40 Trillion Tokens a Day (Yes, More Than OpenAI) | Lin Qiao, CEO of Fireworks
Summary
Host Lucas Biewald interviews Lynn Chow, co-founder of Fireworks AI, discussing the rise of specialized intelligence over general AI. Chow argues for open-source models to prevent a duopoly, citing Fireworks’ massive token processing and cost advantages. They explore how companies use reinforcement learning to fine-tune models on private data, ensuring proprietary control and economic sustainability in the AI landscape.
Topics discussed
Intro: Open intelligence and Fireworks scale
Lynn Chow's journey from LinkedIn to Meta to Fireworks
Motivation: Building specialized intelligence infrastructure
Fireworks metrics and the rise of customized models
Use cases: Coding agents and vertical-specific co-workers
Training stack: RLHF, SFT, and customer engagement levels
RLHF mechanics: Rewards, feedback loops, and human judgment
Specialized vs. General AI: Proprietary knowledge as moat
Economics: Cost control, token pricing, and value maximization
Security and the debate over open vs. closed source models
Geopolitics: Chinese open source models and US strategy
Engineering: Rapid model support and bitwise equivalence
Platform strategy: Quality, scalability, and customization
Open source runtime: Why Fireworks keeps its engine closed
Leadership style: Authenticity, flat hierarchy, and transparency
Startup lessons: Marketing, decision making, and future outlook
Listen ad-free on Castria