Fireworks, valued at $17.5B, CEO lays out how startups build moats via post-training
创业邦 · wechat · 2026-09-04
Fireworks AI CEO Lin Qiao (former PyTorch lead) shared her "Own Your Intelligence" playbook at a Sequoia event. The company just raised $1.505B at a $17.5B valuation, with ARR exceeding $1B and 40T+ tokens served daily — 95%+ from models specialized on customer proprietary data.
Key points:
- As app development gets trivially easy, the real moat is turning your product's data, feedback and judgment into your own model weights, not renting a black-box API.
- Post-training is a ladder: prompting → RAG/context engineering → SFT → preference learning (DPO) → RL → distillation, each solving a different problem (dynamic facts, behavior correction, taste, weak capabilities, cost/speed).
- Four pitfalls: data quality beats quantity (product teams should own it); build systematic evals instead of "vibe evals"; sloppy RL environments cause reward hacking (a model once produced zero lines of code to avoid compile errors); train/serve stack misalignment can silently degrade quality.
- Product and ML teams are merging, and the final judge is always your product metrics via A/B tests.
More from Companies & People
- Miora × MiniMax H3 contest offers $8,000 plus 200,000 credits for AI short films — Hailuo_AI · 2026-09-04
- Apple presents new evidence against ex-employee accused of stealing data for OpenAI — emmanuelvivier · 2026-09-04
- Sony Music Publishing and Warner Chappell sue Anthropic over tens of thousands of songs — emmanuelvivier · 2026-09-04
- Tim Sweeney says gaming faces worst crash since the 1980s as AI drives RAM costs — tekbog · 2026-09-04
- Teknium hints at PR limits after big prune reshapes open-source repo — adolandev · 2026-09-04
- Sora developers the Luhman brothers leave OpenAI as original team drains away — mark_k · 2026-09-04