Post-Training Is How You Keep Your Taste | Fireworks CEO Lin Qiao

Summary

Lin Qiao, CEO and co-founder of Fireworks AI, explains why companies should move beyond off-the-shelf APIs and use post-training to encode their own judgment, taste, and domain expertise into models. She outlines the progression from prompting and RAG to SFT, preference tuning, RL, and distillation, and highlights pitfalls like poor data quality, vibe evals, reward hacking, and training-serving misalignment. She cites examples including Cursor, Doximity, Factory, and GenSpark, and argues post-training can cut serving costs 5–10x and enable millions of specialized models.

• Fireworks AI CEO Lin Qiao discusses post-training at Sequoia's Own Your Intelligence event.
• Companies progress from prompting and RAG to SFT, preference tuning, RL, and distillation.
• Data quality, systematic evals, and aligned training/serving stacks are key pitfalls.
• Post-training can cut serving costs 5–10x and help avoid scaling into bankruptcy.
• Cursor, Doximity, Factory, and GenSpark are cited as post-training success examples.
• Coding was the leading AI use case in 2025, with co-work expanding into legal, finance, and other domains.
• Lin Qiao expects millions of specialized models, one application per use case.