Anthropic 联合创始人 Ben Mann 探讨超级智能的时间线、离开 OpenAI 的原因、AI 的生存风险以及如何为后奇点世界做准备。
Anthropic co-founder Ben Mann discusses the timeline for superintelligence, why he left OpenAI, the existential risks of AI, and how to prepare for a post-singularity world.
要点 · TL;DR
Ben Mann 预测 2028 年有 50% 概率出现超级智能，由指数级扩展驱动。 Ben Mann predicts 50% chance of superintelligence by 2028, driven by exponential scaling.
超级智能出现前必须解决 AI 安全；Anthropic 优先对齐。 AI safety must be solved before superintelligence; Anthropic prioritizes alignment.
扩展定律并未放缓；模型发布加速推动进步。 Scaling laws are not slowing; progress is accelerating with faster model releases.
核心观点 · Key points
到 2028 年出现超级智能的概率为 50%，基于指数级进步和缩放定律。 50th percentile chance of superintelligence by 2028, based on exponential progress and scaling laws.
在 Anthropic，AI 安全是最高优先级；对齐必须在超级智能到来之前解决。 AI safety is the top priority at Anthropic; alignment must be solved before superintelligence arrives.
缩放定律持续成立；随着模型发布频率加快，进步正在加速。 Scaling laws continue to hold; progress is accelerating with more frequent model releases.
经济图灵测试：当 AI 被雇佣从事某项工作而不知其为机器时，即通过测试。 Economic Turing test: AI passes when hired for a job without knowing it's a machine.
宪法 AI 通过自我批评和重写来使模型与人类价值观对齐。 Constitutional AI uses self-critique and rewrite to align models with human values.
AI 进步的最大瓶颈是算力（芯片和数据中心），其次是算法和数据。 Biggest bottleneck for AI progress is compute (chips and data centers), plus algorithms and data.
反共识 · Contrarian takes
AI 进步并未放缓；由于后训练改进和更快的发布，实际上正在加速。 AI progress is not slowing; it's accelerating due to post-training improvements and faster releases.
安全研究与能力是凸性的；Claude 的个性直接源于对齐工作。 Safety research is convex with capability; Claude's personality directly results from alignment work.
AI 带来的生存风险为 0-10%，但从事安全工作的边际影响极高。 Existential risk from AI is 0-10%, but marginal impact of working on safety is extremely high.
即使是像 Ben 这样的 AI 研究人员也无法避免被取代；所有工作都将受到影响。 Even AI researchers like Ben are not immune to job replacement; all jobs will be affected.
在后奇点世界，由于物质极大丰富，资本主义可能面目全非。 In a post-singularity world, capitalism may look completely different due to abundance.
公开模型失败案例能建立与政策制定者的信任，这与典型的企业公关不同。 Publishing model failures builds trust with policymakers, unlike typical corporate PR.
本期章节 · Chapters（共 34）
开场与嘉宾介绍Introduction and Guest
超级智能时间线Timeline for Superintelligence
离开 OpenAI 的原因Reason for Leaving OpenAI
对齐风险Risk of Misalignment
与 Meta 的人才争夺Recruiting Battle with Meta
经济影响与未来工作Economic Impact and Future of Work
欢迎与介绍Welcome and Introduction
签约奖金与顶尖人才价值Signing bonuses and value of top talent
扩展定律与感知平台期Scaling laws and perceived plateaus
定义 AGI 与变革性 AIDefining AGI and transformative AI
AI 对就业与失业的影响AI's impact on jobs and unemployment
奇点后的未来Future after singularity
当前 AI 对就业的影响Current AI impact on jobs
面向未来的职业建议Advice for future-proofing careers
使用 AI 模型的实用技巧Practical tips for using AI models
为 AI 未来教育孩子Teaching kids for an AI future
Ben 为何离开 OpenAI 创立 AnthropicWhy Ben left OpenAI to start Anthropic
安全与进展的张力Safety vs. Progress Tension
人格与安全的关联Personality and Safety Connection
宪法 AI 如何运作How Constitutional AI Works
安全为何是 Ben Mann 的核心Why safety is core to Ben Mann
回应质疑与安全行动Response to Skepticism and Safety Actions
下行风险与超级智能Downside Risk and Superintelligence
软件与具身 AI 的风险Risks from Software and Embodied AI
超级智能时间线预测Timeline Prediction for Superintelligence
定义超级智能Defining Superintelligence
AI 的经济影响Economic impact of AI
递归自我改进与对齐Recursive Self-Improvement and Alignment
模型智能的最大瓶颈Biggest Bottleneck for Model Intelligence