Anthropic CEO Dario Amodei 回归节目，讨论新模型 Claude 3.7、与中国的 AI 竞赛，以及他对 AI 未来的希望与担忧，此时 AI 安全的钟摆正摆向加速主义。
Anthropic CEO Dario Amodei returns to discuss the new Claude 3.7 model, the AI race against China, and his hopes and fears for AI's future, as the pendulum swings from safety to accelerationism.
要点 · TL;DR
AI 模型可能在 3-6 个月内达到被滥用于生物武器等的中等风险。 AI models may reach medium risk for misuse like bioweapons in 3-6 months.
美国必须保持对专制国家的 AI 领先地位，以防止军事威胁。 US must maintain AI lead over autocracies to prevent military threats.
无论公众关注或政治趋势如何，AI 风险都在指数级增长。 AI risk increases exponentially regardless of public attention or politics.
核心观点 · Key points
AI 模型在 3-6 个月内将接近被滥用于生物武器等的中等风险。 AI models are approaching medium risk for misuse like bioweapons within 3-6 months.
美国必须在 AI 上保持对专制国家的领先，以防止军事威胁。 The US must maintain an AI lead over autocracies to prevent military threats.
无论公众关注或政治趋势如何，AI 风险都在指数级增长。 AI risk is increasing exponentially regardless of public attention or political trends.
围绕 AI 安全的两极分化阻碍了细致的政策和风险缓解。 Polarization around AI safety hinders nuanced policy and risk mitigation.
人们应准备好在两年内 AI 将改变编程和许多工作岗位。 People should prepare for AI to transform coding and many jobs within 2 years.
反共识 · Contrarian takes
AI 模型并非商品；它们在能力和风格上差异显著。 AI models are not commodities; they differ significantly in capabilities and style.
不先减缓中国的发展，就无法减缓美国的 AI 发展。 Slowing down US AI development is impossible without slowing China first.
AI 最大的威胁不是当前的聊天机器人，而是未来的自主系统。 The biggest threat from AI is not present-day chatbots but future autonomous systems.
AI 可以被稳健地对齐为善良，但这使得纠正错误变得困难。 AI can be robustly aligned to be good, but that makes it hard to correct errors.
即使 AGI 临近，人们也不应剧烈改变生活。 People should not drastically change their lives even if AGI is near.
本期章节 · Chapters（共 26）
开场：周末AI活动与嘉宾介绍Opening: Weekend AI Events and Guest Introduction
Claude 3.7概述与介绍Introduction and Claude 3.7 Overview
模型命名与即将推出的基础模型Model Naming and Upcoming Base Models
Claude 3.7 Sonnet的安全问题Safety Concerns with Claude 3.7 Sonnet
近期中等风险的潜在影响Implications of Medium Risk in the Near Future
创新竞争与模仿Innovation Competition and Copying
AI模型的竞争与差异化Competition and differentiation among AI models
DeepSeek、对华竞争与国家安全DeepSeek, competition with China, and national security
国际竞争与安全International competition and safety
对巴黎AI行动峰会的失望Disappointment with AI Action Summit in Paris
地理脱节与AGI时间线信心Geographic disconnect and confidence in AGI timeline
两极分化阻碍理解Polarization as a barrier to understanding
国家安全的历史编码与当前AI政治Historical coding of national security and current AI politics
证据呈现不当阻碍风险缓解Poorly presented evidence as enemy of risk mitigation
当前琐碎应用影响AI安全认知Perception of AI safety due to current trivial applications
公众对AI风险与收益的看法Public perception of AI risks and benefits
今年AI的积极面Upside of AI this year
控制模型的难度Difficulty of controlling models
社会对AI的反应Societal response to AI
对IT岗位的影响与职业建议Impact on IT jobs and career advice
Hat GPT环节介绍Hat GPT segment introduction
Perplexity的浏览器与风投基金Perplexity's browser and venture fund
Meta裁员后高管奖金Meta's executive bonuses after layoffs
苹果在英移除加密功能Apple removes encryption feature in UK
时尚美容中不切实际的AI图像Unrealistic AI images in fashion and beauty