Skip to content

All updates

Today 98 items
Nov 3Tue
  1. IT之家 · AI 筛选41

    应对服务业务增长压力，苹果有意通过第三方应用寻找新商机Chinese

    面对监管趋严与手游低迷带来的增长压力，苹果计划通过第三方应用及 AI 合作寻找服务业务新收入来源。9 月财季 App Store 业绩同比基本持平，虽然第三方应用付费订阅量创下两位数增长新高，但受政策冲击其 8 月美国佣金收入下滑 18%。苹果预计 12 月财季服务业务将增长 8% 至 9%，市场亦关注其未来分发第三方 AI 智能体抽佣的潜力。

  2. Claude Code 更新32

    Claude Code v2.1.317 发布Chinese

    Claude Code 发布 v2.1.317 版本，强化了安全检查、权限控制与管理功能。新版本在 Web 或 Shell 工具外发邮箱前增加确认，为 Agent SDK 回调引入 onFailure: "block" 阻断机制，并在 Claude apps 网关提供成本分析 API。更新还修复了桌面端 MCP 服务器断连、自动模式报错以及慢速 token 刷新导致退出等问题。

  3. 爱范儿 · AI 筛选37

    奔驰发布新款长轴距 E 级：全面支持华为手机，42.38 万元起Chinese

    奔驰发布 2026 款长轴距 E 级，售价 42.38 万元至 50.98 万元，首次由中国本土团队主导座舱开发。新车车机接入豆包大模型，实现 0.25 秒唤醒速度、98% 多音区识别准确率，并能通过声纹记住最多 6 位家庭成员。此外，座舱新增支持华为 HiCar 并保留苹果 CarPlay，方向盘按键可联动小艺助手，地图帧率达 60 帧/秒。

  7. IT之家 · AI 筛选29

    鸿蒙智行发布问界新 M8 快速交付公告：最快 1 周内交付新车Chinese

    鸿蒙智行问界汽车发布快速交付公告，问界新 M8 大定锁单后最快 1 周内交付新车，即刻下订可享 5000 元订金抵 10000 元尾款权益。该车将于 11 月 13 日上市且预售价 38.98 万元起，搭载华为乾崑智驾 ADS 5 系统，提供 Max、Max+ 与 Ultra 三个版本共 9 款配置。目前该车展车与试驾车均已抵达全国门店。

  8. IT之家 · AI 筛选51

    OpenAI 发布 LifeVal 基准测试 AI 日常任务处理能力Chinese

    OpenAI 推出名为 LifeVal 的新评估基准，用于衡量 AI 模型协助普通人处理照护、健康、家庭行政和个人财务等日常事务的能力。该基准参考时间使用调查及 3,995 名受访者的实际诉求设计任务，重点考察建议是否达到可执行与可交付状态。目前 LifeVal 主要用于其内部追踪模型表现，尚未公开完整任务集与复现代码。

  10. TechCrunch AI70

    Y Combinator CEO Garry Tan reveals he spends up to $200,000 a year on tokens to run AI agents

    Y Combinator CEO Garry Tan revealed on a podcast that he spends $100,000 to $200,000 annually on tokens, heavily utilizing AI agents to handle work, coding, and personal tasks.

    Why it matters: The piece illustrates the actual expenses and depth of data integration for heavy agent users, offering readers a glimpse into the real-world operational costs and workflows of fully autonomous personal assistants.

  11. TechCrunch AI32

    Apple 或尝试通过第三方 AI 应用增加收入Chinese

    Apple 在财报电话会议中透露，正探索借助第三方 AI 应用及合作伙伴关系提振服务收入。CFO Kevan Parekh 表示 AI 能推动用户换机升级，投资者则推测其或通过在 App Store 分发第三方 AI 智能体并抽取佣金盈利。受多国监管和诉讼挑战影响，Apple 预计 12 月季度服务业务增长为 8% 至 9%。

  12. C114 通信网 · AI与算力（网页）35

    中国电信智能成果亮相2026世界互联网大会乌镇峰会Chinese

    中国电信在2026年世界互联网大会乌镇峰会亮相，集中展示全链路国产化智能云体系与关键核心成果。其智算总规模超过119EFLOPS，率先部署四大全国产液冷万卡集群，星辰大模型在真实训练场景利用率超98%。现场还展出了首款AI原生应用TeleAgent、全国Token套餐及接入7台量子计算机的天衍平台。

  13. C114 通信网 · AI与算力（网页）33

    中国电信在2026年中国工业互联网安全大赛中斩获一等奖等6项大奖Chinese

    中国电信在2026年中国工业互联网安全大赛决赛中选派12支队伍参赛，最终斩获一等奖2项、二等奖1项、三等奖3项，获奖总数与等级居行业首位。在AI安全能力上，其“星辰·见微”安全大模型在Cybersec-Eval测评排名第一，Agentic SOC达到L4级自动化水平且威胁研判效率提升15倍。

  14. Hacker News · AI43

    OpenAI 考虑将数据中心更名为“AI 计算园区”Chinese

    OpenAI 正在筹备一项公关宣传活动以引导公众认知，其中包括提议将其数据中心重新命名为“AI 计算园区”（AI Compute Campuses）。此举旨在影响美国中期选举及 2028 年总统大选前围绕 AI 的公众辩论。OpenAI 全球事务首席官 Chris Lehane 表示，随着对就业、经济和环境的担忧增加，AI 将成为竞选的核心议题。

  15. Hacker News · AI68

    OpenAI Hires Former U.S. Army Secretary as Military Looks to Expand AI Adoption

    OpenAI has hired a former U.S. Secretary of the Army as the U.S. military looks to expand AI adoption. The move reflects how frontier AI companies are strengthening ties with the defense and national security sectors.

    Why it matters: This high-profile appointment reflects the strategic footprint of frontier model makers in defense and security, offering a glimpse into how AI technology is converging with military operations.

  16. IT之家 · AI 筛选47

    王兴兴：宇树未来将重点攻克机器人大脑技术，拟面向全球寻求合作Chinese

    宇树科技创始人王兴兴表示，公司未来将重点攻克机器人大脑技术并突破物理 AI，在推进自身研发的同时面向全球寻求合作。其未来布局还包括提升大规模量产能力、推进硬件适配的物理 AI 小模型落地，以及布局超大规模算力集群。财务方面，宇树科技今年前三季度实现营收 18.17 亿元，归母净利润 3.87 亿元，同比大增 270.54%。

  18. Ollama 更新29

    Ollama v0.40.14-rc0 发布Chinese

    Ollama 发布 v0.40.14-rc0 候选版本，带来多项工具适配与稳定性修复。该版本针对 Claude Code 启用了模型输出限制，使 ChatGPT 断开失败支持重试，并在配置前增加 ChatGPT 路由检查。此外，更新还优化了 MLX 长 top-k 行的分块排序，支持服务端解码云端网页搜索响应，并不再覆盖 OpenAI 模型的默认采样器设置。

  20. IT之家 · AI 筛选22

    中科宇航力鸿二号拟 2027 年一季度开展亚轨道百公里级回收验证Chinese

    中科宇航计划在 2027 年一季度由力鸿二号可重复使用运载器开展亚轨道百公里级回收验证，推动力箭二号运载火箭具备回收复用能力。力鸿二号起飞重量 60 吨，最大飞行高度超 100km，可载 7 人或 3 吨货物。其首飞任务将搭载太空制药装置、无人值守机器人及太空旅游可穿戴监测设备进行在轨验证。

  22. 爱范儿 · AI 筛选28

    早报｜苹果第四财季营收增长 12%，服务收入再创新高 / 《影之刃零》72 小时销量突破 400 万套 / 曝三星 Fold9 回归屏下摄像头Chinese

    苹果公布 2026 财年第四季度业绩，单季营收同比增长 12% 达 1144.77 亿美元，净利润 307.11 亿美元，其中服务营收 315.73 亿美元创历史新高。同时，OpenAI 透露拟近期发布下一代 GPT，其科研实验曾由约 1 万个 AI 智能体协作完成；Anthropic 第三季度营收则增至接近 170 亿美元。

  25. Hacker News · AI79

    Google Discloses LLMJacking Attacks: Stolen AI API Keys Cost a Single Enterprise Nearly $1 Million

    The Google Threat Intelligence Group (GTIG) and Mandiant have disclosed multiple "LLMJacking" attacks targeting enterprise LLM access, with one victim racking up nearly $1 million in cloud bills before the unauthorized access was revoked.

    Why it matters: The report details the complete chain of threat actors stealing LLM credentials and reselling them for profit, offering practical incident response and auditing guidance for enterprises to guard against unexpected cloud bills.

  27. IT之家 · AI 筛选66

    苹果与谷歌下架多款暗藏深度伪造功能的违规 AppChinese

    苹果与谷歌等平台近日集中下架多款违规应用，这些应用以宠物写真、语言学习等日常服务为掩护过审，随后通过服务端切换隐藏界面提供性暴力深度伪造功能。该黑产网络涉及 400 多个关联域名并依托社交平台广告导流，在 Google Play 累计下载超 35 万次。目前苹果已下架 19 款涉事应用，Meta 也已处置 1.16 万个关联广告账户并封禁 209 款应用的广告投放。

  29. Hacker News · AI60

    AI Facial Recognition False Matches Spark Shoplifting Accusations, Leaving UK Shoppers in Data Rights Limbo

    The Facewatch facial recognition system deployed by several UK retailers has generated frequent false matches, leading innocent customers to be wrongfully accused of shoplifting, ejected from stores, or subjected to bag searches. An investigation by civil liberties group Big Brother Watch reveals that once misidentified, an individual's biometric data is shared across a nationwide watchlist spanning multiple retailers, forcing affected shoppers to spend months navigating convoluted processes to clear their names. As the technology provider and retailers pass the buck over liability, calls are mounting for the disclosure of alert volumes and false-positive rates, as well as independent third-party audits of system accuracy.

    Why it matters: The report highlights the dilemmas of cross-store cascading repercussions caused by biometric false positives in retail settings, illustrating the tangible harms of treating algorithmic alerts directly as evidence of wrongdoing.

  30. NVIDIA Developer Blog61

    Diagnosing Distributed AI Workload Failures with NVIDIA Attribution

    NVIDIA showcased how to diagnose cascading failures and pinpoint root causes in distributed AI workloads using the NVIDIA Attribution tool. In distributed environments, an anomaly in a single GPU or network fabric often cascades into NCCL errors across multiple ranks, generating massive downstream noise where conventional exceptions reflect only where the workload halted rather than where the issue originated. Through concrete examples, the piece demonstrates how to run diagnostics based on workload evidence to help troubleshoot and isolate the initial fault.

    Why it matters: Addressing the challenge of troubleshooting massive cascading errors triggered by a single point of failure in distributed training, the post demonstrates how to accurately locate the root cause using NVIDIA Attribution.

  32. IT之家 · AI 筛选30

    IT早报 1103：奥尔特曼称个人电脑浪费内存遭全网声讨，苹果 2026 财年净利润 1321 亿美元，宇树登泰山大赛夺冠Chinese

    OpenAI CEO 奥尔特曼因称“把 DRAM 装在个人电脑中是资源浪费”遭全网及马斯克声讨。大模型与硬件财报方面，Anthropic 第三季度营收近 170 亿美元并已递交 IPO 申请，苹果 2026 财年归母净利润达 1321.75 亿美元。此外，首届 CMG 世界机器人登泰山大赛收官，宇树机器人斩获 1 金 2 银，其 G1 开发版因过热退赛。

  33. IT之家 · AI 筛选72

    日本拟修订汽车安全认证标准，为 L4 与 L5 级自动驾驶车辆量产铺路Chinese

    日本政府计划修订汽车安全认证标准，允许没有方向盘和踏板的 L4 与 L5 级自动驾驶汽车按统一标准接受认证，新规计划从 12 月起适用。载客车辆仍须进行假人碰撞测试，货运车辆可免除该测试但须确保系统故障时能避免事故。日本政府目标到 2030 财年让至少 1 万辆 L4 级公交车和出租车投入道路运营。

  34. Buzzing HN · 中文精选78

    Anthropic's revenue growth momentum plateaus as monthly figures show signs of a slowdown

    Anthropic's Q3 revenue reportedly reached $16.8 billion, but monthly estimates indicate that revenue growth slowed significantly between July and September. Its revenues from Q1 through Q3 were $4.8 billion, $11.6 billion, and $16.8 billion, respectively, with the Q3 quarter-over-quarter increase narrowing to $5.2 billion from $6.8 billion in the prior quarter. Considering an annualized run rate that had already reached $65 billion in July, growth across August and September was very limited.

    Why it matters: The article breaks down monthly revenue estimates and sequential changes, helping readers observe the knock-on effects of slowing growth among leading foundation model developers on compute infrastructure expectations.

  35. Hacker News · AI60

    Training Small Models for Robotic Temporal Logic Planning via Reinforcement Learning and Tool Use

    By leveraging tool use and GRPO reinforcement learning, the authors successfully trained Qwen3.5-9B to perform robotic hierarchical path planning based on Signal Temporal Logic (STL).

    Why it matters: The article demonstrates that combining geometric tools with teacher distillation prevents small models from falling into a zero-reward trap caused by geometric computation errors during RL-based planning.

  36. Hacker News · AI74

    Claude Code triggers over-refusals after new restrictions, prompting some developers to downgrade models

    Anthropic modified Claude Code's system prompt to guard against copyright infringement, leading Claude Opus 5.5 to refuse tasks such as recreating existing games, reverse engineering, and software analysis, which has forced some developers to downgrade to older model versions to resume their workflows.

    Why it matters: The article presents a real-world case of managed agents triggering over-refusal due to system prompt tweaks, where downgrading to older models became a practical workaround for developers to keep their workflows running.

  37. IT之家 · AI 筛选42

    苹果承认 iPhone Duo 折叠屏手机交货延迟，供货紧张何时缓解仍难以预测Chinese

    苹果在财报电话会议上承认新款折叠屏手机 iPhone Duo 面临供不应求与交货延迟，供需平衡恢复时间目前难以预测。该机在开售数小时内售罄且部分订单已排至明年发货，苹果为此加强了与钛金属供应商 Perryman 的合作以提升产能。此外，AI 需求激增推高了存储芯片成本，也令硬件厂商面临普遍的零部件涨价压力。

  38. IT之家 · AI 筛选76

    Tech sell-off weighs on Anthropic IPO valuation expectations, with institutions advising a cut to $1.5T–$1.6T

    Hit by last week's tech rout and doubts over AI returns on investment, institutional investors have told Anthropic and its advisors that its valuation for an IPO later this month should be lowered to between $1.5 trillion and $1.6 trillion, down from the nearly $2 trillion previously discussed in the market.

    Why it matters: The original report reveals the downward revision in institutional valuation expectations for Anthropic amid the tech pullback and ROI concerns, offering insight into how public markets are pricing frontier foundation model assets.

  39. IT之家 · AI 筛选53

    OpenAI 向 2028 美大选潜在候选人发备忘录呼吁重视 AI 治理与基建议题Chinese

    OpenAI 首席全球事务官 Chris Lehane 私下向多位 2028 年美国总统大选潜在候选人发送备忘录，指出 AI 及其配套基础设施将成为下届竞选的关键议题。他在备忘录中警告反对建设 AI 数据中心可能在政策争论中处于不利地位，敦促候选人提出既不放任也不过度垄断的折中治理路线，并向选民解释如何维持技术优势与确保人类主导权。