Fireworks 创始人 Lynn Quo 认为，AI 的未来在于基于专有数据的专用私有智能，而非仅仅通用模型。
Fireworks founder Lynn Quo argues that the future of AI lies in specialized, private intelligence derived from proprietary data, not just generalized models.
要点 · TL;DR
专用智能将取代通用人工智能，每家公司都将拥有自己的模型。 Specialized intelligence will dominate over AGI, with every company owning its own model.
三年内 token 成本将下降 10 倍，推动使用量增长 100 倍。 Token costs will drop 10x in three years, driving 100x usage growth.
开源模型实现定制和控制，对企业采用至关重要。 Open models enable customization and control, crucial for enterprise adoption.
核心观点 · Key points
每家公司都将拥有自己的专用智能，这是必须的，而非可选的。 Every company will own their own specialized intelligence as a must-have, not optional.
Token成本将在三年内下降10倍，推动100倍的使用量增长。 Token costs will drop 10x in three years, driving 100x usage growth.
开放模型支持定制和控制，对企业采用至关重要。 Open models enable customization and control, crucial for enterprise adoption.
未来将有数百万个专用模型，每个应用或用例一个。 The future will see millions of specialized models, one per application or use case.
由于AI扩展成本，产品市场契合与持久业务现在是两个不同的概念。 Product-market fit and durable business are now separate concepts due to AI scaling costs.
公司必须优先优化增长，然后才是利润率，以避免规模扩张导致破产。 Companies must optimize for growth first, then margins, to avoid scaling into bankruptcy.
反共识 · Contrarian takes
单一模型统治一切的AGI是错误的；专用智能将占主导。 AGI as a single model ruling all is wrong; specialized intelligence will dominate.
前沿模型公司可能被高估，因为开放模型可处理90%的企业工作流。 Frontier model companies may be overvalued as open models handle 90% of enterprise workflows.
推理层并非商品；专用部署和定制创造持久价值。 Inference layer is not a commodity; specialized deployment and customization create lasting value.
硬件折旧周期加速，使自建与购买决策取决于时机。 Hardware depreciation cycles are accelerating, making build vs. buy decisions timing-dependent.
主权模型是必要的；国家和公司必须拥有自己的智能基础设施。 Sovereign models are necessary; countries and companies must own their intelligence infrastructure.
对大多数公司来说，造芯片为时过早；工作负载模式过于动态，无法编码到硬件中。 Chip building is premature for most; workload patterns are too dynamic to encode in hardware.
本期章节 · Chapters（共 37）
开场与投资故事Opening and investment story
Fireworks聚焦推理与私有智能Fireworks' focus on inference and private intelligence
私有数据vs通用人工智能愿景Private data vs AGI vision
专业智能与公司独特性Specialized intelligence and company uniqueness
开放vs前沿模型：成本与定制Open vs. Frontier Models: Cost and Customization
国家安全与中国开源模型National Security and Chinese Open-Source Models
企业专业化：Harvey vs LagoraSpecialization in Enterprise: Harvey vs. Lagora
专有知识与专业化Proprietary knowledge and specialization
政府拥有模型Government ownership of models
路由与专业化Routing and specialization
路由层的价值Value of routing layer
Cursor与强化学习基础设施Cursor and RL infrastructure
与Cursor合作及分布式训练创新Partnership with Cursor and distributed training innovation
物理制造的扩展瓶颈Scaling bottlenecks in physical manufacturing
全栈vs专业化策略Full-stack vs. specialization strategy
Jensen在AI堆栈中的角色Jensen's role in the AI stack
AI编码工具支出占开发者薪资比例Spend on AI coding tools as percentage of developer salaries
成本降低与使用增长Cost reduction and usage growth
数据中心部署挑战Data Center Deployment Challenges
利润率与硬件折旧Margins and Hardware Depreciation
优化增长vs毛利率Optimizing for Growth vs Gross Margin
市场成熟度与自建vs购买Market Maturity and Build vs Buy
主权模型与独立性Sovereign models and independence
为何不造芯片Why not build chips
最大瓶颈Greatest bottleneck
收入增长与速度Revenue growth and speed
雇佣George HuHiring George Hu
为何现在雇佣传奇运营者Why now for hiring a legendary operator
快问快答：对增长速度的看法转变Quick fire round: changed mind on growth speed
从Jensen Huang学到的最大教训Biggest lesson from Jensen Huang