Conference Presentation, Lecture
Post-Training Is How You Keep Your Taste | Fireworks CEO Lin Qiao
- Linh Nguyen (CEO/Co-founder of Fireworks) defines the industry shift from building on "rented" off-the-shelf APIs to "owning" intelligence to preserve unique business judgment, taste, and domain expertise.
- Application development has accelerated due to GenAI, reducing the need for large engineering teams and long timelines, which intensifies competition and incentivizes deeper model customization.
- The "ownership of intelligence" lifecycle consists of three steps: curating/generating high-quality proprietary data, fine-tuning existing models to own the weights, and establishing a serving loop with A/B testing.
- A recommended progression for post-training adoption moves from zero-shot/few-shot prompting to Retrieval-Augmented Generation (RAG), then to Supervised Fine-Tuning (SFT), Preference Optimization/DPO, and finally Reinforcement Learning (RL) for deep domain expertise.
- Different post-training techniques address specific model deficiencies: RAG for dynamic/fact-based data, SFT for structural or behavioral corrections, Preference Tuning for product taste, RL for specialized tasks, and distillation for cost/performance efficiency.
- Key challenges in post-training include ensuring high data quality (requiring product team involvement), converting subjective judgment into systematic evaluation metrics, and avoiding "reward hacking" in RL environments.
- Critical alignment is required between training and serving stacks to prevent precision loss due to differences in numerical libraries and optimization methods.
- Cursor exemplifies early post-training success, leveraging user data to develop models (e.g., Composite 2/2.5) that compete with frontier labs while maintaining control over the model supply chain.
- Niche vertical implementations show significant success: Doximity (healthcare) passed the Stanford Harbor Clinic Safety Benchmark; Factory (security) achieved top benchmark results; GenSpark achieved 5x–10x cost reduction compared to frontier models.
- Post-training adoption drivers include "Frontier Agent Builders" needing bespoke harnesses, incumbents facing CFO pressure to reduce high inference costs, and specialized model operators seeking to codify unique logic.
- The speaker predicts a future of "millions of specialized models," with one model propagation per unique use case, enabled by post-training becoming the primary vehicle for specialized intelligence post-product-market-fit.
- Timing for post-training is dictated by the transition from product-market fit to scalable business; it becomes critical once high-quality user data is collected and unit economics threaten to collapse under high API costs.
- Reward signals are described as "code" derived from multi-dimensional rubrics (e.g., blending scores for candidate aptitude and speed in recruiting), where the specific blending logic represents the company's "secret sauce."
- Fireworks offers tiered engagement models ranging from lowest-level API access for teams with deep research expertise to Training SDKs and researcher deployment support for teams seeking iterative control without full system customization.
- ROI justification for post-training centers on encoding unclonable product judgment and reducing inference costs by 5x–10x, allowing startups to scale traffic without "scaling into bankruptcy."