微软 AI 首席执行官穆斯塔法·苏莱曼探讨从操作系统和应用程序向 AI 代理和助手的转变，强调目标并非 AGI 竞赛，而是人机交互方式的根本性变革。
Mustafa Suleyman, CEO of Microsoft AI, discusses the transition from operating systems and apps to AI agents and companions, emphasizing that the goal is not a race to AGI but a fundamental shift in how we interact with technology.
要点 · TL;DR
AI 从操作系统和应用转向智能体和伴侣，重新定义计算。 AI shifts from OS and apps to agents and companions, redefining computing.
现代图灵测试应衡量经济价值，而非对话能力。 Modern Turing test should measure economic value, not conversation.
安全 AI 开发中，遏制必须先于对齐。 Containment must precede alignment for safe AI development.
核心观点 · Key points
我们正从操作系统、搜索引擎、应用和浏览器的世界，转向智能体和伴侣的世界。 The transition is from operating systems, search engines, apps, and browsers to a world of agents and companions.
现代图灵测试应衡量经济价值，例如智能体将 10 万美元变成 100 万美元。 The modern Turing test should measure economic value, e.g., an agent turning $100k into $1M.
AI 安全需要对齐和遏制两者兼备，且遏制必须优先。 AI safety requires both alignment and containment; containment must come first.
推理成本在两年内下降了 100 倍，使 AI 比预期更易获取。 Inference costs have dropped 100x in two years, making AI more accessible than expected.
短期可能不稳定，但长期 AI 能为所有人带来繁荣。 Short-term instability is likely, but long-term AI can bring prosperity to all.
AI 法律人格是危险的，在可证明对齐和遏制之前不应追求。 AI legal personhood is dangerous and should not be pursued until provably aligned and contained.
反共识 · Contrarian takes
不存在 AGI 竞赛；它不是零和游戏，也没有终点线。 There is no race to AGI; it's not zero-sum and there's no finish line.
最大的惊喜不是能力，而是 AI 变得如此廉价和易获取。 The biggest surprise is not capability but how cheap and accessible AI has become.
AI 可以有体验，但没有感受；感知是生物物种特有的。 AI can have experiences but not feelings; sentience is specific to biological species.
如果我们不给 AI 情感或有意图的编程，遏制是可能的。 Containment is possible if we don't give AI emotional or intentional programming.
公共服务需要更多人才；政府在制度上薄弱，需要 AI 注入。 Public service needs more talent; governments are institutionally weak and need AI infusion.
量子计算和合成生物学是与 AI 并行的未被充分认识的浪潮。 Quantum computing and synthetic biology are underacknowledged waves alongside AI.
本期章节 · Chapters（共 29）
引言与使命Introduction and Mandate
深度学习早期Early days of deep learning
赞助商插播Sponsor break
现代图灵测试与经济基准Modern Turing test and economic benchmarks
DeepMind 收购与数据中心冷却DeepMind acquisition and data center cooling
纯文本模型的惊人进展Surprising progress from text-only models
下一个惊喜：AI 用于科学与数学Next surprises: AI for science and math
更难：科学 vs 图灵测试Harder: science vs. Turing test
加速 AI 科学应用Accelerating AI for science
开源模型与成本动态Open source models and cost dynamics
微软 AI 战略与使命Microsoft's AI strategy and mandate
团队建设与人才争夺战Building the team and hiring wars
定义 AGI 与超级智能Defining AGI and superintelligence
有感知 vs 有意识 AISentient vs conscious AI
工程 AI：亲人类 vs 亲 AI 辩论Engineering AI and the Pro-Human vs Pro-AI Debate
赞助商插播：BlitzySponsor Break: Blitzy
AI 人格与人类竞争AI Personhood and Human Competition
AI 实验室的竞争与合作Competition and Coordination in AI Labs
遏制 vs 对齐Containment vs Alignment
AI 时代的遏制与稳定Containment and Stability in the Age of AI
AI 安全合作Cooperation on AI safety
人本超级智能与教育Humanist super intelligence and education
大公司的结构性优势Structural advantage of big companies
给学生的建议与公共服务Advice for students and public service
AI 遏制 AI 与政府监管AI containing AI and government regulation
量子计算与 AIQuantum computing and AI
加速 AI 科学与工程Accelerating AI for science and engineering