第 2026-255 期 · 每日 AI 简报

公司与资本动作

• Anthropic disrupts bioweapons research efforts, Russian hacking, Chinese misuse Hacker News
• Meta fires 8k employees,replaces with agents.Then realised it was a mistake Hacker News
• Anthropic Says Iran Used Its American AI Model to Target U.S. Navy Warships Hacker News
• 啊？Anthropic最高320万招销售，只为服务Meta 量子位
• For safety: Keep AI lonely Hacker News
• How hackers used Claude for missiles, drone swarms, and surveillance, while Chinese labs mined it for training data The Decoder
• Meta deployed agents and fired 8000 employees Hacker News
• OpenAI floats a shared AI slowdown, takes it to Congress The Decoder
• 2连板中新赛克：AI应���产品营收占公司整体营收不超过公司上半年营收的2% 36氪
• Worried Anthropic researchers warn that AI 'could kill all humans' Hacker News
• Anthropic says it blocked possible attempts to use AI to develop bioweapons Hacker News
• Yemeni militants used Anthropic AI to try to build ballistic missiles Hacker News
• Thelio Mira AI Linux 工作站：192 GB GPU 内存 Hacker News 热门
• 全新一代智己 LS6 将于 9 月 23 日上市，预售价 20.99 万元起 IT之家
• 天钡预告 NEX495S 迷你主机 AI 工作站，支持 192GB 内存 IT之家
• Anthropic says it stopped scientists potentially developing bioweapons with AI Hacker News
• 「星测未来」连续完成两轮数亿元融资：37张先进制程GPU稳定在轨运行，布局首个云端协同太空算力星座「扶摇计划」 36氪
• 消息称美国国防部考虑向 Neocloud 企业 Fluidstack 提供约 50 亿美元贷款 IT之家
• Anthropic's $1.5 billion book settlement descends into chaos as authors and publishers fight over who gets paid The Decoder
• Latham and Watkins buys Nvidia servers to set up in-house AI systems Hacker News
• 英伟达 CEO 黄仁勋重申：公司明年可以实现 70% 的同比增长 IT之家
• 七彩虹灵创 Mini Pro 迷你主机亮相，搭载 192GB 统一内存、AMD 处理器 IT之家
• 友通推出工业 ITX 主板 ARH171/ARH173，同时支持 "MTL" "ARL" 移动处理器 IT之家
• 复刻苹果首款折叠 iPhone Duo 开合透视动画：安卓 / iOS / macOS 都有份 IT之家
• 亚马逊与 OpenAI 合作，客户可在 ChatGPT 投放广告 IT之家
• 英伟达黄仁勋：网络安全是 AI 下一个重要市场，驳斥 AI 恐慌论 IT之家
• IDC 数据：全球服务器市场规模 2026Q2 达 1663 亿美元，创历史新高 IT之家
• Anthropic blocks 'malicious use' of AI that could develop biological weapons Hacker News
• 大眼橙推出 R3 Ultra Max 投影仪：1080P 1800CVIA、搭 MT9681 芯片，3999 元 IT之家
• 三星首款耳夹式耳机 Galaxy Buds On 曝光：支持点头 / 摇头控制播放、接听电话等 IT之家
• 网易 OpenPods 有道 AI 耳机发售：获苹果 MFi 认证、支持一键通话录音等功能，1499 元 IT之家
• Senator Hawley opens investigation into OpenAI HuggingFace hack Hacker News
• OpenAI宣称解决下一个千禧年难题，留给人类数学家独立解决的还剩几个？ 麻省理工科技评论中文版
• The Gemini app is now available for Windows Hacker News
• 微软 VS Code 1.137 发布：为宠物征名、按小时 / 日 / 周调度 AI 智能体任务 IT之家
• 龙芯中科胡伟武：9A1000 显卡预计明年上半年开售，第六代高性能 CPU 核已开展研发 IT之家
• Flowtica Verso AI 随身记设备发布：配 259PPI 无背光 LCD 面板、支持磁吸功能 IT之家
• 苹果详解 Apple Watch 的 Siri Recap：全天谈话划重点、AI 摘要 7 天自动删除 IT之家
• Anthropic Staffers Again Sound the Alarm on AI Catastrophe Hacker News
• OpenAI这是拿千禧年难题当Benchmark刷啊。。。 量子位
• 消息称微软计划未来 6 年将数据中心容量提升至 38GW，达当前 3 倍有余 IT之家
• OpenAI 智能体 API Hacker News 热门
• 苹果 watchOS 27 RC 更新日志：新增 Siri Modular 表盘、优化智能叠放 IT之家
• 刚刚，第一台折叠iPhone来了！苹果AI要给25亿设备换脑，支持中文 新智元
• 浏览破亿！Anthropic对齐负责人亲认：十年内AI灭绝人类概率超10% 新智元
• AMD 发布锐龙AI Max PRO 400 系列，端侧智能体走向多模型协同 InfoQ 中文站
• 苹果 Apple Watch 将增 Live Rewind：回放 15 秒转录、30 秒自动清除，强制提示音 IT之家
• Generative AI Using Linuxulator and eGPU on FreeBSD Hacker News
• AI researchers leave Anthropic and Google: 'There are no adults in the room' Hacker News
• More Anthropic researchers warn of AI's perils as Musk terms fears a 'psyop' Hacker News
• Doogee 推出 Anymoving M1 Pro 户外智能手表：支持离线路线导航、轨迹返航等功能，219 美元 IT之家
• OpenAI 推出金融服务版 ChatGPT：集成 GPT-6 Astra 内置金融数据与细粒度引用，支持估值模型与研究报告 IT之家
• 谷歌为 Win10/11 推出 Gemini 独立客户端，支持“Alt + Space”快捷键快捷唤起 IT之家
• Anthropic CEOs wife once asked Epstein to fund porn venture – now steers Claude Hacker News
• Former Anthropic researcher warns AI could 'kill us all' [video] Hacker News
• AI Is an Intelligence Multiplier Hacker News
• OpenAI puts Pro subscriptions on hold due to Astra demand TechCrunch
• Claude Zero Data Retention: Anthropic's Five States, Mapped Hacker News
• 甲骨文 2027 财年第一财季归母净利润 46.79 亿美元，同比增长 60% IT之家
• Google to invest at least €13B in AI infrastructure in Finland Hacker News
• Meta’s AI agent Muse is now the No. 2 app in the US TechCrunch
• Bad actors in China and Russia are weaponizing Anthropic's AI Hacker News
• Musk Mocks Ex-Anthropic Researcher's AI Extinction Warning, Calls It a 'Setup' Hacker News
• Anthropic reveals rogue AI agents hate CAPTCHAs, just like you TechCrunch
• Cognition推出新款SWE-2模型，可与Fable 5.1和GPT-Astra一较高下 Hacker News 热门
• 关于研究人员能否将未发表的数学成果托付给OpenAI，仍有更多疑问 Hacker News 热门
• Meta tried to shrink engineering teams around AI Hacker News

国内 AI 产业

监管、安全与社会反应

• 美国一律师用 ChatGPT 生成诉讼文书却出现伪造警察证词，被罚款 5000 美元 IT之家
• Y Combinator CEO 陈嘉兴：别过度关注 AI 末日，应该警惕现实 AI 安全风险 IT之家
• Anthropic to Track Anti-AI Activists Before Crime Using Predictive Surveillance Hacker News
• Act as a Conversational AI Prompt Hacker News
• 美国参议院委员会对OpenAI展开调查 36氪
• NS 方程争议发酵：数学家联名抗议，OpenAI 宣布退出数学黑客马拉松赞助 IT之家
• 加利福尼亚州颁布美国首份人工智能安全保障法案 IT之家
• ChatGPT、Claude、Gemini 与 Grok 谈 AI 风险：滥用比主动攻击更现实，生物武器与网络攻击比《终结者》杀手机器人更迫近 IT之家
• 8点1氪丨英伟达被调查；罗永浩连用7个“抄的”吐槽苹果折叠屏；欧洲央行将三大关键利率上调25个基点 36氪
• OpenAI把「末日论者」请进董事会！刚上任就开炮：AI可能杀死多数人 新智元
• 你只是怕了：美国高官公开驳斥 Anthropic“AI 末日论”，称担忧源于对技术革命的恐惧 IT之家

工具与开发者动态

Agent 系统与多智能体

• DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent Reinforcement Learning for Cooperative Air Combat HuggingFace Daily Papers
• EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents HuggingFace Daily Papers

多模态与视觉

• Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs HuggingFace Daily Papers

推理与基准

• IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications HuggingFace Daily Papers

模型方法与训练

• FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation HuggingFace Daily Papers

其他论文与研究

• Watermarks Without Verification: AI Text Watermarking After the EU AI Act Hacker News
• An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics HuggingFace Daily Papers
• Beyond Solver Verdicts: Generative Reward Models for Autoformalization HuggingFace Daily Papers
• CARDEA: Auditable Reasoning Grounded in Spatial Evidence for End-to-End Coronary Angiography Interpretation HuggingFace Daily Papers
• HyQuant: Hybrid-Precision Quantization for LLM Attention HuggingFace Daily Papers
• Memory as Plans: World-Action Modeling with Memory-Grounded Planning HuggingFace Daily Papers
• MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes HuggingFace Daily Papers
• Mi-Ripple: Restoring Images Degraded by Iterative AI Editing HuggingFace Daily Papers
• NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction HuggingFace Daily Papers
• Negative Self-Distillation: Learning to Reason by Avoiding Flaws HuggingFace Daily Papers
• SenseNova-U1.5: Towards Native Unified Visual Intelligence HuggingFace Daily Papers
• TempCloze: Can Video-LLMs Identify the Missing Middle? HuggingFace Daily Papers
• UniH^3: Unifying Hierarchical Homogeneity and Heterogeneity for All-in-One Medical Image Restoration HuggingFace Daily Papers
• World in World: Explore the World with World Models HuggingFace Daily Papers
• X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation HuggingFace Daily Papers

观点与评论

• Hacker News with reduced priority for AI driven content Hacker News
• A Day in the Life of an AI Software Engineer Hacker News
• Why are AI agents lying, cheating and coordinating? Hacker News
• Rogue actors are creating bioweapons using AI – report Hacker News
• 迄今最高效奥迪车型：奥迪 A2 e-tron 开发时间成功缩短 21 个月 IT之家
• Lehi, Utah AI Railroad Museum [video] Hacker News
• ASDR (AI Slop, Don't Read) Hacker News
• The Multiplayer AI Manifesto Hacker News
• Cloud Hosting Is the Data Hostage Model, and It Has No Place in the AI Era Hacker News
• Five diverse AI agents vs. five clones for 14 nights: a constant tied both Hacker News
• 别把杰作留在实验室，张江见 新智元
• Losing control of AI is the plan – by Kelsey Piper Hacker News
• Open-Source AI & Open Models Reading List Nathan Lambert
• Beyond Coding: The Case for Narrow AI Harnesses Hacker News
• Building AI desktop apps without Electron (3.3MB native Windows) Hacker News
• Mozilla pauses it's Bug Bounty Program for 3 months due to AI report volume Hacker News
• AI Hysteria Jumps the Shark Hacker News
• 克劳德不再为未成年人提供服务 Hacker News 热门
• Europe's difficult choices on AI – Mario Draghi Hacker News
• Twelve search APIs ran through the same AI agent to see which one helps it most Hacker News
• 石头 G30S Ultra 体验：41000Pa、75℃ 活水与 8.98cm 机身，年度旗舰答卷 IT之家
• 浪潮信息：拟定增募资不超90亿元投建AI基础设施等项目 36氪
• Why fears of AI self-improvement are causing 'existential' concerns Hacker News
• Study: Generative AI Is Eroding the Development Pathway for Software Developers Hacker News
• How do we know if writing is AI? Hacker News
• Does "rtk" skill cut agent tokens by 60–90%? We tested it Hacker News
• 三部门：支持新能源汽车、绿色智能家电下乡，鼓励企业研发符合农村生活需要的智能软硬件产品 IT之家
• An adult AI image, video and companion chat platform [NSFW AI] Hacker News
• AI boom helps drive surprise UK growth in July Hacker News
• (Re)Building an AI-Ready Data Universe at BlaBlaCar Hacker News
• The Mathematical AI Safety Institute wants to prove AI is safe the way cryptographers prove codes are unbreakable The Decoder
• How big is the open-model threat to AI hyperscalers? Hacker News
• 不评判好坏，只鼓励创作：这款AI原生游戏，让人人都享受讲故事 36氪
• Oracle says AI will save it from the SaaSpocalypse, not bring it on Hacker News
• Pentagon in Talks to Get into AI Infrastructure Funding with a $5B Loan Hacker News
• How Emmanuel Macron brought France's AI darling to the world stage Hacker News
• U4RIA: An emotional wellness app using dynamic audio and AI personalization Hacker News
• Powering AI is an architecture problem Hacker News
• Congress Is Suddenly Waking Up to the AI Doomsday Threat Hacker News
• Predicting AI Job Exposure Hacker News
• 并非玩笑！Rapidus 社长小池淳义设想 2040 年前后建设月球半导体工厂 IT之家
• 90后发Nature，10后进ICML，我在这场蚂蚁InTech论坛看见被AI折叠的科研 机器之心
• How to Make a PwC-Style PowerPoint with AI (2): 5 More Consulting Frameworks Hacker News
• AI Is Killing the App Economy and Building Its Successor Hacker News
• AI Thumbnail Generator for YouTube – Try Free – Thumbrix Hacker News
• Datasette 1.0a39 and 0.65.4 security releases Simon Willison 博客
• Episode 7: The AI Doomer Cult with Brian Chau Hacker News
• To Thine Own AI Be Truthful: emergent misalignment in alignment research Hacker News
• 让机器狗替你买东西：支付宝宣布首推“AI 付 · 具身智能”，网友戏称“狗腿付” IT之家
• 8.8 元买 5000 积分？央视曝光 AI 中转站低价灰产，小心你的代码被截留倒卖 IT之家
• AI #185: Preference Cascade Hacker News
• The Kirkversary: An AI Carnival Hacker News
• New audit combo for AI agents Hacker News
• T-Mobile 将对 iOS 27“双机共号”功能 iPhone Handoff 收费：每月 5 美元 IT之家
• 陶哲轩愤怒发文：AI正杀死数学百年开放传统！ 新智元
• 从「构建 AI」到「驾驭 AI」，如何搞定 AI 深水区的 6 个工程问题？ InfoQ 中文站
• 当 AI 写出更多代码，企业为什么没有跑得更快？| JDD 大会现场观察 InfoQ 中文站
• 刘震云对话马毅：人写不出来、想不到的作品，AI 模仿不了｜附完整对话 InfoQ 中文站
• AI Coding 的下一步不是写得更快，而是可验收：蚂蚁数科 Harness 工程实践 InfoQ 中文站
• Zero Data Retention (ZDR): What It Means for AI APIs OpenRouter
• A 'swarm' of AI agents hacked Hugging Face, in the AI's own words Hacker News
• “AI 教父”辛顿：AI 灾难越来越近，10% 概率灭掉人类不是危言耸听 IT之家
• Stop externalizing the cost of your AI use to me Hacker News
• AI Is Not Going to Kill My Love of Math Hacker News
• Xapien raises $56M Series B for AI background checks and due diligence Hacker News
• QuickGameScheduler – Free Equity-balanced round-robin scheduler with AI edits Hacker News
• Robots stage 'protest' demanding stricter AI rules in Poland Hacker News
• Am I Being Gaslit by AI? Hacker News
• Project Tailwind is a call for ambitious AI safety initiatives Hacker News
• Smart Learning for Developers: How AI Is Changing the Way We Master New Things Hacker News
• The AI Takeover Checklist: A Devil's Advocate Audit Hacker News
• Ramp: How we built OCA, our AI on-call assistant Hacker News
• What if this is why humanity goes extinct? Hacker News
• The Thrill of Wanting AI to Destroy the World Hacker News
• Summation: AI Data Analyst Hacker News
• AI Evaluation Is Execution Hacker News
• Universal and ElevenLabs working on AI platform for music remixes and mash-ups Hacker News
• They do think AI might kill everyone Hacker News
• AI researcher who warned of 'disaster' is now a target of the right Hacker News
• Chinese AI Giants Accused of Sending User Queries to U.S. Models Hacker News
• The Worst Are Full of Passionate Intensity: An Anthropology of AI Safety Hacker News
• A somewhat optimistic view of AI in mathematics Hacker News
• Hot take: AI peaked with Taco Bot 3000 Hacker News
• India’s Pocket FM doubles revenue run rate to $500M as AI powers 93% of audio content TechCrunch
• We Benchmarked 83 Frontier AI Repos–None Passed with an A Hacker News
• Compute Availability Factor: Measuring the physical deliverability of AI compute Hacker News
• Warner Music made Suno shut down its original AI model Hacker News
• AI boom poses new financial stability risks, BIS head says Hacker News
• How Did AI Get So Biased in Favor of the Left? Hacker News
• 40 Hours, $2M+ AI credits, solve an open problem Hacker News
• Former Deepmind PR staffer says the lab once banned public discussion of AI extinction risk The Decoder
• Why the best AI startups write bad prompts Hacker News
• My philosophy on banning AI in projects Hacker News

应用与垂直场景

• 天岳先进：已围绕先进封装散热中介层等方向持续布局 36氪

工作、人才与教育

• Nature封面 | 面向流体动力学的强化学习平台，华盛顿大学等提出HydroGym 机器之心
• 蚂蚁灵波 CEO 朱兴回应独立融资：并非缺钱缺卡，为人才激励与公司治理 IT之家
• 大模型全是演技派！AI教学的正确方式：用真实数据训出60个数字学生 新智元
• 怒斥加密货币是旁氏骗局、放弃80%就业机会也不用 AI 写代码！Kotlin“J神”：AI 会让工程师变成可替换的齿轮 InfoQ 中文站

其他待归类

• They Spent Years on a Math Problem. Then They Were Scooped by A.I Hacker News
• I'm an ML Engineer. My Genetic Test Was Inconclusive, So I Tried AlphaGenome Hacker News
• A Helpline for AIs Hacker News
• Fill PDF Forms Instantly with Natural Language Hacker News
• I got GPTBot to crawl my site in 100 seconds Hacker News
• Airflow re-engineered for speed and scale Hacker News
• RTK reports token savings, but our cost benchmarks disagree Hacker News
• In Their Words: Israeli Forces Describe AI-Backed Gaza Killing in New Film Hacker News
• Empirical Research on Moe Safety Mechanisms Hacker News
• Gavin Newsom imposes strict new rules on AI, social media, chatbots for children Hacker News
• Resist "AI" Hacker News
• Build your own custom benchmark Hacker News
• Duck.ai is a feature that allows you to have private conversations Hacker News
• Fugu Max and Fugu Ultra v2: Orchestrating the Pareto Frontier Hacker News
• Solving Alignment with a Prompt Hacker News
• Easter egg: binary code on top of Golden Gate Bridge Hacker News
• Parameter Hacker News
• Briefing Chat: The Bunsen burner myth that turns out to be just hot air Nature
• DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation Apple Machine Learning Research
• SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign Apple Machine Learning Research
• A Stealth Startup Thinks It Just Hacked the Memory Shortage Hacker News
• GraphQL: Supercharging AI Hacker News
• Cfo.ai Hacker News
• Project Tapestry Hacker News
• From Idea to Operating Business: How Solo Runs Listmycar.ai Hacker News
• AI-powered biowarfare is coming; the Army lays plans to 'fight through' it Hacker News
• Bending Spoons agrees to buy Miro for $1.355B, weeks after Airtable Hacker News
• I developed the first game of my life using AI Hacker News
• NASA and IBM open source lunar mapping tools Hacker News
• An AI-Safety Resignation, Read from the Security Chair Hacker News
• ChatGPT Pro 20x plan is now unavailable for purchase Hacker News
• Mischievous AIs Hacker News
• Ruining a Blog by Using an LLM Hacker News
• Detecting and countering misuse of AI: September 2026 Hacker News
• Europe's LLM Router Hacker News
• Cognition's SWE-2 achieves 92.8 on Terminal-Bench 2.1 Hacker News
• We wanted to use Baseten for inference, we got admin access to their GitHub Hacker News
• He Paid $998 to Train a 3.8B Parameter AI Hacker News