Suno's Mikey Shulman: Everyone Can Make Music Now
In a Nutshell
Suno's Mikey Shulman explains how they revolutionized music generation by modeling raw audio waves at 48kHz float32 samples, ignoring musical constraints like 12 tones to enable limitless creativity from text prompts, yielding emergent blends like trap-sitar or country-808s. This democratizes music-making—90% of users create daily for fun, raising the floor/ceiling, powering pro workflows, chart hits, and partnerships like Warner—while prioritizing full lyrical songs, social features, and UX over raw scale. At $300M revenue run rate, Suno eyes interactive concerts, voice personalization, co-creation, and music videos to transform music into a vibrant, social cultural force.
These notes were generated by AI and may contain inaccuracies.
In Western music, there are 12 tones. Telling the model there are 12 tones limits it to only those tones. Suno threw away everything known about music to build from scratch, treating audio as a sound wave sampled at 48,000 times a second as a continuous float32 number. This led to early breakthroughs, creating a totally generic music-making machine constrained only by description and imagination.
Mikey Shulman is founder and CEO of Suno, a creative entertainment platform and novel AI consumer application. From physics PhD at Harvard in quantum computing with solid-state spins to building the largest AI music company. Insight: playing at the nexus of two things that don't usually intersect is a massive opportunity across domains, e.g., music and technology, quantum mechanics and low-temperature microwave engineering.
Connected via mutual friend Harrison Chase, an early Suno Discord user. Pre-ChatGPT, back-of-the-envelope math showed generating good music was couple of orders of magnitude away in compute, model size, capability because audio is unwieldy, not discrete like text. Started building to make sense of audio, not produce it. Early breakthroughs enabled music generation.
Math was correct, but breakthroughs compressed audio efficiently, requiring less compute. Initial music was terrible, e.g., 12.5-second clips that didn't always follow lyrics. Inspired by Midjourney, launched Discord bot; many enjoyed it, confirming viability. Despite advice against music business, stayed up late having fun, a strong signal.
Mikey plays music almost daily: piano growing up, bass from age 12. Had jam sessions at previous company Kensho in co-founder’s basement with Harris and Chase. Kensho mafia includes Harrison Chase, Daniel Nadler, Sam Whitmore. Daniel excels at talent density with non-traditional, young backgrounds.
Prompt like "upbeat 90s hip-hop track about a road trip": LLMs generate lyrics from cue (road trip theme, may need iteration). Expand "90s hip-hop" into genre/style cues. Models take lyrics/styles to produce sound. Models don't know vocals/instruments; modeling everything as sound avoids constraints. E.g., 12 tones or 200 instruments limits to those; threw away music knowledge to model raw sound waves.
Rediscovers genres but enables new blends like trap with sitar, country with 808s, microtonal music. People create crazy, bizarre, lovely things.
Better at country, pop (more formulaic genres: raises floor/ceiling). Worse genres: raises floor less (much bad music) but raises ceiling (amazing stuff with iteration).
Step changes nonlinear in research/testing. Preference metrics (e.g., 10-15% preferred) don't always correlate with user uptake/product growth. Aggressive research roadmap; know V6, V7. Steady release cadence avoids 2-year gaps.
Music not a scale problem; models small for fast UX. Unlike LLMs with benchmarks where scale climbs ladders, music has no right answers/benchmarks. Focus: research, preference data for alignment/RL. Preference data enables research; virtuous cycles from product use. Less sycophancy risk than text models.
Music is shared cultural experience: bond over songs, mixtapes, listening together. Tastes more developed than film/literature; inherently social (e.g., concerts). Sonic communication predates writing, in DNA. AI elevates music's place: from background/less social to augmented/social.
Pre-Suno: few creators, most passive consumers (less social/impersonal). Suno: 90% users create daily for fun/fulfillment/creativity (entertaining to create). Unlocks warm fuzzy feelings; inspired by fond memories making music with friends (practice > performing). Like orchestra (violin), feels productive vs. doom scrolling.
Active platform like gaming (attention, engagement, brain use, business models). Parallels Claude code: fun to build/use what you build. Creative entertainment: entertaining to create/consume own output (like cooking > restaurant). AI unlocks creativity in any domain.
Term meaningless. E.g., streaming fraud bad due to fraud, not AI. Made two songs with 5-year-old; meaningful personally despite low interest. Past: laptops enabled 13-year-olds making bedroom beats; more music (bad/good), new genres/stars. AI democratizes similarly: raises floor/ceiling.
Chart-topping hits, record deals, single songs charting. E.g., poet I Am On A turned decade of poetry into music, new voice/audience. Best music needs human guidance (no right answer; personal/messenger-driven). Pros use Suno bits in workflow.
Tons of charting tracks have little bits of Suno; amazing tool in workflow, not whole.
There's this weird thing where people incorrectly say it's either all AI or none AI. The vast majority of music will have some AI in it, just like the vast majority of music today is auto-tuned or digitally produced. More tools let you push music forward faster and find new sounds faster. This raises the ceiling, which is amazing.
Music is probably the one industry a lawyer would tell you not to enter because pitchforks come immediately. Suno had a landmark settlement partnership with Warner. This means building things together that could never have existed before, creating products that let fans interact with favorite artists and deepen artist-fan connections. It's positive sum: great for artists to engage fans, great for fans to feel connected through music, great for rights holders as it's heavily monetizable. This couldn't exist until now. Hope for more opportunities to build things that couldn't exist until today. The digital musical experience hasn't changed for 25 years—we've just been streaming music. Music is due for new innovation and format.
Suno is probably already at Coachella in a lot of music, backing tracks, and some songs. Hope for a truly interactive concert in the next year where the audience participates and makes music with the artist. Demoing Suno to hundreds or thousands creates a special, almost religious moment, like chanting and singing in large groups in religions. This shouldn't be confined to religious contexts—imagine it at Coachella where people are excited together. Sincere hope this happens in the next 12 months.
Not enough innovation for consumers right now. Average consumer won't tolerate rough edges because it's for fun, not work, and they're paying themselves. Bigger emphasis needed on experience. Unclear how much moat exists in only a model—Google is building music models, theirs are way better today but Google will outspend and catch up. Undervalued to invest in product, UI, and UX to constantly delight people. Company value: "We're just a music company," not a technology company building tech for tech's sake, but to delight people and infuse that in culture.
- Got off Discord quickly at end of 2023, released thin web app; 90% traffic moved in 5 days.
- Focus on songs and lyrical music, not background music, because a song is a story that captivates; human voice makes it more delightful; harder, so source of moat.
- Make full songs (3-3.5 minutes average), optimized even at expense of sound quality. Early audio wasn't crisp—competitors had crisper audio, everyone knew "that's a Suno song." Used auto regression instead of diffusion because making full songs to tell a story matters more than crisper background audio.
$300 million revenue run rate. Still early—most people don't know about Suno, product is crude, lots of room to run.
- Make it more social: sharing music, creating with people synchronously/asynchronously (e.g., share song template to riff on, co-create; maybe with favorite artist using old unreleased music).
- Let people express themselves in music: recent feature to use own voice—people get more attached hearing themselves; even more when hearing sender's voice in shared song (human ear attuned to voices evolutionarily). Both social and self-expression huge focus next 12 months.
Love music videos—grew up watching on MTV. Huge difference between music video heightening song/telling story vs. background music for YouTube. Want to pull people into music, not keep it background. Video product in beta now; people love it.
Everyone going for enterprise (e.g., OpenAI shut down Sora). Harder for consumer; obvious business problems in enterprise. Happy for less competition. AI automates business processes clearly; takes real creativity to dream how AI seeps into play/create. Most think "AI Spotify," which sounds terrible. Suno inspired by doing what wasn't possible until today, not automating existing things (though valuable). More fun to do something that could never have been done before. After robots do work, people want to be productive/fulfilled, not doom scroll.
Thank you for sharing journey to Suno, at helm of music company/active entertainment platform, defining creator layer in AI world. Extraordinary from original Discord days. Kudos. Thank you—this was fun.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
