always try to reproduce a small part of your neocortex on our servers, which is very hard with a click stream of skips. Now, when you tell us what is in your mind, it gets easier to approximate you as a person. So this is really the first time that we have that signal. One way I like to think about it is when we do user research, we do two things.

When we think about Spotify, we always try to reproduce a small part of your neocortex on our servers, which is very hard with a click stream of skips. Now, when you tell us what is in your mind, it gets easier to approximate you as a person. So this is really the first time that we have that signal. One way I like to think about it is when we do user research, we do two things.

We do quantitative testing, A-B test, but before that, we do qualitative testing. We interview a few people deeply to understand the need. Then we build a product and then we A-B test to see if we're right. The promise of generative AI is really a deep, ongoing, qualitative user research with almost 700 million users all the time. It sounds big, but if you squint at it, that's kind of what it is.

We do quantitative testing, A-B test, but before that, we do qualitative testing. We interview a few people deeply to understand the need. Then we build a product and then we A-B test to see if we're right. The promise of generative AI is really a deep, ongoing, qualitative user research with almost 700 million users all the time. It sounds big, but if you squint at it, that's kind of what it is.