Firework CEO Lynn 解释了为何后训练是公司将独特判断和数据嵌入 AI 模型、超越现成 API 的关键。
Lynn, CEO of Firework, explains why post-training is key for companies to embed their unique judgment and data into AI models, moving beyond off-the-shelf APIs.
要点 · TL;DR
后训练将独特判断和数据嵌入模型，打造超越产品市场契合度的持久商业护城河。 Post-training embeds unique judgment and data into models, creating durable business moats beyond product-market fit.
从提示到强化学习的进阶映射人类学习，从事实到品味再到专长。 Progressing from prompting to RL mirrors human learning, from facts to taste to expertise.
数据质量和系统化评估至关重要；后训练可降低成本 5-10 倍，实现盈利扩展。 Data quality and systematic evals are crucial; post-training can cut costs 5-10x, enabling profitable scaling.
核心观点 · Key points
后训练是公司通过将独特判断和数据融入模型来构建持久业务的关键。 Post-training is key for companies to build durable businesses by baking their unique judgment and data into models.
从提示到RAG到微调到强化学习的进展，反映了人类从事实到品味再到专业知识的学习方式。 The progression from prompting to RAG to fine-tuning to RL mirrors how humans learn, from facts to taste to expertise.
数据质量比数量更重要；产品团队是特定用例数据质量的最佳评判者。 Data quality matters more than quantity; product teams are the best judges of data quality for their specific use cases.
评估至关重要；应将主观评估转化为系统化评估，以确保可重复的质量。 Evals are essential; vibe evals should be converted into systematic evaluation to ensure repeatable quality.
对齐训练和服务栈至关重要，以避免因数值差异导致的质量下降。 Aligning training and serving stacks is critical to avoid quality drops due to numerical differences.
在实现产品市场契合后开始后训练，此时你拥有有意义的数据并需要经济高效地扩展。 Start post-training after product-market fit, when you have meaningful data and need to scale cost-effectively.
反共识 · Contrarian takes
产品市场契合不再保证持久业务；需要后训练来构建护城河。 Product-market fit no longer guarantees a durable business; post-training is needed to build a moat.
不仅是初创公司，现有企业也面临规模扩张导致破产的风险；CFO因成本而阻止AI发布。 Incumbents, not just startups, face scaling into bankruptcy; CFOs block AI launches due to cost.
后训练可以降低5-10倍成本，使初创公司能够盈利扩展并与前沿模型竞争。 Post-training can reduce costs 5-10x, enabling startups to scale profitably and compete with frontier models.
奖励工程类似于软件工程，降低了后训练的门槛。 Reward engineering is similar to software engineering, lowering the barrier to entry for post-training.
未来可能出现数百万个专用模型，每个应用用例一个，而不仅仅是少数前沿模型。 The future may see millions of specialized models, one per application use case, not just a few frontier models.
应用克隆很容易；将数据融入模型比应用本身更强大。 Application cloning is easy; baking your data into a model is a stronger moat than the app itself.
本期章节 · Chapters（共 11）
介绍与设置Introduction and Setup
行业转变与深层模型需求Industry Shift and the Need for Deeper Models
拥有你的智能Owning Your Intelligence
从提示到强化学习的演进Progression from Prompting to Reinforcement Learning