Open Source Startup Podcast E181: Why Multimodal Is the Future of AI Data Workloads
Sep 9, 2025
Chang She, Co-Founder and CEO of LanceDB, dives into the transformative power of multimodal AI, highlighting its applications in autonomous vehicles. He shares insights about the revolutionary Lance format that achieved a staggering 9,000% performance gain in real-time analysis. The conversation also covers the trust-building process in open source and the importance of integrated workflows beyond traditional vector databases. Looking ahead, he discusses emerging trends in AI, like audio infrastructure and the challenge for vector databases to evolve or face irrelevance.
AI Snips
Chapters
Transcript
Episode notes
Hunan Dinners Sparked LanceDB
- Chang She and co-founder Lei conceived LanceDB while struggling to manage large image/video datasets during dinners at Henry Chong's Hunan restaurant.
- Their pain with existing tooling for video and image data sparked the decision to rebuild the data foundation from scratch.
Foundation First For Performance
- Performance at scale depends first on the storage foundation, then system optimization, then developer experience.
- Changing the underlying format raises the system's speed-of-light and unlocks much higher ceilings.
9,000% Speedup In A POC
- An early customer using protobuf and Python analytics saw analysis slower than real time; one second of data took >1s to analyze.
- Converting to Lance format and Rust-based queries produced a ~9,000% analytics speed improvement for their scenario mining.