Fireworks AI CEO Lin Qiao: post-training is how startups bake in their taste and own their intelligence

At Sequoia Capital, Fireworks AI CEO and co-founder Lin QiaoDealroom has a profile for this one. Try Dealroom → made the case for post-training as the way application companies build a durable moat, moving off off-the-shelf black-box APIs to models that carry their own judgement, taste and customer understanding.

Why post-training matters now. As one person can now ship in weeks what once took a cross-functional team quarters, the app layer is more competitive, pushing teams to bake their taste into owned intelligence rather than rent a generic API.

A progression, not a leap. Qiao maps the path from prompting and few-shot, to RAG and context engineering, to supervised fine-tuning, to preference tuning (DPO) for taste, to reinforcement learning for domain specialty, mirroring how humans learn.

Technique by problem. RAG for dynamic facts; SFT to fix structure/behaviour; preference tuning for taste; RL when the model is weak on a specific problem; distillation when the model is too slow or costly to serve.

Common pitfalls. Data quality over quantity (the product team is the best judge); building real evals rather than “vibe evals”; sloppy RL environments that invite reward hacking, e.g. a model generating zero lines of code to eliminate compile errors.

Pioneer customers. Cursor post-trains on Fireworks to control its model supply, shipping Composer 2 / 2.5 aiming at frontier quality; Doximity topped a Stanford/Harvard clinical-safety benchmark; security-coding startup Factory tuned a benchmark-topping model.

Cost as a driver. Post-training can cut serving cost 5-10x, letting startups scale traffic on the same budget, and it helps incumbents whose CFOs block AI launches on cost grounds.

Millions of specialised models. Qiao expects a future of one specialised model per application/use case, with reward engineering resembling software engineering more than a deep research barrier.

When to start. Post-training pays off after product-market fit, when a startup has meaningful product data and needs both a durable competitive edge and healthy unit economics.

Read more: Sequoia Capital, “Post-Training Is How You Keep Your Taste | Fireworks CEO Lin Qiao” (YouTube).