

UBS AI Podcast - CEO Series - Ep. 2 (Lin Qiao, CEO of Fireworks)
May 15, 2025
Lin Qiao, CEO and Co-Founder of Fireworks, shares insights from her journey leading the development of PyTorch at Meta. She discusses the shift in AI focus from pre-training to inference, highlighting the challenges and strategies in building a generative AI SaaS platform. The conversation explores generative AI's transformative impact on enterprise productivity, emphasizing effective partnerships. Qiao also delves into the importance of customization in enhancing efficiency, and the future of AI, advocating for open-source models and their potential to lower costs and drive innovation.
AI Snips
Chapters
Transcript
Episode notes
Lin Qiao's PyTorch Journey at Meta
- Lin Qiao led PyTorch development at Meta during the company's transition from mobile first to AI first in 2015.
- PyTorch became the main framework for cutting-edge AI model research and production at massive scale.
Inference Outscales Training
- Inference demand will scale far larger than training since user-facing applications scale with world population.
- Lin's company, Fireworks, focuses on inference over training because inference latency and cost efficiencies are critical challenges.
Model-System Co-Design for Inference
- DeepSeek innovated by co-designing model architecture with system constraints, improving inference efficiency.
- Fireworks builds on this by supporting many specialized open models integrated for complex problem solving.