Technology, Connected - Can AI Make Music Without Stealing?

Episode Date: February 5, 2026

What if you could use AI to make your own music without stealing other people's beats, rhymes and melodies? Unlike platforms trained on scraped catalogs, Overtune’s AI is built on licensed music, starting with ~20,000 loops produced in-house. Producers can submit stems voluntarily, creating a clean foundation for ethical training and attribution.The platform uses vector-based audio embeddings to measure how much each stem contributes to a generated track. This enables automated attribution and proportional royalty distribution when songs are commercialized. Contributions are weighted mathematically, with clear thresholds to credit primary and secondary influences while avoiding excessive fragmentation Please enjoy the show.Cheers,Mark and JeremyPS: Subscribe so other curious minds like you can find our channel.Other ways to connect with us:⁠Listen to every podcast⁠Follow us on ⁠Instagram⁠Follow us on ⁠X⁠Follow Mark on ⁠LinkedIn⁠Follow Jeremy on ⁠LinkedIn⁠Read our ⁠Substack⁠Email: [e-mail]

I understand that your product, like your core product was actually trained on licensed music to start off with. Is that accurate? That's accurate, yes. Can you talk us through it? Yeah. So we took a hard stance against Suno and Udio, especially as, you know, we were learning what they were doing. And first of all, we don't even agree with their product on a fundamental level, okay? I don't believe music is a one-shot creation process with a prompt.

I downloaded it over last Christmas. It was fun with my cousins and my sister where, you know, hey, create a song about our grandfather who escaped Italy, came to Montreal, married somebody, whatever. I'm like, yeah, we're laughing, but like that the fun is over after a month. Like it gets, I thought it got boring fairly quickly. And it's because of that lack of human in the loop, right? Okay. So our model, our training model. Yes, we have around 20,000 loops all created by our in-house producers.

but we figure out a way that, you know, we think we want to try to disrupt the industry in another way where any producer who creates beats can submit their stems or their loop pack to our model. And it goes back to the mathematics of how these allems work. And I'm going to get a little bit technical here, but it's because I want people to understand, like, you know, this stuff is not impossible, that we have the mechanics to do it.

Language can be converted into math, which we call vectors using deep embedding systems. right? And then, you know, once there's an output, you can match the output to the distance from the original vectors to figure out which vectors had the biggest impact on the output. The amazing thing is that with audio generating systems, the process is kind of very similar, where all of the musical stems get converted into what we call audio latent vectors. Now, when you ask it for with a prompt and it converts everything into, you know, a waveform, ultimately, and I can go into the details of that if you're interested,

but I'll avoid it for now. But ultimately, what can happen is that we can measure the distance between all of the origin vectors to the new song creation, figure out which artists contributed the most to it. How accurate is that? I would say there's no more accurate system in the entire world. We're talking about vector mathematics.

We're able to ultra-precisely to infinite flotation points know how close the origin is to, a new generation. So I make the new arm end break. I don't have to worry about my work being ripped stolen or used without my permission. And I can collect all the vulnerabilities from those number one hits which come. And that's essentially what you're doing for the. That's exactly great.

So if a generation is made up of primarily four artists, then if that song goes commercial, that slice of that stem in terms of the royalties, it's going to get divided by those four artists, for that stem across all the stems that are used for the song in terms of like their time length and all that. So everything is measured again mathematically to ensure that proportionally everything is fairly distributed. But the answer is yes, everybody would get their fair, their fair remuneration for their work. Yes. All right. So we're famous for silly thought experiments. So let's try to break this down a little bit. So the three of us are on your platform. I've got a baseline that I put out there. You've got a drum line that you put out there. And Mark has this nice

little vocal stem, a cheeky little vocal stem. How about that? All right, so these things are kind of out there for people to use and come together. All right. So say there's a fourth person that is out there kind of looking, all right, I want to create this bead. I'm not necessarily super musician, but I like putting things together and messing around. So they grab my baseline, they grab your drumbeat, they grab Mark's vocal stem, and they add a few other things to it, and they put out a song. This thing goes to distribution and it goes to Spotify and gets a million plays on Spotify. What does it look like for all four of us, just from a high level. To be frank with you, it's something that we're still toying with.

So even like I have numbers now, it's not going to be probably what it ends up being at the end, right? But for this thought experiment, we're just going to go with easy numbers. But ultimately, for us three, it's a volume play. We are imagining, you know, there's almost 8 billion people on the planet. This platform allows, you know, four billion of them to experiment with music and try stuff. Now suddenly, if we have, you know, of 4 billion people using a music platform, and again, super high numbers, not expecting that, but you get the picture. Just a million people using our music for their stuff.

Suddenly we have $10 per, you know, commercially available song. And suddenly it just starts trickling and then suddenly it's like a giant wave depending on how good or how trending or whatever fact is. The more people are using your stamps, the more it's going to make you some money. But ultimately, that's what it's going to come down to, right? So it's realistically a volume play. I'm more interested in, like, how all these vectors are tracked. Like, you explained it very succinctly, I think better than anyone else has explained, hey, step one, step two, step three.

Here's how it works. But, like, what does the final automation of attribution look like in this world? So the prompt itself is going to dictate from which vectors, it draws on. And typically, it's going to find the middle of several that are similar. So in that respect, it's not going to just be my good, like in a world where, you know, this AI model just has our tracks, yes, but in the real world, it's actually closer to like, you know, dozens of, of contributing factors, right? But I also want you to imagine all of the infinite points in between all those vectors and the infinite arrangements of sounds you can get. So,

in one aspect, no one is going to get the same sound, which is kind of magical about it, right? But when we're talking about, you know, the simplicity and distilling it to its vectors, it's because we're imagining sometimes a vector in 2D space, but the reality is that these vectors exist in like almost 8,000 dimensions, right? So it's tracked across every type of property

you can possibly imagine. So that being said, every generation is going to be different. the proportionality of it is going to change. There's always going to be one vector that's closest because you're going to say like, hey, give me melancholic Afro beats with whatever it is. And there's going to be...

With cheeky British vocals. With cheeky British vocals. Yeah, yeah, yeah, exactly. And it's going to land closer to someone than someone else. And ultimately, so that's going to be, you know, the primary. And then we sort of have to like cut it down where it's like, hey, well, everybody between, you know, 70 and 90% those are the secondaries in from 10 to 100.

Like it's like gravity, right? Where Pluto exerts a gravitational force on the earth, but it's so small you wouldn't really count it, right? So like the last 10% are influencing it a little bit, but not so much to the point where it makes a big impact. It's like in our model right now, those are cut off from the commercial attribution. So it's really about crediting the primaries and secondaries first

and then because, you know, even like a soulful, the jazz drumbeat is going to have an impact on somebody asking for like a hardcore gangster rap song, right? So it's like, where do you make that cutoff? What's the thresholds? How do you make these decisions? And these are questions we tackle every day. It sounds wonderful. And but I also, an excuse if I'm not even standing this right, this is exclusively to overtune. How is it possible to backdate this vector technology? Is it possible? Is it possible? to solve the current problem of new music being ripped and stolen to create these new

and AI generated songs. We've seen some in the country world, haven't we? Was it breaking rust, velvet sundown? Do you see a way to solve for what's already happened? I think that's more of a culture question in the United States than it is anything else. This culture in the states where, you know, let's break it first and then ask for permission later. Uber did it. Airbnb did it. And, you know, these companies eventually go on to become like these mega unicorns, public companies and all that, right?

So I think, like, I'm trying to figure out what to say because I don't want to throw any companies under the bus in a way that can put me in a weird position. Does the technology exist? Will it ever exist? Oh, do you solve for that problem? No. Honestly, what I think has to happen is that they're going to have to rebuild their model from the ground up using new stems and craft everything at the stem level instead of like this one shot large generative model where you crunch all this data in one place and then, you know, hope for the best. My answer to that is no. I don't think there is an easy or elegant way to do it. I really think it requires a complete reconstruction of the technology to consider attribution to consider at the stem level to consider like all of the instruments. that individually make up an entire song and then having that, you know, fit into like a four bar loop. It's pretty complex and technical and these models are just not there yet. I use the analogy a couple of weeks ago is like trying to pull the drops of tea out or drops of cream after you put it in your tea and stir it up. It's really hard to unmix that stuff. It's impossible.

Well, not impossible. Nothing's impossible. But yeah, it's exactly akin to that.