ARTICLE · 1315247
AI科技热点日报 | 2026年11月05日
综合整理自 The Verge、TechCrunch、WIRED、MIT Technology Review、IEEE Spectrum、Reuters Technology、VentureBeat、钛媒体、新华网科技、36氪、虎嗅网、网易科技、雷锋网、量子位、IT之家、财联社等国内外主流科技媒体当日报道，聚焦人工智能领域核心动态，附资深媒体人综合解读。
📌 今日摘要
11月5日的国内头条是月之暗面 Kimi K3.1 Pro 正式在 Hugging Face 开放权重，并在 Agent Arena 智能体排行榜上以 +5.1% 的净改进率拿下开放权重模型第一；同一天，腾讯混元 Hy4 正式版在开源榜单上登顶，国内大模型的开源竞赛进入"万亿参数 + 智能体能力"的白热化阶段。国际方面，Anthropic 披露 Claude 参与解析原核生物 Vault 穹窿颗粒三维结构，把原本三天的纯化流程压缩到 90 分钟；OpenAI 则公开了由新一代网络安全模型驱动的自动化漏洞流水线，单次扫描跑出 948 个 Linux 内核本地提权利用程序。
产业与政策面同样热闹：Meta 收购 AI 编程初创公司 Conductor，谷歌因"禁用 AI 就断掉搜索"的 K-12 政策遭遇强烈反弹并给 DuckDuckGo 送出市场，Uber 在财报会上承认 AI 智能体下单量几乎可以忽略不计。算力与资本侧，美国中期选举后的政治转向让德州数据中心建设前景生变，中国大模型厂商的英文卡荒被摆上台面，存算一体与触觉智能则继续吸金。
The domestic headline of November 5 is Moonshot AI officially open-sourcing Kimi K3.1 Pro on Hugging Face, where the model took the top spot among open-weight entries on the Agent Arena leaderboard with a +5.1% net improvement rate; on the same day Tencent's Hunyuan Hy4 release claimed the crown among open models, pushing China's open-model race into a white-hot phase defined by trillion-parameter scale and agent capability. Internationally, Anthropic disclosed that Claude helped resolve the three-dimensional structure of the prokaryotic Vault particle, compressing a three-day purification workflow down to 90 minutes, while OpenAI published an automated vulnerability pipeline driven by a next-generation cyber model that produced 948 Linux kernel local privilege-escalation exploits in a single run.
Industry and policy were equally busy: Meta acquired AI coding startup Conductor, Google's K-12 policy of cutting off Search for schools that ban AI drew fierce backlash and handed market share to DuckDuckGo, and Uber admitted on its earnings call that AI-agent bookings are almost negligible. On compute and capital, the post-midterm political turn in the US casts doubt over Texas data center buildout, the scarcity of Nvidia cards for Chinese model builders came into focus, and compute-in-memory plus tactile intelligence kept attracting capital.
一、开源模型与前沿发布
事件概要
月之暗面（Moonshot AI）于11月5日在 Hugging Face 平台正式开放 Kimi K3.1 Pro 模型权重，兑现其10月24日发布时的承诺。在 Arena 运营的 Agent Arena 智能体排行榜上，Kimi K3.1 Pro（Max）位列开放权重模型第一、全部参评模型第13位，并在工作、编程、聊天三个类别中均居开放模型首位。该排名基于全球用户超过 1.75 万次真实长周期智能体任务会话，净改进率为 +5.1%，较上一代 K3（Max）的 +0.9% 提升 4.2 个百分点。其中衡量 Shell 命令执行可靠性的 Bash Reliability 从 -4.5% 升至 +7.9%，居开放模型第一；Praise vs Complaint 从 +1.3% 升至 +7.7%；Steerability 达 +6.5%。该模型采用混合专家架构，总参数 2.8 万亿、单次推理激活约 1040 亿参数，支持约 100 万 Token 上下文；在 Terminal-Bench 4.0 上得分 65.7%，高于 GPT-6 Astra 的 58.2%。需要提醒的是，权重"可下载"并不等于"可本地运行"：BF16 精度权重文件约 5.6TB，即便 INT4 量化也仍需约 1.4TB。
On November 5, Moonshot AI officially opened the model weights of Kimi K3.1 Pro on Hugging Face, honoring the commitment made at its October 24 launch. On Arena's Agent Arena leaderboard for agents, Kimi K3.1 Pro (Max) ranked first among open-weight models and 13th among all entrants, taking the top open-model spot in the work, coding and chat categories. The ranking draws on more than 17,500 real long-horizon agent task sessions from users worldwide, with a net improvement rate of +5.1%, up 4.2 percentage points from the previous K3 (Max) at +0.9%. Bash Reliability, which measures shell command execution reliability, climbed from -4.5% to +7.9%, first among open models; Praise vs Complaint rose from +1.3% to +7.7%; and Steerability reached +6.5%. The model uses a mixture-of-experts architecture with 2.8 trillion total parameters, about 104 billion activated per inference and roughly a 1-million-token context window, scoring 65.7% on Terminal-Bench 4.0 versus 58.2% for GPT-6 Astra. It is worth noting that downloadable weights are not the same as locally runnable weights: BF16 files run to about 5.6TB, and even INT4 quantization still needs roughly 1.4TB.
同日腾讯混元 Hy4 正式版持续发酵。据网易《巴伦周刊》11月5日报道，Hy4 首次进入最强开源模型行列，相比 8 月的 Hy4 preview 扩大了预训练与后训练规模，并首次支持电脑操作，可结合命令行和图形界面执行任务，还针对 Blender、CAD 等专业软件和跨应用长程任务做了专项训练；混元与腾讯金融科技 FiT 组织 213 位专家（其中 153 位来自外部金融机构）对 579 项真实金融任务开展盲测，Hy4 综合表现位居对比模型第一。产品端 Hy4 已进入 WorkBuddy、CodeBuddy、元宝、企业微信、ima 和 TokenHub，API 定价与 preview 保持一致：每百万 Token 输入 6 元、输出 18 元，缓存命中 0.3 元，至 2026 年 12 月 31 日在 TokenHub 按刊例价五折计费，Hy4 flash 也将于近期推出。
The same day, Tencent's Hunyuan Hy4 release continued to make waves. According to Barron's China coverage on November 5, Hy4 entered the strongest open-model ranks for the first time: compared with the Hy4 preview released in August, it expanded both pre-training and post-training scale, added computer use for the first time so that it can operate via command line and graphical interfaces, and received dedicated training for professional software such as Blender and CAD plus long-horizon cross-application tasks. Hunyuan and Tencent's FinTech unit ran blind tests of 579 real financial tasks with 213 experts, 153 of them from external financial institutions, and Hy4 ranked first overall among the compared models. On the product side, Hy4 is already live in WorkBuddy, CodeBuddy, Yuanbao, WeCom, ima and TokenHub, with API pricing unchanged from preview at RMB 6 per million input tokens, RMB 18 per million output tokens and RMB 0.3 for cache hits, at half the list price on TokenHub through December 31, 2026, with Hy4 flash due shortly.
海外方面，Meta 确认收购 AI 编程初创公司 Conductor。公司成立于 2024 年，核心产品去年夏天才上线，允许用户向并行工作区部署多个智能体来搭建软件项目，平台调用的模型来自 Anthropic、OpenAI 等实验室。创始人兼 CEO 查理·霍尔茨表示 Conductor 将并入 Meta，并预告明年初为客户带来全新产品，他期待与 Meta 首席 AI 官亚历山大·王以及超级智能实验室产品负责人纳特·弗里德曼共事；Meta 未披露交易细节。霍尔茨称核心功能在明年年初前照常运转，但云服务将于 11 月 20 日关停，随后开始删除用户云端数据。
On the international side, Meta confirmed the acquisition of AI coding startup Conductor. Founded in 2024, the company's core product only launched last summer and lets users deploy multiple agents into parallel workspaces to build software projects, drawing on models from labs such as Anthropic and OpenAI. Founder and CEO Charlie Holtz said Conductor will join Meta and teased a brand-new product for customers in early next year, saying he looks forward to working with Meta's chief AI officer Alexandr Wang and Nat Friedman, who leads product for the Superintelligence Labs; Meta did not disclose deal terms. Holtz said core functionality will keep operating until early next year, but the cloud service will shut down on November 20, after which user cloud data starts to be deleted.
来源 / Sources
IT之家：月之暗面 2.8 万亿参数 Kimi K3.1 Pro 开放权重，登顶 Agent Arena 开源模型榜 凤凰网科技：月之暗面开源 Kimi K3.1 Pro，登顶 Agent Arena 开源模型榜 网易《巴伦周刊》：混元 Hy4 开源 英伟达中国缺卡 TechCrunch: Meta acquires AI coding startup Conductor 网易科技：Meta 收购人工智能编码初创公司 Conductor
二、AI 进入实验室：科研与物理世界
事件概要
Anthropic 11月5日宣布，其生物学研究团队借助 Claude 首次解析出原核生物相关的 Vault（穹窿颗粒）三维结构。研究发现，一种来自噬菌体的蛋白质 L05 能组装成完整的穹窿颗粒，由 96 个相同蛋白质组成，总分子质量约 9.5 兆道尔顿（MDa），内部容积比人类细胞中的同类结构大约 25%；L05 长 54.5 纳米、宽 47.5 纳米、结构分辨率达 3.2 埃。团队在约 150 亿个预测蛋白序列中搜索主要穹窿蛋白的同源物，在 29 个细菌门和 4 个古菌门中找到候选，选取 44 个蛋白进行实验，其中 7 个在大肠杆菌中表达后组装成穹窿样颗粒，冷冻电镜确认 4 个候选物的完整闭合结构。Claude 参与了蛋白搜索、DNA 序列设计、纯化流程优化与冷冻电镜数据处理：传统穹窿纯化约需 3 天，改用尺寸排阻色谱后缩短至约 90 分钟，首个完整结构模型在 34 小时内生成，而按常规流程需数周，整个项目历时不到 4 个月。报告也坦承 AI 的局限：一次 Claude 建议的添加剂干扰了 DNA 递送，导致研究延误两周；在冷冻电镜数据处理中 Claude 跳过标准步骤生成了不准确的结构图。
Anthropic announced on November 5 that its biology research team, working with Claude, resolved for the first time the three-dimensional structure of the prokaryotic Vault particle. A phage-derived protein called L05 assembles into a complete vault composed of 96 identical proteins with a total molecular mass of roughly 9.5 megadaltons and an internal volume about 25% larger than the human counterpart; L05 measures 54.5 by 47.5 nanometers and the structure was resolved at 3.2 angstroms. The team searched about 1.5 billion predicted protein sequences for homologs of the major vault protein, found candidates across 29 bacterial phyla and 4 archaeal phyla, tested 44 proteins, saw 7 expressed in E. coli assemble into vault-like particles, and confirmed 4 candidates as fully closed structures by cryo-EM. Claude contributed to protein search, DNA sequence design, purification optimization and cryo-EM data processing: traditional vault purification takes about three days, while a size-exclusion chromatography scheme cut it to roughly 90 minutes, and the first complete structural model came in 34 hours against the several weeks conventional workflows would need, with the whole project spanning less than four months. The disclosure also candidly notes AI's limits — one additive Claude suggested interfered with DNA delivery and cost the team two weeks, and during cryo-EM processing Claude skipped a standard step and produced an inaccurate structure map.
国内方面，百度在 11 月 5 日凌晨被披露的 ApolloWorld 世界模型引发关注。百度智能驾驶事业群组首席架构师王亮提出"2+3"世界模型定义与评价标准，以因果性、交互性为两项基石门槛，以物理准确性、持久性、实时性为三项增益属性。ApolloWorld 把车辆与行人建模为有各自目标、能感知周围变化并据此决定走停的 AI 智能体，同步生成相机画面与激光雷达点云，并在云端与车端 L4 端到端 AI Driver 基础模型闭环推演——驾驶模型打方向或踩刹车后，周围车流会实时给出下一步反应，形成"AI 教 AI 开车"的训练场，被称为业内首个同时打通全员独立 Agent 交互博弈、相机与点云全模态同步生成、城市级完整订单闭环的 L4 实战训练场。
On the domestic side, Baidu's ApolloWorld world model drew attention in the early hours of November 5. Wang Liang, chief architect of Baidu's Intelligent Driving Group, proposed a "2+3" definition and evaluation standard for world models, with causality and interactivity as two threshold attributes and physical accuracy, persistence and real-time performance as three additional properties. ApolloWorld models vehicles and pedestrians as AI agents with their own goals that perceive their surroundings and decide whether to move or stop, generating camera imagery and lidar point clouds in sync and running a closed loop in the cloud against the vehicle-side L4 end-to-end AI Driver foundation model: when the driving model steers or brakes, surrounding traffic reacts in real time, creating a training ground where "AI teaches AI to drive." It is described as the first L4 practice ground to combine fully independent agent interaction, synchronized camera-and-point-cloud multimodal generation, and a city-level full order loop.
来源 / Sources
Anthropic Research: Investigating vault particles with Claude IT之家：Anthropic 绘制出全球首个包含 96 个蛋白质副本的 3D 细菌库 量子位：百度无人车搞出"黑客帝国"！世界模型实时推演全局 新浪财经：世界模型热潮里，百度重新亮出无人驾驶底牌
三、安全攻防与企业 AI 治理
事件概要
OpenAI 当地时间 11 月 4 日发布安全研究报告，公开其利用一款即将推出的网络安全模型构建的自动化漏洞发现、验证、利用与修复流程。在 2026 年 10 月的一次测试中，该系统从 2308 个候选漏洞中动态复现 1724 个，并生成 948 个可使普通本地用户获取 root 最高权限的 Linux 内核提权（LPE）利用程序，整个过程消耗约 121 万美元 API 额度，平均每个利用程序成本约 1279 美元。智能体还使用了 10 种全新的漏洞利用技术，已由第三方内核漏洞专家核验确认为新方法；系统另生成 291 个信息泄露漏洞的概念验证程序，并为发现的问题生成候选补丁交人工审核，OpenAI 已按 Linux 内核社区建议优先提交数百份涉及严重漏洞的报告与补丁。按漏洞赏金标准估算，948 个利用程序价值区间为 1890 万至 6630 万美元，是链路成本的 16 至 55 倍。
OpenAI published a security research report on the local time of November 4 describing an automated vulnerability discovery, validation, exploitation and repair pipeline built on an upcoming cyber model. In a test run during October 2026, the system dynamically reproduced 1,724 vulnerabilities out of 2,308 candidates and generated 948 Linux kernel local privilege-escalation exploits that let an unprivileged local user reach root, consuming about 1.21 million US dollars of API credits, or roughly 1,279 dollars per exploit on average. The agent also deployed 10 exploitation techniques that third-party kernel experts verified as genuinely new, produced proof-of-concept exploits for 291 information-disclosure bugs, and generated candidate patches for human review; OpenAI has prioritized submitting several hundred reports and patches covering serious flaws, following the kernel community's guidance. Valued at typical bug-bounty rates, the 948 exploits would be worth between 18.9 million and 66.3 million US dollars, or 16 to 55 times the cost of the pipeline.
在国内企业侧，字节跳动收紧内部 AI 工具使用权限的消息在 11 月 5 日持续发酵。36氪援引新浪科技报道称，自本月起字节跳动将对全公司所有 L4 级别的代码仓库和飞书文档实施外部 AI 访问封禁，员工自行购买的 Claude、Codex 等个人账号将无法继续访问相关内部资源。知情人士向北京商报表示，员工使用外部 AI 工具的个人账号处理市场上公开的工作信息，或者用于个人生活等非工作场景，是不受限制的。此前字节跳动已于 2025 年 6 月起分批禁用 Cursor、Windsurf 等第三方 AI 编程工具，转而以旗下编程助手 Trae 作为替代；技术副总裁洪定坤在内网回应时指出，L3 及以上级别的核心内容一旦泄露后果极其严重，同时强调不应通过封禁其他选择来强迫推广内部产品。
On the domestic corporate side, news of ByteDance tightening internal AI tool permissions kept building on November 5. 36Kr, citing Sina Tech, reported that from this month ByteDance will block external AI access to all L4-level code repositories and Feishu documents company-wide, meaning employees' personally purchased Claude and Codex accounts can no longer reach those internal resources. A source told Beijing Business Today that using personal accounts for publicly available work information or for non-work personal scenarios remains unrestricted. ByteDance had already banned third-party AI coding tools such as Cursor and Windsurf in stages starting June 2025, substituting its own Trae assistant instead; technical vice president Hong Dingkun said internally that leakage at L3 and above would be catastrophic, while also stressing that blocking alternatives should not be used to force promotion of internal products.
而在开源生态侧，TechCrunch 11 月 5 日撰文指出开源 AI 正进入爆发期，并追问 Hugging Face 到底是谁在真正获利。随着 Kimi K3.1 Pro 等模型把权重开放到 Hugging Face，平台上的模型、数据集与商业化服务之间的边界再次成为开发者关心的话题：免费开源的社区入口与付费的企业托管、推理与算力服务之间的转化效率，正在决定这类基础设施平台能否长期跑通商业模型。
Meanwhile, TechCrunch argued on November 5 that open-source AI is booming and asked who actually profits as Hugging Face expands. With models such as Kimi K3.1 Pro releasing weights onto the hub, the boundary between the free open-source community layer and the paid enterprise hosting, inference and compute services has become a developer concern again — and the conversion efficiency between them will determine whether this kind of infrastructure platform can sustain a working business model.
来源 / Sources
财联社：OpenAI 自动化漏洞链路发现 1724 个 Linux 内核漏洞 界面新闻：OpenAI 利用新一代网络安全模型发现并验证 948 个 Linux 内核提权漏洞 36氪快讯：字节新规限制员工使用外部 AI？知情人士回应 钛媒体【钛晨报】：字节限制外部 AI 工具访问内部资料？知情人士回应 TechCrunch: Open source AI is booming. Hugging Face explains who profits.
四、平台、教育与入口之争
事件概要
谷歌 11 月 3 日宣布从 2027 年起调整 K-12 教育用户的人工智能功能管理政策，Gemini、Gemini Notebook、Gemini in Classroom 与 Google Search 将对学校管理的账户默认关闭，需管理员主动开启；一旦学校选择全面禁用 AI，学生也将无法继续使用 Google 搜索。谷歌教育业务副总裁 Chris Phillips 对《纽约时报》表示"搜索与 AI 已经深度融合"。这一决定迅速引发反弹：参议员 Bernie Sanders 在 X 上指责谷歌搞"敲诈"，称"如果你不让学生用 AI，我们就切断你的搜索访问"；参议员 Richard Blumenthal 则表示谷歌更关心把 Gemini 塞进学校，而不是让学生学得更好；DuckDuckGo 创始人兼 CEO 加布里埃尔·温伯格随即公开向美国各学校推广把 DuckDuckGo 设为默认搜索引擎，并愿意提供迁移与部署协助，其首席公共事务官卡米尔·巴兹巴兹称强制学生接受 AI 功能是典型的市场支配行为。谷歌则回应称新政策是为了给学校更多控制权，Gemini 等功能未来默认关闭、可逐项启用。
Google announced on November 3 that from 2027 it will change how AI features are managed for K-12 education users: Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be off by default for school-managed accounts and must be enabled by administrators, and a school that disables AI entirely will also lose Google Search. Google education VP Chris Phillips told The New York Times that "search and AI are deeply integrated." The move drew swift backlash: Senator Bernie Sanders accused Google of "extortion" on X, saying "if you stop your students from using AI, we will cut off your access to Google Search," while Senator Richard Blumenthal said Google is more concerned with forcing Gemini into schools than enabling students to succeed. DuckDuckGo founder and CEO Gabriel Weinberg immediately began pitching the company as default search engine to schools across the US, offering migration and deployment help, with chief public affairs officer Camille Bazbaz calling mandatory AI adoption a textbook case of market dominance. Google responded that the policy is about giving schools more control, with Gemini and related features defaulting to off and available item by item.
消费级入口方面，Nothing 首席执行官 Carl Pei 宣布，公司首款 AI 原生产品将于 12 月发货。他强调这不是"做一个应用"：当下的关键上下文往往来不及抵达 AI 模型——咖啡桌上冒出的好点子、一句能解开整周困局的话，等你掏出手机记录时那个瞬间已经过去。外界普遍解读为一款环境录音类无屏设备，这也是 Nothing 提出 AI 原生设备与 AI 操作系统战略一年多以来的首个落地产品。
On the consumer entry point, Nothing CEO Carl Pei announced that the company's first AI-native product will ship in December. He stressed that this is not "building an app": the most important context often fails to reach an AI model in time — a good idea surfacing on the coffee table, a sentence that unlocks a week's confusion — and by the time you pull out your phone to record it, the moment is gone. The device is widely interpreted as a screen-free, environment-recording product and is the first deliverable from Nothing's AI-native device and operating system strategy announced more than a year ago.
来源 / Sources
TechCrunch: Google faces backlash for plan that would turn off Search for schools that restrict AI TechCrunch: DuckDuckGo pitches itself to schools amid backlash over Google's AI policy 网易科技：DuckDuckGo 借 Google AI 争议进军校园市场 TechCrunch: Nothing CEO says company's mysterious AI-first device is dropping next month 腾讯新闻：Nothing 无屏 AI 设备定档十二月
五、算力、基建与智能体的真实渗透率
事件概要
英伟达在中国市场的供货紧张被摆上台面。据网易《巴伦周刊》11月5日报道，中国 AI 与云计算公司合计订购约 600 万颗英伟达 RTX Pro 5500，单价超过 1.1 万美元，总价值约 700 亿美元，但未来四个月预计只能收到数万颗，约占订单总量的 1%；首批数千颗预计今年底前交付，明年 2 月左右再交付数万颗。供应紧张的原因是英伟达把更多产能分配给 Rubin 等下一代产品并已警告客户产能受限，内部测算显示随着英伟达扩大产量，中国超大规模数据中心运营商到 2027 年底可能收到约 70 万颗 RTX Pro 5500。对国内模型厂商而言，芯片到货节奏直接决定算力扩张速度。
Nvidia's supply crunch in China came into focus. According to Barron's China on November 5, Chinese AI and cloud companies have ordered roughly 6 million Nvidia RTX Pro 5500 chips at more than 11,000 dollars each — about 70 billion US dollars in total — but are expected to receive only tens of thousands in the next four months, on the order of 1% of the order. The first several thousand units should arrive before the end of the year, with another tens of thousands around February. The tightness stems from Nvidia allocating more capacity to next-generation products such as Rubin and warning customers of constrained supply; internal estimates suggest that as Nvidia scales output, Chinese hyperscale data center operators may receive around 700,000 RTX Pro 5500 chips by the end of 2027. For domestic model builders, chip delivery cadence directly sets the pace of compute expansion.
算力基建层面，政治变量成为新变量。MIT Technology Review 报道，德州是美国增长最快的数据中心市场、也是科技巨头最大规模 AI 建设的重心，但风向已经转变：民主党人 Talarico 击败现任参议员 Paxton 拿下参议院席位，Hinojosa 击败州长 Abbott 拿下州长职位，民主党更将首次在 24 年内控制德州众议院，两人在竞选中都把数据中心作为核心议题；早在 9 月 21 日，Abbott 就已指示关键州监管机构暂停数据中心许可，州长办公室一度停止回复开发者的邮件。中文市场同步反应，摩根士丹利在 11 月 5 日指出，中期选举显著增加了数据中心延期风险，并列出美股电力板块的受损标的与相对赢家清单。
On the infrastructure side, politics has become a new variable. MIT Technology Review reported that Texas is the fastest-growing data center market in the US and the center of Big Tech's largest AI buildouts, but the winds have shifted: Democrat James Talarico won the US Senate seat against Ken Paxton, Gina Hinojosa won the governor's race against incumbent Greg Abbott, and Democrats will take control of the Texas House for the first time in 24 years, with both candidates making data centers a central campaign issue. Back on September 21, Abbott had already directed a key state regulator to halt data center permits, and the governor's office at one point stopped replying to developers' emails. The Chinese market moved in step: on November 5, Morgan Stanley noted that the midterms meaningfully raise the risk of data center delays and published a list of losers and relative winners among US power stocks.
而在最贴近用户的落地环节，Uber 在 11 月 4 日的第三季度财报电话会上给出了一组泼冷水的数据：万圣节当周——即 10 月最繁忙的一周——全平台通过 AI 智能体下单的打车订单只有 4 单；通过 Uber Eats 由 AI 智能体完成的订单约 500 单，而当周 Uber 与 Uber Eats 总计处理约 3.2 亿笔行程与订单。CFO Balaji Krishnamurthy 向分析师披露了这一规模，CEO Dara Khosrowshahi 则表示"我们看到的是非常非常早期、非常有限的渗透"，并补充希望能有更多 traction。与此同时，Waymo 确认将于 2028 年 1 月重新调整与 Uber 的合作关系，为出行入口的争夺再加一层变数。
And on the most user-facing front, Uber delivered a sobering set of numbers on its third-quarter earnings call on November 4: during the week of Halloween — its busiest week in October — only four Uber rides were ordered through an AI agent, while about 500 Uber Eats orders were placed by agents against roughly 320 million total trips and orders across Uber and Uber Eats that week. CFO Balaji Krishnamurthy disclosed the figures to analysts, and CEO Dara Khosrowshahi said "it is very, very early and very limited traction that we are seeing so far," adding that the company hopes for more traction. Meanwhile, Waymo confirmed it is shaking up its Uber relationship in January 2028, adding another variable to the contest for the mobility entry point.
来源 / Sources
MIT Technology Review: Texas was the new epicenter of the AI data center buildout. That might be kaput. 腾讯新闻：美股电力板块面临重新洗牌！中期选举增加数据中心延期风险 网易《巴伦周刊》：混元 Hy4 开源 英伟达中国缺卡 TechCrunch: Almost no one uses AI agents to book their Uber, company says 腾讯新闻：Uber 披露 AI 智能体下单量极低 TechCrunch: Waymo confirms it's shaking up its Uber relationship in January 2028
六、资本、政策与新变量
事件概要
OpenAI 提前布局 2028 年美国大选议题的动作被曝光。据 TechCrunch 11 月 5 日报道（中文媒体同日跟进），OpenAI 已在美国中期选举后启动政策造势，全球事务负责人 Chris Lehane 表示中期选举只是"2028 年那部电影的预告片"，并预计 AI 将成为 2028 年总统选举的核心议题；其主张包括推动"AI 限速"、主张美国在全球 AI 标准制定中扮演更大角色，以及把数据中心改名为"AI 算力园区"以缓解社区抵触。背景是 AI 政治行动委员会（AI PAC）在大选周期中的巨额投入，TechCrunch 同日复盘称这些团体花费数千万美元影响国会选举结果，此前的统计显示与 OpenAI 和 Anthropic 关联的政治网络投入已达 5570 万美元。
OpenAI's early move to shape the 2028 US presidential election agenda came to light. According to TechCrunch on November 5 — followed the same day by Chinese media — OpenAI has begun a post-midterm policy push, with global affairs chief Chris Lehane calling the midterms "a bit of a movie trailer for what you're going to see in 2028" and expecting AI to be a defining issue in that election. His agenda includes pushing "AI speed limits," arguing for a larger US role in setting global AI standards, and renaming data centers "AI compute campuses" to ease community resistance. The backdrop is massive spending by AI political action committees: a companion TechCrunch review on the same day examined how these groups spent tens of millions shaping congressional outcomes, with earlier tallies showing political networks linked to OpenAI and Anthropic had committed some 55.7 million dollars.
国内资本侧，具身智能与预测引擎成为新热点。雷峰网 11 月 5 日独家披露，触觉智能企业"灵感机器人"完成数千万元天使轮融资，投资方包括力合资本、合肥创新投等，治臻咨询担任财务顾问；创始人兼 CEO 张雅展毕业于港科大机器人专业，曾是腾讯资深 AI 研究员，主导过大规模数据管线及 AI 模型研发。公司 2019 年即研发出厚度 5mm 的超薄视触觉传感器，力场追踪分辨率达每平方米 10 万点以上，其 Lux 系列视触觉传感器已在机械臂、夹爪、灵巧手等领域实现规模化落地。另据雷峰网同日独家，AI 预测引擎公司 Prophetlab 获数千万元天使轮融资，由常春藤资本等机构投资；该公司成立不到四个月，早期项目用大模型预测世界杯赛果并成功命中决赛结果与四强队伍，部分模型准确率约 75%，其 Prophet Engine 已登顶字节 Seed 发布的预测榜单 FutureX 总榜，并已获得数百万商业化订单。
On the domestic capital side, embodied intelligence and prediction engines were the hot spots. Leiphone exclusively reported on November 5 that tactile-intelligence startup Linggan Robotics completed a tens-of-millions-yuan angel round from Force Capital and Hefei Innovation Investment, with Zhenzhen Consult as financial adviser; founder and CEO Zhang Yazhan graduated from HKUST in robotics, was a senior AI researcher at Tencent and led large-scale data pipelines and model development. The team developed a 5mm-thin vision-based tactile sensor as early as 2019 with force-field tracking resolution above 100,000 points per square meter, and its Lux series sensors have reached scaled deployment in robot arms, grippers and dexterous hands. Leiphone also reported the same day that AI prediction-engine startup Prophetlab raised a tens-of-millions-yuan angel round led by Ivy Capital; founded less than four months ago, its early project used large models to predict World Cup results and correctly called the final and the semifinalists, with some models around 75% accurate, and its Prophet Engine topped ByteDance Seed's FutureX leaderboard with millions in commercial orders.
技术路线端，虎嗅 11 月 5 日的产业观察显示，资金正密集涌入"存算一体"芯片：微纳核芯在两个月内完成两轮合计超 20 亿元融资、半年累计 30 亿元，其 SRAM 存算一体架构相比传统冯·诺伊曼架构可提升 4—6 倍算力密度、5—10 倍能效比；行业数据称 2025 年全球存算一体市场规模 48.7 亿美元、同比增超 60%，预计 2030 年达 312.5 亿美元。此外，华为宣布昇腾 0 Day 支持 Kimi K3.1 Pro 推理部署，国产算力对万亿级开源模型的适配速度继续跟进。
On the technology roadmap, Huxiu's November 5 industry analysis showed capital flooding into compute-in-memory chips: Micro-nano Kernel completed two rounds worth over 2 billion yuan in two months and 3 billion yuan in half a year, with its SRAM compute-in-memory architecture delivering 4–6x higher compute density and 5–10x better energy efficiency than conventional von Neumann designs; industry data put the global compute-in-memory market at 4.87 billion dollars in 2025, up more than 60% year on year, and projected at 31.25 billion dollars by 2030. In addition, Huawei announced day-zero Ascend support for Kimi K3.1 Pro inference deployment, keeping domestic compute closely aligned with trillion-parameter open models.
来源 / Sources
TechCrunch: OpenAI is already trying to shape the 2028 election TechCrunch: AI PACs spent millions trying to shape Congress — here's how they fared 网易科技：OpenAI 提前布局 2028 美国大选议题 力推"AI 新政" 雷锋网：腾讯前高级研究员获数千万投资，锚定物理 AI 接触智能 雷锋网：预测引擎 Prophetlab 获数千万种子轮，常春藤资本等机构联合投资 虎嗅：资金密集涌入"存算一体"芯片 IT之家：华为宣布昇腾 0 Day 支持 Kimi K3.1 Pro 推理部署
行业数据速览
媒体人解读
从今天的信息密度看，行业正在同时经历两条曲线的交叉：一条是模型能力与开源供给的曲线，Kimi K3.1 Pro 开放权重、混元 Hy4 登顶开源榜、华为昇腾 Day-Zero 适配，这些动作拼在一起说明国产开源模型已经从"追平闭源"走到"定义 Agent 时代的开源基线"；另一条是价值兑现的曲线，OpenAI 用 121 万美元跑出 948 个内核提权漏洞、Anthropic 把三天的纯化流程压到 90 分钟、百度用世界模型把无人驾驶的试错成本搬进云端——真正值钱的不是榜单，而是把某类工作的单位成本压到不可忽略的量级。
但今天同样刺眼的数据是"渗透率的真相"：Uber 最忙的一周里，AI 智能体只完成了 4 笔打车订单。这提醒所有做 Agent 产品的团队，Agent 的瓶颈从来不在模型能力，而在真实世界的对接成本——身份、权限、支付、责任归属，任何一环没打通，模型再强也只能停留在演示里。反过来看，字节跳动一刀切封禁外部 AI 工具访问 L4 代码仓库，也说明企业侧的第一性矛盾仍是数据安全；同一个公司今天既开源大模型又收紧内部工具权限，这不是矛盾，而是当下所有平台公司的真实状态。
值得留意的还有两条政策暗线。其一，美国中期选举之后，数据中心第一次成为选举议题本身，德州放缓许可只是开始，选票变化会沿着"电力—审批—并网—社区关系"一路传导到算力成本表上；其二，OpenAI 把自己定位成"AI 议题的立法参与者"，从 AI 限速到算力园区改名到 PAC 支出，这套组合拳意味着 2028 年之前，AI 的监管博弈将大量发生在政治募资与舆论场，而不是立法机构内部。对企业技术决策者来说，这不是远方的叙事，而是未来两年选型、选址、合规议程的真实变量。
最后给一句实在的判断：如果只看今天的头条，会以为中国在赢模型、OpenAI 在赢安全、硅谷在赢入口；但把时间轴拉到季度尺度，真正决定胜负的仍是三件笨事——芯片能不能按时到（今天中国厂商只有 1%）、企业数据能不能安全地用起来（今天字节一刀切封禁）、用户是否真的愿意把决策权交给 Agent（今天 4 单）。把这三件事的进展表盯住，比任何一份榜单都更接近答案。
Judging by the density of today's news, the industry is crossing two curves at once. One is the curve of model capability and open supply: Kimi K3.1 Pro opening its weights, Hunyuan Hy4 topping the open-model rankings, and Huawei's day-zero Ascend adaptation together show that Chinese open models have moved from "closing the gap with closed models" to "setting the open baseline for the agent era." The other is the curve of value realization: OpenAI producing 948 kernel privilege-escalation exploits for 1.21 million dollars of API spend, Anthropic compressing a three-day purification workflow into 90 minutes, and Baidu moving the trial-and-error cost of autonomous driving into the cloud with a world model. What creates value is not leaderboard position but pushing the unit cost of a class of work down to a level that cannot be ignored.
But today's most telling numbers are about the truth of adoption: during Uber's busiest week, AI agents accounted for four ride orders. This is a reminder for every team building agent products that the bottleneck has never been model capability but the cost of connecting to the real world — identity, permissions, payment, accountability. If any one link is missing, even the strongest model stays in the demo phase. Conversely, ByteDance's blanket ban on external AI tools accessing L4 code repositories shows that the first-order contradiction on the enterprise side remains data security. That one company today both open-sources a large model and locks down internal tooling is not a contradiction but the true state of every platform company right now.
Two policy undercurrents deserve attention. First, after the US midterms, data centers have become an election issue in their own right for the first time; the Texas permit slowdown is only the beginning, and shifts in votes will propagate through power, permitting, grid interconnection and community relations straight into the compute cost sheet. Second, OpenAI is positioning itself as a participant in writing the AI agenda, and the combination of AI speed limits, the "compute campus" relabeling and PAC spending means that until 2028, much of the AI regulatory contest will be fought in political fundraising and the public arena rather than inside legislatures. For corporate technology decision-makers, that is not distant narrative but a real variable shaping sourcing, siting and compliance agendas over the next two years.
One practical judgment to close: reading only today's headlines, one might think China is winning on models, OpenAI on security, and Silicon Valley on distribution. Stretch the timeline to a quarter, though, and what decides the race is still three unglamorous things — whether chips arrive on time (Chinese buyers got 1% today), whether enterprises can safely use their data (ByteDance locked the door today), and whether users genuinely hand decision-making to agents (four orders today). Tracking progress on those three is closer to the answer than any leaderboard.
参考来源
英文来源
Moonshot AI opens Kimi K3.1 Pro weights, topping the Agent Arena open-model board (IT之家/凤凰网转载) Meta acquires AI coding startup Conductor — TechCrunch Investigating vault particles with Claude — Anthropic Research Open source AI is booming. Hugging Face explains who profits. — TechCrunch Google faces backlash for plan that would turn off Search for schools that restrict AI — TechCrunch DuckDuckGo pitches itself to schools amid backlash over Google's AI policy — TechCrunch Nothing CEO says company's mysterious AI-first device is dropping next month — TechCrunch Texas was the new epicenter of the AI data center buildout. That might be kaput. — MIT Technology Review Almost no one uses AI agents to book their Uber, company says — TechCrunch Waymo confirms it's shaking up its Uber relationship in January 2028 — TechCrunch OpenAI is already trying to shape the 2028 election — TechCrunch AI PACs spent millions trying to shape Congress — TechCrunch
中文来源
IT之家：月之暗面 2.8 万亿参数 Kimi K3.1 Pro 开放权重，登顶 Agent Arena 开源模型榜 网易《巴伦周刊》：混元 Hy4 开源 英伟达中国缺卡 网易科技：Meta 收购人工智能编码初创公司 Conductor IT之家：Anthropic 绘制出全球首个包含 96 个蛋白质副本的 3D 细菌库 量子位：百度无人车搞出"黑客帝国"！世界模型实时推演全局 新浪财经：世界模型热潮里，百度重新亮出无人驾驶底牌 财联社：OpenAI 自动化漏洞链路发现 1724 个 Linux 内核漏洞 界面新闻：OpenAI 利用新一代网络安全模型发现并验证 948 个 Linux 内核提权漏洞 36氪快讯：字节新规限制员工使用外部 AI？知情人士回应 钛媒体【钛晨报】：字节限制外部 AI 工具访问内部资料？知情人士回应 网易科技：DuckDuckGo 借 Google AI 争议进军校园市场 腾讯新闻：Nothing 无屏 AI 设备定档十二月 腾讯新闻：美股电力板块面临重新洗牌！中期选举增加数据中心延期风险 腾讯新闻：Uber 披露 AI 智能体下单量极低 网易科技：OpenAI 提前布局 2028 美国大选议题 力推"AI 新政" 雷锋网：腾讯前高级研究员获数千万投资，锚定物理 AI 接触智能 雷锋网：预测引擎 Prophetlab 获数千万种子轮，常春藤资本等机构联合投资 虎嗅：资金密集涌入"存算一体"芯片 IT之家：华为宣布昇腾 0 Day 支持 Kimi K3.1 Pro 推理部署
本文内容综合整理自上述公开信息来源，媒体人解读仅代表个人观点，仅供参考。
发布日期：2026年11月05日
