lin qiao

Post-training transforms generic foundation models into proprietary assets that encode a company's unique judgment, taste, and domain expertise, creating a defensible moat that pure application-layer cloning cannot replicate.

The AI stack is stratifying: chips (Nvidia) → infrastructure (Fireworks) → models (open/closed) → applications. Fireworks occupies the specialized intelligence platform layer, offering per-workload model tuning + inference optimization (quality, speed, cost) rather than commodity inference. This layer captures value as enterprises demand control over their intelligence.

Open models have crossed a quality threshold where they solve 90% of enterprise problems at 15x lower cost; they are far easier to tune with small proprietary datasets, letting companies "hill climb" to better-than-generalist performance on specific tasks — driving a structural shift from renting closed APIs to owning customized intelligence.

By distilling frontier-model capability into smaller specialized models, post-training slashes serving costs 5-10x, allowing both startups and incumbents to scale AI features profitably instead of having CFOs block launches due to unsustainable API bills.

As AI moves to production scale, companies hit cost walls with closed APIs (CFOs block rollout). The solution: own open-weight models, customize with proprietary data, control inference. This mirrors the SaaS era where every company built its own software stack — now every company will own its intelligence stack.

Companies should progress from prompt → RAG → SFT → preference tuning → RL → distillation to bake unique judgment into models, creating durable moats and reducing serving costs 5-10x versus frontier APIs.

Building effective reward functions (rubrics, multi-dimensional scoring, anti-hacking) requires the same logical reasoning as software engineering, forcing convergence of product and ML teams where product managers become the ultimate arbiters of data quality and model behavior.

The industry is shifting from a few generalist frontier models to millions of domain-specific models (legal, finance, healthcare, coding, security), each post-trained on proprietary product data and reward signals, making post-training the new standard after product-market fit.