Suno's Mikey Shulman: Everyone Can Make Music Now

0:00In Western music, there are 12 tones. If you tell the model there are 12 tones, it will only ever produce those 12 tones. You will be forever limited. And so for us, it was all about, let's throw away everything we know about music. And let's try to do this from scratch. And it's like, it's just a sound wave. It's just sampled at 48 ,000 times a second. And it is a continuous, you know, float 32 number. And let's figure out how to model that. And that was a lot of the early breakthroughs that we had to make. But once we did, we realized that it is a totally generic music-making machine. And now you are only constrained by what you can describe and your imagination.

0:54I'm delighted to welcome Mikey Shulman. Mikey is founder and CEO of Suno, which is building a music company or a creative entertainment platform and has been one of the most novel consumer applications I've seen out of AI. And I'm very, very excited to ask you about your journey and what's ahead for Suno. So thank you for joining us today. Thank you for having me. I'm excited. Okay, awesome. I want to start with your background because it is very, very unexpected. You went from a physics PhD at Harvard, I think quantum computing with solid state spins, to building the largest AI music company in the world.

1:30What insight connected those two things for you? You know, I don't know how. I don't know. Like on paper, I guess I have no business building a consumer entertainment company, but a lot of people went from physics into AI just like 30 years ago. A lot of people went from physics into quantitative trading. I'll be honest, though. Like I was an okay physicist only, and there are a lot of better physicists, including one of my co-founders. And I think what I mostly learned is playing at the nexus of two things that don't usually play together is just a massive opportunity in all domains. It can be music and technology.

2:08It can be quantum mechanics and low temperature microwave engineering, or it could be whatever else you're going to do. You and I got connected in the very early days of Suno when our mutual friends, Harrison Chase, was one of the earliest suno discord users and he was having far too much fun making songs in your discord maybe tell us about the early days of suno how did it how did it come together did you did you set out to build a music company um originally we thought this would actually be too hard um and it's because uh you have to rewind this is like pre the chat gpt moment um we did some like back of the envelope math.

2:50We knew we loved audio, but the back of the envelope math told us that actually producing good music, making good music, generating good music, was like a couple of orders of magnitude away in terms of compute and model size and capability. And it's because music, sound in general, is like very unwieldy. It's not in discrete bits like text is. And so we actually started building a company that was all around using the same technologies to make sense of audio, not to produce it. And very happily, pretty early on, we had the right breakthroughs and we realized, oh, we actually can make music. You're pretty good at math.

3:29What'd you get wrong with your back of the napkin math then? The math was right. We just had some breakthroughs that said like, it's actually, you don't need that amount of compute. You can make the right technological breakthroughs to, if you want to think about it, basically just compress audio really, really efficiently um and that worked a hell of a lot better than we anticipated so it was like a very nice being wrong moment um not all being wrong moments are so pleasant and um to be clear at the beginning the music was terrible but we still uh stayed up late it was good he was only the first 10 users i think he he thought he thought he was pretty good certainly before we put it on discord the music was very terrible before we put it on discord we could make like 12 and a half second clips that wouldn't always listen to the words you ask them to sing.

4:19But we had so much fun doing it. And we thought other people might have fun doing it. And so we kind of took the example of Mid Journey and we said, it's really easy to put a Discord bot out and see, will people enjoy it? And we put it out there and a hell of a lot of people enjoyed it. And that was a really confirmatory moment for us. And so a lot of people told us not to build a music company. It's not the easiest business to work in. Speech is really big. There's a lot of great business use cases for building speech technologies. But when you are staying up late playing with the thing and you don't want to go to sleep, it's like a really good sign that that is what you are meant to be doing.

4:55And so that's what we did. I love that. Are you a musician? I am. I play almost every day. I grew up playing a lot of piano and ended up picking up a bass around age 12 and playing a lot more of that. Okay. So personal and passion point. That's awesome. You know, the revisionist history, which is true, is that we used to have jam sessions at our last company in one of my co-founder's basements. And it's true. We had a lot of fun there. It's not why we started the company. Again, we thought it would be too hard to do this. It was just fun. Meaning at Kensho? At Kensho, yes. Where I met the great Harrison Chase.

5:31The Kensho mafia is like pretty unparalleled. It's Harrison, but also Daniel Nadler, Sam Whitmore you there's a lot of you there's a lot of us I just credit Daniel with that honestly Daniel is like I think the best object lesson in what talent density can do for a company and it was a lot of people with non-traditional backgrounds it skewed very young but he was great at finding people and great at convincing them to join I love that okay so walk us through what happens when like somebody types upbeat 90s hip hop track about a road trip You get the prompt in. What happens? What is the model doing to be able to pass something back to the user that seems like it's quite special?

6:14In some way, it's actually pretty simple. A prompt like that, you have to figure out what are the words of this song. And we use various LLMs to do that, to make the lyrics. And it's taking basically the cue there is road trip. And so like, what should this road trip be about? And it will probably get it wrong because you didn't give us enough information, but that's actually OK. you can iterate on it. And then you said 90s hip hop. And we tried to expand that out into a set of cues that the model can really understand what is the genre, what is the style of this music. And then you put those things together.

6:46You have a lot of lyrics, you have a lot of styles. And we have our models that are trained to take in all of that information and just produce sound. The amazing thing here is that the models don't know that there's vocals and instruments. It doesn't know what kind of instruments there are. Very early on, it was actually quite obvious to us that the more musical knowledge we give the model, the more constrained it will be in a bad way. And so we actually just model everything as sound. And that's what made it so hard, but ultimately that's what make these things so powerful. So just to be concrete about it, in Western music, there are 12 tones.

7:21If you tell the model there are 12 tones, it will only ever produce those 12 tones. You will be forever limited. And if you tell the model there's 200 instruments those are the only sounds that you'll ever be able to make and you won't get the next skrillex using suno and so um for us it was all about let's throw away everything we know about music and let's try to do this from scratch and it's like it's just a sound wave it's just sampled at 48 000 times a second and it is a continuous you know float 32 number and let's figure out how to model that and um that was a lot of the early breakthroughs that we had to make but once we did, we realized that it is a totally generic music making machine.

8:03And now you are only constrained by what you can describe and your imagination. That's so cool. Have you found that we've basically just rediscovered, you know, the existing genres of music and the 12 notes? And like, have you, I guess, independently seen just that same behavior emerge when you're trying to, I guess, learn music from first principles? Or have you seen like a different set of capabilities emerge? No, the amazing thing is now we see new things emerge that you never would have thought of. And so most of the time, this looks like blending genres that really have no business going together.

8:37And so you'll get, I don't know, Trap with a sitar in it, or you'll get Country with 808s in it, or whatever it is. And again, this is like really empowering people to do the things that are in their heads. And it's in a way that would not have been possible without a technology like this or would have been really, really hard. We see microtonal music. It is really inspiring to go and just look at all of the crazy things that people are making. A lot of them sound like genres you know, and a ton of them sound like totally strange and bizarre and lovely. That's awesome. That's really cool. Are there certain genres that you're finding your model is better at in certain genres where you're worse at?

9:18Definitely. We are, I mean, I attempt not to say like good and bad about music other than, you know, it's sampled well, like the full bit depth or full sampling rate. But to the extent that you can make such generalizations, we're like very good at country, we're very good at pop music. And I think the cartoon maybe to have in your head is that there are some genres that are somewhat more formulaic than other genres. And so perhaps we're better at those. But I have some sneaking suspicion that for those, it's as much raising the floor as it is raising the ceiling. And for the things where we're less good at it, we've not raised the floor, and so we make a lot of bad music.

10:06But we have also raised the ceiling. And if you're willing to go for long enough, you'll find amazing stuff. That's so cool. So, Suno V5 seems like it was a real step change in quality. What goes into one of the step changes? You know, it's really hard to predict when the step changes happen because it's really nonlinear in both the research inputs. But actually, it's not even linear in how much, like, our testing says the model is better. And so just as an example, we can measure how much one model is preferred to another model. And you may come up with it's 10 % preferred or 15 % preferred. And you can take two different models and one is 10 % preferred or 15 % preferred.

10:50And the uptake on the other end, how much our users actually love it and use it or how much the product grows when you release it, won't necessarily be all that correlated with what the preference signal is. And it's because music is messy and there's lots of other things that go into it. But to take a huge step back, like we have a pretty aggressive research roadmap. And in some weird way, we're always working on this thing, you know, like, we know what v6 and v7 are at some point, there's lots of things that you want to have your model do, there's lots of improvements that you want to make. And it's almost an arbitrary cutoff of saying like, okay, this is the break, this is what we're going to call v5.5.

11:27And everything that comes after is going to go into the next models and almost just to keep it on a steady cadence of when we release things. because what you would hate to have happen is we don't release stuff for like two years and we try to make, you know, the music model to save humanity. And that's going to come out in two years and we're going to do nothing before then. Yeah, totally. How much of, you know, each of these improvements do you think is just a function of scale, scaling compute, scaling data, and then getting a lot of human preference data back? How much are you guys doing, I guess, more novel research?

12:00Music is really not a scale problem. The models are pretty small. for a variety of reasons. And I think people will often incorrectly take what they know from LLM land, where models are giant and scale helps a ton, and apply it to music. And I think the cartoon that I have in my head is that in LLM land, there's all of these benchmarks. And you can quibble about which ones are flawed and which ones are good. But these benchmarks exist. And scale is actually a pretty efficient way to climb up the ladder and just keep doing better and better on the benchmarks um in music there are no right answers there are no benchmarks um and so scale is somewhat less helpful in solving it it's like a it's a messier problem in many ways aligning models to creative human tastes you and i are not going to agree on every song you and i aren't even going to agree i'll just defer to whatever you say i don't have i mean i i don't think you want to do that um but so um the the and the models not being that big actually lets us um get you the music quicker, which turns out to be really important for good UX.

13:04And so I think a lot of this boils down to research and preference data. And so we gather preference data that lets us align models to what our users like. A really underappreciated thing is how much this preference data actually lets us do research. Like without the scale of preference data that we have, we wouldn't even be able to develop the techniques that we are using. And so there are really some virtuous cycles there in how the product itself keeps getting better just by virtue of having people use it. Interesting. And I guess you can use the human preference data in a much stronger way than the text models because they're all worried about sycopency, right?

13:45And for you, I guess that's much less of a challenge. 100%. 100%. And so I think, yeah, there's just a tremendous amount of edge comes from our ability to understand it, do research on it, and then RL that back into our models. That's awesome. Okay, I'm going to switch gears a little bit and talk about music as a consumer phenomenon. And, you know, you've mentioned consumer creative entertainment platform at the beginning. I want to dig into what that means. Maybe starting with, like, seems like music is just like a cultural social phenomenon of, you know, I like this song, I send it to my friend, you know, it's a scarce resource, we bond over liking that song, and having the mixtapes, listening to it together, et cetera.

14:32And to me, music has always just been like this shared cultural experience. That's what it is. Do you agree with that? And then if so, what does AI music mean for that? I agree with that very strongly. Music has a very different place in culture than other media in a variety of ways. One is actually people's tastes are far more developed in music than they are in other media. Like everybody has taste in music in a way that most people don't have taste in film or literature. And the other thing is that music is actually inherently a much more social medium. And if you think about it, how going to a concert is an inherently social thing, even though you're only really looking at the performers, and it's because of the people around you, in a way that, let's say, going to a movie in a movie theater isn't quite as elevated as it would be compared to an empty movie theater, for example.

15:18And so I think this is actually largely that humans communicating sonically through our mouths and ears. And therefore, music is like a much earlier method of communication than writing. It's like much more in our DNA, I think, music compared to other things. I'm obviously biased. I obviously love music. I'm not sure. I think people assume that, oh, you're just going to have like AI powered Spotify and it's going to dehumanize it and music is going to get terrible. That seems to me to be obviously wrong. I don't think you're going to make a better Spotify just by powering it with AI. And the thing that's really interesting is actually how can we not just change, but elevate the place of music in culture.

16:01And music has this other funny thing that by virtue of being so ubiquitous, it ends up being in a background a lot. And the thing that's amazing is that AI can be used to actually change that and to augment how music is perceived in society and in culture, augment how it is used socially, because it's actually become less social in the last 30 years. And so that is the corner of the universe that we play in and that we are really excited about. Do you see yourselves as today, and I guess when you look at your users, are people more creators of music or are they more consumers of music or both?

16:37This is the crazy thing about Suno. Before Suno, basically everybody was a consumer of music. You know, compared to the 8 billion people on the planet, there are very few people who make music. and the rest of us consume it. And that's fine. It tends to cater to passivity. It tends to cater to making it less social and more impersonal. And the crazy thing about Tsuno is that in any given day, 90 % of the users are going to create something. And the thing that's hard to wrap your head around is you're not creating it to go bring it elsewhere, by and large, to do something with it. People are creating music for the fun and enjoyment and fulfillment that comes with being creative.

17:15And so that, the creation, is actually the entertaining bit. And that is the big step change. It's like everybody in the world is creative. Being creative makes you feel a certain way. This is like in our DNA. And we are basically using technology to allow everybody to feel those warm and fuzzy feelings. A lot of like the inspiration for me personally for doing this comes from just remembering like the fondest memories that I've had or some of the fondest memories that I've ever had or making music with my friends. Not even performing in bands. It's like practice was so much fun. And you get really close to people making music.

17:47And it's because it feels really good to be productive in a way that doom scrolling your favorite app for an hour does not feel so good when you're done. I was an orchestra kid, so I was not doing nearly as cool music as you, but I totally agree. What did you play? Violin. Do you still? Yeah. Oh, excellent. I have perfect pitch. And let's just say I'm definitely not playing 12 tones now. So my ears bleed when I play. But I totally agree with you. Okay, so that's crazy. So you're very much like a – this is not a replace Spotify type platform. This is a creator platform. You're turning people into creators.

18:30This is an active entertainment platform. It almost feels more like gaming than it feels like music listening or music as we currently understand it. 100%. I mean, I will get shit inside of the music industry for saying that there's a lot to learn from gaming. But there's a lot to learn from gaming. And it's not just about how well it monetizes. It is how it grabs your attention and how it pulls you in and makes you use your brain. There's also things about the business model about gaming that music has to learn from. But I think I understand why it is in some sense taboo because music is art and people don't think of gaming as art.

19:11But I think it's silly to say that there's nothing to learn from it. Yeah, yeah, that's awesome. Okay, so it's like a self-expression slash active entertainment platform, some parallels to gaming. And some parallels to even like cloud code, right? Absolutely. So I think the thing that's amazing about making music is that you feel good and fulfilled and you enjoy making it and then you listen to it. And there are parallels. So that's what we call creative entertainment. The entertaining part is being creative. It's not that you are being creative for the sake of bringing the piece of content somewhere else.

19:47I think you see that in cooking. People like to cook even though they can get a better meal at a restaurant. And it's because it is fun to cook and it is fun to consume what you make. And I think a lot of what makes Cloud Code or any of the other platforms so special is that it's fun to build things. And it's fun to use what you build. And even though most of the things that I build are definitely not meant to be hosted in AWS and used by millions of people, I actually enjoy the act of building and I enjoy the act of using the thing that I built. And so I predict that like in 10 or 20 years, there will be way more of these creative entertainment things all over the place.

20:24And it's because that's actually finally possible. Like that is the thing that AI unlocks. It unlocks lots of intelligence things too, but it actually lets everybody be creative in almost any domain. Yeah. I'm guessing you have an opinion on this. What do you think of the word slop? I do. I do have an opinion. I actually I mean, my answer is usually it's thrown around without any meaning. And I don't know. I don't know what people what people mean by that. So, like, sometimes in music, it goes all the way to say, like, there is some streaming fraud. And so people think streaming fraud is slop.

20:58I don't really understand that. Like the fraud part is bad, but the fact that the music was made by AI is an implementation detail. I made two songs with my five-year-old yesterday. Is that slop? In the sense that 99.999 % of the planet has no interest in hearing that. Sure. But that's really meaningful to me. And so if you call that slop, I'm not sure I care. It's an interesting question, though, right? This has happened before, at least in music, where when way more people start to be able to produce something, people get afraid that it's just going to flood all of our ears and all of the platforms with more content.

21:39And this happened when people started to be able to make music on their laptops. You had a lot of 13-year-olds making beats in their bedrooms. And you fast forward to today, that seems like obviously a good thing. Yeah, there's way more music. It means that there's way more, quote unquote, bad music. but it also means that there's way more great music and there's new kinds of music that get made and there's new kinds of stars that get made and I see no reason why way more people making music again would be any different from that I love that so we talked about the floor the slop floor, the non-slop floor what about the ceiling?

22:16tell us a little bit about the most incredible things people have been able to do with Suno I think you guys have had some chart-topping hits now maybe talk a little bit about that um we have had some chart topping hits we've had people sign to record deals we've had people make single songs that chart um and that's amazing and i think about that as that is a new creator um coming with a new perspective that resonates very strongly with people and so that is obviously the ceiling going up my favorite example is an iam onay um who it's the the stage name of um a poet who took all of her beautiful poetry that she had been writing for like a decade and started to make music out of it and found an entirely new voice and an entirely new audience to resonate with her art and like yeah i think this is fantastic this is this is um people connecting right these are this is like the the most personal thing in the world and when you go listen to the music you realize it's extremely personal like the best music will always require human guidance and it's because again music has no right answer you like a piece of music because of how it sounds and because of the messenger who delivers it and we will find new messengers with new sounds and we already are.

23:27And that to me, it's like, obviously the ceiling is going up there. The other thing that's really cool is even if they're like, I just know that there are tons of charting tracks that have little bits of Suno in them. They're not entirely Suno. And it's because for the professionals, it's also just an amazing tool to use as part of your workflow. It's not your whole workflow. And so there's this weird thing where I think people incorrectly say it's like either all AI or non AI. and the vast majority of music will have some AI in it, just like the vast majority of music today is auto-tuned or is digitally produced.

24:00And again, more tools let you push music forward faster, let you find new sounds faster. To me, this is like obviously the ceiling going higher. That's amazing. Okay, so you chose to go after music, which is like probably the one industry that a lawyer would tell you not to go after because the minute you're there, you have pitchforks coming after you. So you just, I think you just had a pretty landmark settlement partnership with Warner. Can you tell us more about that and what you think that means for the future of collaboration with the existing professional music industry? Absolutely. I think, just to back up, you know, I think people incorrectly assume that we hate the existing music industry and especially we hate the record labels.

24:48People also expect me to say like, oh, the record labels are cooked. I think that's obviously wrong. They're some of the most culturally important institutions in the world. They understand music and they understand music culture. They cultivate and grow stars that resonate with billions of people. And the way I see it, it would be a real shame if there were two worlds of music, if there were like an AI world of music and a non-AI world of music. One is it makes no sense because most music will have some AI in it. But the other is it's just bad for the end user to think about having to separate these things in your head and to have to go to different platforms to have effectively similar usage patterns or interactions.

25:28And so what I'm most excited about doing with Warner is actually building things together that could never have existed before and building products that let fans interact with their favorite artists and really deepen the artist-fan connection in ways that are just positive sum for everyone. It's great for the artists. They get to engage with their fans. It's great for the fans. They get to feel like they are engaging with their favorite artists through music. It's great for the rights holders. Obviously, this is like a heavily monetizable thing. And it's something that literally could not have existed up until like approximately right now.

26:02And my sincere hope is that going forward, we find way more of these opportunities of things to build together that couldn't have existed until today. And just to say it out loud, like the digital musical experience has basically not changed for 25 years. We've just been streaming music for 25 years. And I think music is like due for a new innovation and a new format. And so that's what we're here to do. When are we going to see Suno at Coachella? You probably have already. It's like it's probably in a lot of the music. It's probably in a lot of the backing tracks. I'm sure some of the songs...

26:35But I mean like a main stage, you know, like a consumer participation thing. I hope that at some point in the next year, we see a truly interactive concert where the audience is actually able to participate and make music with that artist. One of the coolest parts of my job is if I go and I demo pseudo to an audience of hundreds or even a thousand people is making a song with that many people all at once. And it's a very special moment. And it's like almost religious, you know, like a lot of religions will do like chanting and singing in large groups. And why shouldn't why does that have to be like confined only to a religious context?