微软 CEO 萨提亚·纳德拉讨论从单一模型向生态系统战略的转变、AI 训练中干净血统的重要性以及推理模型的新前沿。
Microsoft CEO Satya Nadella discusses the shift from single models to ecosystem plays, the importance of clean lineage in AI training, and the new frontier of reasoning models.
要点 · TL;DR
AI 的真正价值在于用户获得的独特成果，而非基准分数。 True AI value is measured by unique user outcomes, not benchmarks.
私有评估是企业改进前沿模型的关键知识产权。 Private evals are key IP for companies to improve frontier models.
微软的平台承诺让每家公司都能在前沿运营。 Microsoft's platform promise enables every company to operate at the frontier.
核心观点 · Key points
真正的评估是当人们能做他们看重的独特事情，而不是基准分数。 The true eval is when people do unique things they value, not benchmark scores.
每家公司都应有私有评估，以便在前沿模型上爬山而不泄露轨迹。 Every company should have private evals to hill climb on frontier models without leaking traces.
平台承诺是让每家公司都能用自己的智能在前沿运作。 The platform promise is enabling every company to operate at the frontier with their own intelligence.
公司必须提供切实利益，才能赢得社区对AI基础设施的许可。 Companies must deliver tangible benefits to earn community permission for AI infrastructure.
下一个大创业公司可能建立新大学或教学法，以创造经济机会。 The next big startup could build a new university or pedagogy for economic opportunity.
反共识 · Contrarian takes
我们的工作是构建做Azure网络的智能体系统，而不是自己去做。 Our job is to build the agentic system that does Azure networking, not do it ourselves.
私有评估可能是最大的知识产权，比模型或数据更有价值。 Private evals may be the biggest IP, more valuable than models or data.
按用户定价是预算确定性的产物，不是唯一模式；按用量定价将增长。 Per-user pricing is an artifact of budget certainty, not the only model; consumption pricing will grow.
大多数人喜欢基于结果的定价，直到他们有了结果；然后他们更喜欢按用户定价。 Most people love outcome-based pricing until they have an outcome; then they prefer per-user.
在智能体世界中，通才角色将获得最大回报，而非专才。 The generalist role will see maximum returns, not specialists, in the agentic world.
我们在15个月内建造的Azure容量超过了前15年。 We built more Azure capacity in 15 months than in the first 15 years.
本期章节 · Chapters（共 16）
怀疑与切实利益Skepticism and Tangible Benefits
引言与生态战略Introduction and Ecosystem Strategy
训练策略与MAI模型Training Strategy and MAI Models
规模法则反思与部署Reflections on Scaling Laws and Real-World Deployment
编码之外的用例Use cases beyond coding
企业级概念框架The harness concept for enterprise
自研产品与平台赋能First-party products vs platform enablement
微软的平台承诺Microsoft's platform promise
前沿智能与平台价值Frontier Intelligence and Platform Value
企业vs供应商代理与SaaS演进Enterprise vs Vendor Agents and SaaS Evolution