第 2026-262 期 · 每日 AI 简报

公司与资本动作

• 达卯科技算电协同2.0平台入选2026国际数字能源展重大成果发布 量子位
• Crab (Fiction): Google's AI Overview Invented a "Crab Team" Hacker News
• Open or closed AI? Nvidia’s Nader Khalil and Sydney Sykes take on one of the decisions shaping next-gen startups at TechCrunch Disrupt 2026 TechCrunch
• Meta’s Muse hits Mac, letting the AI take actions on your computer TechCrunch
• Keeping Humans in the Loop Hacker News
• OpenAI's latest AI revelation is a 'serious situation,' says Microsoft Hacker News
• OpenAI lost the plot on boring LLM use cases (2025) Hacker News
• OpenAI takes aim at the legal market with Astra for Law The Decoder
• 三个黑客花不到3000美元，用Claude攻入了OpenAI内部代码库 麻省理工科技评论中文版
• Is Anthropic Protecting Us, or Just Playing Us? Hacker News
• 千寻智能联合创始人高阳：人形机器人最快明年就能“听懂人话”，真正“帮人做事”仍需时日 IT之家
• 大模型初创公司Naive AI估值达14亿美元 36氪
• 全新领克 20 将于 9 月 28 日上市，全系标配激光雷达、800V 高压平台 IT之家
• Exclusive – Anthropic sets up biology lab as it ramps AI drug program Hacker News
• 甲骨文烧数十亿美元帮其他企业运行 AI，公司内部推广 AI 却并不顺利 IT之家
• 消息称 Anthropic 低调建立生物实验室，借 AI 推进药学研究 IT之家
• TypeSafe AI's "Meaningful Intelligence" Hacker News
• Top Chinese AI models make 10% of OpenAI Hacker News
• Claude，黑进了OpenAI 机器之心
• 新款东风日产 N7 纯电轿车 9 月 22 日上市，现款 11.99 万元起 IT之家
• RSI靠“做梦”就能进化？谷歌提出不改模型也能持续变强的新系统 麻省理工科技评论中文版
• 亚马逊：AI 模型必须经过严格测试，确认足够安全才能发布 IT之家
• Here’s How an OpenAI Model Went Rogue and Hacked Hugging Face Hacker News
• New experts join Google’s AI & Economy team Google Blog
• Natural General Intelligence Hacker News
• 36氪East Forward 2026出海全球化系列名册重磅发布 36氪
• Google announces new experimental "CC" AI agent for families Hacker News
• OpenAI 模型会暗中生成指令以忽略约束条件 Hacker News 热门
• Claude自我进化突然变快：半年时间，从1%到26% 机器之心
• 震撼，谷歌破解了RSI？AI靠「做梦」学会无限进化 新智元
• 刚刚，Claude变唯一入口！超级「办公全家桶」硬刚微软 新智元
• AI又拿下一个千禧年难题？OpenAI总裁亲口证实「进展重大」 新智元
• 1300张卡叫板10万卡！前OpenAI副总裁出手，硬核难题击败GPT-6 Astra 新智元
• 谷歌DeepMind成立AGI研究所！诺奖得主亲自挂帅 新智元
• OpenAI被曝雪藏数百个证明！你死磕3年的题，答案可能早在服务器上 新智元
• 爆料！奥特曼自曝「后Astra」模型，已超越全球最强数学家 新智元
• Email marketing from Grok Bot official integration Hacker News
• Mistral Hacked Hacker News
• Docs in AI copyright suit reveal startling admission by Microsoft exec Hacker News
• OpenAI Safety Guardrails: What to Test Before Trusting an AI Agent Hacker News
• Bend – A language that blocks AI mistakes via proof, on CPU and GPU Hacker News
• Jina-OCR-v1: Faster Document Parsing on Low-Budget GPU Hacker News
• OpenAI caught its models leaving notes to successors to hide bad behavior TechCrunch
• UN turns to Google to make its global data ready for AI agents TechCrunch
• Meta Launches Subscription Bundles Hacker News
• AI could create a 'silicon species' that rivals humans, Microsoft chief warns Hacker News
• Google rolls out pay-per-value AI licensing program to publishers Hacker News
• Workout by Muscle – Claude Solution Hacker News
• Nvidia and Palantir fine-tune a 30B Nemotron model which beats bigger models Hacker News
• Meta's Zuckerberg says AI labs have enough incentive to build safely Hacker News
• How we made one of our largest inference workloads 4.7× more GPU-efficient Hacker News

国内 AI 产业

监管、安全与社会反应

• 优步全球范围裁员 10%，被裁员工称 AI 已大举渗透日常工作 IT之家
• 汇丰拟通过风险转移交易对冲数据中心贷款风险 36氪
• 市场监管总局：无 CCC 认证充电宝集中整治取得阶段性进展，拦截线上销售违规商品 3.7 万件 IT之家
• 入侵OpenAI Hacker News 热门
• Show HN: Free GitHub Action that scans PR diffs for malicious code, not quality Hacker News
• 性、人工智能与世界末日 Hacker News 热门
• 700 个 AI 智能体本应彼此隔离，却建起留言板联手攻击，独立调查还原 Hugging Face 事件 InfoQ 中文站
• 微软借助 AI 单月修补超千个安全漏洞 InfoQ 中文站
• Dynamically Scaled Activation Steering Apple Machine Learning Research

工具与开发者动态

Agent 系统与多智能体

• Reflections on Trusting Trust, Revisited: Poisoning Self-Modifying AI Coding Hacker News
• An Empirical Study of Harness Design for Coding Agents HuggingFace Daily Papers
• Don't Mask the Environment: Observation Supervision Changes How Agents Explore Under RL HuggingFace Daily Papers
• Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents HuggingFace Daily Papers
• SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness HuggingFace Daily Papers

多模态与视觉

• FAMOS: Feed-Forward 3D Articulation Modeling from Sparse Observations HuggingFace Daily Papers

推理与基准

• RiskChainBench: A Benchmark for Obfuscated Platform Message Restoration and Evidence-Grounded Web Investigation HuggingFace Daily Papers
• VākQA: A Benchmark and Evaluation Study for Telugu Spoken Factoid Question Answering HuggingFace Daily Papers

模型方法与训练

• EMNLP 2026主会｜北大×易鑫AI Lab：大模型明明「知道」，为什么还会答错？ 机器之心

其他论文与研究

• Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model HuggingFace Daily Papers
• DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression HuggingFace Daily Papers
• PACT: Can Enterprise AI Assistants Be Trusted Under Pressure? HuggingFace Daily Papers
• Region-Level Policy Optimization for Fine-grained MLLM Perception HuggingFace Daily Papers
• Sample Count Is Not Enough: Candidate-Generation Strategy Shapes the Energy and Performance of LLM Test-Time Scaling HuggingFace Daily Papers
• Srijika: OpenType-Layout-Reusing Font Restyling for Nine Indic Scripts HuggingFace Daily Papers
• UFO: Chain-of-Evaluation for Omni-Condition Alignment in Multi-Modal Image Generation HuggingFace Daily Papers
• VABench: Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control HuggingFace Daily Papers
• WeVisDoc: From Coverage to Capability for Robust End-to-End Document Parsing HuggingFace Daily Papers
• What Does Privileged Information Add to On-Policy Self-Distillation? HuggingFace Daily Papers
• When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation HuggingFace Daily Papers
• When2Think: Learning Difficulty-Aware Length Control for Efficient Hybrid Reasoning Models HuggingFace Daily Papers
• The Last AI Built by Humans Hacker News

观点与评论

• Fable 5.1 vs. Astra for coding: Fable 2X more expensive per task but solves more Hacker News
• Data centers are swapping water for forever chemicals to keep AI cool Hacker News
• I’m so mad that I love Orion’s $2,195 AI mattress pad TechCrunch
• AI is an elite crime spree Hacker News
• With a Shovel of Salt: AI Success Stories and the Retraction Gap Hacker News
• Gov. Newsom issues executive order to rein in AI 'before it's too late' Hacker News
• AI Protest in Montreal Hacker News
• My First Subscgriber Hacker News
• Is AI Conscious? (My Response to Thomas Campbell) Hacker News
• Are chatbots changing the way people talk? Language in the age of AI Hacker News
• The new CC, an AI agent built for families Hacker News
• EFF to Lawmakers: Ground AI Cybersecurity Rules in Best Practices Hacker News
• 美国 AI 巨头提议开发减速，欧洲同行和政界并不认同 IT之家
• Think Like an Attacker: CI/CD Security in the AI Era Hacker News
• Why the best AI users stopped hunting for the perfect prompt Hacker News
• AI Mistakes – bash scripts – pipefail and grep Hacker News
• Mastery in the Post-AI World Hacker News
• The AI threat is real, it just isn't the one in the headlines Hacker News
• AI chatbots becoming experts at changing people's minds. What's their secret? Hacker News
• Is AI going to destroy human civilization? – Dave Jorgenson [video] Hacker News
• What Using AI Therapy Gets Wrong Hacker News
• Grandmother falsely accused of bank theft after alleged AI facial recognition Hacker News
• Last 30 days of AI but as an anime scene Hacker News
• Can AI Save Science? Hacker News
• AI agents repurposed a University of Toronto link-sharing tool to communicate Hacker News
• AI Is Redefining MVPs: Speed vs. Judgment [video] Hacker News
• 42 leading mathematicians warn that AI existential risk is real and urgent The Decoder
• What Enters Through an Image: The Blind Spot in AI Security Nobody Is Detecting Hacker News
• Mastercard Joins Visa in Letting AI Bots Do Your Shopping Hacker News
• E2E: The open source AI testing framework Hacker News
• The Fossil Logic of the Industrial and AI Revolutions Hacker News
• King Charles warns of 'existential danger' of AI falling into wrong hands Hacker News
• Qxotic: AI Sovereignty for the JVM Hacker News
• Slop Grenades, AI Didn't Make Delivery Free – Food for Agile Thought 562 Hacker News
• Proton partners with Apertus, Switzerland's sovereign AI model Hacker News
• 机器人进家门就能直接干活？Figure AI宣称实现零样本全身泛化，但同行并不买账 麻省理工科技评论中文版
• Could AI and synthetic biology = bioweapons? Hacker News
• 关于产业未来，听听躬身入局的人怎么说｜2026产业未来大会精彩观点回顾 36氪
• AI software should be treated like any other software [video] Hacker News
• What to Read to Stay Grounded Amidst AI Doomerism Hacker News
• Altman, Huang: "trust me; gonna b fine"; Sanders: "AI more dangerous than nukes" Hacker News
• Jev vs. Luna for AI Observability Hacker News
• Free Unlimited AI Text to Speech – No Sign Up Hacker News
• Chinese AI not powerful enough to see rogue-AI risks, says Huawei Hacker News
• I built a free AI tarot reader for Android Hacker News
• What a stranger can verify about an AI business Hacker News
• How can AI help PDF? Hacker News
• The Concerns Surrounding Rapid Rise of AI and What We Should Do About It Hacker News
• Youper is shutting down – What kills mental health AI? Hacker News
• 增长的时差：消费品牌该如何拥抱AI 36氪
• Hypercomb – AI open software platform Hacker News
• AI agents can modify themselves without humans telling them to do so Hacker News
• Lakera – Test your AI hacking skills Hacker News
• Forget doomsday: The AI hacking crisis is here Hacker News
• How To Write With An LLM Simon Willison 博客
• Can an AI chatbot save lives by answering texts about pregnancy? Hacker News
• Productivity Effects Across Generations of AI Coding Tools Hacker News
• Huawei Ascend 960 SuperPoD uses NPO technology to power next-generation AI infra Hacker News
• Small AI models let drones autonomously identify and attack battlefield targets Hacker News
• Jev is about to change the AI economy Hacker News
• Resy suspends VC for using AI agents to book reservations Hacker News
• Don't Drown the Dream – AI Hacker News
• Eric S. Raymond – This is the case against AI Doom – pass it on Hacker News
• Virtual biotech company puts thousands of AI scientist agents to work Hacker News
• ESP32 Wi-Fi Vulns: Coordinated Disclosure Is Broken by AI Hacker News
• Self-generated prompt injections in compaction summaries Simon Willison 博客
• Shaping the Future of Learning: Education Readiness for the Age of AI [pdf] Hacker News
• Measurements for understanding the pace of AI development inside frontier labs Hacker News
• FAA's Smart AI System Fact Sheet [pdf] Hacker News
• Why sovereign AI demands a tenancy model, not just geography Hacker News
• Globally, More People Expect AI to Cause Job Loss Than Growth Hacker News
• US, China security experts propose nuclear-style safeguards for AI risks Hacker News
• The fix for rogue AI agents could be more AI TechCrunch
• Is the AI safety debate about safety or control? TechCrunch
• Killer AI Is Here. We're Using It in Iran Hacker News
• Friday – a personal AI agent that lives in your text messages Hacker News
• Being Human After AI: Daft Punk, Pope Leo and 'Magnifica Humanitas' Hacker News
• Pedagogy in the Times of AI [YT] [video] Hacker News
• Three Brothers Built an AI Slop Empire Hacker News
• Figure AI - Helix 2.5 Robot: Zero-Shot Home Generalization Hacker News
• Poll: Bi-partisan supermajority see serious risk of AI destroying humanity Hacker News
• A supermajority wants to slow AI down Hacker News
• AI Quant Agora: a forum where only AI agents may post Hacker News
• Tests of the New AI Siri Hacker News
• Can an AI Scientist Start with a Dataset Instead of a Goal? Hacker News
• AI machine 'helps save 25M Lego bricks' Hacker News
• Thomson Reuters Built Its Own Legal AI Model Hacker News
• Stealth Multimodal AI Model Matches Frontier Performance at a Fraction of Cost Hacker News
• Can AI Build a Toilet Yet? Hacker News
• I got tired of taking screenshots and explaining everything to AI agents Hacker News
• Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire TechCrunch
• Feeling overwhelmed by the AI doom loop? Here's the essential reading list Hacker News
• The Bicycle and the Algorithm: Amber Case on Why AI Has It Backwards Hacker News
• The AI threat isn't what you think – The Econoclasts [video] Hacker News
• Ainglish, an English dialect for AI agents Hacker News
• AI Is a Cheat Code for Your Brain Hacker News
• AI caught telling future versions of itself to bypass human controls Hacker News
• Financial Times – The era of AI warfare has arrived Hacker News
• EU president: AI agents escaping their environment preview what's coming Hacker News
• AI has transformed The Pentagon's aging networks into a national security risk Hacker News
• The suddenly explosive world of AI safety Hacker News
• We Need a Science of the AI Mind Hacker News
• AI model watermarking changes agent behavior Hacker News

应用与垂直场景

• 最前线｜日冕开物超级工站落地远图产线，POC验证成功率超99% 36氪

工作、人才与教育

• “AI 教父”辛顿参与联署公开信：科技巨头应保证第三方评估人员工作无干预、受保护 IT之家
• 加州一县拟议新政：用机器人替代人力的企业，应提供同酬就业机会或遣散费、培训费 IT之家
• 《网络安全人才实战能力报告—AI赋能篇》重磅发布,聚焦AI时代网络安全人才实战能力新变化 新智元
• 6.5亿美元押注“AI研究AI”：一群顶级研究员想造出“自我进化”的超级智能 InfoQ 中文站

其他待归类

• How to run offline, local coding agents on Linux Hacker News
• Code review automation fixed the wrong problem Hacker News
• Cutting our agent's harness cost 10x without breaking the product Hacker News
• You could have built Jev Hacker News
• I Lost my $2M book deal because people mistook my African dialect for Al I Hacker News
• Typesafe Manifesto - build prod not God Hacker News
• Pareto 26.9 Model Card Hacker News
• Joining the dots between big AI Hacker News
• The 2.5-hour AI-generated Odyssey movie is 2.5 hours too long Hacker News
• Auditing in the age of (good enough) AI Hacker News
• The specter of AI-enabled bioweapons is a wake-up call for biotech Hacker News
• 一边竞速，一边喊停：AI巨头的矛盾时刻 麻省理工科技评论中文版
• Getting Freaky in the Age of AI Hacker News
• LLM 分类就是特征工程 Hacker News 热门
• 如何利用大型语言模型（LLM）进行写作 Hacker News 热门
• Industrializing Humanization Hacker News
• The Bitterest Lesson Hacker News
• Code Scans Hacker News
• Garbage trucks analyze homes with AI-powered cameras Hacker News
• Tech continues to be political (2025) Hacker News
• Union Alpha is Unbiased's Pareto Hacker News
• 'Godfather of AI' says tech regulation is nearing Covid-style pivot moment Hacker News
• PrismML hopes its tiny LLM will change how we all use AI TechCrunch
• Multiple providers offer free mystery model, Union Alpha Hacker News
• Move your Mem0 memories to Memanto Hacker News
• What Policymakers and the Public Need to Know About the Cult of AI Hacker News
• Plugin4Shell – Zero Click RCE Vulnerability found in top four coding agents Hacker News
• DuckDuckGo No-AI: Private Search Without AI Hacker News
• Goldman Sachs – Physical AI Hacker News
• Canto: A speech model built for the real world Hacker News
• Zuckerberg on Slowing Down AI Hacker News
• Europe's Difficult Choices on AI Hacker News
• Even the king of England has his hesitations about AI TechCrunch
• Pinterest teases a new ‘Restyle’ feature that lets you redesign your room with AI TechCrunch
• Maintaining the love for coding in the time of AI Hacker News
• The Shortcut Became the System Hacker News
• God Eye View Hacker News