Ep 35: CEO of Suno Mikey Shulman on Future of Music with AI, Tactics for Model Eval and Solving the Blank Canvas Problem

Last week, Suno announced $125M in funding, marking a significant milestone in their journey to reshape the music creation landscape. On this week's episode of Unsupervised Learning, we caught up with Suno's founder, Mikey Shulman, to dive into their approach to multiplayer music collaboration, how they got Suno to be so fast and the future of digital concerts (and workshopped new intro music for the pod 👀).

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jacob Effron?

  AI is completely changing the music industry and no product has been more viable out there than Suno. With over 10 million users having generated songs and a recent fundraiser $125 million at 500 million. Mikey Shulman, the CEO of Suno, is really at the epicenter of this ecosystem. We had a fascinating conversation about where Suno is going, the future of AI and audio.

• 0:17Speaker 01

  I try to think very hard, like, what should the future of music look like?

• 0:20Jacob Effron?

  And some of the product and infrastructure challenges they faced as they've scaled to their massive user base. I think folks are brilliant and will enjoy this. Definitely stay tuned for Rashad and my debrief at the app. Mikey, thanks so much for coming on the podcast.

• 0:29Speaker 01

  Oh, it's my pleasure. I'm excited to be here.

• 0:31Jacob Effron?

  Yeah, I mean, I guess to start out, I'm sure no one's ever asked you this before, but what's your favorite piece of music that's been created on Suno?

• 0:37Speaker 01

  Favorite? That's a tricky one. One of two things. There's a piece that is an Estonian-language German lead that, for some reason, touches me in ways that I cannot really describe. Don't speak Estonian. Not the genre of music that I'm particularly into, but I don't know, it does it for me. And so I've probably listened to that one 100 times.

• 1:02Jacob Effron?

  How much has the rest of the Suno user, are you like 100 of the 100 listens or is this a pretty popular one?

• 1:06Speaker 01

  Maybe 100 of the 500 listens. It's not, yeah, it's not the most popular song ever, but for some reason, it did it for me. There was an incredible amount of care taken by the user who made that to really craft it in a way that I guess was the vision that they had. And for some reason, that vision resonated very strongly with me for reasons I can't entirely explain.

• 1:27Jacob Effron?

  It's awesome. I thought that's the best music a lot of times.

• 1:29Speaker 01

  I think so. I think so. Yeah, I just love it. I can't, I don't have a ton more to say of why I haven't really figured it out yet. But that's, I guess, the beauty of music is like, it kind of touches you. It makes you feel that is why we all love music. And so that one touches me and makes me feel in a certain way.

• 1:44Jacob Effron?

  Yeah. How about of your own creations?

• 1:46Speaker 01

  Oh, I do a lot with my son. And so there's a ton of songs that are about him in different fantastical situations.

• 1:58Speaker 01

  One in particular recently of him driving his Zamboni, which three-year-olds are definitely not allowed to do in the state of Massachusetts. But, and I think the reason there, I think that those resonate so strongly with me is a little more obvious, which is that it is the experience of crafting music with someone, in particular, your son in this case, that it is that whole experience that I'm really drawn to and not only the final product. And I think there's actually something to that where.

• 2:24Speaker 01

  Right now, music is really mostly made in service of the final product, of the final song, of the final album, and without a lot of care for the journey. But I think making music with other people is some of the most, if not the most enjoyable moments of my life. And so I think that's really why I love those songs. It is the moment in time that they mark that I will never get back. So I'll go with one of those.

• 2:45Jacob Effron?

  That's amazing. Do you get feature requests from your three-year-old son? Are there things that he wishes the product would do?

• 2:50Speaker 01

  Not a ton. Mostly song requests. And then a lot of, "I don't like that one," for reasons that he can't entirely explain. So we just do them again and again until we get them right.

• 3:00Jacob Effron?

  That's awesome. I love that. You'll have to send him the Zamboni one after. That feels...

• 3:04Speaker 01

  Yeah, we can play it.

• 3:05Jacob Effron?

  That's a fun... I feel like young kids have such wild imaginations to be able to just actualize that.

• 3:10Speaker 01

  There's something, I think, very special about kids. And I don't know if it is their view on the world, or the fact, maybe, they don't have shame yet, but the way that they just interact with music and love to rock out with, you know, spoons for drumsticks, or pretending to play the violin, or the trumpet, or the tuba, or whatever it is. And it is a form of play that adults lose, for whatever reason. And that's too bad. And I think part of what we are trying to do is bring that back to adults.

• 3:41Jacob Effron?

  It's so interesting. I thought that's a common theme in a lot of these AI creator tool companies, where the idea that we all had this imagination at one point, and the ability to just kind of dream up fantastical things, and at some point, maybe we were told, "We're not a great artist or musician," or "You have to be this level," or "Use these complex tools." And now to kind of democratize access to that inner kid within all of us is pretty cool.

• 4:02Speaker 01

  Yeah. And bring it out. I think we talk a lot internally about, like, paint by numbers as being an experience that people actually liked. And somehow, we've lost touch with sort of a guided creative experience in some way. And I think, again, that is really too bad. People... I think humans are really hardwired to enjoy these experiences. And there are a lot of, let's say, cultural reasons why these kind of get beaten out of us. And I'm excited to see what happens when we bring these back.

• 4:35Jacob Effron?

  It'll be cool to see, as your son gets older, too, what he... As the evolution of... You probably can track a lot about aging just through the evolution of the songs that you guys generate.

• 4:43Speaker 01

  100%. I see this very much as, like, I... People have asked me before, will you go back and listen to that song in the future, whether it's the Zamboni song I made with my son or the Estonian lead? And the answer is 100%. And people, I think, are a little bit incredulous, but I think, it is totally obvious to me that I'm going to go back and look at pictures in my phone from this time period. And so why wouldn't I listen to those songs? The same way, I think we all have songs in our lives that we listen to, and they bring us back to a certain moment, for whatever reason, they are tied to that moment. And I think this is a way to just amplify that.

• 5:17Jacob Effron?

  I mean, I guess, digging into Suno itself, I'm curious, like, how do the people who best use Suno today use it? Obviously, I feel like you have these two class of users, right? You have folks that are just kind of prompting from scratch, and that's me, I'm the basic user. And then I think you obviously have folks that are doing more advanced things in the product. Like, what does the ideal Suno user look like today for you?

• 5:39Speaker 01

  It's a good question. I think, I won't say best user, I won't pass judgment there. But there's kind of, right now, two categories. So one is you, and you said "basic", and I prefer to think of this more as casual. And that isn't to say casual is not using it a lot, but the use cases are what we like to call "soundtracking your life". I am narrating my life musically as happy and sad and funny and memorable things happen to me. And music is a way to tell stories. It is the way humans have been telling stories for a really long time, or a way humans have been telling stories for a really long time. And so there are a lot of people making songs about Starbucks getting your name wrong, or a lot of music gear showing up at your house for unexpected reasons. And these are

• 6:24Speaker 01

  funny moments that you will relive. And then there's the second, I would say, more power user, who are people who have found this as an amazing creative outlet, who are enjoying the process of making music. They enjoy the final product also, they enjoy the song, but they are getting lost in the process. And they're sitting down, and they have a sound that is in their head, and they have a song that they are trying to make, and they have a story that they're trying to tell. And it may take them two hours, but they will tell that story, and they will get that out. And I think it's amazing to see the same tool being able to be used for such radically different use cases. It's quite fulfilling.

• 6:57Jacob Effron?

  And for that latter group, like, what have you learned about the way they end up using it? Because I imagine you release this out in the wild, and it's like, who knows what some of the more sophisticated users will do with it?

• 7:05Speaker 01

  A really confirmatory thing for us has been that there are a lot of people with great taste and great ideas in their head musically and without the means to do it. And that we don't have to be constrained, let's say, by the way things are done now, which is to say you're going to have some complicated piece of production software like Ableton or Pro Tools or Logic or even GarageBand. And those are all amazing tools, but they just have very steep learning curves and that there are processes that can be reimagined with certain technological breakthroughs that will let people make music in completely different ways. And that's been really great to see come through, that people will actually enjoy the process of making it and use it as a great outlet.

• 7:49Jacob Effron?

  Well, I'm super interested. I mean, I think you had like 10 million people make songs or some crazy number. I feel like a lot of AI products have this like blank canvas problem where someone shows up for the first time and they're like, what is this tool? Or like, how does this work? And if you can get past that first part, you kind of create your first, you know, song, you experience the magic of it and it kind of, it rocks and makes intuitive sense. Like, how have you thought about getting like, you know, teaching people how to use the product or getting around that, like, you know, just the initial, like, where do I begin with a prompt? I feel like so many products have this problem, just a blank screen and kind of maybe being overwhelmed of where to begin.

• 8:19Speaker 01

  100%. You know, I don't even think we've solved it, really. I think a lot of people will come and they will see a blank text box or they will see a text box with an idea in it. And that idea may or may not be applicable to them. And I think, you know, the way it is now is we can have a suggestion there. The way it will be in the future, I think, will hopefully be far more intuitive. And just as an example, we did a Valentine's Day kind of experience for people where you can make a song for your Valentine. And there, the nice part there is that the reason to make a song is sort of obvious. I'm making it for my loved one. And then we can kind of guide you through that process. And I think it's all about getting people to realize that you don't need a reason to make a song.

• 9:03Speaker 01

  Like, those reasons already exist the same way those reasons exist to send somebody a text message or to take a picture. And that there's more muscle memory in understanding that music is a first-class citizen, honestly, in the world of communication and multimedia and storytelling. So I think there will be a lot of stuff that improves, actually, in the future for that cold start problem, if you will, of the blank canvas. And one thing I will say is that it's actually funny that In the text world, LLMs are seen as a great way to get over that writer's block and that somehow that doesn't quite exist yet musically. So I don't know if the amazing thing will be hum the melody that's in your head,

• 9:46Speaker 01

  tap on the table for the beat that's in your head, what's the last picture you took in your phone, what's on your mind, but like much more intuitive and not like, what would you like to make a song about?

• 9:56Jacob Effron?

  Yeah, I think I've heard you talk about this before where obviously today you're prompting via text, but like there's so many other ways you could prompt into the model. And you know, yeah, it's always like sometimes it's seeing something beautiful that makes us think of a piece of music or it's, you know, to your point, humming something. And I, I think it's really interesting to think about the different ways, different entry points that you could have into the underlying music model.

• 10:15Speaker 01

  I think the, the fact that so many AI tools, including ours, are still fairly text-driven is just a sign of how early it still is and how much room there is to grow and impact people and figure out what are the beautiful and intuitive ways people want to interact with this. And I think, you know, it took a long time to figure out how humans would want to interact, let's say with music on their phones, you know, and so it's totally not obvious to me that we're going to have figured this out, you know, in the first year of, of doing this. But I am really excited for what happens when you can really remove that barrier for people to making songs to see what they come up with.

• 10:54Jacob Effron?

  Yeah. What are some other, I mean, the Valentine's Day one makes so much sense, you know, why not create something, a piece of music for, you know, for a specific kind of reason? Are there other things that jump, that you guys think about of like, oh, this would be other ways we could prompt those first initial uses of the platform?

• 11:10Speaker 01

  Yeah, I think, I think all things that are not expressly describing the thing you want to make a song about, because that's the really hard thing, because people don't necessarily know immediately. So it can just be describing your mood, that can still be text-based, but I think there's lots of things around how do we express ourselves through visuals, through sounds, through, you know, there's, there's a ton of people who are building beautiful music, taking sounds from, you know, just their everyday world, they're clinking glasses and they're making TikTok videos about the process of making hit songs out of those things. And I think. Everybody has that. Everybody finds sounds interesting and why can't we turn all of these sounds into beautiful music that inspires us? And so we're working on a lot

• 11:53Speaker 01

  of those things. It should be not too much longer before we have them, yeah.

• 11:57Jacob Effron?

  I love that. And then obviously, you know, you described this kind of two different types of users. You know, as you think about the more advanced folks that are spending those two hours, like, you know, sitting down and really remixed or creating a song, how do you think about like the set of features that you want to build for that set of folks? And, you know, where the kind of product goes there?

• 12:14Speaker 01

  For the power users, I think it is all about letting people have control over the music to get the sounds that are in their heads. But importantly, that is not, I think, the way it is done today, because those tools are very hard to learn. So it is not like stem out the drums and swap out the snare drum to a clap. That's not how most people are going to think about this. But instead, it is finding all of the ways to let people really pour themselves into it. And I think the real problem with the text box is not even the writer's block of like, what am I going to put in that text box? The real problem is it is a sterile way to describe what you are after. And the real way would be to pour your heart out singing, or to make a montage of a few pictures that you

• 12:59Speaker 01

  find very moving, or even to figure out how to mood board sounds from your everyday to inspire the model. And so I think if we can let people really pour themselves into it, they will enjoy the journey of making music a lot more. And I'm quite certain that when you enjoy the journey of making music, the recipient of that music enjoys it a lot more. And so that's really like a virtuous cycle for the whole experience of making music, for both sides of that conversation.

• 13:24Jacob Effron?

  I've heard you talk a lot about before, obviously, I think you've mentioned some of the most joyous experiences in your life were creating music with other people and this whole kind of shared experience of it. And obviously, I think one thing that's super powerful about Suno is the ability to share songs among like a small group of friends or something. And I think we were talking before we started recording that like our fearless producer Rashad, you know, he ships all this stuff to my house and I have all these boxes for our home podcast studio. And it was just so ridiculous, I had to create him a country song of like, why are there so many boxes that are just like on showing up to my house? And I think that is such a powerfully social experience. I'm wondering how you think about that in the context of the actual creation? Because I think today, I imagine, most of your users are kind of single player creators, like, you know, creating something and sharing it out. How do you think about like, you know, you obviously talked about playing with other people in this process of co-creating.

• 14:09Jacob Effron?

  Where does that kind of go in the Suno platform?

• 14:11Speaker 01

  Yeah, I think that is going to be a huge focus in the future is everything, for lack of a better word, everything multiplayer, everything about making music with other people. And that can be synchronously, like you're literally making music together at the same time. That can be asynchronously. So like, maybe you send somebody half of a song and they're going to finish it. Or maybe you send someone a musical idea and they pass it back to you modified in some way. And I think that, again, music is really a conversation. And so it is entirely reasonable to assume that, just like, you know, real conversations in natural language happen synchronously or asynchronously. They happen IRL, they happen over Slack, that there's going to be lots of ways that that musical conversation happens.

• 14:53Speaker 01

  I think we will eventually have things like, things that will mimic a jam session, basically, which is like, it is a synchronous conversation of I'm going to express some ideas and you're going to react to them. And then you're going to riff on them and I'm going to take it back over. And it is entirely fluid. And the same joyful experience that you, that someone can have around that can be recreated for people who are not experts at playing an instrument. And so, and it's because everybody has taste. And so you could, for example, have some interesting lyrics and have some interesting sounds. And I can take them and I can riff off of that and I can adapt it in some way. And I can extend it or I can add to it and then I can pass it back to you.

• 15:34Speaker 01

  And at the end, we may have really enjoyed one hour of working together and that was a really fun hour. And maybe we also came out of it with a five-minute song that we will almost certainly revisit many, many times. And so I think, like you said before, and I've said this, like, I think the most fun I have ever had on this planet is jamming with my friends. And I think, I see no reason that we can't give that to people. And I'll just be clear, like, I'm a mediocre musician, I am. That doesn't make me enjoy jamming with my friends any less. And so I don't think I see any kind of technical hurdles to giving people that, that kind of enjoyment.

• 16:10Jacob Effron?

  Totally. And it's a great, it's really interesting to think like what. You know, I'm sure there'll emerge a set of people that are really skilled and talented at this, and that becomes, you know, part of like these musicians of the future. And you can imagine even just watching people riff live, you know, obviously, it's it's both a social experience with your small group of friends, but you can also imagine it totally being something that people want to see happen live as well for the folks that are really talented at it.

• 16:30Speaker 01

  100%. Let me tell you, one of the coolest nights in in my life in the last couple of months was realizing that there was a Twitch streamer with a fairly big following, streaming to a football stadium sized group of people, just streaming himself for two hours doing Suno. And this was amazing. This is a this is a digital concert of sorts. But the amazing thing there is Twitch will let all of the streamers, sorry, all of the viewers, micropay to interact with the streamer. And so now we've actually gone from like a more sterile digital concert to an interactive one. And I think it was just so powerful to me to realize that like, that was bringing together a whole football stadium of people the same way that going to see a football game will bring that stadium full of people

• 17:14Speaker 01

  together. And so, and that's where the technology is now. And like, when we make that kind of a first class citizen, that kind of group experience, it will get so much more powerful.

• 17:23Jacob Effron?

  Totally. Now, you've got my mind racing about like, you know, at sports stadiums, they could totally just play Suno-generated songs about what's been happening in the in the game, rather than the the classic, you know, fire up the team ones.

• 17:33Speaker 01

  100%. And how much cooler would it be if some of the fans can have input in there, and maybe some of the players can have input in there? I think that'll be wild. And if you make a song with your favorite athlete, you will feel good, I promise.

• 17:45Jacob Effron?

  I guess moving on to some of the other decisions you made on the on the product side, my understanding today, like, obviously, everyone's kind of thinking about pricing AI products. And I think you guys currently, what you have kind of a free, you know, set of songs, and then if for your power users, you kind of charge as they generate more and more songs. How have you thought about pricing here, you know, relative to the cost you guys have? And how do you think about like, the business models, you know, and how that how that evolves?

• 18:07Speaker 01

  Yeah, I think I think I don't have a ton to say here for the following reason. And I don't think anyone really knows how this is all going to shake out. I think we are, it is so early, and the way that people are going to enjoy this is going to change. And so we're kind of like actively not trying to innovate on the business model until we really do the important product innovation.

• 18:31Speaker 01

  I'll say something maybe a little mean, which is to blame you and

• 18:34Jacob Effron?

  You can blame me.

• 18:35Speaker 01

  For this.

• 18:35Jacob Effron?

  Hopefully not me personally.

• 18:36Speaker 01

  Not you personally. Well, maybe you personally a little bit. I think it is the same people building and investing in AI companies as it was building and investing in SaaS companies a decade ago. And we and everyone else in this game has just kind of almost blindly adapted SaaS pricing to a world that it's not necessarily super relevant to or super. It's not tailor made for it. Like, it's not zero marginal cost to make a song. And so, and I think it's entirely reasonable to have done this, but it's probably not what it's going to look like in 10 years. And so, yeah, I think, and it's also not obvious to me that the way this is for music will

• 19:19Speaker 01

  be the way this is for text will be the way this is for video. I think it's going to be so product and use case dependent to figure out how people are actually going to feel like they're getting the right value out of it for their money. So I'm excited to see where it goes. It's like not kind of top of mind for us right now, at least.

• 19:37Jacob Effron?

  Another thing I'm super curious how you do is like model evaluations. Like, I imagine, you know, because there's something, obviously, I mean, I imagine there's objective things like, you know, probably I've heard you talk before about like the underlying quality of the audio and, you know, but like, obviously, music can be said as subjective thing, too. How do you do eval? And like, no, I mean, you imagine you guys are working on new models. Like, how do you know they're better? Like, what is the internal process you have?

• 20:00Speaker 01

  Yeah, it's a great question. Unfortunately, it is much harder than the other domains. And in the other domains, it's not straightforward either. So there will be automatic metrics for things like audio quality, and they're pretty flawed. And, you know, we'll measure them. We make sure they get better. But ultimately, one thing we say at the company is aesthetics matter, which is to say, squishy metrics are going to exist or intangible metrics are going to exist, and there's going to be judgment calls to be made. And it is important that we have people who truly deeply love music and who enjoy listening to it to be able to evaluate it, because very often, things do come down to subjective things. And I think this is very different from text where there are, let's say, benchmarks that all

• 20:44Speaker 01

  have to do with reasoning and how correct was your output. And music does not have a correct answer. Ultimately, the test for quality is how much do our users love the music that comes out of it. And I think that is super dependent on a bunch of stuff. That is the song quality, like were these like moving and touching songs. But that was also like, did the model produce a song that I wanted it to produce? It's like the amount of control that you have over the music turns out to be a really important metric. And unfortunately, there's like nothing good automatic there. So unfortunately, a lot of human eval. Maybe

• 21:28Speaker 01

  things will get better in the future. I'm actually not quite sure it will, actually.

• 21:31Jacob Effron?

  It's so interesting. I feel like, you know, in a completely different space, we had Beyond the CTO Sourcegraph on, and he was saying that like, look, we spent all this time trying to do evals, but the reality is the best thing we can do is just like release these newer models out and just see like our metrics. You know, we do code completion and we do, you know, other suggestions. And it's like, you can tell pretty quickly, like users will let you know, like whether they like a model or not. And you can tell by what they accept, or how much time they spend in the platform. And I think like, you know, it's interesting to see that happening across the board.

• 21:58Speaker 01

  I'll tell you, I'll tell you, you know, we have the benefit of a, of a large and engaged user base who will tell us these things, both implicitly from we ship a new model and how much does it get used and how much are people liking it? And if we give them choices between models, how much are they choosing model A over model B? But then also explicitly, we have an amazing Discord community of people who will just tell us like, yeah, I think something's wrong. And so grateful to have that community to kind of tell us stuff, or else you can be flying blind sometimes.

• 22:26Jacob Effron?

  Let's say, you know, something is wrong with the model. Like, how do you actually go about like fixing some of these like weird, you know, so someone says, oh, I generated this, you know, or this class of problems that show up in the model, you know, just not as familiar with, with like a music model. Like, how would you actually go about, you know, oh, this thing we don't love about the underlying model, we'd love to make it better in, you know, X use case or whatnot?

• 22:46Speaker 01

  It is unfortunately very case dependent. I don't think there are great rules because everything is so subjective here. And so, you know, as an example off of the top of my head, maybe we ship the model that tended not to end songs reliably enough. And so like, there would just be like way too long outros or like a bunch of silence at the end of these things. And then usually after you diagnose it, in those cases, the fix is pretty straightforward. But other times not. So I, yeah, I don't think I have anything like great to say here. Unfortunately, this is the cliche thing of like, in machine learning, the more you actually spend time looking at your data, the better you're going to understand your model. And like, that doesn't seem to be going away anytime soon.

• 23:26Jacob Effron?

  Is there anything that models can't do today that you're like, oh, I just wish we could figure out like, you know, X thing or Y thing on the, on the music side?

• 23:32Speaker 01

  Gosh, a ton. You know, I think mostly around all things iterative, you know, sometimes you get something out that was wrong, and you know exactly how it was wrong, and you don't have a way to express like, do that, but change whatever thing X. Like, I think that's a, I think that's a big one. There will be lots of other axes of control that we try to give people. But right now, as an example, doesn't really listen to you for how fast you want the music. You know, these are like very loose guidelines for the model, and I think we can do a lot better. You know, BPM is one of those things that has an answer. It's either like, you either got it, or you didn't. And you can quantify objectively, this is how far off I was. And so for things like that, I think, I think we'll see improvements

• 24:17Speaker 01

  pretty quickly.

• 24:18Jacob Effron?

  Yeah. What are the metrics that you track, like, you know, to know, hey, you know, we're moving in the right direction here? Like, what are kind of the North Star metrics for CINO?

• 24:25Speaker 01

  I think it's like, just how much people enjoy the product. And so that's like very, very high level. It's like, you know, how many users are making songs? How many users are coming back day after day to make songs? How, what is the probability that you will exhaust your free tier? Like, that would be a sign of, even if you don't want to pay, that is a sign that you just enjoyed the last several minutes of your life. And then I think increasingly, it is like, did you share it? Did you get something shared with you? Because this is kind of a social game, ultimately.

• 24:58Jacob Effron?

  As I think about the product experience, one thing that is compelling to anyone that uses it, it's just, it's so fast. Like, you really do get, it's amazing, I think it's one of these magical experiences, the first time you use it, where it's like, you know, you're with your three-year-old, you have the same body idea, and like, it's literally there in seconds. Any tips you can share for builders about like, how you've gotten this so fast?

• 25:16Speaker 01

  I think we have to get a lot faster, actually. I mean, I'll share some tips in a second, but like, you know, Spotify is an amazing product and has done us the rather unfortunate favor of when you click play on Spotify, it is really fast and they have spent a lot of time optimizing that. And so that has set a really high bar for us for just people expecting, I click go and I want to hear that song instantly and we are not nearly as fast as Spotify is. Of course, we have to make the song. So, you know, I think one is just caring about it, you know? It's like just caring to try to improve these things and just realizing that with every passing 100 milliseconds, someone is increasingly likely to click away and do something else.

• 26:00Speaker 01

  And it makes the journey of making a song so much worse if you go and you do something else. You don't get sucked into that journey. So I think to a very large degree, it's just caring about it. We use transformers, so we don't make a secret about that. These things are autoregressive, which means that people have probably figured out that we are streaming you the song while we are still making the song and that helps a lot, let's say, compared to a diffusion-based model. So yet another reason we like transformers, but I think, I really think these things do matter. At the end of the day, you are trying to deliver an experience to somebody around the joys of music and the user doesn't care what model architecture you use.

• 26:44Speaker 01

  The user cares, did you deliver me a song quickly and did it make me feel a certain way? And so just focusing on that is important.

• 26:49Jacob Effron?

  Yeah. And then I guess on like the infrastructure side, I'm curious, like you're one of the few companies that has just experienced like this insane spike of usage. And so I feel like, I'm curious, anything you just learned about the infrastructure required to support that and like where we are kind of today and the development of tools for a company like you to, you know, be able to just, you know, seemingly overnight get to 10 million folks generating songs.

• 27:11Speaker 01

  Um, yeah, we're still really early, so I think the tool chain here will definitely evolve. You know, we've been pretty choosy with exactly where we want to innovate and where we want to build stuff and where we want to buy stuff. We make, no secret, we are, we are big fans and customers of Modal who help deploy jobs onto GPU infrastructure very, very easily in a way that is just exceedingly developer-friendly. I'm not a paid chil here, I just really like the product.

• 27:39Jacob Effron?

  I am a paid chil because we're investors, but I appreciate you saying.

• 27:42Speaker 01

  No, and what that team has built is Really, an incredible piece of infrastructure. And so we're like very grateful not to have had to innovate there. You know, I think the other thing just to realize is because music and audio in general is behind images and text, a lot of problems are figured out for us by the image and text communities and open source communities. That has been really, really fortunate for us. You know, if you were on the cutting edge doing text and you needed to serve stuff really well, and then you basically needed to write your own continuous batching infrastructure, and then something like VLLM comes out and shows you how to do it, and that just happened kind of before we needed it.

• 28:27Speaker 01

  And so that's just one example of these things. I think like at a higher level, things will get easier and easier. And in some sense, it's like not, it's a rather unfortunate place to be out in front because you do have to figure out what, you know, what are the things that are going to be important. And so I think just for people building, just be very deliberate on where you want to innovate and where you want to not innovate.

• 28:49Jacob Effron?

  I love to talk about audio more broadly, too, because obviously you started off building a general audio model. I think your heart was clearly always on the music side. But, you know, obviously, it's been a big, a big few weeks in audio with GPT-4O and, you know, the ability to, I feel like now lots of folks' minds have gone racing of like different ways to incorporate audio into AI applications. So I'm curious like what opportunities you see on the application side.

• 29:11Speaker 01

  Yeah, I'm so glad that this wave is happening right now for people to realize that audio is and should be a first-class citizen in this world, and just to realize that it is how the vast majority of human communication happens. So I think we just don't realize how impactful it will be to be able to interact with systems like you do another human being. And that is The quote unquote obvious ones like customer service. And that is the the less obvious ones, like there are lots of places where we think typing is the most efficient way to interact with it, maybe like code, and it's not.

• 29:53Speaker 01

  And so I'm like exceedingly open minded what this can do. Just like, I have a horse in this race in the sense that I am one of the worst readers you will ever meet in this life. And just seeing, and so I love doing things face to face, and I really hate like email and Slack as a result. But just seeing the ability to start using the most expressive form of communication in other domains, I think for me personally is going to be like kind of amazing. So, yeah, I don't know. I'm like very open minded. I don't think I have any like great, great insight here to tell you like where you should

• 30:36Speaker 01

  be looking, where you should be investing your dollars, but it's going to be big.

• 30:39Jacob Effron?

  Well, if you didn't have Suno, is there like something you'd be building right now?

• 30:41Speaker 01

  Yeah, it might not even be in music. I don't know, like Suno, I mean, I love music, but I think, look, actually, AI is really crowded in some sense, right? And so I think it's probably something like really not sexy. I have always, I have always had this idea that there's a lot of home infrastructure that is just terrible and that smart home technology is just terrible. Being able to interact with your house over voice would just be so much more powerful. You know, I can tell Alexa to turn off the lights, but that's like, that is like very crude. And like, why can't I tell Alexa to like cook something for me, you know? So, yeah, I think one in particular thing in in home infrastructure that is just like

• 31:26Speaker 01

  a peeve of mine is like plumbing. It's like very opaque. It's very manual. There's no telemetry into your home's plumbing system. I have always thought the hardware was there, you know, 30 years ago to be able to do this stuff, and I don't really know why people aren't doing it. So forced to say I'm going to build another company, it's probably like-

• 31:45Jacob Effron?

  The classic like music startup to plumbing.

• 31:47Speaker 01

  Yeah.

• 31:47Jacob Effron?

  Home pivot.

• 31:48Speaker 01

  Smart sewer pipes.

• 31:48Jacob Effron?

  That's a story that we've seen play.

• 31:49Speaker 01

  Smart sewer pipes, that's what I want to do.

• 31:51Jacob Effron?

  I love it. I mean, obviously, you have kind of built a lot on the audio model side. Where do you think the like general audio model space goes? I mean, obviously, you've had companies like 11 Labs that have been really successful. And there's always been this question of like, will OpenAI just kind of subsume that market? Like, are there rooms for lots of audio models out there? Or like, do we kind of converge on a few?

• 32:09Speaker 01

  I don't know if I'm the right person to answer this question. I'll give you my opinion, which is that increasingly, everybody is going to try to do multimodal stuff. And I think it's going to be actually longer than people realize until audio is treated as more than just an interface into an LLM. Like, GPT-4O is still fairly, it's an extremely impressive interface to the LLM, and I love it, don't get me wrong. And same with 11 Labs, it is an extremely impressive model. But ultimately, these things are interfaces into a sense-making layer that is done entirely in text. And I think it's going to be much longer until these things are actually kind of incorporated into one another.

• 32:52Speaker 01

  And so I think that means that there's a lot longer until, if at all, until there's any consolidation into one giant model. I'd throw the question back to you, you know, as someone who looks at the whole space, like text or otherwise, like, is how much consolidation happens? And how much kind of single-purpose stuff do you think ends up happening?

• 33:11Jacob Effron?

  Yeah, I mean, it's a question we think about a lot. I mean, I think a lot of the dynamics are also determined by just like the broader tech ecosystem, where there's a lot of people that are very incentivized to have a horse in this race and not just, you know, there not to be one provider that happens to be tied up with one of the major tech companies and not the others. And so I think in some senses, you're going to see other model types willed into existence almost because you need, you know, I mean, you're seeing this in the LLM space, right? Like, Snowflake, you know, like, has announced, or there's rumors they're buying an LLM. Databricks, obviously, has been super focused on this, you know, Amazon, Google, Microsoft, they all need an offering, right? And so I think you'll see that in code, you'll see that in like even video, any category that kind of like takes off. I think there's, you know, there's a lot of folks that want something to work there.

• 33:55Jacob Effron?

  And I think that actually will be a counterbalance against us converging on like one single model type.

• 34:00Speaker 01

  That's probably good, just given, I think, how diverse the applications will be when we think about it's not all just like business-y stuff sitting in front of my computer. It's not all logic. It's not all music. It's like, it's lots of stuff. There's going to be robotics. There's going to be healthcare. There's going to be like weird, random things like, you know, what does the LLM for chip design look like? Like, how do we put more AI into that? You know, so I think, oh, gosh, that makes me, that makes me really excited.

• 34:28Jacob Effron?

  Obviously, we got you, I guess, a day or two after you announced this, this big fundraiser. And so, one, congratulations.

• 34:34Speaker 01

  Thank you.

• 34:35Jacob Effron?

  And two, I'm curious. So you raised $125 million. Usually, I think of raises of that size, you know, being foundation models that are super compute intensive. Like, is that mostly to train models? Or like, how do you think about what you go do with this, this now large, large sum of cash you got?

• 34:50Speaker 01

  Yeah, it's, it's kind of scale everywhere. So I think it is certainly to train big models. I think music models will not require quite the same level of compute as the biggest text models do in the near future. But that doesn't mean that you don't need an incredible amount of care to do these things. And I, you know, something that's easy to forget is that music is very far behind. We don't even know the right way to model it. Like, in text, more or less, everybody is doing very small variants on the same thing, and it's really impressive and it's working. Don't get me wrong. Like, I'm very glad everybody's doing that. The right way to model music isn't even worked out.

• 35:27Speaker 01

  And then I think increasingly, I'll say that models will get bigger. Maybe they don't, they don't get quite as big because it's not all just a question of scale and get the right answer. There's like taste involved. There's control involved. I do think, though, that it gets increasingly expensive to train these things, not just for the flops, but for the research that has to go into it, for the very specialized data that you need to put into these things, for the axes of control that you need to put into these things. And so that's definitely going to be a big part of it. Obviously, running a company, you need to hire the best people. So that requires a bunch of stuff. So, yeah, I think I am really excited about.

• 36:10Speaker 01

  We try to think very hard of like, what should the future of music look like? And basically, the way we think about deploying the capital is not exactly train this model, it's not exactly hire this team, but it's like, how do I more quickly pull forward the future of music that we envision and deploy capital toward that? So there's going to be a lot of stuff coming.

• 36:26Jacob Effron?

  I love that. What are like the music model capabilities that would be meaningful to you? Like, I don't know if there's like, you know, a few milestones that you think about like, oh, it'd be amazing one day if like, you know, I, yeah. And like, I think about, you know, we try on the image side, we try the same, you know, when we were evaluating these companies, we'd always try the same prompt. We'd be like, it'd be amazing if one day this prompt was just like nailed perfectly. I don't know if you have any equivalent things where you think about on the music side. You know, you've alluded to some of it earlier, but I'm curious, like, are there some capabilities that would kind of just jaw drop you if you got to the point where they started to work?

• 36:57Speaker 01

  There are definitely some. I don't, I don't want to give, I don't want to give too much away. I think like, at a very high level, I think people care a lot about, like, when is someone going to have a model that makes a three and a half minute pop song that is like indistinguishable from what a human being would make in a recording studio? And I understand why that seems to be the benchmark. To me, that actually is not that interesting of a, of a milestone because just because this is not a fact-based thing with objective answers, actually, the ceiling is quite a bit higher than that, right? Music is all about how it makes you feel. And so I don't even think we know how to measure it, but like, three and a half minute

• 37:42Speaker 01

  pop song, indistinguishable from what a human would do in a recording studio, just to me feels like a somewhat artificially constrained.

• 37:49Jacob Effron?

  Yeah, it'd be kind of disappointing if that's all we got.

• 37:51Speaker 01

  Yeah, exactly. And so, um, yeah, there's, there's going to be a lot of stuff coming in that direction. What, what do I want, like personally? This is just a me thing, not even a product direction thing, because we're not doing this right now. I want a Vision Pro app because I want to play air guitar with a band. And there's a tremendous amount of technology that would have to get built, like doing stuff in real time enough synchronously to make that a reality would be very, very difficult. But ultimately, I think that is the goal. Like, this is an enjoyable game, and I don't mean, I say game not to cheapen it. I say game because games are things that humans have been doing for millennia. Games are really, really enjoyable. The video game industry is huge.

• 38:31Speaker 01

  And I want people to be able to interact with this in the most intuitive way possible. And so I think that's really like-

• 38:39Jacob Effron?

  That's so cool. The music would just respond, like, you know, you're strumming slowly and the music's kind of responding to that.

• 38:43Speaker 01

  Totally. Maybe, maybe actually it's not even air guitar. Maybe I'm a conductor and I have a whole symphony, you know? Like, how amazing would that be? I would play that for hours. We can't do that yet.

• 38:52Jacob Effron?

  I was like, let us know when the Vision Pro app-

• 38:54Speaker 01

  Yeah, exactly.

• 38:54Jacob Effron?

  I'm excited for that.

• 38:55Speaker 01

  You'll be the first, you'll be the first to know that-

• 38:57Jacob Effron?

  Yeah, let me know when I can conduct. That seems, that seems super fun. I'm curious, like, you know, how you contextualize what's happening in kind of the broader AI music market. Like, obviously, you know, you've had a few competitors pop up. Like, do you feel like this is a space where, you know, there'll be different kinds of businesses? Or like, you know, if you were to fast forward 10, 20 years, is there kind of one big AI music player or is there kind of, you know, different niches here?

• 39:20Speaker 01

  I think this market is ultimately really, really big. I think a whole bunch of technologies, not only AI, but obviously AI has the power to greatly expand the market for music, which is smaller than it should be right now. I would say, I'm like, the same way I'm like really excited OpenAI is doing more and more voice because it's showing people that this is a real thing. I'm like very excited that there are people coming into the space because this is a real thing. This is going to grow. I think it's so green field right now that I can't imagine that, there's so much to do that I can't imagine that there's only one company. I'm like very excited that people are starting to do what we do, don't, you know?

• 40:02Speaker 01

  But like, people will build amazing tools meant for professional artists. It's just not something we're focusing on.

• 40:08Jacob Effron?

  Yeah.

• 40:09Speaker 01

  And those will be great businesses and those will definitely increase the quality of music that gets made and the amount of music that gets made. It's just not what we focus on. People will go after background music. They already are. Not what we focus on, but that is a huge business if you just think about like how many YouTube videos have background music in them, right? So I think there's just going to be a lot of stuff here. The thing that I am personally the most excited about, though, is the experiences that we can build for the average person. Like when we think about how do we, what is the future of music? How do we expand? How important music is in people's lives? How do we expand the joy that people have around music? It becomes very obvious framed like that. A, we want to build for the general consumer. I want like my mom doing this.

• 40:53Speaker 01

  My mom is not a musician. And B, we should just directly build them the experiences that we think they're going to really enjoy.

• 41:00Jacob Effron?

  What's going on in like the IP partnership space? Like I feel like in the image space, a lot of folks have kind of tied up, you know, data sets with, you know, to get kind of copyright approvals. I haven't seen as much happening in like the music world. Like I don't know if maybe I missed the Taylor Swift announcement, but I don't think there's been like massive partnerships yet on that side. Like, how do you, you know, I feel like everyone always talks about the Spotify versus Napster approach in the, in the creative spaces. Like, where kind of are we in that today?

• 41:25Speaker 01

  We are so early. And I think Spotify and Napster is actually a pretty reasonable analogy. And I think there will be Spotify's and there will be Napsters and there will be people who are excited to work with the industry and grow that. And then there will be people who are excited to work against the industry. And I think that will be uphill. We are obviously very excited to work with the industry. Unlike specific artist partnerships, I think this is a weird one. And it's something that we have, we work very closely with some artists, but when people hear artist partnership, they think like, can I just make new Charlie Puth songs or can, and?

• 42:06Speaker 01

  This is just an opinion. It is something that we have always stayed away from because doing it without the artist's express consent is obviously illegal. But also, again, just an opinion, I think these are very viral moments, and I think these are not as big a part of the future of music as people think. And to me, the analogy is like, the first time you played with GPT, you probably remember it. You probably made Shakespeare sonnets about podcasting, or Shakespeare sonnets about VC investing, and then the third-

• 42:36Jacob Effron?

  Those are my two interests, so that had to be one of those two.

• 42:39Speaker 01

  Yeah, exactly. And then like, the third time you did that, you were like, okay, this is a little silly. That doesn't mean GPT isn't wildly helpful. You probably use it every day, right? It's just not for that. And to me, this feels like the same analogy. I think the future of letting people enjoy the process of creating music is not necessarily letting them enjoy making a new song by Charlie Puth, or not necessarily having them have Taylor Swift sing "Enter Sandman." But instead, it is allowing them to go on this journey about talking to things that, about the things that are relevant to them. And so, we have always just been like, basically ignoring that kind of stuff, and letting other people

• 43:20Speaker 01

  do it. And it's a flash in the pan, as far as I'm concerned. It is very viral, and then ultimately kind of fades into the background.

• 43:27Jacob Effron?

  Yeah. I'm struck at super interesting, like, you may end up actually end up in a world where those types of platforms onboard people to the possibilities, like to your point, like the Shakespeare sonnets of ChatGPT, where like, many of the first times people used it, and then once they like, just get exposed to that tech to begin with, they'll kind of realize some of the, you know, they'll stay for some of your use cases.

• 43:45Speaker 01

  I think that's right, but I think it would be much more straightforward to just show them the first time the amazing use case. So instead of enjoying making the Shakespeare sonnet, and then going to do the, I don't know what your favorite use case for GPT is, like, wouldn't it be swell if you were just like, oh, hey, you have that like, messy email where someone pasted a bunch of data, and you wanted to format it as JSON, and then like, you could just paste that in and get the JSON out immediately, and then like, just give people that musical experience. So that's- It's like, yeah, you heard some interesting sounds while you were making coffee this morning. Like, put them here. Let's make that, let's build a song together and you will have the best 10 minutes of your life.

• 44:23Speaker 01

  And that seems like a far more straightforward entree into the art of making music than making some weird cover song. Just, just an opinion, you know, people can disagree.

• 44:33Jacob Effron?

  I love that. Well, we always like to end our interviews with kind of a quick fire round where we get your, your quick take on overly broad questions that should probably be their own independent.

• 44:42Speaker 01

  Okay, let's do it.

• 44:42Jacob Effron?

  We asked the same, same thing.

• 44:43Speaker 01

  Wait, before you start, you know, I'll tell you, because you mentioned Lang Chan and Harrison Chase before, so I'll, I'll take the opportunity to plug him. He's amazing. When we used to work together, he would have.

• 44:54Jacob Effron?

  Yeah, what, okay, what was in the water back at Kentsch? Like, maybe I'll have to ask you that. Like, why are there so many amazing AI founders?

• 45:03Speaker 01

  Kentsch is a, the, the talent density there was just amazing. And the, the founder, who is a close friend, just, I guess, I think maybe I've modeled myself on this, just found so much personal enjoyment working with, with great people. You know, I, I think at Kentsch, I, I like, was so happy to be the dumbest guy in the room and, and I continue to do that at Suno. So like, you know, it's, it's a, it's, if you have no ego, it is really fun being around people who are just really, really capable. And Harrison is one of them. He, he had this thing where he would just be like, he, he called them hot takes. It's like, I'm just gonna tell you something very fast. It's like only 80% true, but you have to let the, that last 20% go.

• 45:44Speaker 01

  Because there's like that nugget of truth is like really, really good. And so, uh, let's see how many hot takes we can, we can get there.

• 45:51Jacob Effron?

  Amazing. All right. Well, then I'm, I'm excited with that, with that preamble. I guess the first is just like one thing that's overhyped and one thing that's underhyped in the AI world today.

• 45:59Speaker 01

  Oh, um, okay, I'll give you a hot take. I think open source is overhyped.

• 46:05Jacob Effron?

  Interesting. Why?

• 46:05Speaker 01

  Yeah, I think, um, just, open source software is amazing. Open source AI is really tough just because of the barriers to doing it right now. And so it's not obvious to me that you don't want much more explicit financial incentives to make the best stuff.

• 46:24Speaker 01

  And I'll caveat, like, that is why you see the best open source stuff coming out of Meta, right, who has the financial resources to do what they need to do.

• 46:31Jacob Effron?

  Right. You just think that as compute costs increase for each subsequent generation of these models, it's going to be super hard for open source to catch up without a business model?

• 46:38Speaker 01

  Certainly in the short term.

• 46:39Jacob Effron?

  Yeah. And what about underhyped? I don't know if there's anything that's underhyped in AI today. It all feels pretty...

• 46:43Speaker 01

  There's a lot. Can I say music? You're not going to let me get away with that one.

• 46:46Jacob Effron?

  I mean, I'd hope you believe that, so.

• 46:48Speaker 01

  I think music is underhyped. I think music is just underhyped. Forget even in AI, just as a part of people's lives. I'm going to go with that.

• 46:55Jacob Effron?

  I love that. What's been like the biggest surprise in building AI features for Suno? So maybe like one thing you thought would work that just didn't, or one thing that you're like, I don't know what we're doing here, and it actually ended up working out super well?

• 47:05Speaker 01

  I think letting people feel the downstream impacts of their music has worked out really, really well. And so that is letting people feel proud of what they made, and that is proud on a micro level of making it really easy to just text somebody a song, and proud at a macro level of somebody getting a good enough song that they hit our showcase. I'll tell you just like a funny anecdote which surprised us. We, you can edit your song title, like if you, if you mess it up by accident, you type in the wrong thing, you can go back and edit it. And what we found is that when people would hit our trending page, they would go and edit their song titles to put their names in the song titles because it feels good to feel proud of it. And I think that's something to lean into, not

• 47:49Speaker 01

  to shy away from, like, people should feel proud of what they do. So I think that, that worked out, that worked out better than expected.

• 47:57Speaker 01

  Something that didn't work out great, at the beginning, we owned some of our own hardware. Actually, the first thing we did when, when we left Kentsho was buy a couple of GPU boxes that are now just collecting dust. So, so didn't really realize the scale of what would be required, and you can't run that at home.

• 48:12Jacob Effron?

  What's something like you've changed your mind on since, since building Suno? Maybe something, a hypothesis you had about the business in the beginning that you're like, yeah, that was wrong.

• 48:20Speaker 01

  I think one of the biggest things I got wrong was Discord. So I thought we would be on Discord for much longer than we did. And don't get me wrong, I love Discord and we have an amazing community there. But I looked at something like Midjourney that is only on Discord and getting just a tremendous amount of usage and thought that we would be there much, much longer. And then I'll give you like one interesting number is last November, we shipped a very thin, underpowered web app that didn't even have all of the functionality of the Discord. And it took five days before 90% of the usage was on the web.

• 48:59Speaker 01

  And it is, I think for building experiences, we are not, we are not building a music model. We are building experiences and the experience has to be all, the entire experience needs to be pleasant. And so Discord is just, it's a messaging platform and it's amazing, but it is not the right place for an all encompassing music experiment. And so I got that one totally wrong.

• 49:24Jacob Effron?

  Yeah. What if, I mean, I feel like there's so many AI consumer apps that are navigating Discords and community management, like anything you've learned around that where, you know, how you guys kind of engage with the, with the broader community that you think you've gotten better at?

• 49:37Speaker 01

  I'm so fortunate we have this community. It's like that we've gotten better at. It is, I'm just, yeah, it's, in some sense, it's simultaneously a lot easier and a lot harder to listen to 300,000 people. There's a lot of signal that's there. It is maybe a little bit harder to cut through some of the noise at 300,000 people. But I think we, the product would not be where it is today without that community and the leaders and moderators of that community. Like it is just such an underappreciated resource. And so maybe I'll go back to the last thing and I said, like, I thought we would be on Discord forever. If we had never launched our product on Discord, our Discord would be actually a lot smaller. And then we would not have that incredible resource. And so I'm not sure I would change very much from how we did it. It was just a lot of learning.

• 50:21Jacob Effron?

  Yeah. Which AI startup outside of music and audio are you most excited about?

• 50:25Speaker 01

  Oh gosh, uh, it's got, I want some physical hardware. I, I like, one day I will have-

• 50:31Jacob Effron?

  Is that the physics PhD in you?

• 50:32Speaker 01

  It might be. One, it's missing soldering stuff, I suppose. I want some piece of physical hardware. I don't think I have a company in mind per se. I think, I mean, I mentioned Modal just because as a, as someone who builds on top of it, I think that developer experience is like amazing. Yeah, I don't know. I don't need to. I don't. I've already plugged some friends' companies. I'll shut up now.

• 50:55Jacob Effron?

  Okay, well, I definitely, as the second to last thing, I think I'm going to pioneer this, which is, I feel like we're obviously going to have to use and very joyfully use Suno-generated music for the intro and outro of this episode.

• 51:05Speaker 01

  Excellent.

• 51:05Jacob Effron?

  I'd love for you to create the intro song now.

• 51:08Speaker 01

  Okay, let's do it.

• 51:09Jacob Effron?

  I'd love to see you do it. What should we, what should we, what should we kick this episode off with?

• 51:12Speaker 01

  I don't know. It's got to be about unsupervised learning, right?

• 51:15Jacob Effron?

  That could be about, it could really be about, you know, about anything here, but I'd love to see the guru of Suno, you know?

• 51:21Speaker 01

  I'm, you know, I'm not great at it. Another thing that the community is amazing for is teaching us how to use these models, you know? It's like, it's like crowdsourcing prompt engineering, basically. So, I don't know, let's say a bouncy trap song about unsupervised learning. Let's see what comes out there.

• 51:41Speaker 02

  Unsupervised learning, no inhibitions. I'm a machine, no doubt I'm on a mission. Shredding data like it's my composition. Unleashed on the beat, I'm a lyrical beast.

• 51:52Speaker 01

  Oh, I see.

• 51:53Speaker 02

  Get ready for the feast. Unsupervised, I'm breaking all the rules, finding hidden gems.

• 51:59Jacob Effron?

  Well, I feel like folks are going to want to pull on a bunch of threads we talked about here. I know they're going to want to go find that song we just generated, too. And so what is the best way for folks to learn more about, you know, the work you're doing at Suno?

• 52:10Speaker 01

  Yeah, definitely come, come use the product, suno.com. Very, very easy to just start making songs. Takes, takes a second. We're hiring, primarily on the East Coast, but also some remote. So, New York, Cambridge, Massachusetts, and everywhere in the world.

• 52:27Speaker 01

  And then, yeah, you know, we're always interested to hear, like I said, it's just so early in this game, we're always interested to hear people's ideas of what the future of music is. And that is total novices, that is experts, that is everyone in between. I think at this point in the arc of things, there's just so much to do, and there's no way we think about everything. And so, like, we're exceedingly open-minded to what, what can be possible.

• 52:54Jacob Effron?

  Totally. Well, I have to say, as a Boston native, I think it is so cool that, like, one of the most innovative companies in AI is being built in, out of Cambridge.

• 53:01Speaker 01

  Thank you. I don't know why it's so, why it's so surprising to people, but it is. I think, you know, there's obviously, one of the world's best music schools is in Boston. Some of the world's best universities are in Boston. So, to me, it's, it's, it's not shocking, but I think...

• 53:15Speaker 03

  Totally, I will take the love. Let's see if we can get you to move back to Boston.

• 53:19Jacob Effron?

  Oh, hey, may it be the first of many. But Mikey, thank you so much. This has been an amazing conversation.

• 53:23Speaker 03

  So great to be here. Thanks for having me.

• 53:25Jacob Effron?

  Rashad, I was so happy to have the creator of the product of, you know, that song that I basically had to send you after you tried to set up a podcast. Essentially full studio.

• 53:37Speaker 03

  Should we give people the full context on what happened?

• 53:39Jacob Effron?

  We're gonna have to put this song in the debrief.

• 53:41Speaker 03

  I want to play it right now here. I sent it to my girlfriend.

• 53:45Jacob Effron?

  You sent, what, it must have been like a dozen boxes to my house and they all just showed up one day?

• 53:49Speaker 03

  No, no, no. Let's backtrack. You said you want a good studio setup in your apartment so that we could record virtual episodes of this podcast and Vital Signs. And I said, okay, but like podcast equipment is very bulky. Are you cool with me sending a ton of shit to your place? And then I sent you a bunch of Amazon boxes and then your wife got mad at you and then we started.

• 54:09Jacob Effron?

  Are you saying it didn't deserve the full devastating country song about how hard it was to have so many packages just out and around my house?

• 54:17Speaker 03

  I've actually listened to that song for like a week.

• 54:25Jacob Effron?

  It's just so good.

• 54:27Speaker 03

  I think country is the move for Suno.

• 54:28Jacob Effron?

  Country is usually where I go with, uh, with Suno. Hip hop's not bad either, but my country is usually my go-to.

• 54:38Speaker 03

  We will link it with all the other references. The same bonus song and everything. Uh, he was awesome. I like that guy.

• 54:44Jacob Effron?

  Uh, Mikey's awesome.

• 54:45Speaker 03

  I think I caught him say like, this is the most exciting thing, or I'm so excited about like, whatever. He's just such an optimist.

• 54:52Jacob Effron?

  Yeah. Also now he's a physics PhD. He's like a Renaissance man. He's, he's obviously super humble, but incredibly smart.

• 54:59Speaker 03

  Yeah. He also, uh, shelled Modal for us, which I thought was pretty cool.

• 55:03Jacob Effron?

  Yeah. I mean, did I know that Suno was a big Modal customer? Yes. Maybe a little bit of leading the witness, but.

• 55:10Speaker 03

  He went ham on investors a little bit blaming investors for, what was he trying to say about like why VCs are messing with the like pricing models for these tools?

• 55:18Jacob Effron?

  I think that like he's basically like, look, you know, all these VCs in the last decade have learned, you know, the SaaS pricing model. That's what they expect to see in the description and all the metrics of VC have been optimized around that. And like, it might not make sense in the context of AI.

• 55:33Speaker 03

  And so what is, what is a founder to do in that situation? Like they're getting pressure from the board because the board's like, we need to be making revenue, but at the same time.

• 55:40Jacob Effron?

  I don't even know it's about pressure. I think it's more just about a pricing model. And I think his point is, you know, in the SaaS world, you basically don't have much marginal costs. Like, once you give someone a subscription and, you know, there's not, like them using the tool doesn't add a lot to your underlying cost.

• 55:55Speaker 03

  I see, I see.

• 55:56Jacob Effron?

  But the difference is, you know, if you give someone a Suno license and they run it 100,000 times, like you pick up a pretty hefty bill.

• 56:01Speaker 03

  Right, right, right.

• 56:02Jacob Effron?

  And on the compute side. And so I think that, you know, I appreciate his honesty. Like, everyone I think is trying to figure out like, what is the right pricing model, given the amount of times you use these models are exactly, you know, correlated to your underlying cost.

• 56:15Speaker 03

  Yeah, 100%. So he said they were intentionally not innovating on pricing, which I feel like we've heard from a couple of our guests as well.

• 56:22Jacob Effron?

  I think like it's so clear from hearing him talk that he feels like they're in like the top of the first inning of like this entire thing where, you know, the behavior they want is kind of you hear the sound and you spontaneously create something and you're, you're jamming with friends and they're obviously seeing glimmers of it. Like I loved how excited he got about that Twitch stream and seeing someone, you know, creating live and people watching. There's early signs that there's, there's fit there. But I feel like he thinks this is so far from the end state of where it goes. Like, why kind of figure out the, the pricing, you know, now? And clearly, I think that, you know, the success of the company and the funding they've raised, there's lots of people that feel similarly.

• 56:55Speaker 03

  What did you think about the Twitch streamer thing? Did you know what he was talking about?

• 56:59Jacob Effron?

  I didn't personally watch that, but I love that. I, you know, your mind gets racing when you hear him talk about like, you know, this future of co-creation and this idea that like, you know, the same way we go to hear jazz musicians riff off each other, like people that are really good at kind of creating these songs from scratch or riffing. You can even imagine comedians incorporating this into like their routines. Like, it's, it's cool. You know, I think to his point, like maybe music just becomes more of a fundamental, an easy thing to, you know, to incorporate into lots of.

• 57:25Speaker 03

  He made a couple of references to Charlie Booth. Have you seen the video he's talking about where he like, Jimmy Fallon, I think is interviewing him and he brings out like the, the music producer. Desk thing and then he records himself like hitting a mug with a spoon and it makes a little like ding sound and then he like pitches that down and then he makes like a whole beat on it and it the video is like super, super viral. So I mean, there's definitely something there with like the live creation process of

• 57:53Jacob Effron?

  Totally and just like how much fun you can tell he so visceral lives his own product. Like he loves, I think that that whole, I think we've all been a part of, you know, those of us who I grew up playing music and it's totally true. Like there's nothing more fun than kind of being in one of those jam sessions with your friends regardless of where it goes.

• 58:08Speaker 03

  Yeah. Were you good? Are you good?

• 58:09Jacob Effron?

  Uh, I wouldn't say I'm good. No. Um, I think, uh, Mikey was saying on the way out that he has like a disappointed piano teacher and I think my piano teacher was disappointed.

• 58:18Speaker 03

  You asked him about metrics, like what metrics is he tracking? And he was basically like, how much do people love the product? Uh, it was basically

• 58:25Jacob Effron?

  But it's interesting, right? Cause there's so many different things you could optimize for. He talked, uh, if I remember correctly, there's obviously a number of creations and there's, you know, do you come back and do you kind of blow through your free credits? But optimizing for that can also be pretty different than optimizing for sharing and like, is this kind of a social experience?

• 58:39Speaker 03

  Yes.

• 58:40Jacob Effron?

  And I think, uh,

• 58:41Speaker 03

  Like the Wordle effect.

• 58:43Jacob Effron?

  Yeah. I haven't heard a coin like that, but I like that.

• 58:46Speaker 03

  Yeah.

• 58:47Jacob Effron?

  Yeah. I think there's like this interesting, um, I feel like it's a very good single player product right now. And I love, I mean, clearly you can tell like top of mind for him is how do you make this like this collaborative, you know, thing you want to share with your friends type experience?

• 58:58Speaker 03

  Yeah. Multiplayers.

• 59:00Jacob Effron?

  Because honestly, the most fun, I, you know, I create a fair amount of student music, but you and I just laughing over our song is like, you know, it's way more fun than just like the, the random thing I sit on top of.

• 59:08Speaker 03

  Yeah. A hundred percent. Uh, he had a hot, a hot take about open source that it's uh, overrated.

• 59:16Jacob Effron?

  You know, I, I think, uh, what was behind that point, which, which makes sense is that basically, you know, uh, I think he thinks the state of the art models are going to continue to be super compute intensive and cost a lot. And, you know, unless you have a business that can support, you know, putting that amount of compute, it's not totally clear yet what the business model of a lot of these open source companies are. And so, um, you know, that's obviously an intention with the fact that it feels like if you just stay six months behind state of the art, the cost of training these models is actually quite low and, and not super expensive. And so, um, you know, I think, you know, since we've started this podcast, I feel like we always ask, you know, closed versus open. And I know the reality we all feel is like they're both going to be in markets. Like, uh, you know, I'm not sure one is, is just going to eat the other.

• 59:54Speaker 03

  People always say, I feel like the takeaway I always hear is that like GPT is gonna is is ahead. So it's going to stay ahead in that and in that way, it's going to sort of like own most of the market. Do you subscribe to that? Like, is that how you look at it?

• 1:00:08Jacob Effron?

  Well, I think that like people are constantly finding, we're still like the early days of just finding use cases for like these LLMs. And so I think that basically, you know, as we discover use cases, inevitably, the open source models get good enough to whatever we discovered like a year ago with like GPT-3, or even, you know, now we're discovering with GPT-4. In six, 12 months, the open source models will be good enough to do whatever those things are, whether it's writing, you know, summarizing your emails or generating marketing copy or some of these early use cases that we've seen today. Of course, a year from now, GPT-5, whatever the latest state of the art models, there'll be a whole new set of use cases they can do. And so I think the case for these closed source models is just the most economically valuable stuff will always be at the frontier and what the state of the art models can do. And but all the while, while those things are being discovered,

• 1:00:54Jacob Effron?

  you know, a year later, they're, you know, potentially possible in the in the open source world. So they both end up being pretty big markets.

• 1:00:59Speaker 03

  Yeah.

• 1:01:00Jacob Effron?

  That was a fun one. I've been looking forward to getting Mikey on for a while. He's been the inspiration of a lot of, a lot of our, I mean, I think our title track here is generated on Suno. And we obviously introduced our LPs to some of the Suno music.

• 1:01:14Speaker 03

  Yeah. I think this is our first live demo also.

• 1:01:17Jacob Effron?

  Yeah, it's probably my favorite, my favorite AI product that I use day to day. It's most fun, that's for sure. That was unsupervised learning. A huge thanks to Mikey for just a fantastic episode. Definitely stay tuned next week. We'll have another great entrepreneur or leader in the AI space. And if you're really enjoying this, please share with a friend. We're trying to get the word out, continue to grow the great listener base we have. So appreciate you tuning in and we'll look forward to seeing you next week.