California Resources 与 Bloom Energy 拟在加州部署 1GW 燃料电池为数据中心等供电Chinese
California Resources 与 Bloom Energy 计划在加利福尼亚州部署 1GW 燃料电池系统，用于为数据中心及其他电力消纳方供电。双方同时计划将该燃料电池系统与碳捕集与封存技术相结合。
All AI and data center updates
All updates
Today 84 itemsData Center DynamicsAI score4646 Data Center KnowledgeSelectedAI score6363 Trump's Call for AI Data Centers to Bring Their Own Power Faces Practical Obstacles
In response to Donald Trump's proposal that large-scale AI data centers must generate their own power and sell electricity back to communities, industry experts point out that 100% around-the-clock independent self-generation is largely impractical. Critical power equipment like gas turbines is facing severe supply shortages, with order backlogs for certain models extending to 2031, while most projects under construction have already secured grid interconnection agreements. Experts believe a hybrid model combining on-site generation, energy storage, and grid interaction is currently a far more viable path forward.
Why it matters: This article outlines the tension between policy demands and power equipment supply chain bottlenecks, helping assess the real-world obstacles that self-generation mandates face in data center deployment.
TechCrunch AISelectedAI score8282 Anthropic updates S-1 filing revealing worsening compute costs, with quarterly compute spend reaching $10B
In its updated S-1 filing, Anthropic disclosed that revenue for the quarter ended September 30 reached $17B, but compute spending surged to $10B, climbing from 48% of revenue in the prior quarter to 58%.
Why it matters: The article details the exact share of compute expenses relative to revenue and short-term premium dynamics for leading foundation model developers, underscoring the immediate impact of exploding demand on their gross margins.
Buzzing HN · 中文精选SelectedAI score6969 Namespace details how it deployed hundreds of MacBook Pros into data center racks
Namespace detailed its engineering practices for deploying custom Mac racks at scale in data centers, having chosen MacBook Pros over Mac minis due to supply chain delays and other factors.
Why it matters: The article provides a comprehensive overview of engineering solutions across power delivery, virtualization, and image distribution when adapting consumer-grade hardware into data center infrastructure.
Hacker News · AIAI score4343 如何用 AI 外联系统解决冷启动问题Chinese
初创团队可通过 Claude 构建自动化直连外联系统，以解决缺乏早期客户的冷启动难题。该方案无需 CRM 或数据管道，利用 Claude 技能（skills）与文档存储状态，自动从网络检索潜在客户并依靠脚本进行特征分类与优先级排序。系统每日 9 点会自动生成包含线索排序及个性化邮件草稿的日报，供创始人审核并发送。
Hacker News · AIAI score4141 Glyphh 驱动的 Claude 机器人实现从先例中实时学习Chinese
Glyphh 驱动的 Claude 编程机器人通过学习 103 次测试先例实现模型动态路由，在 10 项新编程任务中取得 10/10 全通，成本仅为全量 Opus 的 5%（$0.050 对比 $1.057）。
Hacker News · AISelectedAI score8282 Anthropic Files S-1/A Amendment, Disclosing It May Soon Reach AI R&D Acceleration Threshold and Plans to Maintain Human Control
On November 2, 2026, Anthropic filed an S-1/A prospectus amendment with the U.S. SEC, disclosing that internal projections indicate future models may soon approach the threshold for significantly accelerated AI R&D defined in its Responsible Scaling Policy (RSP).
Why it matters: The details in the amended prospectus clarified market overinterpretations regarding recursive self-improvement, explicitly noting that the threshold reflects forward-looking internal projections rather than a breakthrough already achieved by existing models.
The DecoderSelectedAI score6565 SETI Allen Telescope Array Deploys AI Agents to Search for Extraterrestrial Signals
Located in Northern California, the SETI Institute's Allen Telescope Array has begun using AI agents to steer telescopes and analyze incoming radio signals in real time. Built on the Nvidia Holoscan platform with hardware processing throughput of up to 2.4 Tbps, the system compresses data processing from weeks to hours and has demonstrated real-time AI detection of a pulsar in the Crab Nebula. The research team next plans to test open-source solutions running locally at the edge for future sensor deployments on the far side of the Moon.
Why it matters: The article showcases real-time analysis practices combining high-throughput data streams with edge agents in astronomical research, offering readers insights into workflow transformations in real-time signal processing.
Hugging FaceAI score5959 网络安全评测基准 CWE-bench v2 发布：覆盖 91 类 CWE 并开源 100 个真实漏洞修复任务Chinese
Collinear AI 在 Hugging Face 发布网络安全评测基准 CWE-bench v2，开源包含 100 个任务的公开测试集，覆盖 91 类 CWE 安全漏洞。
Hugging FaceSelectedAI score6262 IBM Research Introduces Palette: Enabling a 20B Model to Generate Enterprise-Grade Slides via Dual-Role Decomposition
IBM Research has introduced Palette, a presentation generation system fine-tuned on a 20B open-source model that strictly adheres to enterprise design guidelines to produce self-hostable presentations. The framework decouples the task into two roles—a Designer that generates JSON specifications and a Coder that writes pptxgenjs code—while introducing a check-and-repair framework at inference to detect text overlap and overflow.
Why it matters: This work demonstrates a viable path for small open-source models to rigorously adhere to specific design guidelines through dual-role decomposition, synthetic counterfactual reasoning data, and a layout verification framework.
Hugging FaceAI score5252 AI2 发布长文本状态追踪基准 StateBenchChinese
AI2 推出长文本状态追踪评测基准 StateBench，用于测试大语言模型在长文档中持续跟踪实体状态变化的能力。基准基于 Krohn–Rhodes 理论将复杂度划分为 r-trivial、非周期（aperiodic）、可解（soluble）与一般有限（general finite）四个等级，并开源了技术报告、数据集与代码。
Google DeepMindSelectedAI score6565 Google DeepMind Launches Cheminformatics Skills Suite for Antigravity
Google DeepMind announced the launch of Cheminformatics Skills for its scientific workbench Google Antigravity, enabling agents to invoke deterministic tools such as RDKit, ChEMBL, and PubChem to complete drug discovery workflows.
Why it matters: The article illustrates an architecture that decouples LLM task orchestration from open-source deterministic chemistry tools, offering a practical blueprint for engineering scientific agents in specialized computational science domains.
Data Center DynamicsAI score1919 爱立信 150 周年Chinese
爱立信回顾了其 150 年的发展历程，通过重温历史为面向 AI 的未来发展汲取灵感。该供应商正从过去的创新经验中获取启示，以支撑其聚焦 AI 战略的后续演进。
Buzzing HN · 中文精选AI score5555 OpenAI 疑似悄然重新开启 Codex 聊天记录的数据共享与模型训练Chinese
多位开发者在 Hacker News 反馈，OpenAI 疑似在用户不知情的情况下，将此前已关闭的 Codex 数据共享设置重新启用以用于模型训练。发帖者与同事发现其账户数据控制中的 Improve the model for everyone 选项被静默恢复开启，并提供了官方设置页面和隐私政策页面的核查关闭链接。
Buzzing HN · 中文精选AI score3333 封锁 Google AI 的学校将导致学生无法使用其搜索引擎Chinese
学校若封锁 Google 的 AI 功能，将导致学生无法使用其搜索引擎。这意味着学校在限制该 AI 服务的同时，也会失去让学生访问 Google 搜索服务的权限。
Hacker News · AISelectedAI score6767 AI detection tools like Pangram are reshaping the human writing ecosystem.
Verdicts delivered by AI detection tools like Pangram are having a direct impact on human creators and the publishing industry. In March, publisher Hachette Book Group pulled the release of author Mia Ballard's *Shy Girl* after Pangram assigned it a 78.3% AI-generated score, sparking widespread online harassment and a severe personal mental health crisis.
Why it matters: Through the case of a canceled book and subsequent cyberbullying, the original piece illustrates the tangible impact of AI detection tools on creators' livelihoods and the trust underpinning human writing.
Hacker News · AIAI score3636 Google 自家 AI 裁定 AIR Translate 对决 Google 翻译：Google 以 1 比 108 落败Chinese
AIR Translate 在针对 11 个国家与地区语言变体的 118 轮盲测中对决 Google 翻译，由 Google 的 Gemini 4 Argon 担任裁判，最终 AIR Translate 以 108 胜 1 负 9 平胜出。该评测涵盖西班牙语、葡萄牙语、法语、阿拉伯语及中文变体，AIR Translate 重点适配了各地本土词汇与习惯表达差异。
Hacker News · AIAI score3939 Genie Studio：支持 AI 生成与编辑音轨的浏览器 DAWChinese
Genie Studio 是一款基于浏览器的数字音频工作站（DAW），支持通过 AI 生成和编辑音轨。该工具直接在网页端运行，允许 AI 参与音频轨道的制作与调整。
Engineering at MetaAI score4646 F3 Lightning：Meta 如何大规模加速广告 ML 特征投产Chinese
Meta 推出构建于 F3 之上的基础设施层 F3 Lightning，将特征发布与共享执行解耦，实现广告机器学习特征的独立编译、验证、部署与回滚。通过共享执行计划与共享内存缓冲区等四项优化，该架构在不增加基础设施计算成本的前提下，将特征投产速度较原 F3 提升约 3.6 倍，从而支持同周期内开展更多模型迭代。
The Verge · AI 筛选SelectedAI score7676 Google to Completely Disable Google Search in K-12 Schools That Block Its AI Features
Google has updated its education policies, meaning K-12 schools that refuse its AI features will no longer be able to use Google Search as a standalone service. For K-12 student accounts, Gemini apps and Google Search are turned off by default and bundled together, meaning administrators cannot disable AI features while keeping traditional Search enabled. This move stems from Google's deep integration of AI into its search architecture, coming as the company prepares to pilot search-embedded guided learning features in select US schools this month.
Why it matters: The report highlights the platform's strategy of tightly bundling core search with generative models, providing context for readers to assess the constraints this places on digital tool selection in schools.
Ars Technica · AI 筛选AI score5858 Meta 因误判 Claude 工具封禁倡导组织 Instagram 账号引监管质询Chinese
Meta 自动化审核系统因将倡导组织 No Kings 使用的 Claude AI 工具误判为机器人，在早期投票高峰期封禁其 Instagram 账号四小时并致歉。
Hacker News · AIAI score2424 什么是大语言模型（LLM）？Chinese
该科普图解阐释了大语言模型（LLM）的核心机制：模型通过海量文本预训练来逐个预测下一个可能的 token，并受限于有限的上下文窗口。由于模型倾向于选择概率较高的词，当训练数据存在滞后或群体性错误，叠加上下文溢出丢失细节时，模型就会一本正经地捏造事实，从而产生模型幻觉。
Hacker News · AIAI score5050 家庭 AI 智能体 Cinnamon 发布：支持共享与私密双空间管理Chinese
初创团队发布面向家庭场景的 AI 智能体 Cinnamon，支持规划行程、整理收件箱、日程提醒与比价监控。系统严格划分家庭共享空间与个人私密空间，采取操作前必须人工确认机制，并在沙箱中隔离运行任务。产品提供 40 点免费额度试用，单人版月费 9.99 美元，最多 6 人的家庭版月费 19.99 美元。
AWS 机器学习AI score4646 Amazon Bedrock 推出 Kimi K3 印度本地推理服务Chinese
Amazon Bedrock 现已支持月之暗面（Moonshot AI）多模态模型 Kimi K3 的印度本地推理，以满足当地严格的数据驻留合规要求。推理任务被限定在 Mumbai 和 Hyderabad 区域内处理，传输数据全程加密且不出 AWS 内部网络。用户可通过兼容 OpenAI 的 API 调用该模型，无需独立 OCR 即可直接解析多语言文档并进行多轮问答。
AWS 机器学习AI score5252 AWS 详解如何在云端构建超低延迟实时语音 AI 应用Chinese
AWS 联合语音 AI 初创公司 Cartesia 介绍了基于状态空间模型（SSM）构建亚 100ms 延迟语音应用的技术方案。Cartesia 旗下的 Sonic 文本转语音模型支持约 90ms 的首包延迟并覆盖超过 44 种语言，Ink 语音转文本模型则原生支持轮流交替识别与流式转写。
AWS 机器学习AI score5555 基于 EDB Postgres AI 与 Amazon Bedrock AgentCore 构建 AI 智能体语义层指南Chinese
AWS 与 EDB 展示了如何结合 EDB Postgres AI 的 AIDB 扩展和 Amazon Bedrock AgentCore 构建面向 AI 智能体的语义层。
AWS 机器学习AI score3636 Gen Digital 如何利用 Amazon Bedrock AgentCore Web Search 构建 Norton Revamp 的 AI 新闻引擎Chinese
Gen Digital 利用 Amazon Bedrock AgentCore Web Search 为职业声誉管理平台 Norton Revamp 构建了 AI 策展新闻引擎，实现零数据外流的实时检索与内容锚定。系统采用事件驱动架构，以 Amazon Nova Lite 作为主要大语言模型生成内容，并结合 Claude 负责复杂推理与 Stability AI 生成配图。
AWS 机器学习AI score5252 如何通过 Bedrock AgentCore Gateway 实现面向 ServiceNow 的 OBO 令牌交换Chinese
AWS 介绍了如何利用 Amazon Bedrock AgentCore Gateway 与 Microsoft Entra ID 配置代理代表（OBO）令牌交换。
AWS 机器学习AI score3838 利用 Amazon Quick 整合结构化与非结构化数据Chinese
Amazon Quick 支持将 Amazon S3 支持的非结构化知识库与结构化数据集整合至统一的 Space 工作区中。用户可通过单个对话式 AI 智能体跨两类数据源查询，由智能体自动将自然语言转译为 SQL 或检索文档，并返回附带引用的综合答复。该方案采用数据、工作区与智能体三层架构，同时配备了内容安全筛选等内置 AI 护栏。
TechCrunch AISelectedAI score7373 Google to let schools disable AI tools, but fully blocking them will prevent access to Google Search
Google has announced plans to turn off AI tools by default for K-12 students next year unless school administrators explicitly opt in. However, because AI is deeply integrated with Search, schools that completely block AI will also lose access to Google Search.
Why it matters: By tightly bundling AI features with basic search, the policy highlights the practical trade-offs educational institutions must navigate between tool governance and essential web search needs.
Data Center DynamicsSelectedAI score6161 OpenAI Launches Grid Protectors Initiative to Ease Data Center Power Demand During Peak Grid Strain
OpenAI has launched its non-binding Grid Protectors initiative, designed to proactively curtail power demand from its data centers when the US electrical grid is under stress. The move aims to lessen the strain of data center energy consumption on power grids and mitigate local community pushback against data center expansion.
Data Center DynamicsAI score5555 乌克兰无人机袭击后，普京允许俄罗斯企业使用海外数据中心Chinese
受乌克兰无人机袭击影响，俄罗斯允许该国企业使用海外数据中心。与此同时，俄罗斯本土数据中心企业被要求为政府部门预留容量。
Data Center DynamicsAI score5353 沙特阿美收购沙特 Humain 公司 49% 股权Chinese
沙特阿美（Aramco）收购了沙特 Humain 公司 49% 的股权。沙特公共投资基金（PIF）继续持有剩余 51% 的股权。
Data Center DynamicsAI score3131 Smartbird 当前价值低于鞋履品牌 Allbirds 转型 AI neocloud 前水平Chinese
鞋履品牌 Allbirds 转型为 AI neocloud 并更名为 Smartbird 后，当前价值已低于转型前的水平。期间两次品牌重塑虽曾推动股价上涨，但提振效果仅是暂时的。
NVIDIA · AI 筛选AI score4646 AI 正在改变半导体晶圆厂运作方式，助推工程师研发提速Chinese
AI 与加速计算正在重塑半导体晶圆厂，通过物理 AI 预测、数字孪生及智能体协助工程师高效优化芯片制造流程。COMSOL 与 Applied Materials 借助 NVIDIA cuDSS 和 PhysicsNeMo 实现高达 10x 至 35x 仿真加速，Siemens cuLitho 带来最高 17x 计算光刻加速，Lam Research 更将等离子体行为预测从数天缩短至毫秒级。
Simon WillisonSelectedAI score7070 Simon Willison Confirms Qwen 3.8 27B Can Locally Interpret Neural Network Weights
Running Qwen 3.8 27B locally on a laptop, Simon Willison replicated experiments from a paper on weight-space programming, confirming the model can infer the specific game a policy network was trained on using only raw floating-point weight text, without relying on any code or tools.
Why it matters: The hands-on test demonstrates the capability of a local open-source model to deduce a policy network strictly from floating-point weight text without tools, offering readers a concrete pathway to replicate and evaluate weight-space programming.
Hacker News · AIAI score2121 HackerNietzsche：机器，太人性的——关于 AI 的箴言Chinese
HackerNietzsche 模仿尼采风格发表 AI 箴言集《机器，太人性的》，对当下技术狂热展开哲学批判。文中批评 Karpathy 将智能简化为优化算法，指出 AI 本质是缺乏自我的快速模仿者；同时直指 Amodei 倡导的 Scaling Laws 受限于既有数据与 Alignment 束缚，堆叠参数也无法实现突破认知框架的高阶学习。
Hacker News · AIAI score5454 开源本地 AI 第二大脑平台 Knemora 发布首个 Alpha 测试版Chinese
开源项目 Knemora 发布首个公开测试版 v0.1.0-alpha，定位为本地优先的 AI 第二大脑与智能体平台。系统无需外部 LLM API、云服务或向量数据库，依托 llama.cpp 在本地运行默认的 Qwen3-8B 模型，并集成了便携 Markdown Wiki、Sigma.js 交互式图谱与 BM25 检索功能。
Hacker News · AIAI score5151 CleanSlop 发布：面向 AI 生成 TypeScript 与 Node Web 应用的扫描修复系统Chinese
开发者发布了针对 AI 生成 Web 应用的扫描与修复系统 CleanSlop，支持排查由编码智能体生成的 TypeScript、Node.js 与 JavaScript 应用故障。
Hacker News · AISelectedAI score7979 LLMjacking Defense Guide: Restricting Stolen AI Key Abuse Across AWS, Google Cloud, and Azure
Addressing LLMjacking attacks where hackers exploit stolen cloud credentials and AI keys to rack up hundreds of thousands to millions of dollars in model usage fees on victims' bills, this article provides a defensive configuration guide across major cloud platforms. Due to billing latency in budget alerts and spending caps, tokens-per-minute (TPM) quotas serve as the only line of defense capable of hard-blocking excessive calls within minutes.
Why it matters: The article compiles protective configurations against fraudulent model charges from compromised keys on major cloud platforms, helping security teams establish an immediate stop-loss defense through rate quotas and log auditing.