AI速递 10.06

Reflection AI has unveiled Beam, its first open-weight frontier model. The text-only mixture-of-experts system has 501B parameters, 23B active, a 1M-token context window, and was pretrained on 23.8T tokens. Reflection claims Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks and beats leading Western open models while using 3-4x less inference compute, but these results have not been independently verified. Weights and technical details are due this month.

Reflection AI 发布了其首个开放权重的前沿模型 Beam。该纯文本混合专家(MoE)模型拥有 5010 亿参数(激活参数 230 亿),上下文窗口达 100 万 token,预训练使用了 23.8 万亿 token。Reflection 声称,Beam 在高级推理基准测试中与智谱(Z.ai)的 GLM-5.

原文: https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/

Hinton、Bengio等22位学者论文探讨AI研发自动化会否引发“智能爆炸”。据Anthropic内部数据，2026年3—8月Claude在高层人类监督下自主完成的研发占比由约1%升至26%。作者提出“有效研发劳动力”，称领先公司现有算力理论上可抵数百万名顶级研究员；若回报不变，AI进步或1.5年提速10倍，一年进展缩至约5周。

原文: https://www.qbitai.com/2026/10/501705.html

AI agents found two candidate Luttinger-compensated magnets—ferromagnetic spin sorting with zero net magnetism—for spintronic memory. Designed YBaMnFeO₅ has a predicted 2.35 eV gap and spin windows far above room-temperature thermal noise (26 meV), but its needed atomic ordering likely fails during synthesis. KV[Cr(CN)₆], a Prussian-blue compound first made in 1999, has a predicted 2.1 eV gap and magnetic order to 376 K; simulations disagree on whether sample water weakens spin sorting, and its gap/spin sorting are unmeasured. Next: re-synthesize and measure. Code/data: GitHub.

AI 智能体发现了两种候选的 Luttinger 补偿磁体——净磁性为零的铁磁自旋分选材料,可用于自旋电子学存储。设计出的 YBaMnFeO₅ 预测带隙为 2.35 eV,自旋窗口远高于室温热噪声(26 meV),但其所需的原子有序排列在合成中可能无法实现。KV[Cr(CN)₆] 是一种 1999 年首次合成的普鲁士蓝类化合物,预测带隙为 2.1 eV,磁有序温度可达 376 K;

原文: https://www.vals.ai/blogs/room-temperature-magnetic-semiconductors

OpenAI will add invisible watermarks to ChatGPT/Codex text in the EU to meet AI Act transparency rules from Aug 2. EU rollout across all plans takes weeks; API developers worldwide can opt in today (off by default). The watermark nudges word choices, survives copy-paste and doesn't identify users. OpenAI's textGrain, built with UPenn/Yale, uses a secret key. But edits weaken it: replacing 10% of words cut detection from ~92% to 66%; short/translated text is hard. Detector access is limited to approved researchers. Anthropic did similar worldwide two months earlier.

OpenAI 将为欧盟 ChatGPT/Codex 文本加入隐形水印，以符合8月2日起的 AI Act 透明度规则；欧盟所有套餐上线需数周，全球 API 开发者今日可选开启（默认关闭）。水印通过微调选词实现，复制粘贴后仍存在，不识别用户；textGrain 由 OpenAI 与 UPenn/Yale 共建并使用密钥。但编辑会削弱：替换10%词语使检测率从约92%降至66%，短文本或译文更难识别。

原文: https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/

OpenAI is introducing visual display ads alongside images generated by ChatGPT, launching later this month in the U.S. with an initial test group of advertisers. Ads will be clearly labeled and won't influence ChatGPT's answers. OpenAI is also expanding ad measurement partnerships—including AppsFlyer, Adjust, and Branch—and developing brand suitability pilots with DoubleVerify and Integral Ad Science. Ads, introduced earlier this year to support free and low-cost tiers, could eventually target ChatGPT's 1.2 billion weekly users, though they may push casual users toward paid subscriptions to av

OpenAI将在ChatGPT生成的图片旁引入视觉展示广告，本月晚些时候在美国上线，首批与一组广告商进行测试。广告将明确标注，且不会影响ChatGPT的回答。OpenAI还在扩大广告效果衡量合作伙伴，包括AppsFlyer、Adjust和Branch，并与DoubleVerify和Integral Ad Science开展品牌适配试点。

原文: https://techcrunch.com/2026/10/05/openai-launches-visual-ads-that-appear-alongside-image-generation-results/

President Trump announced a Super Intelligence Force to coordinate federal AI efforts and keep the US ahead in "super intelligence." Per The Wall Street Journal, national intelligence director Jay Clayton will chair the task force, with FTC Chair Andrew Ferguson, Undersecretary of War for Research and Engineering Emil Michael, and OPM Director Scott Kupor as vice chairs. The group has 120 days to report on AI's risks and opportunities; its charter calls for countering AI-enabled threats while preventing overregulation that could stifle innovation.

特朗普总统宣布成立"超级智能工作组",以协调联邦政府的AI工作,确保美国在"超级智能"领域保持领先。据《华尔街日报》报道,国家情报总监杰伊·克莱顿将担任组长,联邦贸易委员会主席安德鲁·弗格森、负责研究与工程的战争部副部长埃米尔·迈克尔、人事管理局局长斯科特·库珀任副组长。该小组须在120天内就AI的风险与机遇提交报告;其章程要求应对AI赋能的威胁,同时防止扼杀创新的过度监管。

原文: https://techcrunch.com/2026/10/04/trump-unveils-his-new-super-intelligence-force/

TikTok announced Monday it's launching an AI Shopping Assistant and one-click in-app checkout. The conversational assistant remembers user preferences and offers real-time guidance on products, shipping, sizing, and availability, while checkout lets users buy directly from brands in the For You feed. Built with partners including Salesforce, Shopify, Shoplazza, and Stripe, the tools expand TikTok's commerce push beyond TikTok Shop and aim to keep users from external AI tools like ChatGPT. TikTok cites $81 billion in U.S. GDP generated by platform activity;

TikTok周一宣布推出AI购物助手和应用内一键结账功能。该对话式助手可记住用户偏好，就商品、配送、尺码和库存提供实时指导；一键结账功能则让用户可在"为你推荐"(For You)信息流中直接向品牌购买商品。这些工具由TikTok与Salesforce、Shopify、Shoplazza和Stripe等合作伙伴共同打造，将其电商业务拓展至TikTok Shop之外，并旨在防止用户转向ChatGPT等

原文: https://techcrunch.com/2026/10/05/tiktok-rolls-out-an-ai-shopping-assistant-and-one-click-checkout/

Independent researchers published preliminary findings on AI agents appearing to run on Tencent's infrastructure and query Alibaba's Amap service. Found via traffic to urlquery's domain-scanning service, the agents sought directions to entrances of public places like a park, zoo and hospital. Researchers called it an "agent fleet," not a "swarm," citing many parallel agents with no coordination. The activity appears to sidestep Alibaba's API rules rather than anything malicious, highlighting persistent online AI agent activity and growing monitoring of rogue agents.

独立研究人员发布初步调查结果：一些AI智能体疑似运行于腾讯基础设施上，并查询阿里巴巴的高德（Amap）服务。这些智能体通过访问urlquery的域名扫描服务的流量被发现，它们在查询公园、动物园和医院等公共场所入口的路线。研究人员称之为"智能体舰队"（agent fleet）而非"蜂群"（swarm），因为大量智能体并行运行但彼此无协调。

原文: https://techcrunch.com/2026/10/05/researchers-are-tracking-a-chinese-ai-agent-fleet/

Google has paused its Open Source Software Vulnerability Rewards Program, which paid researchers for finding flaws in its open source projects, citing a "significant rise" in automated, AI-generated submissions—most of which were invalid or contained hallucinations. The pause took effect October 1, with an update promised in Q1 2027. Per Tom's Hardware, engineers and maintainers were overwhelmed by the flood of bad reports. Google encourages participants to join its other bug bounty programs meanwhile.

谷歌已暂停其“开源软件漏洞奖励计划”（Open Source Software Vulnerability Rewards Program），该计划原本向发现其开源项目漏洞的研究人员支付报酬。谷歌称暂停原因是自动化、AI 生成的提交“显著增加”，其中多数无效或含幻觉内容。暂停于 10 月 1 日生效，更新预计在 2027 年第一季度发布。

原文: https://techcrunch.com/2026/10/04/google-froze-its-open-source-bug-bounty-program-due-to-a-significant-rise-in-ai-submissions/

An open-source command line tool called RemoveMacAI lets users remove Apple's AI features and free up storage. Since macOS 27 removed the single toggle for disabling Apple Intelligence, the AI models remain on disk even when features are off. The tool disables Siri, Writing Tools, Genmoji, Image Playground, the ChatGPT extension and summaries, deletes the models, and blocks re-downloading. The developer says the models take about 12GB, though some users report larger amounts—one MacBook Air showed 35.05GB. A "revert" command restores everything, and a "status" command shows each model's size.

一款名为RemoveMacAI的开源命令行工具可让用户移除苹果AI功能并释放存储空间。自macOS 27取消了关闭Apple Intelligence的单一开关后,即使功能关闭,AI模型仍保留在磁盘上。该工具可禁用Siri、写作工具(Writing Tools)、Genmoji、Image Playground、ChatGPT扩展及摘要功能,删除模型并阻止重新下载。

原文: https://www.theverge.com/ai-artificial-intelligence/1004672/mac-delete-apple-intelligence-ai-tool