Summary

Shulman’s biggest structural call: models will eventually disappear into product. “At some point — I don’t know if it’s version four or version five — there will be a last model release that is released as a model, and everything else is just product releases.” Harry agrees nobody will pay more for a better model in five years; Shulman goes further — you won’t even know what model is under the hood, because “this is a product game at the end of the day.”
Music AI needs more than text-style scaling. Text benchmarks are objective; music is taste, so “scale is not the answer to all the problems” and models stay relatively small. Suno’s edge is not architecture (it’s a Transformer) but audio tokenization — “take everything from the open source text community for how to make and scale models” — plus a talent thesis of hiring economists (natural experiments) and physicists, because “AI is an empirical discipline and whoever can run more high-quality experiments quickly will win.”
The market-size thesis: music should work like a video game — interactive, social, and paid for like Fortnite, in an industry “so much bigger than the music industry.” Suno charged from day one, users are “paying to create the music — this has nothing to do with listening,” and the paywall doubles as a research instrument. GPU burn is the biggest cost “by far… a few times payroll.”
On the June 2024 lawsuit, a notably candid admission: “we know that there are some copyrighted works in our training data — that’s not illegal, it’s stock standard for the industry, it’s what every AI company does.” If Suno loses at trial, “the company’s not dead but it’s obviously not good for us” — though he thinks a loss unlikely, and frames the fight as the music industry’s “fixed pie mentality” versus growing the pie to gaming-industry size.
The new business models are the tradeable part: artist-licensed personal models (an Ariana Grande who opts in gets unlimited supply of herself — though “she might not even be allowed to do that” under her contracts), prompt-level royalties Spotify Radio never pays, Patreon-style forks of a teenage creator’s model, and the likely Timbaland partnership — not paid in cash, with him as an adviser — that “gives cover to up-and-comers” since most artists “behind closed doors admit that they use and love Suno.”
Distribution scorecard: he thinks TikTok wins short-form discovery and Spotify wins the algorithmic Discover Weekly, and he thinks things may break on TikTok then become famous on Spotify. YouTube beats Spotify on engagement “as measured by the ARPU for ads.” Discovery is already far more algorithmic than people realize — popularity “is a product of recommendation algorithms,” not just quality. Suno refuses the vast majority of gen-AI video companies asking for an API: background music “is not making music more valuable.”
Against the consensus: “everybody believes AI is inevitable and I don’t think it is — we need to go and do it.” His two dystopias if the West sits back: offshore actors enabling unlimited artist impersonation (“endless Ariana Grande songs without giving a cent to Ariana Grande”), and hyper-personalized antisocial streams that know your heart rate and mood — “just hitting that nerve in your brain, it’s like a drug.” Bonus contrarian call: quantum computing “will be amazing but you still should not do it” — it needs government money, not venture, because “I don’t think you’re going to get your money back in a reasonable time.”

Deep dive

1. Scale won’t solve music — taste will, and models stay small

Shulman’s core technical divergence from the LLM consensus: text scaling works because problems are objective — “I want to get a better SAT score, I want to do better on this benchmark” — but “music is totally subjective, and so scale is not the answer to all the problems. The models stay relatively small,” requiring other techniques to give models good taste.
Suno runs a Transformer and is “not shy about that” — the moat is audio representation, not architecture: “it’s not obvious how to tokenize audio, but if you spend night and day doing that and basically take everything from the open source text community for how to make and scale models, that’s a pretty good recipe.”
Aligning models to human taste rather than objective truth is the recruiting pitch: “there’s nowhere else to do it.” Suno’s usage volume enables heavy A/B testing, but he flags real uncertainty — “it’s totally not obvious that the same techniques used to align LLMs to weird human taste should be the same techniques we use to align music models.”

2. The last model release — after that, everything is product

The episode’s sharpest prediction: “at some point — I don’t know if it’s version four or version five — there will be a last model release that is released as a model, and everything else is just product releases,” because “people aren’t going to care what powered the thing — you’re just going to care that the music made you feel a certain way.” Harry’s parallel call: in five years nobody pays more for access to newer models.
His swipe at the industry’s default interface: “OpenAI is amazing — they did every AI company a huge disservice because everybody thinks that just like empty text box is now the right interface, and it is for ChatGPT and it is incorrect for basically everything else.”
He rejects the prompt era entirely — “I hope in 6 or 12 months we’re not using the word prompt… we shouldn’t be guiding you, we should be listening to you.” When Harry cites prior guests (likely Bsky and Gustav) saying they’d measure candidates by prompt quality: “if it took a really complicated prompt and 1,600 iterations, that’s a failure of my product, not a failure of the candidate.”

3. Music should be a video game — that’s the market-size argument

The framing that anchors the whole thesis: video games are interactive, engaging, better with friends — “nobody half plays video games the way people put on music in the background.” Make music that engaging and people pay for it like they pay for games, in an industry “so much bigger than the music industry… it seems just crazy that music should not be as engaging as Fortnite.”
Harry’s economist pushback — infinite supply reduces price — gets a clean concession-and-reframe: “it reduces the average value of any given piece of content, but it greatly increases the value of music to society.” The tagline: “we’re not making music, we’re making musicians.”
Why so few people make music today: it’s like running — “it is hard to run, it is painful to run… most people drop out of that pursuit.” Building for a billion people means removing that pain, not making current creators 10% faster.

4. Charging from day one turned the paywall into a research instrument

Suno charged from its Discord-bot launch, against “the traditional wisdom of Silicon Valley of just give the product away for free, scale scale scale” — “we didn’t want to be a novelty item.” Beyond offsetting GPU burn, the paywall helps identify whom to interview: without it, “I’m kind of lost in the desert” on who found the product valuable. “Maybe we would be a little bit bigger, but we would also have a worse product.”
The success metric is unusual: did you hit the paywall on day one — “even if you didn’t go through it, you just enjoyed the last 10 or 12 minutes of your life.” The fraction that does is “fairly high.”
Time-to-wow is measured obsessively: they can inject artificial latency and watch satisfaction drop — “10 seconds is worse than eight.”
Cost structure, stated plainly: GPU spend is “the biggest thing by far… a few times payroll,” including a large research cluster. And the ML team will scale more slowly than software, with sublinear returns: “research is not something you can solve with scale — you can just throw more bodies and get comparatively more results.”

5. The RIAA lawsuit: an admission, and an engineers-vs-lawyers theory of the fight

On the June 2024 suit, careful but direct: “we know that there are some copyrighted works in our training data. That’s not illegal. It’s stock standard for the industry — it’s what every AI company does.” The suit “wasn’t totally shocking” — every AI company gets sued, “everybody in music gets sued.”
His framing of the waste: citing likely Andrei Shleifer’s growth research — “more engineers equals more growth, more lawyers equals less growth” — against a music business with “such an embedded fixed pie mentality.” What he wanted from the labels: “talk to us first, probably… deciding to sue first and ask questions later seems to me to be inefficient.”
Downside honestly sized: if it goes to trial and Suno loses, “the company’s not dead, but it’s obviously not good for us — I don’t think that’s likely.” He also refuses founder bravado about litigation: founders who claim to “feed off the pain… are lying. It’s going to cost money and time.” His counter to the labels’ logic: “even if you could make Suno-like companies go away, do you really want that? What if the music industry can be as big as the gaming industry?”

6. The new business models: licensed artist models, prompt royalties, and likely Timbaland as adviser

The Ariana Grande scenario — exclusive licensed models with full artist control — gets an unambiguous yes, with a contractual twist: “she might not even be allowed to do that… she doesn’t own all of her music, but she owns her name and likeness.” Today Suno blocks her name in prompts (“Suno is for original music, not for impersonating people”), but a super-fan model would be “the equivalent of fan fiction… way more valuable and way more engaging than an AMA with her.”
Harry’s Dean Lewis example lands the sharpest gap in today’s economics: Spotify Radio plays “songs like his” and he does not get paid — versus a future where an artist opts into being promptable and takes a cut. Shulman: “I would love to see a future where you can.” The broader point: “right now people are paying to create the music — this has nothing to do with listening,” and the capped stream-share pie “does not properly account for” interactions like thousands of fans remixing likely Timbaland in a contest — “so much cooler than meeting them backstage.”
The likely Timbaland deal itself: “we don’t pay him in cash” — he is an adviser who was a fan before they met (“you can’t fake this stuff”). Its real function: “you have someone who has made it, who does not need to do this, publicly saying yeah I use Suno and it’s awesome — this gives cover to up-and-comers,” since “the vast, vast majority of artists I speak to behind closed doors admit that they use and love Suno.”

7. Spotify vs TikTok vs YouTube — and why Suno won’t sell background music

The scorecard: “he thinks TikTok wins on short-form video discovery and Spotify wins on the algorithmic Discover Weekly” — but he thinks things may break on TikTok first and then become famous on Spotify. The bigger misconception: “people don’t realize the extent to which the music that is popular is a product of recommendation algorithms” — not only of quality.
YouTube beats Spotify on engagement “as measured by the ARPU for ads” because video engages; he hopes Spotify’s video push works, since so much listening is deliberately ignored — maybe 100% of the student engineers he’s interviewed in the last 2 years code to music whose “whole point is not to pay attention to it.”
The API refusal is a strategy statement: gen-AI video companies frequently ask for Suno music behind their videos and “the answer is always no” — “being the background music for your video is not making music more valuable. You can go get Epidemic Sound music for very, very cheap.”
Pop got boring for structural reasons — songs shorter and “much more homogeneous in melodies, harmonies, song structures,” a product of digital production tools, streaming, and TikTok, with Olivia Rodrigo cited as structuring songs for the algorithm. His verdict withholds judgment: “it’s not obviously good or bad, it just is.”

8. “AI is inevitable” is a cop-out — and two dystopias if builders sit back

His contrarian quick-fire answer: everyone — incumbents and disruptors — says “we know it’s coming,” incumbents so they don’t “sound like Luddites,” AI people as “a security blanket.” Both are “basically just putting your hands up… if we just say it’s inevitable, it’s not going to come and happen. We need to go and do it.”
Dystopia one: an offshore group “not bound by US laws” enabling permissionless impersonation — “endless Ariana Grande songs without giving a cent to Ariana Grande.” Already possible: “you saw this with the Drake ghostwriter thing — and it’ll be good, and it’ll just get better and easier.”
Dystopia two: hyper-personalization as a local optimum — an app that knows your texts, mood, and heart rate, streaming “endless music that only you’re going to like, that is just hitting that nerve in your brain — it’s like a drug. And this is extremely antisocial.” Harry’s song for his girlfriend is the counter-case: made for someone, “so much more social than what exists now.”
Where creation goes: “increasingly taste is the only thing that matters in art, and skill is going to matter a lot less” — the rock star → DJ → playlist-curator progression continues into “I can’t play the piano, but I am very good at picking through the Suno music — and that makes me a creator.” The closing analogy: photography pre- and post-Instagram — each image worth less, far more people making a living, “in aggregate this is a much, much better future.”

9. Operator scar tissue: Discord, remote work, focus, and reading VC-speak

The confessed mistake: staying a Discord bot too long because “I look at Midjourney, they’re printing money.” A thin web app shipped in November took 90% of traffic in five days — “there’s no world in which you can say that I got that right.” Discord’s 400k members now serve a different job: community and feedback.
Remote work is allowed by exception, despite the cost: “exceptions are a sign of judgment,” and as you scale, per-person judgment calls get “very, very hard.” A CEO friend’s line he quotes: “I know remote work is killing my company — I just can’t get people back.”
Hardest part of the job is focus: “at any given time there are 30 things we could do that would be very impactful, and we have to choose three.” The named casualty: the API everyone asks for — “that doesn’t really build the future of music that I want to do.” Related hiring principle: “good judgment is more important than good skills — we know how to assess skills and we don’t know how to assess judgment, so everybody’s teaching for the test.”
On the “little over 125” raised: “capital is a weapon” with step changes, not a 10% balance-sheet top-up. And a decoded piece of VC-speak he wishes he’d known: “we reserve a good portion of our fund in case you need it” really means “I’m going to take my pro rata and there’s nothing you can do about it — in the bad cases I’m not going to give it to you. That’s built as very founder friendly and it’s actually the opposite.”

10. Quantum, talent arbitrage, and first-order thinking

The spiciest side-call: quantum computing’s promises are incredible “but you still should not do it” — basic physics research remains, “all of the companies that sprang up around this are not doing all that well,” and it needs “massive government intervention,” not venture: “I don’t think you’re going to get your money back in a reasonable time.” Harry’s objection — that’s a tragedy of the commons — stands unresolved; Shulman holds that joining quantum startups now is bad career advice.
The talent arbitrage, since Suno won’t match OpenAI pay: he sees economists and physicists as strong ML hires — economists for first-principles reasoning and natural experiments in data-poor environments, experimental physicists because “AI is an empirical discipline, and whoever can run more high-quality experiments quickly will win.”
His deepest worry is cognitive, not technological: “people are not good at judging first-order effects” — at his MIT Sloan course, critics call ChatGPT the end of education when the first-order effect is “every person in the world has a median competent tutor… this is obviously amazing for education.” The second-order fix is on instructors: “I need to change what I teach people,” because employers will let you use GPT anyway.
On Harry’s AI-winter worry (a bank event where “the ROI is not here”): “I live in a bubble in Cambridge, you live in a finance bubble in London — those people are going to potentially have a disillusionment, but is the whole world going to be like that? Most people are still not expecting anything.”

Meta Goes MAGA Mode | EP 117 Why Open a Bookstore Today? A Conversation with Feng Rusong’s 丁永勋