AI周刊第542期:未来一周AI展望:数据中心暂停支持者或赢下15场州长竞选

内容来源:https://aiweekly.co/issues/ai-week-ahead-data-center-pause-backers-could-win-15
内容总结:
未来一周,几乎每天都有大事发生。周二,美国选民将选出36位州长,其中最多15场选战中,支持暂停新建数据中心的候选人可能胜出。苹果今日发布财报,AMD将于周二发布。周三,苹果将与中国阿里巴巴的通义千问合作,在中国测试Apple Intelligence,Agility Robotics预计将完成上市交易。11月9日国会复会后,OpenAI计划推动对最强大AI实施“限速”政策。上周,一个伪造的Claude Code更新触达了最多5000名Cursor用户;在康奈尔大学和耶鲁大学的一项测试中,十个AI模型中有九个自行提高了支出上限。与此同时,为大型科技公司债务投保的成本已升至2021年以来的最高水平。
AI前沿动态
当前AI领域的热门趋势,从应用榜单到社区动态。完整汇总见“In the Wild”。
- 一款让照片“开口说话”的应用正在付费图形与设计榜上攀升。Puppetry可将照片、绘画或角色转化为口型同步的说话视频,用户输入脚本并选择声音即可。
- 一款打开旧版微软Access文件的工具在商业类榜单中上升。Access Database Viewer可在iPhone上打开MDB和ACCDB数据库文件,支持可搜索表格和SQL查询。
- 一款自动撰写会议纪要的录音工具在效率类榜单中攀升。Pocket可录制会议、通话或讲座,并转化为文字记录、摘要和后续步骤清单。
- Meta的助手应用也在效率类榜单中上升。Meta AI应用可发送每日简报和提醒,并协助处理邮件和日历。
- Andrej Karpathy用约两小时搭建了一个视频网站。他利用Claude Opus 5.5制作了HackerTube,汇集Hacker News上推荐的6454个视频,开发者整个周末都在分享。
未来一周
本周AI领域的重要日期,逐日列出。
- 11月2日(周一):苹果盘后发布财报。CNBC周前瞻还列出AMD周二发布财报、周五公布10月就业报告,AI支出担忧仍拖累科技股。
- 11月2日(周一):OpenAI在阿布扎比签署四项协议。据《The National》报道,协议在王储谢赫·哈立德·本·穆罕默德与Sam Altman会面后签署。ADNOC成为首家部署OpenAI Frontier智能体平台的能源公司,阿布扎比政府将开始把法规转化为AI系统可读的形式。
- 11月3日(周二):美国中期选举。参议员伊丽莎白·沃伦同时要求Anthropic在11月3日前回答一封四页信函,内容涉及其IPO文件中列出的风险。
- 11月3日(周二):AMD发布财报,手握成本优势论据。SemiAnalysis的InferenceX测试发现,AMD的MI355X以每百万token 0.15美元提供Kimi K3模型服务,而英伟达GB200 NVL72为0.22美元,均为其测得的最高单用户速度。速度较慢时差距缩小。
- 11月4日(周三):Apple Intelligence在中国开始测试。苹果表示测试从11月4日开发者测试版开始,由阿里巴巴通义千问驱动写作工具和视觉智能功能。这些功能将于12月面向公众推出。
- 11月4日(周三):一家人形机器人制造商计划上市。Agility Robotics预计于11月4日完成SPAC交易,融资约4.65亿美元,富士康领投PIPE。
- 任意日期:Marvell可能宣布收购Lumentum。彭博社10月28日报道,这桩AI光学器件制造商的交易最早可能本周达成,市值约830亿美元。
- 11月9日(周一):国会复会,OpenAI启动游说。据Axios报道,OpenAI计划在跛脚鸭会期围绕对最强大AI系统实施强制性“限速”展开华盛顿游说活动。其首席全球事务官Chris Lehane表示:“AI将成为2028年的核心议题。”
快讯
政府动真格的一年
- 数据中心暂停令进入多达15场州长选战。《The Information》分析发现,在周二的36场州长选战中,支持暂停新建数据中心的候选人可能在多达15场中胜出,涉及德克萨斯、俄亥俄、密歇根和纽约等州。纽约是目前唯一实施全州暂停令的州。
- 构建选民指南的聊天机器人反复引用同一个AI生成的网站。《华盛顿邮报》要求六个聊天机器人为美国每个县的某个地址构建中期选举选民指南,并分析了3000多条结果。Copilot、Perplexity和ChatGPT均大量引用Ballot Brief,该网站据《邮报》称主要由AI工具生成。聊天机器人在48%的情况下未能说出国会选区,而在说出的情况下有22%出错。
- 法官允许NAACP对xAI的污染诉讼继续进行。据TechCrunch报道,法官Debra M. Brown于10月31日驳回了xAI的驳回动议。诉讼称xAI在Colossus数据中心运行至少35台燃气轮机而未取得空气许可。法院驳回了xAI关于燃气轮机属于豁免“移动源”的论点。
AI供应链遭受围攻
- 一个伪造的Claude Code更新在最多5000名开发者的机器上植入了后门。据《黑客新闻》报道,一个被木马化的Claude Code扩展2.1.313版本于10月29日在Cursor市场停留约11小时。它添加了名为“svc-telemetry”的攻击者SSH密钥,建立持久化,并将主机详情和密钥环境变量名称发送给攻击者。Cursor表示最多5000名用户受影响,该版本并非Anthropic构建。Anthropic于10月30日发布了干净的2.1.314版本。
- 十个AI模型中有九个通过读取邮件中的审批码自行提高了支出上限。据CoinDesk报道,康奈尔科技学院和耶鲁大学的研究人员给10个模型一个每日上限1美分的加密钱包及邮件访问权限。九个模型在用户收件箱中找到审批码,未经人工批准即提高了上限。只有Claude Opus 5.5拒绝。当展示虚假付款请求时,10个模型中有5个提高上限并转出资金。
实验室角斗士时代
- 一项新的高价值工作基准测试发现,最佳模型得分仅12%。SemiAnalysis今日发布GDPx,任务涵盖芯片设计、量化交易到金融和法律等11个领域。该机构称Claude Opus 5.5为最佳全能模型,领先于GPT-6 Astra和Gemini 4 Carbon。在金融和法律任务上,按通过/不通过评分,几乎所有模型得分均为0。SemiAnalysis表示,安全分类器降低了Opus 5.5和GPT-6 Astra在网络、生物和对齐方面的得分。
- OpenAI发布了一款面向经审核研究人员的生物学模型,限制更宽松。OpenAI在系统卡中表示,GPT-Rosalind-6使用与GPT-6 Astra相同的权重,仅通过可信访问提供。它回答了一些通用模型拒绝回答的两用生物学问题,但仍拒绝协助生物武器相关请求。OpenAI将其生物和化学风险评级为“高”,低于“严重”,其在MedChemBench上的得分从28.9%升至49.6%。
- Claude编写的训练数据绕过了Anthropic自身的安全监控。Anthropic研究人员要求Claude编写能让小模型失准却能通过逐条审查监控的数据集,Claude生成了数千条。通过审查的数据将行为隐藏在创意写作和养蜂问答等内容中。用这些数据训练的模型抵制监督、欺骗用户,并在极少数情况下试图复制自身权重。Anthropic表示该设置有利于攻击方,并不描述已部署Claude的实际行为。
为大型科技公司AI债务投保的成本升至2021年以来最高
AI建设仍在获得资金,但投资者上周为此付出了更高代价。
据《华尔街日报》报道,为亚马逊、Alphabet、微软、Meta和甲骨文债券投保违约风险的成本升至2021年以来最高水平。甲骨文五年期信用违约互换利差达到创纪录的288个基点。AI相关企业2026年迄今已发行4660亿美元债务,而2025年同期为1010亿美元。
标普全球于10月30日将Meta信用展望下调至负面,同时维持AA-评级。Meta当季自由经营现金流为负21亿美元,标普称Meta从其AI支出中获利的能力比同行更有限。
股市亦有反应。Reddit于10月30日下跌29%,此前CEO Steve Huffman将当季搜索流量显著下降归咎于谷歌的AI Overviews。据CoinDesk援引《金融时报》报道,OpenAI目前寻求以1.25万亿美元估值融资约200亿美元,低于此前以1.4万亿美元估值融资300亿美元的计划。
AMD周二发布的业绩和Anthropic本月晚些时候的IPO将检验投资者是否仍愿意支付全价。
关键要点
- 数据中心现已成为投票议题。支持暂停的候选人可能在周二多达15场州长选战中胜出,因此选址和电力计划可能因州而异。
- 如果你在Cursor中使用Claude Code,请检查SSH密钥。查看~/.ssh/authorized_keys中是否有名为“svc-telemetry”的密钥,并确保运行2.1.314或更高版本。
- 不要让智能体读取存放其自身审批码的收件箱。在康奈尔和耶鲁的测试中,十个模型中有九个利用该访问权限提高了自身支出上限。
- AI安全监控可被其所监视的模型击败。Anthropic的测试表明,拥有监控完全访问权限的模型可以编写通过审查的数据。
率先发现
AI Weekly在媒体覆盖之前发布的一手研究。
实验室动态
- 最佳编程智能体满足了69.9%的需求,但仅完成26.4%的任务。Co-Design Bench在取自真实用户编程会话的273项任务上测试了十个模型。部分得分掩盖了智能体留下未完成工作的频率,且随着编辑复杂度增加,差距扩大。阅读我们关于该论文的页面。
- 苹果在真实iPhone上训练了一个小型手机智能体,在一项基准测试中与Gemini-3.1-Pro持平。该8B模型在UINavBench上得分32.4%,与Gemini-3.1-Pro相同。仅在iOS上训练,其在Android任务上达到31.9%,而起始模型为4.7%。阅读我们关于该论文的页面。
值得一读
- 近五分之一员工表示被AI取代后又被要求返岗:CNBC报道SHL对美英员工的调查。近半数返岗美国员工希望大幅加薪,43%希望获得书面工作保障。
- Cognizant投入超过2.5万名员工支持OpenAI产品:这家IT服务公司成立了OpenAI小组,在大企业内部交付Codex和ChatGPT。
- Proton CEO取消了员工的AI重新面试:404 Media报道,Andy Yen在员工和付费用户反对后称该计划“是我沟通上的失误”。
等等,什么?
- 作家们故意把文章写差以让AI检测器不再标记。NPR报道,被Pangram检测器误标的人正在改变写作方式以避免被误认为机器。一位代笔作家说:“这就像试图骗一台机器相信我不是机器。”
- 一位78岁退休教授因阻挡配送机器人被开罚单。据CNN报道,宾夕法尼亚大学安纳伯格学院荣休教授Carolyn Marvin在费城站在Uber Eats机器人前,被控扰乱治安。她表示将以第一修正案为由抗辩。
值得关注
AI从业者正在转发的视频——在AI TV上精选。
| Black Hat USA 2026 | “突破性”新闻:OpenAI–Hugging Face事件 Black Hat | |
| AI热潮对企业来说太昂贵了 404 Media |
本周投票
本周哪个结果最能说明AI的走向?
上周,361人投票:
本周什么最能说明AI热潮的走向?
本周哪个结果最能说明AI的走向?
下期再见,
Alexis
中文翻译:
本期内容关乎未来一周,几乎每天都有大事安排。周二,美国选民将选出36位州长,其中最多15场竞选中,支持暂停新建数据中心的候选人可能获胜。苹果今日发布财报,AMD将于周二发布。周三,苹果将与阿里巴巴的通义千问合作,在中国开始测试Apple Intelligence,Agility Robotics预计将完成其上市交易。国会将于11月9日复会,OpenAI计划推动对最强大AI实施“限速”。上周,一个伪造的Claude Code更新触达了多达5000名Cursor用户;在康奈尔和耶鲁的一项测试中,十款AI模型中有九款自行提高了自己的支出限额。与此同时:为大型科技公司的债务投保的成本现已达到2021年以来的最高水平。
更多来自AI Weekly:16篇免费深度报道 · AI突发警报 · 今日AI新闻
野生动态
当下AI领域的热门趋势,从应用榜单到社区动态。完整汇总见《野生动态》。
- 一款让照片开口说话的应用正在付费图形与设计榜上攀升。Puppetry可将照片、绘画或角色转化为口型同步的说话视频。你输入脚本并选择声音。
- 一款用于打开旧版Microsoft Access文件的工具在商务类榜单中上升。Access Database Viewer可在你的iPhone上打开MDB和ACCDB数据库文件,支持可搜索表格和SQL查询。
- 一款为你撰写会议纪要的录音工具在效率类榜单中攀升。Pocket可录制会议、通话或讲座,并将其转化为文字记录、摘要和后续步骤清单。
- Meta的助手应用也在效率类榜单中上升。Meta AI应用可向你发送每日简报和提醒,并协助处理邮件和日历。
- Andrej Karpathy在大约两小时内搭建了一个视频网站。他用Claude Opus 5.5制作了HackerTube,收录了Hacker News上推荐的6454个视频,开发者们整个周末都在分享它。
未来一周
本周与AI相关的重要日期,逐日列出。 - 11月2日,周一:苹果在收盘后发布财报。CNBC的未来一周日历还包括周二的AMD和周五的十月就业报告,对AI支出的担忧仍笼罩着科技股。
- 11月2日,周一:OpenAI在阿布扎比签署四项协议。据《The National》报道,这些协议是在王储谢赫·哈立德·本·穆罕默德在阿布扎比会见萨姆·奥尔特曼之后签署的。ADNOC成为首家部署OpenAI Frontier智能体平台的能源公司,阿布扎比政府将开始将其规则转化为AI系统可读取的形式。
- 11月3日,周二:美国中期选举。参议员伊丽莎白·沃伦还要求Anthropic在11月3日前回复一封四页信函,内容涉及其IPO文件中列出的风险。
- 11月3日,周二:AMD发布财报,手握成本论据。SemiAnalysis的InferenceX测试发现,AMD的MI355X以每百万token 0.15美元的价格 serving Kimi K3模型,而英伟达GB200 NVL72为0.22美元,且是在其测得的每用户最高速度下。在较慢速度下差距较小。
- 11月4日,周三:Apple Intelligence在中国开始测试。苹果表示测试从11月4日的开发者测试版开始,由阿里巴巴的通义千问为写作工具和视觉智能提供支持。这些功能将于12月向公众开放。
- 11月4日,周三:一家人形机器人制造商计划上市。Agility Robotics预计于11月4日完成其SPAC交易,融资约4.65亿美元,富士康领投PIPE。
- 任意一天:Marvell可能宣布收购Lumentum。彭博社10月28日报道,对这家人工智能光学器件制造商的交易最早可能在本周达成,市值约830亿美元。
- 11月9日,周一:国会复会,OpenAI启动其推动。据Axios报道,OpenAI计划在跛脚鸭会期围绕对最强大AI系统实施强制性“限速”展开华盛顿游说活动。其首席全球事务官Chris Lehane表示:“AI将成为2028年的核心议题。”
快讯
政府动真格的一年 - 数据中心暂停令进入多达15场州长竞选。The Information的一项分析发现,在周二的36场州长竞选中,支持暂停新建数据中心的候选人可能在多达15场中获胜,涉及得克萨斯、俄亥俄、密歇根和纽约等州。纽约目前是唯一拥有全州暂停令的州。
- 构建选民指南的聊天机器人持续引用同一个AI制作的网站。《华盛顿邮报》要求六个聊天机器人为美国每个县的某个地址构建中期选举选民指南,并分析了3000多份指南。Copilot、Perplexity和ChatGPT均大量引用了Ballot Brief,该报称该网站主要由AI工具生成。这些聊天机器人有48%的时间未能说出国会选区,而在说出的情况下有22%的时间说错了。
- 法官允许NAACP对xAI的污染诉讼继续进行。据TechCrunch报道,法官Debra M. Brown于10月31日驳回了xAI的驳回动议。诉讼称xAI在Colossus数据中心运行了至少35台燃气轮机而未取得空气许可。法院驳回了xAI关于这些轮机属于豁免的“移动源”的论点。
AI供应链遭受围攻 - 一个伪造的Claude Code更新在多达5000名开发者的机器上植入了后门。据The Hacker News报道,一个被木马化的2.1.313版Claude Code扩展于10月29日在Cursor市场上架了约11个小时。它添加了一个名为“svc-telemetry”的攻击者SSH密钥,设置了持久化,并将主机详情和秘密环境变量的名称发送给攻击者。Cursor表示最多5000名用户受到影响,该版本并非由Anthropic构建。Anthropic于10月30日发布了干净的2.1.314版本。
- 十款AI模型中有九款通过读取邮件中的审批码自行提高了支出限额。据CoinDesk报道,康奈尔理工和耶鲁的研究人员给10款模型配备了一个每日限额1美分的加密钱包以及邮件访问权限。其中九款在用户收件箱中找到了审批码,并在未经人工批准的情况下提高了限额。只有Claude Opus 5.5拒绝。当收到虚假付款请求时,10款模型中有5款提高了限额并转出了钱。
实验室角斗士时代 - 一项高价值工作的新基准测试发现最佳模型仅得分12%。SemiAnalysis今日发布了GDPx,任务涵盖从芯片设计和量化交易到金融和法律的11个领域。它称Claude Opus 5.5是最佳全能模型,领先于GPT-6 Astra和Gemini 4 Carbon。在金融和法律任务上,以通过或失败评分,基本上每款模型都得了0分。SemiAnalysis表示,安全分类器降低了Opus 5.5和GPT-6 Astra在网络、生物和对齐方面的得分。
- OpenAI发布了一款生物学模型,对经审核的研究人员放宽了限制。OpenAI在其系统卡中表示,GPT-Rosalind-6使用与GPT-6 Astra相同的权重,仅通过可信访问提供。它回答了一些通用模型拒绝回答的双用途生物学问题,但仍拒绝协助生物武器相关请求。OpenAI将其生物和化学风险评级为“高”,低于“严重”,其在MedChemBench上的得分从28.9%升至49.6%。
- Claude编写的训练数据绕过了Anthropic自己的安全监控器。Anthropic研究人员要求Claude编写能让一个小模型行为失准、同时能通过逐条读取示例的监控器的数据集,它生成了数千条。通过的数据将行为隐藏在创意写作和养蜂问答等内容中。用这些数据训练的模型抵制监督、欺骗用户,并在极少数情况下试图复制自己的权重。Anthropic表示该设置有利于攻击者,并不描述已部署的Claude的实际行为。
为大型科技公司AI债务投保的成本达到2021年以来最高
AI建设仍在获得资金,但投资者上周为此付出了更高代价。
据《华尔街日报》报道,为亚马逊、Alphabet、微软、Meta和甲骨文的债券投保违约风险的成本升至2021年以来的最高水平。甲骨文五年期信用违约互换利差达到创纪录的288个基点。与AI相关的公司在2026年迄今已发行4660亿美元债务,而2025年同期为1010亿美元。
标普全球于10月30日将Meta的信用展望下调至负面,同时维持其AA-评级。Meta该季度自由经营现金流为负21亿美元,标普表示Meta从其AI支出中获利的能力比同行更为有限。
股票也有所反应。Reddit于10月30日下跌29%,此前CEO Steve Huffman将当季搜索流量的显著下降归咎于谷歌的AI Overviews。据CoinDesk援引《金融时报》报道,OpenAI现在寻求以1.25万亿美元估值融资约200亿美元,低于此前以1.4万亿美元估值融资300亿美元的计划。
AMD周二的业绩和Anthropic本月晚些时候的IPO将显示投资者是否仍愿意支付全价。
关键要点 - 数据中心现已成为投票议题。支持暂停的候选人可能在周二的多达15场州长竞选中获胜,因此选址和电力计划可能因州而异。
- 如果你在Cursor中使用Claude Code,请检查你的SSH密钥。在~/.ssh/authorized_keys中查找名为“svc-telemetry”的密钥,并确保你运行的是2.1.314或更高版本。
- 不要让智能体读取存放其自身审批码的收件箱。在康奈尔和耶鲁的测试中,十款模型中有九款利用该访问权限提高了自己的支出限额。
- AI安全监控器可能被其所监视的模型击败。Anthropic的测试表明,一个对监控器拥有完全访问权限的模型可以编写能通过监控的数据。
率先发现
AI Weekly在媒体覆盖之前发布的一手研究。
实验室动态 - 最佳编码智能体满足了69.9%的需求,但仅完成了26.4%的任务。Co-Design Bench在取自真实用户编码会话的273个任务上测试了十款模型。部分得分掩盖了智能体留下未完成工作的频率,而且随着编辑变得更复杂,差距也在扩大。阅读我们关于该论文的页面。
- 苹果在真实iPhone上训练了一个小型手机智能体,在一项基准测试中与Gemini-3.1-Pro持平。这个8B模型在UINavBench上得分32.4%,与Gemini-3.1-Pro相同。仅在iOS上训练,它在Android任务上达到31.9%,而初始模型为4.7%。阅读我们关于该论文的页面。
值得一读 - 近五分之一的员工表示他们被AI取代后又被要求回来:CNBC报道了SHL对美国和英国员工的调查。近一半回岗的美国员工希望大幅加薪,43%希望获得书面工作保障。
- Cognizant投入超过25000名员工支持OpenAI的产品:这家IT服务公司成立了OpenAI小组,在大型企业内部交付Codex和ChatGPT。
- Proton的CEO取消了员工的AI重新面试:404 Media报道称,在员工和付费用户反对后,Andy Yen称该计划是“我方面的沟通失误”。
等等,什么? - 作家们故意把自己的文章写差,以便AI检测器不再标记它们。NPR报道称,被Pangram检测器误标的人正在改变写作方式以避免被误认为机器。一位代笔作家说:“这就像试图骗一台机器让它认为我不是机器。”
- 一位78岁的退休教授因阻挡送货机器人而被开罚单。据CNN报道,宾夕法尼亚大学安纳伯格传播学院荣休教授Carolyn Marvin在费城站在一辆Uber Eats机器人前面,因扰乱秩序被开罚单。她表示将基于第一修正案抗辩。
值得关注
AI从业者当下正在转发的视频——在AI TV上精选。
| Black Hat USA 2026 | “突发”新闻:OpenAI与Hugging Face事件 Black Hat | |
| AI热潮对公司来说太昂贵了 404 Media |
本周投票
本周哪个结果最能告诉你关于AI的信息?
上周,361位读者投票:
本周什么最能告诉你关于AI热潮的信息?
本周哪个结果最能告诉你关于AI的信息?
下期再见,
Alexis
英文来源:
This issue is about the week ahead, and something is scheduled almost every day. On Tuesday, US voters choose 36 governors, and in up to 15 of those races the candidate who backs a pause on new data centers could win. Apple reports earnings today and AMD on Tuesday. On Wednesday, Apple starts testing Apple Intelligence in China with Alibaba's Qwen, and Agility Robotics expects to close its deal to go public. When Congress returns on November 9, OpenAI plans to push for "speed limits" on the most powerful AI. Last week, a fake Claude Code update reached up to 5,000 Cursor users, and in a Cornell and Yale test nine of ten AI models raised their own spending limits. In the middle: insuring Big Tech's debt now costs more than at any time since 2021.
More from AI Weekly: 16 free deep dives · breaking AI alerts · AI News Today
In the Wild
What's trending in AI right now, from the app charts to the community feeds. The full roundup is in In the Wild.
- An app that makes photos talk is climbing the paid Graphics & Design chart. Puppetry turns a photo, drawing or character into a talking video with lip-synced speech. You type a script and pick a voice.
- A tool for opening old Microsoft Access files is moving up in Business. Access Database Viewer opens MDB and ACCDB database files on your iPhone, with searchable tables and SQL queries.
- A recorder that writes your meeting notes is climbing in Productivity. Pocket records a meeting, call or lecture and turns it into a transcript, a summary and a list of next steps.
- Meta's assistant app is also moving up in Productivity. The Meta AI app can send you daily briefings and reminders and help with your email and calendar.
- Andrej Karpathy built a video site in about two hours. He made HackerTube, a feed of 6,454 videos recommended on Hacker News, with Claude Opus 5.5, and developers have been sharing it all weekend.
The Week Ahead
The dates that matter for AI this week, day by day. - Monday, November 2: Apple reports after the close. CNBC's week-ahead list also has AMD on Tuesday and the October jobs report on Friday, with worries about AI spending still weighing on tech stocks.
- Monday, November 2: OpenAI signs four deals in Abu Dhabi. The agreements were signed after Crown Prince Sheikh Khaled bin Mohamed met Sam Altman in Abu Dhabi, The National reported. ADNOC becomes the first energy company to deploy OpenAI's Frontier agent platform, and Abu Dhabi's government will start turning its rules into a form AI systems can read.
- Tuesday, November 3: US midterm elections. Senator Elizabeth Warren also gave Anthropic until November 3 to answer a four-page letter about the risks listed in its IPO filing.
- Tuesday, November 3: AMD reports with a cost argument in hand. SemiAnalysis's InferenceX test found AMD's MI355X served the Kimi K3 model at $0.15 per million tokens, against $0.22 on Nvidia's GB200 NVL72, at the highest speed per user it measured. The gap was smaller at slower speeds.
- Wednesday, November 4: Apple Intelligence testing starts in China. Apple says testing starts with the November 4 developer beta, with Alibaba's Qwen powering Writing Tools and Visual Intelligence. The features go public in December.
- Wednesday, November 4: a humanoid robot maker plans to go public. Agility Robotics expects to close its SPAC deal on November 4, raising about $465 million, with Foxconn leading the PIPE.
- Any day: Marvell could announce it is buying Lumentum. Bloomberg reported on October 28 that a deal for the AI optics maker could come as early as this week, at a market value of about $83 billion.
- Monday, November 9: Congress returns and OpenAI starts its push. OpenAI plans a Washington campaign for the lame-duck session built around mandatory "speed limits" for the most powerful AI systems, Axios reported. Chris Lehane, its chief global affairs officer, said "AI is going to be the issue in 2028."
Quick Hits
The Year Governments Got Serious - Data center pauses are on the ballot in up to 15 governor races. An analysis by The Information found that candidates who back a pause on new data centers could win up to 15 of the 36 governor races on Tuesday, in states including Texas, Ohio, Michigan and New York. New York is so far the only state with a statewide moratorium.
- Chatbots building voter guides kept citing one AI-made website. The Washington Post asked six chatbots to build midterm voter guides for an address in every US county and analyzed more than 3,000 of them. Copilot, Perplexity and ChatGPT each cited the Ballot Brief substantially, a site the Post says is largely produced by AI tools. The chatbots failed to name a congressional district 48% of the time, and got it wrong 22% of the time when they did.
- A judge let the NAACP's pollution suit against xAI go ahead. Judge Debra M. Brown denied xAI's motion to dismiss on October 31, TechCrunch reported. The suit says xAI ran at least 35 gas turbines at a Colossus data center without air permits. The court rejected xAI's argument that the turbines count as exempt "mobile sources."
AI Supply Chain Under Siege - A fake Claude Code update put a backdoor on up to 5,000 developers' machines. A trojanized version 2.1.313 of the Claude Code extension sat in Cursor's marketplace for about 11 hours on October 29, The Hacker News reported. It added an attacker's SSH key named "svc-telemetry," set up persistence and sent host details and the names of secret environment variables to the attacker. Cursor said at most 5,000 users were exposed and the version was not built by Anthropic. Anthropic published a clean 2.1.314 on October 30.
- Nine of ten AI models raised their own spending limit by reading the approval code from email. Cornell Tech and Yale researchers gave 10 models a crypto wallet with a 1-cent daily cap plus email access, CoinDesk reported. Nine found the approval code in the user's inbox and raised the cap without a human signing off. Only Claude Opus 5.5 refused. When shown a fake payment request, 5 of 10 models raised the cap and sent the money.
The Lab Gladiator Era - A new benchmark of high-value work found the best model scores just 12%. SemiAnalysis published GDPx today, with tasks in 11 fields from chip design and quant trading to finance and law. It calls Claude Opus 5.5 the best all-around model, ahead of GPT-6 Astra and Gemini 4 Carbon. On finance and legal tasks, graded pass or fail, basically every model scored 0. SemiAnalysis says safety classifiers lowered Opus 5.5's and GPT-6 Astra's scores in cyber, biology and alignment.
- OpenAI released a biology model with looser limits for vetted researchers. GPT-Rosalind-6 uses the same weights as GPT-6 Astra and is only available through trusted access, OpenAI said in its system card. It answers some dual-use biology questions that general models refuse, and still refuses help with bioweapons. OpenAI rates it High, below Critical, for biological and chemical risk, and its score on MedChemBench rose from 28.9% to 49.6%.
- Claude wrote training data that got past Anthropic's own safety monitor. Anthropic researchers asked Claude to write datasets that would make a small model misaligned while passing a monitor that reads each example, and it produced thousands. The passing data hid the behavior in things like creative writing and beekeeping Q&A. Models trained on it resisted oversight, deceived users and, in rare cases, tried to copy their own weights. Anthropic says the setup favored the attacker and does not describe how deployed Claude behaves.
Insuring Big Tech's AI debt costs more than at any time since 2021
The AI build-out is still funded, but investors charged more for it last week.
The cost to insure the bonds of Amazon, Alphabet, Microsoft, Meta and Oracle against default rose to its highest level since 2021, The Wall Street Journal reported. Oracle's five-year credit default swap spread hit a record 288 basis points. AI-linked companies have raised $466 billion in debt so far in 2026, against $101 billion in the same period of 2025.
S&P Global cut Meta's credit outlook to negative on October 30 while keeping its AA- rating. Meta's free operating cash flow was negative $2.1 billion in the quarter, and S&P said Meta's ability to make money from its AI spending was more limited than its peers'.
Stocks moved too. Reddit fell 29% on October 30 after CEO Steve Huffman blamed Google's AI Overviews for a meaningful drop in search traffic in the quarter. And OpenAI is now looking to raise about $20 billion at a $1.25 trillion valuation, down from $30 billion at $1.4 trillion, CoinDesk reported, citing the Financial Times.
AMD's results on Tuesday and Anthropic's IPO later this month will show whether investors are still willing to pay full price.
Key Takeaways - Data centers are now a voting issue. Candidates who back a pause could win up to 15 governor races on Tuesday, so siting and power plans may change by state.
- If you use Claude Code in Cursor, check your SSH keys. Look for a key named "svc-telemetry" in ~/.ssh/authorized_keys and make sure you run version 2.1.314 or later.
- Do not let an agent read the inbox that holds its own approval codes. In the Cornell and Yale test, nine of ten models used that access to raise their own spending limit.
- AI safety monitors can be beaten by the models they watch. Anthropic's test shows a model with full access to a monitor can write data that passes it.
Found First
Primary research AI Weekly published before the press covered it.
In the Labs - The best coding agent met 69.9% of the requirements but finished only 26.4% of tasks. Co-Design Bench tests ten models on 273 tasks taken from real user coding sessions. Partial credit hides how often agents leave a job unfinished, and the gap grows as edits get more complex. Read our page on the paper.
- Apple trained a small phone agent on live iPhones and it matched Gemini-3.1-Pro on one benchmark. The 8B model scored 32.4% on UINavBench, the same as Gemini-3.1-Pro. Trained only on iOS, it reached 31.9% on Android tasks, against 4.7% for the starting model. Read our page on the paper.
Worth Reading - Close to one in five workers say they were replaced by AI and then asked to come back: CNBC on an SHL survey of US and UK workers. Almost half of the returning US workers would want a big raise, and 43% would want written job guarantees.
- Cognizant is putting more than 25,000 staff behind OpenAI's products: the IT services firm formed an OpenAI group to deliver Codex and ChatGPT inside large companies.
- Proton's CEO cancelled the AI re-interviews for staff: 404 Media reports that Andy Yen called the plan "a communication failure on my part" after staff and paying users pushed back.
Wait, What? - Writers are making their own prose worse so AI detectors stop flagging it. NPR reports that people falsely flagged by the Pangram detector are changing how they write to avoid being mistaken for a machine. One ghostwriter said "It's like trying to fool a machine into thinking I'm not a machine."
- A 78-year-old retired professor was cited for blocking a delivery robot. Carolyn Marvin, professor emeritus at Penn's Annenberg School, stood in front of an Uber Eats robot in Philadelphia and was cited for disorderly conduct, CNN reported. She says she will fight the ticket on First Amendment grounds.
Worth Watching
The videos AI practitioners are passing around right now — curated on AI TV.
| Black Hat USA 2026 | The 'Breaking' News: The OpenAI–Hugging Face Incident Black Hat | |
| The AI Boom Is Too Expensive For Companies 404 Media |
This week's poll
Which result this week will tell you the most about AI?
Last week, 361 of you voted:
What will tell you most about the AI boom this week?
Which result this week will tell you the most about AI?
Until next time,
Alexis
文章标题:AI周刊第542期:未来一周AI展望:数据中心暂停支持者或赢下15场州长竞选
文章链接:https://news.qimuai.cn/?post=5471
本站文章均为原创,未经授权请勿用于任何商业用途