Barrchives How To Build AI That Creates Music, with Suno CEO, Mikey Shulman
Jan 29, 2025
Mikey Shulman, Co-founder and CEO of Suno, dives into the fascinating world of AI-generated music. He shares how audio data presents unique challenges compared to text, such as scale and tokenization. Discover why he chose transformer architectures and the importance of rich metadata for music datasets. Mikey envisions a future where anyone can create music, emphasizing collaboration and user feedback. He also reflects on AI's potential to enhance the music economy and deepen artist-fan relationships.
AI Snips
Chapters
Transcript
Episode notes
Use A/B Outputs To Learn Taste
- Use platform feedback loops to learn human preferences by comparing model outputs.
- Serve multiple variants and use user choices to steer models toward better music.
Combine Metrics With Musical Judgment
- Evaluate new models with both quantitative A/B tests and human judgment.
- Recruit people with strong musical taste to judge tradeoffs and decide launch readiness.
A Short Hebrew Regression
- V4 initially produced broken Hebrew output for about a week before a fix.
- Suno detected and patched the language-specific regression quickly after release.