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

Industry Trends & Strategic Shift - Software development resource requirements have collapsed, allowing individuals to build production-ready applications in a few weeks instead of requiring a team of tens of product engineers over multiple quarters[5] - Companies are shifting from off-the-shelf black box APIs to post-training and owning their own intelligence to build durable business moats[6][8][9] - The post-training methodology progresses logically from prompt engineering and RAG to supervised fine-tuning, preference tuning, and reinforcement learning[13][15][17] Cost Optimization & Economic Performance - Post-training specialized models can achieve a 5 to 10 times cost reduction compared to frontier models, enabling startups and incumbents to support significantly higher traffic without scaling into bankruptcy[40][44][62] - Incumbent enterprises face significant cost burdens when deploying AI features, often leading Chief Financial Officers to block AI feature launches until post-training is applied to remediate costs[44][45] Practical Implementation & Challenges - Data quality and the product team's judgment are crucial for successful model tuning, replacing traditional separate ML organization structures with cross-functional convergence[23][24][25] - Transitioning models from training stacks to serving stacks requires strict alignment of calculations, numerics, and optimization libraries to prevent precision loss[31][32] - Successful pioneers like Cursor, Doximity, and Factory have leveraged post-training to match or beat frontier labs on benchmarks for coding and healthcare safety[35][36][37][38]