【聊了什么】 6月12日，美国商务部以国家安全为由，限制 Anthropic/Claude 最新模型 Mythos 5 和 Fable 5 向外国人开放。随后，Anthropic 将这两个刚刚发布、被视为公司最强能力代表的模型全线下架。一个 AI 产品更新，为什么会突然变成国家安全事件？所谓“外国人不能用”，又如何影响到整个 AI 行业？ 本期我们从技术层面解释大语言模型的“护栏”：它为什么需要拒绝回答某些问题，为什么护栏太低会带来安全风险，护栏太高又会误伤正常科研和普通用户。Anthropic 一直主张 AI 监管，但当政府真的出手，监管又立刻变成出口管制、国籍边界、法律灰色地带和政治表态问题。 我们也把讨论推到更大的层面：Anthropic 的价值观、AI 公司与军方合作、硅谷加速主义者的反应、开源模型和 AI 安全之间的矛盾，以及特朗普和桑德斯都提到过的 AI 国有化、主权基金和全民分红。AI 不只是技术产品，也正在成为国家安全、文化战争、劳动力市场和人类社会未来秩序的一部分。 【支持我们】 如果喜欢这期节目并希望支持我们将节目继续做下去： 也欢迎加入我们的会员计划： https://theamericanroulette.com/paid-membership/ 会员可以收到每周2-5封newsletter，可以加入会员社群，参加会员活动，并享受更多福利。 合作投稿邮箱：american.roulette.pod@gmail.com 【时间轴】 01:13 Claude 最新模型被限，Anthropic 全线下架 05:23 技术、政治与意识形态如何交织 05:52 什么是大模型“越狱”和“护栏” 12:37 亚马逊、安全人员与模型漏洞举报 14:01 政府与 Anthropic 各执一词的罗生门 21:47 Anthropic 为什么主动呼吁 AI 监管 23:15 出口管制：美国能否限制外国人使用模型 30:12 Anthropic 的价值观、国家安全和军方合作 41:34 禁令会不会削弱美国 AI 的领先优势 43:40 硅谷加速主义者为何为 Anthropic 被锤叫好 45:03 白宫内部权力变化与 AI 政策走向 49:07 用户数据、文化战争与开源模型争议 53:12 从公司自律到国家入股：AI 巨头国有化想象 54:17 AI 冲击劳动力市场之后，社会如何分配红利 【我们是谁】 美轮美换是一档深入探讨当今美国政治的中文播客。 本期的主播和嘉宾： 小华：媒体人 曹起曈：青椒，政治行为研究者 主播和嘉宾的言论不代表其所在机构或其雇主的观点。 【What We Talked About】 On June 12, the U.S. Commerce Department restricted access to Anthropic/Claude’s latest models, Mythos 5 and Fable 5, for foreign nationals on national security grounds. Anthropic then took both newly released models, widely seen as the company’s most capable systems, offline for all users. How did an AI product update suddenly become a national security incident? And how does a rule that “foreigners cannot use it” affect the entire AI industry? In this episode, we explain the technical side of large language model “guardrails”: why models need to refuse certain requests, why guardrails that are too weak can create safety risks, and why guardrails that are too strict can interfere with legitimate research and ordinary users. Anthropic has long advocated for AI regulation, but once the government actually intervenes, regulation immediately turns into a question of export controls, national boundaries, legal gray areas, and political signaling. We also broaden the discussion to Anthropic’s values, AI companies’ cooperation with the military, reactions from Silicon Valley accelerationists, the tension between open-source models and AI safety, and proposals raised by both Trump and Sanders around AI nationalization, sovereign wealth funds, and public dividends. AI is no longer just a technology product; it is becoming part of national security, the culture war, the labor market, and the future order of human society. 【Support Us】 If you like our show and want to support us, please consider the following: Join our membership program: https://theamericanroulette.com/paid-membership/ Support us on Patreon: www.patreon.com/americanroulette Business inquiries and fan mail: american.roulette.pod@gmail.com 【Timeline】 01:13 Claude’s latest models are restricted, and Anthropic takes them offline 05:23 How technology, politics, and ideology become entangled 05:52 What “jailbreaking” and “guardrails” mean for large language models 12:37 Amazon, security researchers, and reports of model vulnerabilities 14:01 The competing narratives from the government and Anthropic 21:47 Why Anthropic has actively called for AI regulation 23:15 Export controls: can the U.S. restrict foreigners from using AI models? 30:12 Anthropic’s values, national security, and cooperation with the military 41:34 Will the ban weaken America’s lead in AI? 43:40 Why Silicon Valley accelerationists are cheering Anthropic’s troubles 45:03 Shifting power inside the White House and the future of AI policy 49:07 User data, culture-war politics, and the open-source model debate 53:12 From corporate self-regulation to state ownership: imagining AI nationalization 54:17 After AI disrupts the labor market, how should society share the gains? 【Who We Are】 The American Roulette is a podcast dedicated to helping the Chinese-speaking community understand fast-changing U.S. politics. Our hosts and guests: 小华 (Xiao Hua): Journalist, political observer 曹起曈 (Thomas Cao): Assistant professor at the Tufts Fletcher School The views expressed by the host and guests do not represent the opinions of their employers or any affiliated institutions.
