justoneapi-all · JustOneAPI / 一人公司 最新 热点 趋势 / all
综合整理自 TechCrunch、WIRED、IEEE Spectrum、钛媒体、IT之家、PChome、同花顺等当日公开报道，聚焦 AI 资本开支与信用重估、具身智能系统战、自动驾驶第二波、Agent 架构范式、存储与推理成本、模型可信度工程。 📌 今日摘要 需要先说明一个客观背景：2026 年 11 月 1 日是星期日，因此欧美一线科技媒体的当日原创稿量明显偏低，本期国际信源以 TechCrunch、WIRED、IEEE Spectrum 的当日发布内容为主，中文信源则相对饱满——这一期的信息结构本身就是"中国供给侧信息密度高于国际媒体"的一个切面。 本期的主线是**"账本"与"账本之外"**。账本一侧：Anthropic 已向纳斯达克提交 IPO 注册文件，拟用代码"ANT"，招股文件披露近 9000 亿美元的未来算力义务，远超其 1175 亿美元负债与 1120 亿美元现金；标普全球评级把 Meta 的"AA-"长期发行人评级展望由稳定下调至负面，这是四大超大规模云厂商中首家因 AI 基础设施支出被"亮黄牌"的公司；高盛测算四巨头 2026 年合计资本开支将达 7250 亿美元，比 2025 年的 4100 亿美元增长 77%，而市场估值逻辑已从"收入增长"切换为"现金回报"，CoreWeave 单月下跌 37%、Nebius 下跌 35% 首当其冲。账本之外：Meta 与傅利叶智能在三个月内联合做出人形机器人原型 Max（身高约 1.5 米、体重约 48 公斤、搭载英伟达 Jetson Thor），把 Muse 模型从软件拽进家庭；Trucks VC 今年已收到 81 家自动驾驶初创公司的提案，比去年多出三分之二；DeepSeek Harness 团队公开 Super-agent 原型，用双层架构与工具集权限边界，把 Agent 从"单会话"改造成"常驻系统"；在硬件端，长鑫科技前三季度 LPDDR 系列收入 781.91 亿元、占营收 52.02%，而 HBM 对晶圆的乘数占用正把通用内存涨价一路推到手机与笔记本的物料清单上。 The keyword of this issue isthe ledger and what lies beyond it. First, a practical caveat: November 1, 2026 fell on a Sunday, so the volume of original reporting from the first-tier US and European technology outlets was unusually thin — which is why this issue leans on same-day TechCrunch, WIRED and IEEE Spectrum pieces alongside a denser Chinese source flow, and that imbalance is itself a small but real signal about where the industry's daily information production now sits. On the ledger: Anthropic has filed with the Nasdaq to list under the ticker "ANT," and the filing discloses close to $900 billion in future compute obligations against $117.5 billion of liabilities and $112 billion of cash; S&P Global Ratings cut the outlook on Meta's "AA-" long-term issuer rating to negative, making it the first of the four hyperscalers to be flagged over AI infrastructure spending; Goldman Sachs puts the four giants' combined 2026 capital expenditure at $725 billion, up 77% from $410 billion in 2025, while the market's valuation logic has flipped from revenue growth to cash return, with CoreWeave down 37% in a month and Nebius down 35%. Beyond the ledger: Meta and Fourier Intelligence jointly built a humanoid prototype called Max in three months — about 1.5 meters tall, roughly 48 kilograms, running an Nvidia Jetson Thor — pulling the Muse model out of software and into the home; Trucks VC has fielded pitches from 81 autonomous vehicle startups this year, two-thirds more than last year; the DeepSeek Harness team has open-sourced a Super-agent prototype that uses a two-layer architecture and tool-level permission boundaries to turn agents from single-session tools into resident systems; and on the hardware side, ChangXin Memory's LPDDR line posted ¥78.191 billion in revenue through the first three quarters, 52.02% of the total, while HBM's multiplier effect on wafer capacity is pushing memory price inflation all the way down to phone and laptop bills of materials. 一、资本面再校准：Anthropic 招股文件披露近 9000 亿美元算力义务，标普首将 Meta 展望下调至负面 事件概要 11 月 1 日，钛媒体《Edge AI Daily 早报（11 月 1 日）》与《数智周报》先后披露，Anthropic 已向美国证券交易委员会提交 IPO 注册文件，拟以代码"ANT"在纳斯达克挂牌。招股文件中最刺眼的数字不是收入，而是承诺：近 9000 亿美元的未来算力义务，对照其 1175 亿美元负债与 1120 亿美元现金，这家模型公司的资产负债表已先于利润表完成了一次极限扩张。同一份文件披露的另一组数字来自一级市场：OpenAI 最新一轮融资的年化营收信号由 700 亿美元修正至 500 亿美元，其 2030 年算力支出预期也从 1.4 万亿美元削至约 6000 亿美元——这两条数据在同一天被放在同一张表上，构成了本期最直白的产业注脚：算力承诺在缩，收入预期也在缩。 信用市场的反应更早一步。10 月 30 日，标普全球评级将 Meta 的"AA-"长期发行人评级展望由稳定下调至负面（评级本身维持），使其成为四大超大规模云厂商中首家因 AI 基建支出被亮黄灯的公司。标普给出的理由包括 Meta 的 AI 变现能力弱于同行，以及杠杆率最快在 2027 年升破 1 倍（去年约 0.1 倍，警戒线为 1.0 倍），并预计其 2026 年资本支出约 1400 亿美元、2027 年加速至 2250 亿美元以上。现金流侧的背离已经出现在报表里：Meta 三季度营收 646 亿美元、同比增长 26% 并超出预期，但单季资本开支 349.1 亿美元近乎翻倍，经营现金流 327.9 亿美元无法覆盖，自由现金流首次转负至 -21 亿美元（2012 年以来首次），财报后股价下跌约 7%；标普测算其自由现金流缺口 2026 年为 220 亿美元、2027 年扩大到 710 亿美元、2028 年仍有 390 亿美元，即 2028 年之前难以回正。10 月 22 日，Meta 五年期 CDS 利差扩大 7.6 个基点至 114.79 个基点的历史新高，股价当日下跌 7.0% 至 648.85 美元，至 10 月 30 日收于 614.99 美元。Alphabet 维持全年 1950 亿至 2050 亿美元资本开支指引，亚马逊过去 12 个月自由现金流净流出 424 亿美元、2026 年资本开支约 2200 亿美元，两家同样遭到抛售。市场层面，10 月 30 日纳斯达克综合指数下跌 2.6% 收于 26008.86 点，10 月累计下跌 5.3%，标普 500 科技板块跌 6.8%、费城半导体指数跌 15.4%，均创 2025 年 3 月以来最差单月表现。高盛策略师 Amanda Lynam 测算，Meta、微软、亚马逊、Alphabet 四家 2026 年合计资本开支将达 7250 亿美元，较 2025 年的 4100 亿美元增长 77%，估值逻辑由"收入增长"转向"现金回报"。 产业侧的乐观与悲观同期并存。亚马逊首席执行官安迪·贾西在财报电话会上表示，AWS 年化收入已达 1840 亿美元，预计 2027、2028 两年同比增速都将高于 40%，未来 5 到 10 年有望成长为万亿美元营收规模的业务。城堡投资联席首席投资官 Pablo Salame 则发出警告：支撑 AI 革命的巨大资本需求，正与态度谨慎、不愿承担此类风险的贷款机构发生冲突，"AI 浪潮虽会水涨船高，但也会让一些船沉没"。特斯拉首席执行官马斯克最新警告有 200 多家新云公司将大规模倒闭或合并，而摩根大通则表示超大规模数据中心运营商状况良好已是市场共识，头寸调整上，AI 债务市场对基础设施变得更加挑剔。硬件端，国产厂商的业绩已经先行：中兴通讯第三季度营业收入 388.13 亿元，同比增长 33.99%，归母净利润 13.07 亿元，同比增长 394.46%。而在海外，TechCrunch 在 11 月 1 日的报道中把 AI 高管的"不接地气"摆到了台面上：Dazzle AI 首席执行官 Marissa Mayer 因让 AI 工具代写女儿的学校作业描述而受到批评，OpenAI 的 Sam Altman 曾建议父母为孩子上学的路程生成个性化播客，Meta 的扎克伯格则认为通过 AI 满足社交与浪漫需求将越来越能被接受；在 Anthropic 即将上市、华尔街对高额 AI 支出正在失去胃口的当口，这篇报道追问的核心是：这些高管是否已经离普通用户太远。 On November 1, TMTPost's "Edge AI Daily" morning briefing and its weekly "Shuzhi" roundup both reported that Anthropic has filed registration documents with the US securities regulator to list on the Nasdaq under the ticker "ANT." The most striking number in the filing is not revenue but commitment: close to $900 billion in future compute obligations, set against $117.5 billion of liabilities and $112 billion of cash. The model company's balance sheet has already undergone an extreme expansion ahead of its income statement. Two more figures disclosed around the same filing come from primary market: OpenAI's annualized revenue signal for its latest round has been revised down from $70 billion to $50 billion, and its projected 2030 compute spend has been cut from $1.4 trillion to roughly $600 billion. Placed on the same table on the same day, these two data points form the most direct industrial annotation of this issue: compute commitments are shrinking, and revenue expectations are shrinking too. Credit markets moved a step earlier. On October 30, S&P Global Ratings cut the outlook on Meta's "AA-" long-term issuer rating to negative while leaving the rating itself unchanged, making it the first of the four hyperscalers to be flagged over AI infrastructure spending. The stated reasons include Meta's weaker AI monetization relative to peers and a leverage ratio that could break 1.0x as early as 2027, up from roughly 0.1x a year earlier, with S&P projecting 2026 capital expenditure of about $140 billion and acceleration beyond $225 billion in 2027. The divergence is already visible in the filings: Meta reported third-quarter revenue of $64.6 billion, up 26% and ahead of expectations, but with quarterly capital expenditure of $34.91 billion nearly doubling and operating cash flow of $32.79 billion unable to cover it, free cash flow turned negative for the first time at -$2.1 billion, the first negative print since 2012, and the stock fell about 7% after the report; S&P estimates the free cash flow gap at $22 billion in 2026, widening to $71 billion in 2027 and still $39 billion in 2028, meaning it will not return to positive before 2028. On October 22, Meta's five-year CDS spread widened 7.6 basis points to a record 114.79 basis points and the stock fell 7.0% to $648.85, closing at $614.99 on October 30. Alphabet maintained its full-year capital expenditure guidance of $195 billion to $205 billion, while Amazon saw $42.4 billion of net free cash flow outflow over the trailing twelve months and plans about $220 billion of 2026 capital expenditure; both were sold off as well. At the market level, the Nasdaq Composite fell 2.6% on October 30 to 26,008.86, ending October down 5.3%, while the S&P 500 technology sector dropped 6.8% and the Philadelphia Semiconductor Index fell 15.4% — both the worst monthly performance since March 2025. Goldman Sachs strategist Amanda Lynam estimates the four giants' combined 2026 capital expenditure will reach $725 billion, up 77% from $410 billion in 2025, with the valuation logic shifting from revenue growth toward cash return. Optimism and pessimism coexisted on the industry side. Amazon CEO Andy Jassy told the earnings call that AWS annualized revenue has reached $184 billion, that growth will exceed 40% year over year in both 2027 and 2028, and that the business could reach trillion-dollar revenue scale in five to ten years. Castle Investment co-CIO Pablo Salame warned that the enormous capital demands supporting the AI revolution are colliding with lenders who are reluctant to carry that risk: "the AI wave will lift boats, but it will also sink some, and the wave will keep surging." Tesla's CEO has warned that more than 200 new cloud companies could fold or merge at scale, while JPMorgan counters that healthy conditions at hyperscale data center operators have become a market consensus and that positioning adjustments show AI debt markets becoming more selective toward infrastructure. On the hardware side, domestic results have already moved ahead: ZTE reported third-quarter revenue of ¥38.813 billion, up 33.99% year over year, and net profit attributable to shareholders of ¥1.307 billion, up 394.46%. Overseas, TechCrunch's November 1 piece put the out-of-touch problem among AI executives on the table: Dazzle AI CEO Marissa Mayer drew criticism for using an AI tool to draft a school assignment description about her daughter; OpenAI's Sam Altman has suggested parents generate personalized podcasts for their children's drive to school; and Meta's Zuckerberg has argued that meeting social and romantic needs through AI will become increasingly acceptable. With Anthropic about to go public just as Wall Street loses its appetite for heavy AI spending, the central question this piece asks is whether these executives have drifted too far from ordinary users. 来源 / Sources TechCrunch：AI can make tech CEOs' lives easier. What about everyone else? 钛媒体：Edge AI Daily 早报（11月1日） 钛媒体：【数智周报】Anthropic 提交 IPO 注册文件，拟以"ANT"为代码在纳斯达克上市 二、具身智能系统战：Meta 与傅利叶三个月做出 Max 原型，Agility 招股书把 RaaS 成本彻底摊开 事件概要 11 月 1 日，多家中文科技媒体证实，Meta 正与中国机器人公司傅利叶（Fourier）联合研发人形机器人原型，内部代号"Max"。据 IT之家援引 The Information 的报道披露，该原型研发用时三个月，身高接近 5 英尺（约 1.5 米），体重约 106 磅（约 48.1 千克），以英伟达 Jetson Thor 芯片为核心动力，并配备独立的运动控制软件平台。Meta 首席技术官 Andrew Bosworth 在公司内部把这款人形机器人定位为旗下消费级 AI 智能体 Muse 的"物理化身"，使其能在家庭场景中执行实际任务；双方今年 9 月已启动第二代原型 Max 2 的研发，新版本将重新设计躯干结构、改进手部机械结构以提升操控精度、新增投影仪以支持视觉内容呈现，并配备额外传感器以实现上下楼梯功能。目前产品尚未设定上市时间，也没有销售原型机的计划。 同日《通信世界》刊发的一篇深度报道，把这周具身智能赛道几条看似分散的信息放进了同一张产业地图，结论是竞争重心已经从"能不能做出机器人"转向"能不能拿到入口、喂饱数据、稳定交付、算清成本"。在入口端，Meta 把 Muse 智能体与家庭机器人绑定，实质是争夺"大模型—物理入口—真实数据"之间的通道；而 Jetson Thor 出现在 Max 身上，说明具身智能的算力底座正在向边缘侧集中——机器人在本地完成感知、规划、动作控制与安全校验，云端承担训练、仿真与长程任务优化。在产品端，2026 世界机器人大会上，宇树展示了约 49 台机器人完成全自动表演的天坛演示，GD01 载人机甲定价 390 万元起（车高 3 米多、载人后约 500 千克、可变形为四足模式），As2-W 轮足机器人自重仅 20 千克，新一代机器人速度达 12.65 米/秒；宇树创始人王兴兴的判断是"具身智能 ChatGPT 时刻或需 2 到 10 年"。硬门槛则在交付侧：2026 年国内人形机器人累计市场规模约 20 余亿元、对应出货约 1.5 万台，智身科技累计完成约 1.5 万台交付、"钢镚"系列单品出货突破 1 万台，优必选今年将建成万台级产能，京东平台 2026 年至 8 月 20 日人形机器人销售已达万台、同比增长 9 倍。但圆桌嘉宾的判断很直接：机器人"大脑"的可靠性是首要矛盾，当前部分模型宣称成功率达 99%，意味着每执行 100 次任务就可能出现一次重大故障，行业需要向 99.99% 推进，且仍缺少覆盖研发、软件、新产品导入、量产与质量管控的标准化体系。 数据与成本是另外两条主线。10 月 31 日媒体报道显示，一支 50 多人的具身智能数据采集团队在 2026 年 6 月一次性租下 30 套房源，用途不是居住，而是让人形机器人在真实居家环境中练习家务，训练内容从物品抓取、放置扩展到整理房间、冲煮咖啡、清洗器皿，场景从单一操作台延伸到家庭、厨房、餐厅、超市、药店。行业数据显示，触觉数据价格比视觉数据高两倍左右——视觉等视频数据可能在小几百元每小时，触觉数据在大几百元每小时；2026 年前七个月，国内人形机器人传感器领域发生近 20 起融资，合计超过 31 亿元，仅 7 月单月就约 15.6 亿元。11 月 1 日，星衍智能披露了 Auriga 框架，把具身智能的 RSI（递归自我改进）路线争议摆上台面：该框架将 Physical RSI 分为 L1 任务级搜索（本质是"进化答案"）、L2 单次在线适应（"进化过程"）与 L3 系统级能力改进（"进化系统"）三层，选择从 L3 出发，在不针对目标任务预调优、不依赖专家演示、单次执行且不允许环境重置的 Zero-shot 协议下参与 RoboDojo 和 RoboCasa365 评测，以 GPT-6 Astra 为基础模型，在 42 项任务、6300 次测试中考察零样本操作、跨场景迁移与在线恢复能力，执行流程遵循 Preview—Execute—Verify。成本账本则来自即将上市的公司：按计划于 11 月 5 日在纳斯达克挂牌的 Agility Robotics，其 10 月披露的 S-4 与投资者日材料显示，2025 年全年确认营收 178.2 万美元，对应 25 亿美元投前估值，市销率超过 1400 倍，其中约 63.5% 来自关联方，面向非关联方的销售只有 65.01 万美元；信托账户从约 4.24 亿美元缩至约 1.64 亿美元，SPAC 股东赎回六成以上，总募资从"超 6.2 亿美元"落到约 4.65 亿美元。其 RaaS 模型为：一台 Digit 月费约 8500 美元，一年约 10.2 万美元，另加约 2.5 万美元一次性部署费，按五年生命周期累计约 50 万美元收入，对照全负担时薪约 30.5 美元的人工，机器人小时成本约为人工的 46%；但若单机单位时间产出只有人工的 1/5 到 1/15，按同等工作量计算，Digit 当前花费可能是人工的 2 到 7 倍。GXO 佐治亚州 Flowery Branch 站点 2024 年中上线，到 2025 年 11 月累计完成 10 万个周转箱搬运，任务准确率约 98%，折算日均约 196 箱。更关键的指标是日历填满率——按公开数据估算单机日历填满率约 8% 到 27%，中枢大致在 15%，而专用分拣设备可做到八成以上；Digit 5 把续航从约 4 小时降到 90 分钟，但充电时间缩到约 9 分钟，运行充电比从约 2:1 提升到约 10:1，单日有效工作时间可超过 20 小时，续航不再是主要矛盾，任务调度与自主性才是下一轮竞争焦点。Agility 锁定的超过 3 亿美元多年期订单来自单一未披露关联客户，对应约 1000 台 Digit 5，采用三年期 RaaS、服务期摊在四年内，并附带按部署逐步归属的认股权证；GXO 首席执行官也公开表示，公司尚未在人形机器人上实现投资回报。资金面上，2026 年 10 月具身智能赛道发生 33 起融资、总额 101.05 亿元，9 月为 60 起、110.28 亿元，热度有所下降，但灵巧手、触觉传感器、电子皮肤等"身体能力"环节仍在逆势吸金，10 月 29 日一家山东柔性织物触觉企业完成亿元 A 轮融资。 国际市场的资金也在同一时间涌入物理世界。TechCrunch 在 11 月 1 日披露，自动驾驶技术公司 Bedrock Robotics 完成 6 亿美元 C 轮融资，估值达到 41 亿美元，由 CapitalG 与 Emergence Capital 联合领投；同一天，马萨诸塞州制造业加速器 Re:Build 融资 21 亿美元，用于扩大硬科技产品制造产能并孵化工业初创企业，该轮由 General Catalyst 与摩根大通战略投资集团领投，Re:Build 由亚马逊前商业 CEO Jeff Wilke 联合创办，公开把自己定位为"美国的深圳"。这两笔交易放在一起看，与 Meta 三个月造出 Max 的节奏指向同一个判断：硬件本体与真实场景的价值，正在被资本重新定价。 On November 1, several Chinese technology outlets confirmed that Meta is co-developing a humanoid robot prototype with the Chinese robotics company Fourier Intelligence, internally code-named "Max." Citing The Information, IT之家 reported that the prototype took three months to build, stands close to five feet (about 1.5 meters) tall, weighs roughly 106 pounds (about 48.1 kilograms), runs on an Nvidia Jetson Thor chip, and comes with its own motion-control software platform. Meta CTO Andrew Bosworth positioned the machine internally as the "physical embodiment" of the company's consumer AI agent Muse, designed to execute practical tasks in the home; the two parties started work on a second-generation prototype, Max 2, in September, with the new version redesigning the torso, improving the hand mechanism for manipulation precision, adding a projector for visual presentation, and fitting extra sensors for stair-climbing capability. There is no launch date, and no plan to sell the prototype. A deep report published the same day by Communications World laid out what looked like several unrelated pieces of embodied-AI news on a single industrial map, and reached a conclusion: the competitive center of gravity has shifted from "can you build a robot at all" to "can you capture the entry point, feed the model data, deliver reliably, and account for cost." On the entry point, binding Muse to a home robot is really a fight over the channel connecting models, physical endpoints, and real-world data; and Jetson Thor appearing on Max shows the compute foundation of embodied AI concentrating at the edge, with perception, planning, motion control and safety checks running locally while the cloud handles training, simulation and long-horizon task optimization. On products, at the 2026 World Robot Conference, Unitree staged an automated performance of roughly 49 robots at the Temple of Heaven, showed the GD01 manned mech priced from ¥3.9 million (over three meters tall, about 500 kilograms when occupied, convertible to a quadruped mode), presented the As2-W wheeled-legged robot weighing just 20 kilograms, and showed a new generation reaching 12.65 meters per second; Unitree founder Wang Xingxing's judgment is that the "ChatGPT moment for embodied AI may take two to ten years." The hard threshold is on the delivery side: the domestic humanoid robot market reached roughly ¥2 billion in 2026, corresponding to about 15,000 units shipped; Zhishen Technology has completed around 15,000 cumulative deliveries with its "Gangbang" line breaking 10,000 units; UBTECH will build ten-thousand-unit humanoid capacity this year; and on JD's platform humanoid robot sales reached 10,000 units between January and August 20, 2026, up ninefold year over year. But the roundtable verdict was blunt: reliability of the robot "brain" is the primary contradiction. Some models currently claim a 99% success rate, meaning one significant failure per 100 task executions, and the industry needs to push toward 99.99%, while still lacking a standardized system covering R&D, software, new product introduction, mass production and quality control. Data and cost are the other two threads. Reporting on October 31 described a embodied-AI data collection team of more than 50 people that rented 30 apartments in a single stroke in June 2026 — not for living, but so humanoid robots could practice housework in real home environments, with training content expanding from grasping and placing objects to tidying rooms, brewing coffee and washing labware, and scenes extending from a single workstation to homes, kitchens, restaurants, supermarkets and pharmacies. Industry data puts tactile data at roughly twice the price of visual data, where video data runs to a few hundred yuan per hour while tactile data runs to several hundred yuan per hour; in the first seven months of 2026, nearly 20 funding rounds totaling more than ¥3.1 billion occurred in China's humanoid robot sensor sector, with about ¥1.56 billion in July alone. On November 1, Xingyan Intelligence disclosed the Auriga framework, putting an argument over the route of embodied RSI on the table: the framework splits Physical RSI into L1 task-level search (essentially "evolving the answer"), L2 single-pass online adaptation ("evolving the process") and L3 system-level capability improvement ("evolving the system"), choosing to start from L3 and to participate in RoboDojo and RoboCasa365 evaluations under a Zero-shot protocol that avoids target-task pre-tuning, expert demonstrations, repeated execution and environment resets; using GPT-6 Astra as its base model, it examines zero-shot manipulation, cross-scene transfer and online recovery across 42 tasks and 6,300 trials, following a Preview—Execute—Verify execution flow. The cost ledger comes from a company about to list: Agility Robotics, scheduled to list on the Nasdaq on November 5, disclosed in October through its S-4 and investor day materials that it recognized just $1.782 million of revenue for full-year 2025 against a $2.5 billion pre-money valuation — a price-to-sales ratio above 1,400 — with about 63.5% coming from related parties and only $650,100 in sales to non-related parties; its trust account shrank from about $424 million to about $164 million, SPAC shareholders redeemed more than sixty percent, and total proceeds fell from "over $620 million" to roughly $465 million. Its RaaS model prices a Digit at about $8,500 per month, about $102,000 per year, plus a one-time deployment fee of about $25,000, for roughly $500,000 of revenue over a five-year lifecycle; against a fully loaded human hourly rate of about $30.50, the robot's hourly cost is roughly 46% of a worker's — but if a single unit produces only a fifth to a fifteenth of a person's output per unit time, Digit's cost on equal workload may be two to seven times that of a human. GXO's Flowery Branch site in Georgia went live in mid-2024 and by November 2025 had completed 100,000 tote movements at roughly 98% task accuracy, or about 196 totes per day. The more telling metric is calendar fill rate: published estimates put single-unit calendar fill between 8% and 27%, with a midpoint near 15%, while dedicated sortation equipment can exceed eighty percent. Digit 5 cut runtime from about four hours to 90 minutes, but charging time dropped to about nine minutes and the running-to-charging ratio improved from roughly 2:1 to about 10:1, allowing effective work time beyond twenty hours per day — meaning endurance is no longer the main contradiction, and task scheduling and autonomy are. Agility's locked-in order of more than $300 million over multiple years comes from a single undisclosed related customer and covers about 1,000 Digit 5 units on a three-year RaaS with the service period amortized over four years, plus warrants vesting progressively with deployment; GXO's chief executive has also said publicly that the company has not yet achieved a return on investment from humanoid robots. On funding, October 2026 saw 33 rounds worth ¥10.105 billion in embodied AI, against 60 rounds and ¥11.028 billion in September — heat has cooled, yet dexterous hands, tactile sensors and electronic skin keep attracting capital, with a Shandong flexible-fabric tactile startup closing a ¥100 million Series A on October 29. International money is moving into the physical world at the same time. TechCrunch reported on November 1 that autonomous vehicle technology startup Bedrock Robotics raised $600 million in a Series C valuing it at $4.1 billion, co-led by CapitalG and Emergence Capital, and that Massachusetts manufacturing accelerator Re:Build raised $2.1 billion to expand hard-tech manufacturing capacity and launch industrial startups, led by General Catalyst and JPMorganChase's Strategic Investment Group; Re:Build was co-founded by Amazon's former commercial CEO Jeff Wilke and bills itself as "America's answer to Shenzhen." Read together with Meta's three-month prototype, those two deals point to the same conclusion: the value of hardware bodies and real-world scenarios is being repriced by capital. 来源 / Sources IT之家：消息称 Meta 与中国企业傅利叶打造人形机器人原型，号称是 Muse 模型的物理化身 钛媒体：傅里叶帮 Meta 研发机器人原型，打造"Muse 物理化身" 通信世界：具身智能的"系统战"：Meta 抢入口，宇树闯万台，Agility 算成本 TechCrunch: TechCrunch Mobility: A new, bigger AV boom is here 三、自动驾驶第二波：Trucks VC 一年收到 81 家初创提案，端到端架构把入场门槛压到百万美元以下 事件概要 11 月 1 日，TechCrunch Mobility 刊发专题报道指出，自动驾驶行业正经历"第二波、而且更大"的创业潮。专注早期投资的 Trucks VC 合伙人 Puneeth Meruva 表示，该机构 2026 年迄今已收到 81 家自动驾驶初创公司的提案，比去年多出三分之二，数量足以超越十年前的泡沫期。支撑这一判断的是技术底座的变化：过去行业长期试图从"AV 1.0"走向端到端架构的"AV 2.0"，而如今基础模型已经提供了让端到端系统成为可能的技术架构；这类模型在不同环境与场景之间的泛化能力尤其出色，有望让自动驾驶部署变得更便宜、更快。随着英伟达 Alpamayo 与通义千问的通用视觉—语言—动作模型开源，初创公司的起步门槛进一步降低。Meruva 说，Trucks 反复看到 startups 在几个月内就能完成部署，融资额不足 100 万美元。而这 81 家提案中，最出人意料的是纯 robotaxi 业务数量的增长，许多公司直接把 Waymo 当作竞争对手，Meruva 认为小型、细分领域的自主驾驶公司仍有空间——也许不在全球尺度上与 Waymo 竞争，而是在特定区域或特定类型路线上竞争；他预计明年还会有更多初创公司来投递提案。 与之呼应的是硬件侧的国产进展与终端侧的销售数据。11 月 1 日，IT之家报道，2026 年 10 月汽车销量与交付数据陆续公布：比亚迪单月交付 490,168 辆、同比增长 10.97%，继续稳居榜首；上汽集团销售 430,018 辆，位居第二；吉利以 310,183 辆超过奇瑞（282,458 辆）拿下第三。零跑汽车实现连续 4 个月月销超 10 万台，10 月销量 106,127 辆；小米汽车交付量在 10 月重回 50,000 辆；"蔚小理"三家 9 月分别交出 34,246 辆、50,237 辆与 29,025 辆；鸿蒙智行 10 月交付新车 33,537 辆，均价超过 40 万元。同期产业链层面，第二十四届半导体封装测试技术与市场年会确定于 11 月 4 日至 6 日在南京举行，主题为"AI 启新程，封测铸'芯'基"，会议规模约 2000 人，华天科技、长电科技、通富微电、壁仟科技、摩尔线程、寒武纪、长江存储、海光信息、华为、小米汽车、蔚来、阿里巴巴等产业链企业确认参会，先进封装、TGV 玻璃基板等方向是本次大会的交流重点——智能驾驶芯片与整车的放量，最终都要落到封测与先进封装产能的匹配上。 On November 1, TechCrunch Mobility published a feature arguing that the autonomous vehicle industry is in the middle of a second and even larger startup wave. Puneeth Meruva, a partner at early-stage investor Trucks VC, said the firm has fielded pitches from 81 autonomous vehicle startups so far in 2026 — two-thirds more than last year, and enough to surpass the boom years of a decade ago. What supports the claim is a change in the technical foundation: for a long time the industry tried to move from "AV 1.0" to the end-to-end architecture of "AV 2.0," but foundational models now supply a technical architecture that makes end-to-end systems possible. These models are particularly good at generalizing across environments and scenarios, potentially making autonomous vehicles cheaper and faster to deploy. The arrival of powerful open source models such as Nvidia's Alpamayo and Qwen's general-purpose vision-language-action models has lowered the barrier to entry further. Meruva said Trucks consistently sees startups deploy within a few months and with less than $1 million raised. Of the 81 pitches, what surprised him most was the growing number of pure-play robotaxi companies, many positioning themselves as direct competitors to Waymo, and he believes smaller, niche autonomy companies still have room — maybe not competing with Waymo globally, but in specific regions or on specific route types. He expects even more autonomous vehicle startups to pitch the firm next year. Hardware-side progress in China and terminal sales data moved in parallel. IT之家 reported on November 1 that October 2026 vehicle sales and delivery figures have begun to come in: BYD delivered 490,168 units, up 10.97% year over year, remaining firmly on top; SAIC Group sold 430,018 vehicles to take second; and Geely took third with 310,183 units, ahead of Chery's 282,458. Leapmotor delivered 106,127 units in October, a fourth consecutive month above 100,000; Xiaomi Auto's deliveries returned to 50,000 in October; and the "蔚小理" trio recorded 34,246, 50,237 and 29,025 units respectively in September. Harmony Intelligent Mobility delivered 33,537 new vehicles in October at an average price above ¥400,000. On the supply chain, the 24th Semiconductor Packaging and Testing Technology and Market Annual Conference was confirmed for November 4 to 6 in Nanjing under the theme "AI 启新程，封测铸'芯'基," with roughly 2,000 attendees; Huatian Technology, JCET, Tongfu Microelectronics, Biren Technology, Moore Threads, Cambricon, Yangtze Memory, Hygon, Huawei, Xiaomi Auto, NIO and Alibaba are among the confirmed participants, with advanced packaging and TGV glass substrate among the focus areas — because the ramp of autonomous driving chips and vehicles ultimately has to match packaging and advanced-packaging capacity. 来源 / Sources TechCrunch Mobility: TechCrunch Mobility: A new, bigger AV boom is here IT之家：2026 年 10 月汽车销量 / 交付榜出炉：比亚迪超 49 万辆稳居榜首，小米月销重回 5 万辆 IT之家：第二十四届半导体封测年会定档 11 月 4 日至 6 日举办 四、Agent 架构范式切换：DeepSeek 公开 Super-agent 原型，把"单会话"改造成"常驻系统" 事件概要 11 月 1 日 22 时 12 分，PChome 报道，DeepSeek Harness 团队发布架构探索文章《Super-agent 设计探索：迈向更自主的 Agent 系统》，并同步以 MIT 协议在 GitHub 开源原型 dsh-superagent-prototype。文章对现有 Agent 设计的短板给出了相当直接的判断：主流 Agent 仍以单一 Session 为单位，人通常只能同时跟进 3 到 5 个活跃 Session，"单任务上限看模型，多任务上限看人的注意力"；即便依靠 CLI 与 MCP 扩展，也会卡在单会话同步阻塞、上下文混乱、设备离线即停这几个老问题上。Super-agent 的目标是打破 Agent 与单次会话的绑定，成为跨会话保留上下文、并行推进多任务且互不干扰的实体，并把它定义为"连续、高效、主动、持久"四个特征，用五块架构来支撑：双层架构兼顾连续与高效，事件驱动实现主动响应，计算、状态与环境分离加 Watchdog 实现持久自愈。团队也坦承，部分设计已在原型中落地，但仅验证路线而非产品；大规模场景下的上下文边界、A2A 通信、人机协作、安全隐私仍是待解难题，而对于一个替人推进任务的"主动、持久、自治"系统，权限边界与出错时的责任归属恰恰是最该被追问的部分。 同日网易号的一篇技术拆解，把这套设计落到了代码层面。核心是两层，但两层不靠提示词分工，而靠工具集分工：主会话承担 Router 角色，拥有 14 个应用工具，覆盖读取、路由、Ticket 管理、重新评估、评估、通信与退订、配置，其中 orchestrator_route 指令一次可按顺序提交 1 到 20 个决策；关键在于它不直接执行具体任务，甚至可以不配置 Bash 这类工具，只负责人类交互、全局上下文与任务分派。执行器是 Worker 角色，每个 Ticket（一张被派出的工单，对应一个独立会话与一个目录）各自挂 6 个应用工具，专注把单一任务做完；同时它拿不到某些工具——向人提问或展示、修改 profile、在 worker 槽位之外调度工作、启动或消息其他 Session、编辑外部文档，一概不给，Ticket 结束后工具调用直接被拒绝。换句话说，权限边界写在工具集上，而不是写在提示词里，这是整套设计里最不花哨也最容易复刻的一处。代码实现更朴素，几乎把传统后端系统的约束搬进了 Agent 运行时：Engine 是 Trigger、Thread、Decision、Ticket、Assignment 五类聚合的唯一写入者；存储是单个 SQLite 文件，Runtime 先取得目录锁再打开，防止第二个进程迁移或写入；副作用走持久队列，Effects 在事务内入队、提交后才开始执行，同一个 kind/key 只保留一行，失败按指数退避重试；心跳有固定节拍，Passes 每两秒执行一次，同一时间只跑一个 pass，期间到来的请求合并到下一次执行；健康检查按部件分散登记，同一 effect 类型连续失败三次才报告错误。基于同样思路的内部 Super-agent 实例已在持续运行，日常参与问题复现与修复、代码审查响应与 Harness 自身的迭代——群聊里反馈软件问题，即使没人 @ 它，它也能识别潜在缺陷并在后台尝试复现。原型本身的适用范围收得很紧：只面向一个 GitHub 仓库，可选接入飞书，网页面板只显示需要人处理的事。 这条技术路线在同一天也撞上了真实世界的风险提示。WIRED 在 11 月 1 日报道了一起典型的"过度依赖 AI 建议"事件：三名 20 岁的登山者在加州北部沙斯塔山，因准备不足、体力估计过高而未在下午返程、晚上 19:30 才登顶，被迫摸黑下山，次日清晨在山体困难地带迷路，其中一人还摔伤了膝盖；他们事后承认，准备不足与计划失误部分来自依赖谷歌的 AI 聊天机器人 Gemini。美国林务局沙斯塔山荒野攀登与雪崩中心项目经理 Nick Meyers 说："这种人来得多，准备不足也不做研究。但得说他们确实尝试了——我觉得他们基本把脑子关掉了。"同类情况并不孤立，另一位到访者在"禁止露营"标牌附近搭起帐篷，坚持称 AI 告诉她有权这么做。AI 进入野外与高风险场景之后，"陪伴式建议"与"可执行的判断"之间的边界，已经从产品体验问题变成了安全问题。 On November 1 at 22:12 Beijing time, PChome reported that the DeepSeek Harness team had published an architecture exploration article, "Super-agent 设计探索：迈向更自主的 Agent 系统," and open-sourced the prototype dsh-superagent-prototype under the MIT license on GitHub. The article gives a blunt diagnosis of where current agent design falls short: mainstream agents still operate on a single session, a person can typically keep only three to five active sessions in flight, and "the ceiling on a single task is the model, but the ceiling on multiple tasks is the human's attention." Even with CLI and MCP extensions, the system still hits single-session synchronous blocking, context confusion, and stopping dead when the device goes offline. Super-agent aims to break the binding between an agent and one session, becoming an entity that preserves context across sessions and advances multiple tasks in parallel without interference, defined by four characteristics — continuous, efficient, proactive and persistent — and supported by five architectural blocks: a two-layer architecture for continuity and efficiency, event-driven design for proactive response, and separation of compute, state and environment plus a Watchdog for persistent self-healing. The team also concedes that parts of the design have landed in the prototype but validate a route rather than a product; context boundaries at scale, A2A communication, human-machine collaboration, and security and privacy remain unsolved, and for a system that proactively and persistently advances tasks on a person's behalf, permission boundaries and accountability when something goes wrong are precisely what most need to be interrogated. A technical teardown published the same day on NetEase brought the design down to code level. The core is two layers, but the split is not enforced by prompt wording — it is enforced by tool sets. The main session plays the Router role and holds 14 application tools covering reading, routing, ticket management, re-evaluation, evaluation, communication and unsubscribe, and configuration, with the orchestrator_route command able to submit one to twenty decisions in order; the key point is that it does not execute concrete tasks itself and can even be configured without tools like Bash, handling only human interaction, global context and task dispatch. The executor is the Worker role: each ticket — a dispatched work order corresponding to an independent session and directory — carries six application tools dedicated to finishing one task, while being denied others entirely, including asking or showing things to a person, modifying the profile, scheduling work outside worker slots, starting or messaging other sessions, and editing external documents; after a ticket closes, tool calls are simply rejected. In other words, the permission boundary is written into the tool set rather than into the prompt, which is the least glamorous and most readily replicable part of the whole design. The code itself is plainer still, essentially importing traditional backend constraints into the agent runtime: the Engine is the sole writer of five aggregate types (Trigger, Thread, Decision, Ticket, Assignment); storage is a single SQLite file and the Runtime takes a directory lock before opening it, preventing a second process from migrating or writing; side effects go through a persistent queue where Effects are enqueued inside the transaction and only begin executing after commit, one row is kept per kind/key, and failures retry with exponential backoff; heartbeats run on a fixed cadence, with Passes executing every two seconds, only one pass running at a time and requests arriving in between merged into the next run; and health checks are registered per component, with an error reported only after the same effect type fails three times consecutively. An internal Super-agent instance built on the same ideas is already running continuously, participating in issue reproduction and repair, code review response and the iteration of the Harness itself — when a team member reports a software problem in group chat, it can identify the defect and attempt to reproduce it in the background even if nobody @-mentions it. The prototype's own scope is deliberately narrow: one GitHub repository, optional Lark integration, and a web panel that shows only what needs human attention. That same day, this technical direction collided with a real-world risk signal. WIRED reported on November 1 a textbook case of over-reliance on AI advice: three 20-year-old hikers on Mount Shasta in Northern California, underprepared and overestimating their fitness, did not turn back in the afternoon, summited at 7:30 pm, were forced to descend in the dark, and by morning were lost on a gnarly part of the mountain, one of them with a banged-up knee. They later admitted that some of the poor preparation and planning came from relying on Google's AI chatbot Gemini. Nick Meyers, program manager of the US Forest Service's Mount Shasta Wilderness Climbing and Avalanche Center, said: "We get plenty of people who come unprepared or don't do any research at all. To these guys' credit, they did try. But I think they kind of turned their brains off." The pattern is not isolated — another visitor pitched a tent near a "No Camping" sign, insisting AI told her she had a right to do so. Once AI enters the wild and other high-stakes settings, the line between "companionship-grade advice" and "actionable judgment" has shifted from a product-experience question to a safety question. 来源 / Sources PChome：DeepSeek 提出 Super-agent 架构设计：迈向更自主的 Agent 系统 网易：14 个工具只分派、6 个工具只干活：DeepSeek 把 Agent 拆成了两层？ WIRED: Chatbots Are Leading Hikers Astray, With Dangerous Consequences 五、算力与存储成本：推理数据量首超训练量，长鑫 LPDDR 占比过半，HBM 挤占把涨价推到手机 BOM 事件概要 11 月 1 日 18 时 46 分，IT之家援引报道称，中国电子信息产业发展研究院（赛迪研究院，CCID）发布《新域新质创新发展十大趋势》报告，其核心判断是：我国人工智能推理数据量首次超过语料消耗量，标志着 AI 从模型训练研发阶段迈入规模化产业应用阶段。数据显示，截至 2026 年 6 月，我国日均词元调用量超过 175 万亿，相比 2024 年年初增长 1000 多倍；顶尖大模型的迭代周期也从过去的几个月压缩到几周甚至几天，人工智能正从"喂模型"的实验室阶段转入"用模型"的产业落地阶段。产业侧同步铺开：随着"人工智能 + 制造"行动深入实施，我国已累计建成 500 多家卓越级智能工厂，覆盖 31 个省（区、市），涵盖超过 80% 的制造业行业大类；国产算力、大模型已在钢铁、汽车等 20 个行业落地 1000 多个应用场景。 存储环节的数据同样在同一天被刷新。11 月 1 日 21 时 19 分，IT之家报道长鑫科技 2026 年三季报业绩交流会记录：2026 年上半年，DDR 系列产品实现 694.72 亿元收入、占营业收入比例 46.22%，LPDDR 系列产品实现 781.91 亿元收入、占营业收入比例 52.02%；相较上半年，前三季度 DDR 系列收入占比持续上升，主要原因是 DDR5 产品自 2024 年底量产后销量持续增长。公司提到，第五代工艺技术平台是目前量产的最先进平台，已推出 245Ball 和 496Ball 封装的 LPDDR5X 产品，后续会基于该平台逐步升级 DDR5、DDR6 等产品线；在 4F² 垂直沟道 DRAM 架构取得关键突破的基础上，第六代工艺技术平台已完成工程样片的验证。同时公司表示，受 AI 应用快速发展带来的需求驱动，四季度依然收到服务器客户的强劲需求，DRAM 供不应求的态势仍将延续，预计四季度营收继续保持增长趋势。 涨价如何传导到终端，同花顺 11 月 1 日的深度稿给出了一条清晰的链条：HBM 把多颗 DRAM 垂直堆叠以更高带宽配套 AI 芯片，代价是晶圆的乘数占用——同等容量下，HBM 芯片面积是普通内存的 2 至 3 倍，每为 AI 生产 1GB HBM，就要放弃约 3GB 通用内存产出；目前全球约七成先进存储产能已被 AI 数据中心吸收，手机、电脑等消费电子只能争抢剩余三成，而 HBM 还依赖先进封装，良率爬坡艰难，新建专用晶圆厂从动工到量产至少需要两年。公开资料预计到 2026 年全球 HBM 市场规模突破 600 亿美元，SK 海力士与美光已公开宣布 2026 年 HBM 产能全部售罄，客户甚至提前锁定 2027 年乃至 2028 年的供应。传导已经走过合约价走到整机物料清单：上半年通用内存合约价几近翻番、闪存涨幅超过五成，多家投行测算各类存储芯片供需缺口普遍超过 4%、创下十余年来新高；手机存储成本占整机物料成本的比例从 10% 至 15% 升至 30% 以上，部分大存储旗舰接近 50%；12GB 手机内存采购价从 200 元涨至近 600 元，1TB 闪存从 200 元出头升至 600 元。供给侧同样在扩张：SK 海力士通过美国存托股份发行募资 265 亿美元，成为外国企业在美国规模最大的 IPO；三星电子单季营业利润暴增 18 倍。 终端侧则呈现另一条产品逻辑。TechCrunch 在 11 月 1 日的报道中记录了苹果首席执行官 John Ternus 接受 WIRED 记者 Steven Levy 采访时的表态："我们不想为了 AI 而发布 AI。当我们做对了的时候，人们不需要去分辨这件事是 AI 带来的还是传统软件带来的。真正重要的是，我们有没有做出他们真正喜欢的东西。" 这位 9 月正式出任 CEO 的苹果高管承认，外界普遍认为苹果在 AI 上落后——尤其是被推迟的更强版 Siri 直到今年夏天才发布——但他坚持"我很喜欢我们现在的状态"，并表示苹果"绝对会利用 AI"。当被追问 AI 是否会通过智能体让 App 这类形态变得不再必要时，他回应："那也许会让生活变得非常轻松，不过产品的创造是一件非常有人味、非常个人的事情，是我们热爱去做的事。" 就在同一天，TechCrunch 还评测了谷歌 Pixel 11 Pro Fold：机身展开厚度 5 毫米、折叠后 10.1 毫米，比上代轻 19 克，官方称耐久性提升三倍，采用新的玻璃陶瓷盖板与玻纤复合背板，外屏峰值亮度提升至 3600 尼特（上一代为 3000 尼特），但折痕在强光下依然明显。作为端侧 AI 硬件的一环，这类折叠屏旗舰所承载的仍是同一个命题：把生成式能力塞进一块屏幕，代价是供应链与 BOM 的全面吃紧。 On November 1 at 18:46 Beijing time, IT之家 reported that the China Center for Information Industry Development (CCID, known as CCID Research) had released its report "New Domains, New Quality: Ten Innovation and Development Trends," whose central judgment is that China's AI inference data volume has for the first time exceeded training material, marking the industry's move from a model-training R&D phase into scaled industrial application. The data show that as of June 2026, average daily token invocation exceeded 175 trillion, more than a thousandfold increase from early 2024; that the iteration cycle of frontier models has compressed from months to weeks or even days; and that AI has shifted from the laboratory "feeding models" phase into the industrial "using models" phase. Deployment is spreading in parallel: as AI-plus-manufacturing initiatives deepen, more than 500 excellence-grade smart factories have been built across 31 provinces and municipalities, covering more than 80% of manufacturing industry categories, with domestic compute and large models landing more than 1,000 application scenarios across 20 industries including steel and automotive. The memory side of the ledger was refreshed the same day. At 21:19 on November 1, IT之家 reported ChangXin Memory's Q3 2026 results briefing: in the first half of 2026 the DDR product line generated ¥69.472 billion in revenue, or 46.22% of total revenue, while LPDDR generated ¥78.191 billion, or 52.02% of the total; versus the half-year figure, DDR's share of revenue kept rising through the first three quarters, mainly because DDR5 has grown steadily since entering volume production at the end of 2024. The company noted that its fifth-generation process platform is the most advanced currently in mass production, having introduced LPDDR5X in 245Ball and 496Ball packages, with DDR5 and DDR6 product lines to be upgraded on that platform going forward; and that on the strength of the key breakthrough in the 4F² vertical-channel DRAM architecture, the sixth-generation process platform has completed engineering-sample validation. The company also said that driven by rapid AI application growth, it continued to receive strong demand from server customers in the fourth quarter, that the DRAM supply shortage would persist, and that fourth-quarter revenue is expected to keep growing. How the price increase transmits to end devices was laid out clearly in a deep piece from 同花顺 on November 1: HBM stacks multiple DRAM dies vertically to pair with AI chips at higher bandwidth, and the cost is a multiplier effect on wafer capacity — at equal capacity, an HBM chip occupies two to three times the area of ordinary memory, so every 1GB of HBM produced for AI means giving up roughly 3GB of general-purpose memory output. Roughly seventy percent of advanced memory capacity worldwide has now been absorbed by AI data centers, leaving consumer electronics to compete for the remaining thirty percent, and HBM additionally depends on advanced packaging, where yield ramps are hard and a purpose-built fab takes at least two years from groundbreaking to volume production. Public estimates put the global HBM market above $60 billion in 2026, and SK hynix and Micron have both announced that 2026 HBM capacity is fully sold out, with customers locking in supply for 2027 and even 2028. The transmission has moved past contract prices into bills of materials: general memory contract prices nearly doubled in the first half and NAND rose more than fifty percent; multiple banks estimate memory chip supply-demand gaps generally above 4%, a record in more than a decade; the memory share of a phone's bill of materials has risen from ten to fifteen percent to over thirty percent, approaching fifty percent on some large-storage flagships; and the purchase price of 12GB phone memory has gone from ¥200 to close to ¥600 while 1TB NAND has moved from just over ¥200 to ¥600. Supply is expanding too: SK hynix raised $26.5 billion through a US depositary share offering, the largest IPO by a foreign company in the United States, and Samsung Electronics reported an eighteen-fold surge in quarterly operating profit. On the terminal side, a different product logic ran in parallel. TechCrunch reported on November 1 that Apple CEO John Ternus, interviewed by WIRED's Steven Levy, said: "We don't want to ship AI for AI's sake. When we're doing it right, people won't have to think about whether it was AI that made this happen or more traditional software. What matters is, did we build something that they really love?" Ternus, who formally became CEO at the start of September, acknowledged that Apple has often been described as behind on AI, particularly after delaying a more powerful Siri that finally launched this summer, but insisted "I really like where we're at" and that Apple will "absolutely go on leveraging AI." Pressed on whether AI might reshape Apple's ecosystem more dramatically, with agents potentially making apps unnecessary, he replied: "That could make life really easy, I suppose, but the creation of products is a deeply human and personal thing, something that we love to do." The same day, TechCrunch also reviewed Google's Pixel 11 Pro Fold: 5mm thick unfolded and 10.1mm folded, 19 grams lighter than the previous generation, with durability claimed to be three times better, a new glass-ceramic cover and glass-fiber composite back, and an outer display peak brightness of 3,600 nits against 3,000 nits before — though the crease remains clearly visible in bright light. As part of the on-device AI hardware chain, this flagship foldable carries the same proposition: squeezing generative capability into a screen at the cost of a fully strained supply chain and bill of materials. 来源 / Sources IT之家：赛迪研究院：我国人工智能推理数据量首超训练量，截至 6 月日均词元调用量超 175 万亿 IT之家：长鑫科技：2026 年 LPDDR 系列产品实现 781.91 亿元收入，占营业收入比例 52.02% 同花顺：AI 存储的挤占账：HBM 如何把成本推给手机电脑 TechCrunch: Apple CEO says the company doesn't 'want to ship AI for AI's sake' TechCrunch: Pixel 11 Pro Fold: Sturdy rectangle in a wide wide world 六、可信度工程：当 AI 开始批量生产内容，评审、检测与评测标准同时承压 事件概要 11 月 1 日 14 时，IEEE Spectrum 刊发主编 Harry Goldstein 的观点文章《Guarding the Gates of Knowledge》，主题是 AI 批量生成内容对同行评审与知识库可信度的冲击。文章从一个日常场景切入：在社交媒体上滑动信息流时，作者发现约一半内容是 AI 生成的"AI slop"——有经验的用户能一眼识别，不少人无意转发后才发现，甚至有人因为觉得有趣或有助于制造混乱而故意分享。问题在于检测与拦截能力正在被追平。文章引用自由撰稿人兼应用数学博士 Benjamin Skuse 的亲身实验：他尝试让自己用 AI 生成的论文通过 IEEE 的防线。features 与 AI beat 编辑 Eliza Strickland 描述了整个过程——"他基本上就是点了'运行实验'和'写论文'，然后就再没做任何事……他直接把它扔进了出版机器的嘴里。" Skuse 发现，IEEE 目前的人工评审流程与研究完整性检查程序，尚能与使用免费工具的人维持某种猫鼠游戏式的平衡；但更先进的模型不再犯同样低级的错误，这意味着出版商已被卷入一场军备竞赛——一边是越来越逼真的 AI 内容，一边是用于识别它们、却同样在迭代的系统。同样的防守战也在维基百科上演。这个全球最大的人类生成知识库正面临同样的问题，只是人手更少、工具更少：WikiProject AI Cleanup 志愿项目的创始成员 Ellia Lebleu 表示，"维基百科以保存人类生成的信息而闻名，他们想守住这一点。谁知道还能守多久。" 文章最终提出的判断是：如果我们无法共同假定科学记录可信、也无法假定基本事实可靠，那么建立在这两者之上的共识性决策就会被动摇。 而这份压力并不局限于内容领域。在具身智能一线，评测标准与数据质量同样是当前的卡点。《通信世界》11 月 1 日的报道指出，当前部分模型宣称成功率已达 99%，但"每执行 100 次任务就可能出现一次重大故障，无法满足现实使用要求，行业需要向 99.99% 推进"；硬件本体与运动控制"小脑"已趋于成熟，数十至上百台机器人整齐列队行进可以验证硬件一致性，但具身大模型的技术路线尚未收敛，工程化落地仍处在早期，行业缺少覆盖研发、软件、新产品导入、量产和质量管控的标准化体系。数据侧的约束同样具体：视觉可以靠互联网视频喂养，语言可以靠文本喂养，但触觉、力觉和真实操作反馈很难从现成语料中获得；行业数据显示触觉数据价格比视觉数据高两倍左右，视觉等视频数据可能在小几百元每小时，触觉数据在大几百元每小时——这直接决定了 2026 年前七个月国内人形机器人传感器领域近 20 起融资、合计超 31 亿元的资金流向，也解释了为什么一支 50 多人的团队需要一次性租下 30 套真实房源来采集训练语料。同日，星衍智能披露的 Auriga 框架给出了另一种应对思路：与其让模型在固定任务上不断"刷题"，不如把 Physical RSI 推向系统级能力改进——以 GPT-6 Astra 为基础模型，在不针对目标任务预调优、不依赖专家演示、单次执行且不允许环境重置的 Zero-shot 协议下，考察 42 项任务、6300 次测试中的零样本操作、跨场景迁移与在线恢复能力，执行流程遵循 Preview—Execute—Verify，出现偏差时保留已完成进度、从当前状态继续而非重置环境重来。这套设计背后有一个更根本的判断：如果具身智能的进步主要来自"刷题"，每进入一个新场景厂商都要重新采语料、重新训练、重新调参，项目制特征会非常明显；只有当进步能沉淀为系统级能力，机器人才能从单点演示走向跨场景复用。 On November 1 at 14:00, IEEE Spectrum published an opinion piece by Editor in Chief Harry Goldstein titled "Guarding the Gates of Knowledge," on the pressure that bulk AI-generated content places on peer review and on the credibility of knowledge repositories. The article opens with an everyday scene: scrolling through social media, the author found roughly half of the feed filled with AI slop that astute users would call out as machine-generated, often to the embarrassment of people who unknowingly shared it, and sometimes to the delight of posters who shared it deliberately because they found it amusing or useful for sowing confusion. The problem is that detection and interception capabilities are being matched step for step. The piece cites the firsthand experiment of freelance writer and doctor of applied mathematics Benjamin Skuse, who tried to get AI-generated papers past IEEE's defenses. Features and AI beat editor Eliza Strickland described the process: "He basically hit 'run experiment' and 'write paper.' And then he didn't do a damn thing to it. He just launched it into the maw of the publishing engine." Skuse found that IEEE's human-review processes and research-integrity checker programs are currently keeping pace in something like a cat-and-mouse game against people using free tools. But more advanced models no longer make the same elementary mistakes, which leaves publishers locked in an arms race between increasingly plausible slop and the systems aimed at detecting it. The same defensive fight is underway at Wikipedia, the global depository of human-generated knowledge, which confronts the same problem with fewer people and fewer tools: Ellia Lebleu, a founding member of the WikiProject AI Cleanup volunteer effort, said, "Wikipedia is known as a repository of human-generated information, and they want to hold on to that. Who knows how long they'll be able to do that." The article's final judgment is that if we cannot collectively assume the scientific record is trustworthy, or that basic facts are reliable, then consensus-based decisions built on those two things are themselves shaken. That pressure is not confined to content. On the embodied AI front, evaluation standards and data quality are the current chokepoints. A report in Communications World on November 1 noted that some models now claim a 99% success rate, but that "每执行 100 次任务就可能出现一次重大故障，无法满足现实使用要求，行业需要向 99.99% 推进"; that hardware bodies and the motion-control "little brain" have become mature, and that lining up dozens to hundreds of robots in step can validate hardware consistency, but that the technical route for embodied foundation models has not converged and engineering deployment is still early, with the industry lacking a standardized system spanning R&D, software, new product introduction, mass production and quality control. The data-side constraints are equally concrete: vision can be fed from internet video and language from text, but tactile, force and real manipulation feedback are hard to obtain from existing corpora. Industry data puts tactile data at roughly twice the price of visual data, with video data at a few hundred yuan per hour and tactile data at several hundred yuan per hour — which directly explains where money went in the humanoid sensor sector, with nearly 20 rounds worth more than ¥3.1 billion in China's first seven months of 2026, and why a team of more than 50 people had to rent 30 real apartments in one go to collect training material. The Auriga framework disclosed the same day by Xingyan Intelligence offered an alternative response: rather than having models endlessly "brute-force questions" on fixed tasks, push Physical RSI toward system-level capability improvement — using GPT-6 Astra as a base model and examining zero-shot manipulation, cross-scene transfer and online recovery across 42 tasks and 6,300 trials under a Zero-shot protocol with no target-task pre-tuning, no expert demonstrations, single execution and no environment resets, following a Preview—Execute—Verify flow that preserves completed progress and continues from the current state rather than resetting the environment. Behind that design sits a more fundamental judgment: if embodied AI progress comes mainly from brute-forcing questions, every new scene forces vendors to re-collect material, retrain and retune, and the project-based character becomes extremely pronounced; only when progress can be deposited as system-level capability can robots move from single-point demonstrations to cross-scenario reuse. 来源 / Sources IEEE Spectrum: Guarding the Gates of Knowledge 通信世界：具身智能的"系统战"：Meta 抢入口，宇树闯万台，Agility 算成本 钛媒体：Edge AI Daily 早报（11月1日） 行业数据速览 指标 数据 备注 Anthropic 未来算力义务 近 9000 亿美元 IPO 招股文件披露，对照 1175 亿美元负债、1120 亿美元现金 Meta 自由现金流缺口 2026 年 220 亿美元 / 2027 年 710 亿美元 / 2028 年 390 亿美元 标普测算，2028 年前难回正 四巨头 2026 年合计资本开支 7250 亿美元 高盛测算，较 2025 年 4100 亿美元增 77% Meta 信用评级展望 "AA-" 展望由稳定下调至负面 四大超大规模云厂商中首例因 AI 基建支出被降展望 纳斯达克 10 月表现 累跌 5.3%；费城半导体指数跌 15.4% 均为 2025 年 3 月以来最差单月 赛迪研究院推理数据 日均词元调用量超 175 万亿（截至 2026 年 6 月） 较 2024 年初增长 1000 多倍，推理量首超训练量 长鑫科技 LPDDR 收入占比 781.91 亿元，占营业收入 52.02% 2026 年上半年口径；DDR 系列 694.72 亿元占 46.22% HBM 晶圆乘数占用 同等容量下占用面积为普通内存的 2~3 倍 每生产 1GB HBM 放弃约 3GB 通用内存产出 手机存储成本占 BOM 从 10%~15% 升至 30% 以上 部分大存储旗舰接近 50% 12GB 手机内存采购价 200 元 → 近 600 元 1TB 闪存从 200 元出头升至 600 元 Trucks VC 收到的 AV 初创提案 81 家（2026 年迄今） 较去年多三分之二；部分公司融资不足 100 万美元 Meta × 傅利叶 Max 原型 3 个月研发完成，约 1.5 米、48.1 千克 搭载英伟达 Jetson Thor，定位为 Muse 的物理化身 Agility Robotics 2025 年营收 178.2 万美元 对应 25 亿美元投前估值，市销率超 1400 倍 Digit RaaS 定价 月费约 8500 美元 / 年约 10.2 万美元 另加约 2.5 万美元部署费，五年累计约 50 万美元 单机日历填满率 约 8%~27%，中枢约 15% 专用分拣设备可做到八成以上 2026 年 10 月人形机器人融资 33 起、101.05 亿元 9 月为 60 起、110.28 亿元 2026 年 10 月汽车交付（比亚迪） 490,168 辆，同比 +10.97% 上汽 430,018 辆、吉利 310,183 辆、奇瑞 282,458 辆 媒体人解读 这一期最值得反复看的，不是某个模型又涨了几个点，而是三组"账"开始同时被摊开，而且互相印证。 第一组是资本账。Anthropic 把近 9000 亿美元的算力义务写进招股文件，标普把 Meta 的展望下调至负面，高盛说四巨头一年要花 7250 亿美元——这三个数字摆在一起，指向的是同一个变化：AI 产业的估值锚已经从"能力领先"迁移到"现金流可验证"。CoreWeave 月跌 37%、Nebius 跌 35%，跌的不是技术，是折旧与利息的账本；苹果重回市值第一，靠的也不是 AI 成绩单，而是稳定的消费电子基本盘。市场不再为想象力付钱，只为能还的钱付钱。对技术管理者来说，这意味着明年做预算时，"模型能力提升"不再是能写进立项报告的收益项，能写进去的只有单位任务的边际成本下降了多少。 第二组是物理账。Agility 招股书里那组数字极其诚实：2025 年全年营收 178.2 万美元，市销率 1400 倍，六成以上收入来自关联方；RaaS 定价下机器人小时成本约为人工的 46%，但只要单机产出只有人工的五分之一到十五分之一，总成本就是人工的 2 到 7 倍。这才是机器人行业的真实底牌——不是"能不能做出来"，而是"每台机器每年真正干活的时薪是多少"。日历填满率中枢 15% 这一个指标，比任何发布会都更能说明当前的产业化水位。反过来看 Meta 与傅利叶三个月做出 Max 的意义也不在"快"，而在于它把英伟达 Jetson Thor、Meta 的模型和中国的本体工程能力强行捏到了一张牌桌上：模型能力在海外、本体工程在中国、算力平台在英伟达。这是分工，不是排名。 第三组是工程账，也是我认为本期最被低估的一条。DeepSeek 的 Super-agent 原型表面上讲的是"双层架构、事件驱动、Watchdog"，真正有价值的是它把权限边界写进了工具集而不是提示词——Router 拿到 14 个工具但不干活，Worker 拿到 6 个工具但拿不到提问、调度别的会话、改 profile 这些权限，Ticket 结束后调用直接被拒。这是一套可以在两周内复刻的工程约束，却比任何 Agent 框架都更接近生产系统该有的样子。对照着看，具身智能那一侧的 Auriga 框架提出的是同一个问题的另一半：让进步沉淀为系统级能力，而不是每个新场景都重新"刷题"。两者放在一起，其实是在回答同一个命题——怎么让一个自主性更强的系统，跑在一个可审计、可回滚、可追责的地基上。而 IEEE Spectrum 那篇《Guarding the Gates of Knowledge》把反面也摆出来了：一半社交媒体信息流是 AI 生成的，出版商与维基百科都在跑一场追不上的军备竞赛。内容侧如果守不住入口，可信度工程就是无根之木。 给同行的三条具体建议。第一，别再把"接入 Agent"当成一个功能点去做立项，把它当成一次运行时改造——先算清楚 Agent 会拿到哪些工具、谁给它的权限、错了谁负责。DeepSeek 那套工具集隔离的做法，今天就能抄。第二，任何 AI 相关的成本模型，都要把 HBM 对通用内存的挤占和推理调用量的增长放在一起算：日均 175 万亿词元的推理规模下，存储成本已经不是采购细节，而是产品定价的一部分——12GB 手机内存从 200 元涨到近 600 元这件事，会一路传导到你自己的硬件 BOM 和用户的价格预期里。第三，如果你在做机器人或具身智能方向的评估，别用"成功率 99%"这种指标汇报，用"日历填满率"和"每千次任务的人工介入次数"。前者是实验室语言，后者是采购语言，客户只听后者。 最后留一句提醒。本期是星期日，国际媒体的原创供给明显偏薄，我们能拿到的国际内容集中在 TechCrunch、WIRED 与 IEEE Spectrum 三家；这也是过去半年产业信息分布的一个真实切面——真正高密度的产业判断，正在越来越多地从中国的一线供给侧产生。读日报的人应当意识到，当某一周国际源变薄时，不代表世界变安静了，而是信息生产的位置挪了地方。 The most important thing in this issue is not that some model gained a few points, but that three sets of "ledgers" have been opened at once and corroborate each other. The first is the capital ledger. Anthropic wrote close to $900 billion in future compute obligations into its filing, S&P cut Meta's outlook to negative, and Goldman Sachs says the four giants will spend $725 billion in a single year. Read together, these point to one shift: the valuation anchor of the AI industry has migrated from "capability leadership" to "verifiable cash flow." CoreWeave falling 37% in a month and Nebius falling 35% are not declines in technology but in depreciation and interest schedules; Apple reclaimed the top market-cap slot not on an AI scorecard but on a steady consumer-electronics base. The market no longer pays for imagination, only for money that comes back. For technology managers, the practical consequence is that next year's budget will not accept "the model got better" as a benefit line — only measurable declines in marginal cost per task. The second is the physical ledger. The numbers in Agility Robotics' filing are unusually honest: $1.782 million of full-year 2025 revenue against a $2.5 billion pre-money valuation, a price-to-sales ratio above 1,400, and more than sixty percent of revenue from related parties; under RaaS pricing the robot's hourly cost is roughly 46% of a human's, but if a single unit delivers only a fifth to a fifteenth of a person's output per unit time, the total cost is two to seven times that of a human. That is the real card in the robotics industry — not "can we build it" but "what is the effective hourly wage of each machine per year." The calendar fill rate, with a midpoint near 15%, describes the industrialization water level more honestly than any launch event. Read in the other direction, the significance of Meta and Fourier producing Max in three months lies not in the speed but in the fact that it forces Nvidia's Jetson Thor, Meta's models and Chinese body engineering onto the same card table: model capability overseas, body engineering in China, compute platform at Nvidia. That is division of labor, not a ranking. The third is the engineering ledger, and I think this is the most underrated story of the issue. On its surface, DeepSeek's Super-agent prototype is about a two-layer architecture, event-driven design and a Watchdog. What is genuinely valuable is that it writes permission boundaries into the tool set rather than into the prompt — the Router holds fourteen tools but does no work; the Worker holds six but is denied the ability to ask questions, schedule other sessions, or modify the profile; once a ticket closes, tool calls are simply rejected. These are engineering constraints that could be replicated within a fortnight, and they are closer to what a production system should look like than most agent frameworks. Read alongside it, the embodied-AI side of the same issue — the Auriga framework — asks the other half of the same question: how to let progress accumulate as system-level capability instead of re-brute-forcing every new scene. Put together, they are answering one question: how do you run a more autonomous system on a foundation that is auditable, reversible and accountable. And IEEE Spectrum's "Guarding the Gates of Knowledge" shows the other side of the ledger: roughly half of a typical social feed is machine-generated, and publishers and Wikipedia are both running a detection arms race they are not winning. If the content side cannot hold the entry point, credibility engineering is a house built on sand. Three concrete suggestions for colleagues. First, stop treating "adding an agent" as a feature to be scoped, and start treating it as a runtime change — first establish which tools the agent holds, who grants the permissions, and who is accountable when it errs. DeepSeek's tool-set isolation can be copied today. Second, every AI cost model should place HBM's displacement of general-purpose memory alongside the growth in inference call volume: at a scale of 175 trillion daily tokens, memory cost is no longer a procurement detail but part of product pricing — the move of 12GB phone memory from ¥200 to close to ¥600 will flow straight through into your own hardware bill of materials and your users' price expectations. Third, if you evaluate robots or embodied systems, stop reporting "99% success rate" and start reporting "calendar fill rate" and "human interventions per thousand tasks." The first is laboratory language, the second is procurement language, and the customer only hears the second. One last caveat. This issue fell on a Sunday, and international original reporting was visibly thin — what we could reach came down to TechCrunch, WIRED and IEEE Spectrum. That is also a real cross-section of how information production in this industry has redistributed over the past six months: the densest industrial judgment now increasingly originates on the Chinese supply side. Anyone reading a daily brief should realize that when the international sources thin out for a week, it does not mean the world has gone quiet — it means the production of information has moved. 参考来源 英文来源 AI can make tech CEOs' lives easier. What about everyone else?（TechCrunch，2026-11-01） TechCrunch Mobility: A new, bigger AV boom is here（TechCrunch，2026-11-01） Pixel 11 Pro Fold: Sturdy rectangle in a wide wide world（TechCrunch，2026-11-01） Apple CEO says the company doesn't 'want to ship AI for AI's sake'（TechCrunch，美国东部时间 2026-10-31 晚） Chatbots Are Leading Hikers Astray, With Dangerous Consequences（WIRED，2026-11-01） Guarding the Gates of Knowledge（IEEE Spectrum，2026-11-01） 中文来源 Edge AI Daily 早报（11月1日）（钛媒体） 【数智周报】Anthropic 提交 IPO 注册文件，拟以"ANT"为代码在纳斯达克上市（钛媒体） 傅里叶帮 Meta 研发机器人原型，打造"Muse 物理化身"（钛媒体） 消息称 Meta 与中国企业傅利叶打造人形机器人原型（IT之家） 2026 年 10 月汽车销量 / 交付榜出炉（IT之家） 第二十四届半导体封测年会定档 11 月 4 日至 6 日举办（IT之家） 长鑫科技：2026 年 LPDDR 系列产品实现 781.91 亿元收入（IT之家） 赛迪研究院：我国人工智能推理数据量首超训练量（IT之家） DeepSeek 提出 Super-agent 架构设计：迈向更自主的 Agent 系统（PChome） 14 个工具只分派、6 个工具只干活：DeepSeek 把 Agent 拆成了两层？（网易） 具身智能的"系统战"：Meta 抢入口，宇树闯万台，Agility 算成本（通信世界） AI 存储的挤占账：HBM 如何把成本推给手机电脑（同花顺） 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。 ；检索词:一人公司 最新 热点 趋势