AI 资讯日报 · 2026-10-23（第九轮）
本次共新增 71 条 AI 资讯，涵盖 5 个类别，来自 12+ 家信息源。
更新时间：2026-10-23 03:07
产业商业 (22 条)
Anthropic Public Benefit Fund: Investing in the public benefit
Anthropic承诺将部分股份捐赠给Anthropic公共利益基金。
Claude Frontier Academy: $100M to train 10,000 engineers by 2027
Anthropic投资1亿美元用于Claude Frontier Academy计划，目标2027年底培训1万名前沿部署工程师。
Barclays scales Claude to upgrade operations across global banking
巴克莱银行扩大与Anthropic战略合作，将企业级AI系统整合至全球运营。
Donating Agent Skills to the Agentic AI Foundation (70+ platforms)
Anthropic将Agent Skills开放标准捐赠给Agentic AI Foundation，已有70+平台集成。
OpenAI models coming to Google Cloud (GPT-6 Astra/Sol/Luna)
OpenAI将Codex和前沿模型带到Google Cloud，首发GPT-6 Astra/Sol/Luna。
Ask a Scientist: How can AI help doctors spend more time on collaborative care?
Google发布AMIE临床研究：AI如何帮助医生花更多时间在协作护理上。
Microsoft’s Mustafa Suleyman: AI outperforms doctors on hard cases
微软AI CEO Mustafa Suleyman称AI在疑难病例诊断上已超越医生，追求医疗超级智能。
Before the ambulance arrives: How AI is helping the NHS rethink patient care
微软AI帮助英国国家医疗服务体系（NHS）在救护车到达前重新思考患者护理。
Saint-Gobain and Microsoft Discovery build concrete vision for the future
圣戈班与微软Discovery合作构建材料科学未来。
NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations
NVIDIA通过开放模型和机器人技术变革欧洲零售业。
Meta’s Generative Ads Model (GEM): Scaling the Central Model for Frontier Ads AI
Meta推出生成式广告模型GEM，扩展前沿广告AI核心模型。
Mistral Advances AI for Engineering with NVIDIA
Mistral与NVIDIA合作推进工程领域AI应用。
MiniMax M3.1 Flash shows China AI price race has no floor in sight
MiniMax M3.1 Flash表明中国AI价格战尚无底限。
OpenAI acquires startup Orchestra Bio to beef up biotech business
OpenAI收购初创公司Orchestra Bio以增强生物技术业务。
Replit makes OpenAI its default over Anthropic
Replit CEO称不再与Anthropic合作，将OpenAI设为默认。
Tesla starts Optimus robot production in California as profit falls 43%
特斯拉开始在加州生产Optimus机器人，利润下降43%。
Meta to create a new class of shares to fund AI push
Meta计划创建新类别股票以资助AI扩张。
It’s a good time to sell gray paint (AI data center boom)
AI数据中心热潮中，灰色涂料销量暴涨。
Deel launches Pangaea AI models for HR and payroll, betting on specialist models
Deel推出Pangaea HR/薪资AI模型，押注专用模型。
OpenAI invests in brain implants startup to detect cancer
OpenAI投资脑植入物初创公司用于癌症检测。
Disney asked Character.AI to go bankrupt to avoid lawsuits
迪士尼要求Character.AI破产以逃避诉讼。
Mark Zuckerberg agrees to stay in California despite billionaires’ tax ballot measure
扎克伯格同意留在加州，尽管面临富豪税投票提案。
安全政策 (9 条)
Patch the Planet: scaling global response to critical vulnerabilities
Patch the Planet倡议：AI代理发现全球关键漏洞的协同响应行动。
Major US banks locking out AI agents that log in with customer credentials
摩根大通、花旗、美国银行、摩根士丹利等美国主要银行开始限制使用客户凭证登录的外部AI代理。
Chinese hackers spied on thousands of companies after breaching IT giant ConnectWise
中国黑客通过入侵IT巨头ConnectWise监视数千家公司。
Anthropic pitches AI safety bill that may block a model if 50+ people could be killed
Anthropic推动AI安全法案：模型若可致50人以上死亡将被阻止发布。
Someone hacked Switch 2 by using an old game and AI
有人利用旧游戏和AI破解了Switch 2。
OpenAI, Anthropic, Nvidia had record lobbying quarters
OpenAI、Anthropic和Nvidia创下纪录的游说季度支出。
Hackers exploited PHP vulnerabilities days before they were patched
黑客在PHP漏洞被利用前数天已发动攻击。
AI guardrails aren’t stopping criminal hackers
Wired报道：AI护栏未能阻止犯罪黑客。
Design flaw in Meta Muse share links exposes users’ personal conversations
Meta Muse分享链接设计缺陷暴露用户私人对话。
工具产品 (18 条)
Introducing Claude Haiku 5.5 — fastest, most capable small model
Claude Haiku 5.5是最快最强的小模型，适用于摘要、子代理和大规模浏览器应用。
Introducing Claude Sonnet 5.5 — 30%+ faster, up to 30% less cost
Claude Sonnet 5.5性能提升，运行速度提高30%+，成本降低高达30%。
How we built ChatGPT’s novel ad stack for user trust and relevance
OpenAI揭示ChatGPT广告系统构建内幕，围绕用户信任和广告相关性设计。
Gemini 4 Argon: Built to solve your hardest problems
Google DeepMind发布Gemini 4 Argon，专为解决最困难问题而生。
AI agent at Bar-Ilan University helps students catch up after missed lectures
巴伊兰大学的AI代理帮助学生补上错过的课程。
Introducing Microsoft’s latest custom chip: Maia 300
微软发布最新自研芯片Maia 300。
Meta built Forum with AI-Native Engineering
Meta揭示论坛产品如何通过AI原生工程构建。
Voxtral 3 — Mistral’s new speech AI model
Mistral发布Voxtral 3语音AI模型。
Jackrong’s Qwopus V2 slashes AI agent reasoning time on local hardware
Jackrong发布Qwopus V2，大幅缩短本地硬件上的AI代理推理时间。
Microsoft’s PDD Kit makes formal math proofs a gate for AI Rust code
微软PDD Kit将形式化数学证明作为AI Rust代码的门槛。
Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately
IBM Granite Speech 5.0：更准确的转录、翻译和音频属性识别。
Canva releases its first image generation model
Canva发布其首个图像生成模型。
Meet Mori: A cute robot mushroom companion for $649
Mori：649美元的蘑菇形AI伴侣机器人开售。
Anthropic’s answer to Muse: Claude Duet, a proactive AI agent at $20/mo
Anthropic推出Claude Duet应对Meta Muse，月费20美元，主动式AI代理。
星辰意图理解模型Xing4.0-29B-Intent发布——能’做选择题’也能’做填空题’
中电信人工智能科技公司发布开源意图理解专用模型Xing4.0-29B-Intent。
国模Agent第一！MiniMax爆改Flash战力
MiniMax M3.1 Flash在国内Agent基准测试中排名第一。
Apple: Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
Apple公布Mac Studio上的MLX分布式LLM推理方案。
GitHub: 3 Java Champions share Copilot productivity tips
3位Java Champion分享GitHub Copilot提高生产力的技巧。
学术前沿 (15 条)
Google breaks open a bottleneck in quantum computing (magic state cultivation)
Google在量子计算中打破瓶颈——通过magic state cultivation技术。
The Thin Orchestrator: Making AI agents more reliable by giving router less to do
Meta发表AI代理架构研究：精简编排器以提高可靠性。
Now grammar rules alone can trigger a sharp order transition
研究发现局部语法规则可触发急剧有序转变。
How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77%
研究表明，专门化小模型词汇表可将推理成本降低77%。
NVIDIA Kumo Relational — outperforms supervised ML with no training
NVIDIA Kumo Relational模型在关系数据上超越监督机器学习，无需训练。
HGM-Dual: Lineage-Aware Guided Agent Harness Optimization for Long-Horizon Planning
血统感知的代理框架优化，用于长期规划任务。
GuideEvolve: Self-Evolving Agent Framework with Expert Guidance
具有专家指导的自进化代理框架，3Ts机制。
ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination
心智理论连续Token表示，用于多代理协调。
UserSWE: Evaluating Coding Agents in Multi-Turn Conversations with Dynamic Users
评估编码代理在多轮对话中与动态用户交互的能力。
PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents
预测性认知记忆用于自进化长期对话代理（NeurIPS 2026接收）。
SpatialBrain: Spatial Intelligence Growing from Experience
经验驱动的空间智能增长模型。
Events, Not Pixels: Where World Models Get the Return Wrong
世界模型在回报计算上的错误——应当关注事件而非像素。
Recursive self-improvement: Flint model is scary good
Wired：递归自我改进的最新模型Flint表现惊人。
Why AI isn’t ready to tackle our convoluted genomes
Philip Ball：为何AI尚未准备好应对复杂的基因组。
Plan or Retrieve First? Apple’s scale study of tool discovery architectures
Apple发布工具发现架构的大规模研究：先计划还是先检索？
观点 (7 条)
NVIDIA pitches AI factories, but 81% of companies say GPUs run at half capacity
VentureBeat报道：NVIDIA力推AI工厂，但81%的公司表示GPU利用率不足一半。
Only 13% of companies see real returns from AI, Gartner survey finds
Gartner调查显示，仅13%的CIO认为AI带来重大或变革性回报，58%的CIO被要求通过AI实现成本节约。
Anthropic workforce becoming less junior, spending more time in meetings
Anthropic表示其团队初级员工减少，会议时间增加。
Google’s AI overview led people to think they were donating to James Talarico
Google AI概览功能导致人们误以为向James Talarico捐款，实际资金流向了他人。
Wildwood director: AI is antithetical to everything that we are at this studio
莱卡动画《Wildwood》导演Travis Knight称AI与工作室理念背道而驰。
Europe wants data centers, climate goals in the way
Europe面临数据中心需求与气候目标之间的矛盾。
An AI agent joined a listserv of top internet researchers
AI代理加入顶级互联网研究人员的邮件列表，引发混乱。