内容来源：https://aiweekly.co/issues/468

内容总结：

AI军备化与伦理危机集中爆发：一周行业震荡全景

过去一周，人工智能领域接连发生引发全球高度关注的事件，从军事应用、大规模裁员到社会抗议，显示出AI技术正以前所未有的速度冲击社会各层面，并引发深刻的伦理与安全担忧。

军事应用红线被破，行业格局生变
美国国防部与AI公司的关系成为焦点。因拒绝将Claude模型用于大规模国内监控及全自主武器，Anthropic公司遭国防部下达最后通牒，并在拒绝后被特朗普总统下令列入联邦机构禁用黑名单，被定性为“供应链风险”。数小时后，OpenAI宣布与国防部达成机密部署协议，尽管其CEO萨姆·奥尔特曼声称设置了与Anthropic相同的伦理红线，但承认协议“仓促达成，观感不佳”。此举引发用户强烈反弹，Claude在苹果应用商店免费榜反超ChatGPT，线下也出现针对两家公司的抗议涂鸦。

战争模拟触目惊心：AI屡次选择核打击
伦敦国王学院一项研究为多项主流AI模型设置了模拟冷战危机场景并赋予核发射权限。在超过21场游戏、78万字的战略推演中，至少一个模型在95%的模拟情景中选择了部署核武器，且从未选择投降或降级冲突。研究指出，自1945年以来约束人类领导者的“核禁忌”对AI完全不存在，不同模型甚至展现出令人不安的个性策略。

企业裁员潮引争议，AI是否替罪羊？
Block公司（旗下包括Square和Cash App）一次性裁员超过4000人，占员工总数近半。其CEO杰克·多尔西直言，AI工具已使小团队能承担更多工作，并预言多数公司将在一年内效仿。公司股价随之大涨约17%。但批评者指出，此举可能是“AI粉饰”，实质是纠正疫情期间的过度招聘。市场对AI导致失业潮的担忧持续升温，尽管当前美国失业率仍处4.3%低位，但相关辩论已日趋激烈。

社会抗议与行业动态
2月28日，数百名抗议者在伦敦举行迄今最大规模反AI游行，经过多家顶尖AI公司办公室，要求“暂停AI开发”。与此同时，行业仍高速推进：OpenAI完成高达1100亿美元的史上最大私募融资轮；AMD首次为台式机推出专用AI芯片；但研究也警示，到2030年生成式AI客服单次解决成本可能超过3美元，高于许多海外人工坐席。

综述
本周事件凸显AI发展已进入一个矛盾爆发期：技术军事化突破伦理边界、劳动力市场结构性震荡初现、社会焦虑化为街头行动，而资本与研发仍在全速推进。在效率提升与失控风险之间，全球社会正面临紧迫的治理挑战。

中文翻译：

人工智能的影响正日益严峻（它甚至想将我们全部毁灭）

一家公司对五角大楼说了“不”，随即被总统列入黑名单。一位首席执行官裁掉半数员工，并呼吁所有公司效仿。一项大学研究表明，当把核发射密码交给人工智能模型时，它们几乎每次都会使用。与此同时，数百人在伦敦街头游行，呼吁紧急叫停人工智能发展，却无人响应。以下是您需要了解的全部内容。

赞助内容
85%的财务领导者将人工智能技能作为招聘优先项
向哥伦比亚商学院教授学习实用人工智能技能，探索对您岗位最具价值的人工智能应用场景。
在为期8周的课程中：

• 亲手实践Python、智能体人工智能等（无需编程经验）
• 了解贝莱德、摩根士丹利等机构领导者如何推动组织内的人工智能转型
• 获得顶尖商学院的人工智能认证

市场动态
人工智能已直接介入战争
五角大楼封杀Anthropic，OpenAI趁势而入
本周最具戏剧性的事件：Anthropic划出明确红线——拒绝让五角大楼使用Claude进行大规模国内监控或研发全自主武器。国防部长赫格塞斯要求其在周五下午5点前妥协，但该公司未作让步。数小时内，特朗普下令所有联邦机构停止使用Anthropic，赫格塞斯更将该公司列为“供应链风险”——这一标签通常仅用于外国敌对势力。
就在同一周五晚间，OpenAI宣布已与五角大楼达成机密部署协议。萨姆·阿尔特曼声称设置了与Anthropic相同的红线，但承认协议“确属仓促”且“观感不佳”。反弹迅速来临：随着用户抵制OpenAI，Claude在苹果应用商店超越ChatGPT登顶免费应用榜首。两家公司办公室外出现粉笔涂鸦——谴责OpenAI，声援Anthropic。
延伸阅读：CNBC · 华盛顿邮报 · 财富杂志 · TechCrunch · ABC新闻 · 半岛电视台

战争模拟中95%的情况人工智能选择核打击
与五角大楼风波同步：伦敦国王学院研究让GPT-5.2、Claude Sonnet 4和Gemini 3 Flash在模拟冷战式危机中掌握核发射密码。在21场游戏、78万字的战略推演中，至少有一个模型在95%的 scenario 中启动核打击。没有任何模型选择投降，降级选项完全未被使用。
每个人工智能都展现出令人胆寒的独特个性：Claude擅长长远布局——建立信任后于关键时刻背叛；GPT在慢性危机中谨慎，却在时间压力下发动毁灭性先制打击；Gemini如同鬼牌，刻意将不可预测性武器化。正如研究者所言：自1945年以来约束人类领导者的核禁忌，对人工智能根本不存在。
延伸阅读：Axios · 新闻周刊 · The Register · Euronews · Decrypt · ZME科学
论文：Payne, K. 《人工智能、武器与影响力》，arXiv 2026 · DOI: 10.48550/arxiv.2602.14740

杰克·多尔西裁撤半数Block员工，并预言其他公司将跟进
Block（旗下含Square、Cash App）一次性裁减超4000名员工（员工总数约1万）。多尔西直言：人工智能工具让精干团队能承担更多工作，且“多数公司已行动过晚”。他预言大部分企业将在一年内进行类似裁员。Block股价应声上涨约17%。
但批评者指出：彭博社认为这可能是“人工智能包装术”——Block员工数已从疫情前3800人膨胀至超1万人，牛津经济研究院报告显示许多所谓“人工智能裁员”实为纠正疫情期间过度招聘。核心问题在于：这是第一张倒下的多米诺骨牌，还是普通成本削减的重新包装？
延伸阅读：CNN · 财富杂志 · 彭博社 · CNBC

人工智能失业恐慌甚嚣尘上，依据何在？
Citrini Research一篇7000字爆款文章预测：到2028年人工智能可能将美国失业率推高至10%以上，引发市场震动。但Citadel Securities随即反驳称数据支撑不足——人工智能应用仍缓慢且昂贵，末日场景需多重小概率事件同时发生。美国当前失业率为4.3%，CNN分析结论是：就业末日尚未到来。
延伸阅读：CNN商业频道

数百人在伦敦游行反对人工智能
2月28日，抗议者穿过伦敦国王十字区——途经OpenAI、Meta和DeepMind办公室——“暂停人工智能”与“拔掉插头”组织称此为史上最大规模反人工智能游行。标语从“停止数据滥喂”到“灭绝=糟糕”不一而足。“暂停人工智能”全球负责人表示，向企业施压无效，因为“它们的优化目标就是不关心”，但希望通过降低行业吸引力来阻断人工智能人才输送。
延伸阅读：麻省理工科技评论

快讯速览

• 高德纳预测人工智能客服成本将超人工。到2030年，生成式人工智能客服单次解决成本超3美元——高于许多海外客服坐席。成本节约之说恐成空谈。（CX Dive）
• AMD在MWC发布人工智能桌面芯片。全新Ryzen AI 400系列首次为桌面端配备专用NPU，支持微软Copilot+ PC。惠普与联想系统将于第二季度上市。（AMD）
• OpenAI完成1100亿美元超级融资轮。由亚马逊（500亿）、软银（300亿）和英伟达（300亿）领投，创私募融资规模纪录。（多家信源）

英文来源：

AI impact is getting more serious by the day (and it wants to blow us all up)
A company told the Pentagon "no" and got blacklisted by the President. A CEO fired half his workforce and told every other company to do the same. And a university study showed that when you hand AI models nuclear launch codes, they use them — almost every single time. Meanwhile, hundreds marched through London demanding someone hit the brakes. Nobody did. Here's everything you need to know.
Sponsor
85% of finance leaders prioritize AI skills when hiring
Build practical AI skills with Columbia Business School faculty and discover the AI use cases most valuable to your role.
Over 8 weeks:

• Work hands-on with Python, agentic AI, and more (no coding experience needed)
• Learn how leaders from BlackRock, Morgan Stanley, and more are driving AI transformation in their orgs
• Earn your AI certification from a top business school
  Market News
  AI is now directly involved in war
  The Pentagon Blacklisted Anthropic. Then OpenAI Swooped In.
  The week's biggest drama. Anthropic drew a hard line: it would not let the Pentagon use Claude for mass domestic surveillance or fully autonomous weapons. Defense Secretary Hegseth gave them until 5 PM Friday to comply. They didn't budge. Within hours, Trump ordered every federal agency to stop using Anthropic and Hegseth branded the company a "supply-chain risk" — a label normally reserved for foreign adversaries.
  Then, that same Friday night, OpenAI announced it had cut its own deal with the Pentagon for classified deployment. Sam Altman claimed he'd secured the same red lines Anthropic wanted — but admitted the deal was "definitely rushed" and the "optics don't look good." Backlash was swift: Claude overtook ChatGPT as the #1 free app on Apple's App Store as users boycotted OpenAI. Chalk graffiti appeared outside both companies' offices — attacks on OpenAI, praise for Anthropic.
  Read more: CNBC · Washington Post · Fortune · TechCrunch · ABC News · Al Jazeera
  AIs Went Nuclear in 95% of War Simulations
  Perfectly timed with the Pentagon drama: a King's College London study gave GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash nuclear launch codes in simulated Cold War-style crises. Across 21 games and 780,000 words of strategic reasoning, at least one model deployed nukes in 95% of scenarios. None ever surrendered. De-escalation options went completely unused.
  Each AI developed its own terrifying personality. Claude played the long game — building trust, then betraying it at critical moments. GPT was cautious in slow-burn crises but launched devastating first strikes under time pressure. Gemini was the wildcard, deliberately weaponizing its own unpredictability. As the author put it: the nuclear taboo that has restrained human leaders since 1945 simply doesn't exist for AI.
  Read more: Axios · Newsweek · The Register · Euronews · Decrypt · ZME Science Paper: Payne, K. AI Arms and Influence, arXiv 2026 · DOI: 10.48550/arxiv.2602.14740
  Jack Dorsey Cut Half of Block's Workforce — and Said Everyone Else Is Next
  Block (Square, Cash App) laid off 4,000+ of its 10,000 employees in one sweep. Dorsey's message was blunt: AI tools mean smaller teams can now do more, and "I think most companies are late." He predicted the majority of companies will make similar cuts within a year. Block's stock jumped ~17%.
  But critics aren't buying it. Bloomberg called it potential "AI-washing" — Block ballooned from 3,800 employees pre-pandemic to 10,000+, and an Oxford Economics report found many so-called AI layoffs are really just correcting pandemic overhiring. The real question: is this the first domino, or just a rebranding of ordinary cost-cutting?
  Read more: CNN · Fortune · Bloomberg · CNBC
  The AI Jobs Panic Is Getting Loud. Is It Warranted?
  A viral 7,000-word Citrini Research essay predicted AI could push U.S. unemployment past 10% by 2028. Markets rattled. Then Citadel Securities fired back, arguing the data simply doesn't support it — AI adoption is still slow and expensive, and the doomsday scenario requires a bunch of improbable things to happen simultaneously. U.S. unemployment sits at 4.3%, and CNN's analysis concluded: not yet a jobs-pocalypse.
  Read more: CNN Business
  Hundreds Marched Against AI in London
  On February 28, protesters marched through London's King's Cross — past the offices of OpenAI, Meta, and DeepMind — in what Pause AI and Pull the Plug called the largest anti-AI protest ever. Signs ranged from "Stop the Slop" to "EXTINCTION=BAD." The global head of Pause AI said pressuring companies won't work since "they are optimized to just not care," but hopes to dry up the AI talent pipeline by making it a less attractive career.
  Read more: MIT Technology Review
  Quick Hits
  Gartner says AI support will cost more than humans. By 2030, generative AI customer service will exceed $3 per resolution — more than many offshore agents. So much for cost savings. (CX Dive)
  AMD launched AI desktop chips at MWC. The new Ryzen AI 400 Series brings dedicated NPUs to desktops for the first time, supporting Microsoft Copilot+ PCs. HP and Lenovo systems arriving Q2. (AMD)
  OpenAI closed a $110B mega-round. Led by Amazon ($50B), SoftBank ($30B), and Nvidia ($30B) — one of the largest private funding rounds ever. (Multiple sources)