Suno’s CEO Says the Future of Music Is a Conversation, Not a Prompt

Jacob Effron introduces Suno as the breakout AI music product: more than 10 million users have generated songs, and the company recently raised $125 million at a $500 million valuation. He frames CEO Mikey Shulman as being at the epicenter of AI music and says the conversation will cover the future of AI and audio as well as Suno’s product and infrastructure challenges. The episode also includes a host debrief after the interview.

Asked for his favorite Suno creation, Shulman names an Estonian-language track that he has listened to about 100 times, even though it is not the most popular song and is not his usual genre. He attributes its pull to the care the user took to realize a specific vision that resonated with him for reasons he cannot fully explain. He also mentions songs he made with his three-year-old son, including one about his son driving a Zamboni, a vehicle three-year-olds are not allowed to drive in Massachusetts.

For Shulman, the songs made with his son matter because they mark the experience of crafting music with someone, not only the final product. He argues that music is currently made mostly in service of the final song or album, with little care for the journey, while making music with other people has been among the most enjoyable moments of his life. The songs become a way to preserve a moment in time he will never get back.

Shulman connects that experience to the playfulness children have with music—banging spoons as drumsticks or pretending to play orchestral instruments—and says adults lose that form of play for cultural reasons. He and Effron discuss how AI creator tools can democratize access to the imagination people had before they were told they were not artists or musicians. Shulman also cites paint-by-numbers as a guided creative experience people liked, suggesting that such guided creativity has been lost and is worth bringing back.

Shulman divides Suno users into two broad groups. The first is casual users who use the product for “soundtracking your life,” narrating happy, sad, funny, or memorable moments musically. He cites songs about Starbucks getting your name wrong or unexpected music gear arriving at your house as examples of funny moments people will relive.

The second group is power users who treat Suno as a serious creative outlet and enjoy the process as much as the result. They may sit down with a sound in their head and a story to tell, then spend two hours working until they get it out. Shulman says it is fulfilling to see the same tool support such radically different use cases.

For the power-user group, Shulman says Suno confirmed that many people have great taste and musical ideas but not the means to realize them. Traditional production tools such as Ableton, Pro Tools, Logic, and even GarageBand are powerful but have steep learning curves. He argues that technological breakthroughs let teams reimagine those processes so more people can make music in different ways.

Effron raises the blank-canvas problem: new users may not know what to prompt or where to begin. Shulman admits Suno has not fully solved it; a blank text box or a suggested idea may not fit the user, though a Valentine’s Day experience worked because the reason to make a song was obvious. He says the future should be more intuitive—humming a melody, tapping a beat, using the last photo on your phone, describing your mood, or turning everyday sounds into music—because people do not need a reason to make a song any more than they need one to send a text or take a picture.

For power users, Shulman says control should not mimic existing DAWs, where users stem out drums and swap a snare for a clap. Most people do not think that way, and the text box is sterile because it asks them to describe what they want rather than pour themselves into it. He imagines users singing their hearts out, assembling a montage of moving pictures, or building a moodboard of everyday sounds; if people enjoy the journey, the recipient tends to enjoy the music more.

Shulman says a major future focus is “everything multiplayer”: making music with other people synchronously or asynchronously. A user might send half a song for someone to finish, or pass a musical idea back and forth for modification, because music is a conversation. He compares this to a jam session where one person expresses an idea, another reacts and riffs, and the exchange stays fluid; everyone has taste, so non-experts can still have a joyful hour and maybe produce a five-minute song they revisit.

Shulman points to a Twitch streamer with a large following who streamed himself using Suno for two hours to an audience the size of a football stadium. Twitch’s micro-payments let viewers interact with the streamer, turning a sterile digital concert into an interactive one. He sees this as an early version of a group experience Suno wants to make first-class, and he imagines stadiums generating songs about the game with input from fans or players.

Suno currently offers a free set of songs and charges power users as they generate more. Shulman says the company is deliberately not innovating on the business model until it has done more important product innovation, because it is still too early to know how people will want to enjoy AI music. He argues that the same people who built and invested in SaaS a decade ago have adapted SaaS pricing to AI almost blindly, even though making a song has nonzero marginal cost; pricing may also differ across music, text, and video.

On evaluations, Shulman says music is harder than other domains. Automatic metrics for audio quality are flawed, and the company accepts that “aesthetics matter,” meaning intangible judgments are unavoidable. Unlike text benchmarks that test reasoning or correctness, music has no correct answer; the real test is whether users love the output and whether the model gives them control over what it makes. Human evaluation remains central because control is important and there is no good automatic substitute yet.

Suno also learns from a large, engaged user base. When it ships a new model, usage and model-choice behavior provide implicit signals, while an active Discord community provides explicit feedback when something feels wrong. Shulman says that community helps the company avoid flying blind.

When a model misbehaves, Shulman says fixes are case-dependent. One example was a model that did not end songs reliably enough, producing overly long outros or silence at the end; after diagnosis, that fix was relatively straightforward. More broadly, he repeats the machine-learning cliché that spending time looking at data is still one of the best ways to understand a model. He also names missing capabilities: iterative edits like “do that but change X,” and better control over tempo or BPM, where the model currently treats requests as loose guidelines.

Shulman describes Suno’s north-star metrics in terms of enjoyment: how many users make songs, how many return day after day, how likely users are to exhaust the free tier, and whether songs are shared or received. Exhausting the free tier is a positive signal even if the user does not pay, because it means they enjoyed the last several minutes. Sharing matters increasingly because Suno is a social experience.

Effron notes that Suno feels fast, and Shulman says it has to get much faster. Spotify has set a high bar by making playback feel instant, and Suno is not nearly as fast because it has to generate the song. He says every 100 milliseconds makes a user more likely to click away, which breaks the creative journey, so caring about latency is a large part of the work. Suno uses autoregressive transformers and streams the song while it is still being generated, which helps compared with diffusion-based generation; the user only cares that the song arrives quickly and makes them feel something.

On infrastructure, Shulman says Suno is still early and tries to be choosy about where to innovate versus buy. He praises Modal for making it easy to deploy jobs onto GPU infrastructure and for being developer-friendly, and he says Suno is grateful it did not have to innovate there. Because music and audio lag behind images and text, Suno benefits from problems already solved by those communities, such as continuous batching infrastructure; he advises builders to be deliberate about where they innovate.

Shulman welcomes the GPT-4o wave because it makes people realize audio is and should be a first-class citizen; most human communication happens through audio. He says interacting with systems like another human will be impactful in obvious areas such as customer service and in less obvious ones such as code, where typing may not always be the most efficient interface. Personally, he describes himself as one of the worst readers and says he prefers face-to-face communication, so more expressive interfaces in other domains are exciting to him.

Asked what he would build outside music, Shulman jokes that it would be something unsexy. He points to home infrastructure and smart homes as terrible, saying voice interaction with a house is crude today—Alexa can turn off lights, but it cannot cook something—and he is especially bothered by opaque, manual plumbing with no telemetry. He says the hardware has existed for decades and does not know why more people are not building it.

Shulman expects everyone to try multimodal models, but he thinks it will take longer than people realize for audio to be more than an interface into an LLM. GPT-4o and ElevenLabs are impressive interfaces, but he argues the sensemaking still happens in text, so consolidation into one giant model is not imminent. Effron adds that the broader tech ecosystem—companies such as Snowflake, Databricks, Amazon, Google, and Microsoft needing offerings—creates incentives for multiple model types to survive, even if the market also consolidates.

Suno recently raised $125 million, and Shulman says the capital is for scale everywhere: training bigger models, hiring, and pulling forward the future of music the company envisions. He expects music models may not need the same compute as the biggest text models in the near future, but training still requires care, specialized data, research, and control mechanisms. He also notes that music modeling is far behind text—there is not even a settled right way to model music—so the expense is not only flops but the work around the model.

Shulman pushes back on the idea that the key milestone is a three-and-a-half-minute pop song indistinguishable from a human studio recording. He understands why it sounds like a benchmark, but calls it artificially constrained because music has no fact-based objective answer and its ceiling is higher: music is about how it makes you feel. His personal wish is a Vision Pro app for playing air guitar with a band, or conducting a symphony, which would require hard real-time synchronization; he frames music as an enjoyable game humans have played for millennia.

On the market, Shulman says AI music is ultimately very big and very green-field. He is excited that competitors are entering because they show this is real, and he expects different businesses: tools for professional artists, background music for YouTube videos, and other niches, none of which Suno is primarily focused on. Suno’s focus is the average consumer—he wants his non-musician mom using it—and building experiences that expand how important music is in people’s lives.

On IP and artist partnerships, Shulman says it is very early and compares the moment to Spotify versus Napster: some companies will work with the music industry and grow it, while others will work against it and face an uphill path. Suno works with some artists but stays away from making new songs by specific artists without consent; he calls such covers viral moments but not a big part of the future. He compares them to early GPT tricks like Shakespeare sonnets about VC, which were memorable but less valuable than just showing users a genuinely useful first experience.

In the quick-fire round, Shulman credits the talent density at Kensho, where he worked with LangChain’s Harrison Chase, and says he was happy to be the dumbest guy in the room among highly capable people. His hot take is that open source AI is overhyped: open source software is amazing, but open source AI is hard under current compute barriers, and he thinks explicit financial incentives matter—which helps explain why Meta produces strong open source work. He calls music underhyped as a part of people’s lives. He also says a surprising win was letting people feel proud of their music, down to editing song titles on the trending page to include their names; a surprise miss was buying GPU boxes early that now collect dust. He says he changed his mind about Discord: a thin web app shipped last November reached 90% of usage within five days, showing Suno is building an experience, not just a music model.

For the live demo, Effron asks Shulman to generate the episode’s intro song, and Shulman—saying he is not a great prompt engineer—uses the community’s crowdsourced prompt knowledge to make a bouncy trap song about unsupervised learning. The generated track includes lyrics about being a machine on a mission, shredding data, and breaking rules. Shulman then points listeners to suno.com and says Suno is hiring primarily on the East Coast—New York and Cambridge, Massachusetts—as well as remotely, and is open to ideas about the future of music from novices and experts alike.

In the host debrief, Jacob Effron and producer Rashad discuss the country song Effron made after Rashad shipped a dozen podcast-equipment boxes to his apartment, which they treat as an example of Suno’s social fun. They describe Shulman as an optimistic physics PhD and note that he plugged Modal, a Suno infrastructure vendor. They also unpack his critique of SaaS pricing: AI products have real marginal compute costs, so a subscription can become expensive if a user generates 100,000 songs, yet companies are still adapting SaaS metrics.

The debrief returns to co-creation: Effron and Rashad highlight Shulman’s excitement about the Twitch streamer, a viral Charlie Puth clip of making a beat from a mug and spoon, and imagine comedians or musicians incorporating live AI music into performances. They note that Suno is still a strong single-player product but that multiplayer sharing is more fun, which matches Shulman’s product direction. On open source, they frame Shulman’s hot take as a bet that the most economically valuable use cases will appear at the closed frontier first and become possible in open models later; Effron still expects open and closed models to coexist. They close by saying the podcast’s own title track was generated on Suno and that it is Effron’s favorite AI product to use day to day.