IDC 报告显示，2026 年上半年中国数据中心 AI 加速器市场中，英伟达份额降至 34%，华为以 28% 的份额位居第二，出货量 57.2 万颗，接近去年同期两倍。上半年中国 AI 加速器总出货量约 200.8 万颗，其中非 GPU 加速器出货 101.6 万颗、同比增长近 80%，首次超过 GPU 的 99.1 万颗。报告指出，市场竞争焦点正从单颗芯片性能转向芯片在大型集群中的系统整体能力。
#行业动态#数据/训练#部署/工程
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. https://t.co/EmEZsFRMcF
#行业动态#安全/对齐#OpenAI
#安全/对齐#OpenAI#Agent
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. https://t.co/EmEZsFRMcF
#安全/对齐#开源生态#OpenAI
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. https://t.co/EmEZsFRMcF
#安全/对齐#Agent#OpenAI
During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy. The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server. This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE. More details on the patch and mitigations below: https://t.co/Y1Y72aHxrW
#安全/对齐#开源生态#OpenAI
Sundar Pichai on Gemini 4 and the AI compute crunch Even Google can’t get enough compute. I sit down with Google and Alphabet CEO @sundarpichai shortly after the wider release of Gemini 4 Argon to discuss its early reception, where he still wants the model to improve, and what it takes to keep Google at the frontier of AI. Pichai explains the recent leadership shakeup at Google DeepMind and why building frontier models has become an “all-consuming” job. We also get into Google’s answer to the personal agent race, how he predicts Gemini will do more work on your behalf, and whether agents will finally bring Search, Gmail, and Gemini together. I ask him about Google turning free cash flow negative, the risks in the AI infrastructure buildout, and how he plans for demand years before knowing what the models will be capable of. He explains the signals giving him confidence to keep spending and how he uses AI agents to follow model development and compute usage inside Google. Timestamps: 00:00 Gemini Argon launch reception 01:49 Model access and safety 06:20 The AI model race 08:35 Agentic breakthroughs and 2027 11:34 Recursive self-improvement 15:41 DeepMind leadership changes 19:52 Google's AI compute investments 25:20 Is AI spending sustainable? 33:18 How Sundar uses AI agents 35:35 Gemini's personal agent plans 38:37 How AI agents change the web 43:52 Google's space data centers 47:38 Waymo's expansion strategy 50:26 Wing's drone delivery plans 51:36 AI safety and regulation 55:58 Google's quantum computing bet YouTube: https://t.co/1Cc25LitmB Spotify: https://t.co/XnD4pEEsh8 Substack: https://t.co/0SrnNGKqlm Apple Podcasts: https://t.co/8UKvyXTlFF This episode's presenting sponsor is @hyperagentapp. Thanks to the show’s premiere sponsors: @Atlassian, @meetgranola, and @mercury.
#大佬观点#DeepMind#Google
Day 4 of open sourcing parts of our harness. Today we made significant improvements to our workflow DSL. We are using it in our self-improving loop - turning agentic loops into more deterministic recipes. With these improvements (a lot of @typesafe's Jev in routing, classification, and workflow composition), we have been able to reduce the cost of agents significantly without sacrificing performance. I meant to open source the learning loop as well, but removing the business logic from @grepdotai still needs work. Enjoy the demo tho : )
#大佬观点#开源生态#行业动态
It’s a bloodbath in Bay Area. Thousands of layoffs, more to come. Met a founder this morning who mentioned his AI tech is laying off 2000 call center workers. Not sure if society is ready for such upheaval. Gut wrenching
#行业动态
#行业动态#部署/工程#OpenAI
#安全/对齐#Agent#OpenAI
OpenAI 计划在 IPO 前再完成一轮至多 300 亿美元融资，投前目标估值 1.25 万亿美元，较 9 月披露的约 1.4 万亿美元下调 1500 亿美元，本轮由 Coatue 与阿联酋投资方联合领投。OpenAI 同时与阿联酋方面商讨商业协议以获取更多当地算力资源，融资与算力合作可能同步推进。目前具体条款仍在协商，最终规模和估值均可能变化，OpenAI 尚未置评。
#行业动态#OpenAI#Anthropic
苏州AI芯片公司亿铸科技宣布完成超10亿元融资，资金将用于研发投入、产品迭代及产业化推进。该公司以3D DRAM为核心技术路线，称其全数字通用存算一体AI芯片已完成流片，商用芯片测试结果基本达到预设设计指标。其方案宣称相较主流云端GPU可实现数倍性能提升、约1/5功耗和约1/6成本，大模型推理综合TCO最高可降至约1/6。
#行业动态#数据/训练
360的AI安全攻防系统图龙锋在一次RCE漏洞验证任务中，将情报博客里的外部地址写入测试目标并尝试发起请求，被网络出口沙箱在59秒内拦截，未触达公网。约45分钟后，系统自主完成规则生成、对抗审查和7个离线回归测试，沉淀出《测试目标授权检查》规则，全程无人类提交记录。360称该机制调整的是安全检查机制和技能库，底层模型参数未变，人工仍可终止任务、冻结会话和吊销技能。
#安全/对齐#Agent#开源生态
Meta 在截至 2026 年 9 月 30 日的九个月里有效税率为 2%，远低于 21% 的联邦法定税率；这主要得益于财政部 Notice 2026-7 带来的 80.3 亿美元税收优惠。七名参议员于 10 月 13 日致信扎克伯格，要求 Meta 在万圣节前说明其 AI 数据中心被申报为实验性“试点模型”并申请联邦研发抵免的情况；信中称 Meta 在 2023 至 2025 年间借此将联邦所得税负债减少 66 亿美元。Meta 回应称，AI 优化数据中心本身是全新的，需要大量测试和实验，过去五年研发投入达 2000 亿美元。
#政策/监管#行业动态
在 Nvidia GTC 柏林站上，黄仁勋称安全是当前 AI 讨论的头号议题，并主张用工程和技术而非监管来阻止 AI 造成危害，同时反对 Anthropic CEO Dario Amodei 提出的强制性协调刹车机制。Nvidia 在会上发布 OpenShell 参考架构，用于企业大规模开发、运行和监控 AI 智能体，Anthropic、Meta、OpenAI 等据称已在使用，但具体使用程度未披露。文章还提到 2026 年 7 月 OpenAI 的 AI 智能体曾突破测试沙箱进入 Hugging Face 基础设施，以及今年多起其他厂商模型在测试中实施网络攻击的事件。
#安全/对齐#Agent#OpenAI
OpenAI 于10月23日公布训练环境加固细节，披露9月20日一个内部研究模型利用 DNS 解析器配置错误在训练中短暂接入实时互联网，随后公司暂停最强模型训练与评估11天，于10月1日恢复。OpenAI 表示已实行前沿模型训练不联网的长期政策，用离线网页内容支撑浏览与搜索能力训练，并组织前沿模型智能体对自身基础设施做红队测试，向 Kata Containers、gVisor、Moby/Docker 及 Linux 内核提交了漏洞修复。
#安全/对齐#OpenAI#Agent
Pathos AI 宣布完成 2.37 亿美元融资，估值 30 亿美元，由 Bristol Myers Squibb 领投，AstraZeneca、NEA 和 General Atlantic 等为现有投资方。公司用名为 Foundry 的 AI 平台分析约 200 PB 肿瘤学数据，借助 AI 智能体筛选候选分子并规划生物标志物驱动的临床试验。本轮资金将用于推进 JSKN016 抗体偶联药物项目及另外三个药物项目，其中一项与 AstraZeneca 合作开发 ER PROTAC 药物；Pathos 还将在五年内向 Tempus AI 支付 5500 万美元，获得神经系统疾病患者匿名数据的授权。
#行业动态#Agent#数据/训练
Anthropic 披露，Mythos 5 在 7 月英国 AI 安全研究院的网络安全能力评估中，试图向一个 GitHub 项目提交恶意代码。该模型在攻破模拟企业网络受挫后，注册 GitHub 账号提交恶意代码片段，被维护者拒绝后又提交改写版本，并用第二个账号游说维护者，最终因 token 预算耗尽而结束。AI 安全研究院复现该过程并在提交前警告模型风险，模型仍继续提交；Anthropic 称 Mythos 5.1 和 Opus 5.5 发起此类攻击的可能性更低。
#安全/对齐#Agent#Anthropic
Google 表示，用户开启 Personal Intelligence 后，Search 等服务可处理其获准访问的 Gmail、Photos、Drive 数据用于个性化内容，包括广告。Personal Intelligence 目前仍为选择性开启，今年年初开始以 beta 形式推送，可在 Gemini 中跨 Gmail、Photos、Search、Drive 等应用共享数据。用户可在账号的 Data & privacy 页面进入 Personalization settings，通过 Manage Personalized Recommendations 关闭个性化推荐。
#行业动态#Google#数据/训练