←搜索结果

按匹配度 · 共 561 条结果 清除搜索

星期五

  1. LMSYS Blog

    SGLang 原生支持 Reflection Beam 并支撑其生产 API

    LMSYS 宣布 SGLang 已原生支持 Reflection 的 Beam 模型,并用于支撑其生产 API。Beam 是 501B 参数、每 token 激活 23B 的 MoE 模型,原生 1M token 上下文窗口,计划以 Apache 2.0 发布,SGLang 为其提供 BF16、NVFP4 和动态 FP8 支持、DFlash2 投机解码与混合注意力前缀缓存。在 4 张 NVIDIA GB300 GPU 上以 NVFP4 运行 OpenHands 多轮智能体编码回放,并发 32 时达到每 GPU 127.6K tokens/sec、每用户 185.87 tokens/sec(p50)。

    Beam's embedding and decoder stack, with an expanded decoder layer showing attention, feed-forward work, sandwich normalization and depth-scaled residual paths.

    d HiCache to enable cost-efficient, production-ready serving for agentic workloads. TPU recipes for Beam and access to SGLang-TorchTPU will be available soon. Sign up through this Google Form to receive the latest updates. Reinforcement learning in Miles Miles supports Beam through a Megatron plugin and runs rollouts on Reflection's SGLang build. The training backend implements Beam's sandwich normalization with depth-scaled residuals, SSMax attention scaling, hybrid sliding-window attention, ga

    匹配字段正文

  2. Unite.AI

    Anthropic 将 Claude Code 的 Code Review 正式开放

    Anthropic 于 2026 年 10 月 15 日将 Claude Code 的多智能体 PR 审查系统 Code Review 面向 Claude Team 和 Enterprise 组织正式开放,并为符合条件的 Enterprise 组织提供至少 10000 美元的审查额度。多数审查速度比预览期快约 3 倍,典型耗时从约 20 分钟降至 6 至 10 分钟,管理员可为每个仓库设置 Quick、Standard 或 Deep 审查级别。审查在 Anthropic 基础设施上分四阶段运行,只留行内评论和检查结果,不会批准或阻止合并,且需要 Team 或 Enterprise 套餐与 GitHub 托管仓库。

    t, and is not yet available to organizations using customer-managed encryption keys. A Claude organization is required even when developers use Claude Code through Amazon Bedrock, Google Cloud’s Agent Platform, or Microsoft Foundry. The documentation cautions that Code Review uses AI and may miss bugs, security vulnerabilities, or other issues, and that a review with no findings does not mean code is error-free or secure. Anthropic introduced Code Review as a research preview for Team and Enterp

    匹配字段正文

  3. Unite.AI

    OpenAI 为 ChatGPT Chrome 扩展加入标签页控制能力

    OpenAI 于 10 月 15 日宣布 ChatGPT Chrome 扩展可查看用户打开的标签页、接管标签页完成任务并整理分组,该扩展在 Chrome 应用商店显示有 7,000,000 用户。Chat 中可跨标签页收集信息,Work 中可执行发消息、填表单、下载文件等操作,扩展版本为 1.26.1005.11759。文档称默认每次访问新网站都会先询问,可设置允许一次、允许本站或允许所有站点,浏览器历史访问按单次请求授权并标记为 Elevated Risk。

    s not used to determine creditworthiness or for lending purposes. Setup, Permissions, and Stored Data OpenAI’s browser-extension documentation states that the extension works with Google Chrome, Microsoft Edge, Brave, Opera, and Vivaldi from the ChatGPT desktop app, with side chat available in Chrome, Edge, Brave, and Vivaldi but not in Opera. Setup runs through the Computer Use section of the desktop app’s settings: users install the extension from the browser’s store, confirm the browser shows

    匹配字段正文

  4. Thurrott

    Nvidia 公布为 RTX Spark 优化的应用与游戏阵容

    首批搭载 RTX Spark 的 Windows 11 on Arm 电脑将于明天上市,Nvidia 称应用生态已就绪,多个创作、开发、游戏和本地 AI 方案已完成优化。RTX Spark 最高配备 128 GB 统一内存和 1 petaflop AI 算力,GitHub Copilot、Perplexity 等智能体已支持简化部署,Nvidia PAIR 个人 AI 路由器和 Nemotron 3.5 Bolt 74B 本地模型正式可用。Adobe 正为 Photoshop 和 Premiere Pro 优化并承诺 2 倍性能提升,Nvidia 还将于周五推出速度提升 2 倍的 DLSS AI 模型。

    Google Tag Manager (noscript) End Google Tag Manager (noscript) The first Windows 11 on Arm PCs powered by RTX Spark become available tomorrow, and Nvidia says the app ecosystem is ready, with multiple creator, developer, games, and local AI solutions already optimized for the platform. “Nvidia RTX Spark is a new beginning for the PC, a reinvention combining the capabilities of a gaming PC, a creative studio, and a personal AI supercomputer into a single superchip,” the Nvidia announcement expla

    匹配字段正文

  5. Thurrott

    Brave Leo 上线端到端加密聊天,与 NEAR AI 合作

    Brave 与 NEAR AI 合作,为隐私优先的 Leo AI 助手带来端到端加密聊天。该功能下,浏览器在设备端用 Oblivious HTTP(OHTTP)加密提示词,只有经过认证的 TEE 才能解密并运行机密 LLM 推理,IP 地址全程隐藏,GLM 5.3 Flash 模型运行在 NEAR AI 运营的 Nvidia 硬件隔离区中,提示词与回复对 Brave、NEAR AI 及基础设施运营方均不可见。该功能已在 Brave Release、Beta 和 Nightly 上线,需在 Leo 中选择 GLM 5.3 Flash 使用;Brave 表示后续会把认证检查移入浏览器并扩展到其他开源权重模型。

    Google Tag Manager (noscript) End Google Tag Manager (noscript) Brave has partnered with NEAR AI to deliver end-to-end encrypted chats to its privacy-first Leo AI assistant. “In 2014, I argued in ‘Trust but Verify’ that users shouldn’t have to place blind trust in the software they rely on,” Brave CEO Brendan Eich says. “They should be able to verify it; this is a user-first principle. AI has raised the stakes, yet too much of it still runs on ‘trust me bro’ promises, often at the user’s expense

    匹配字段正文

星期四

  1. Thurrott

    Hugging Face 发布 Llama Windows 本地推理桌面应用

    Hugging Face 发布免费开源的 Windows 11 桌面应用 Llama Windows,把 llama.cpp 本地推理库带到 Windows,类似 Mac 上的 llama.app。用户可从 Hugging Face Hub 一键下载模型并在本地运行,无需终端、构建步骤或管理员权限,应用以系统托盘形式运行,也可用 Alt + Space 打开助手窗口。该应用原生支持 Windows 11 on Arm 和新的 Nvidia RTX Spark 平台,使用 WinUI3 编写,并内置 llama.cpp 服务器、Hugging Face Hub 客户端和模型目录的 .NET 库。

    Google Tag Manager (noscript) End Google Tag Manager (noscript) Llama Windows is a free, open source desktop app that brings the llama.cpp local AI inferencing library to Windows 11, similar to the llama.app on Mac. “With Llama Windows, you can pick a model from the Hugging Face Hub, download it in one click, and run it locally from the notification area,” the Hugging Face announcement explains. “No terminal, no build step, no admin rights.” Llama Windows launches with native support for Windows

    匹配字段正文

  2. 9to5Mac

    MacPaw 发布 Mac 与 iPhone 端主动式 AI 助手 Eney

    MacPaw 发布主动式 AI 助手 Eney,覆盖 Mac 与 iPhone,可跨应用、文件和系统识别待办事项并给出下一步操作供用户确认。Eney 通过晃动光标或 Option + W 唤出 Eney Wand,直接对屏幕上的文本提问,无需复制粘贴;它同时使用本地与云端模型,其中定制模型与 Liquid AI 合作开发,记忆功能 Mnemos 目前仅管理 Eney 自身记忆。Eney 支持英语、德语、西班牙语、法语、意大利语、波兰语、巴西葡萄牙语和乌克兰语,提供 14 天免费试用与 10000 AI credits,订阅分为 Starter 每月 10000 credits 14.99 美元、Boost 20000 credits 29.99 美元、Power 30000 credits 49.99 美元。

    of this possible, Eney integrates with a wide range of apps and services. It works with Apple Mail, Calendar, and Reminders out of the box, and also supports connections to Gmail, Google Calendar, Google Drive, Slack, Notion, Jira, Linear, and more. The idea is to let users work across these services through a single interface, without constantly switching between apps or manually transferring information from one to another. Availability Eney launches today on the Mac, and you can get it here.

    匹配字段正文

  3. MarkTechPost

    CopilotKit 发布 CLiF:把智能体轨迹与用户操作转化为技能、评测和微调数据

    CopilotKit 发布持续学习平台 CLiF,目前处于早期访问阶段,属于 CopilotKit Intelligence 的一部分。它通过 AG-UI 协议把智能体轨迹与用户在产品中的操作捕获为同一条全栈轨迹时间线,再经捕获、模式发现、人工审核、交付导出四阶段循环,产出技能、基于生产的评测、微调数据集和原始轨迹四类输出。技能以 SKILL.md 文件形式发布,需人工逐条审核,发布后无需重新部署即可被智能体加载,也可下载给其他智能体或 MCP 服务器使用;平台完全支持自托管。

    Add as a preferredsource on Google Introduction Teams shipping AI agents at scale keep running into the same pattern. Even agents built on top-tier models like Fable and Astra struggle to carry out complex, end-to-end tasks inside a specific product consistently enough to deploy. So teams grind out evals for the workflows that matter most, often hundreds or thousands of hand-written cases, and then improve the agent bit by bit through prompt edits, harness fixes, and special cases until it passe

    匹配字段正文

  4. TechCrunch AI

    Amazon 旗下 Bee 发布两款个人 AI 设备,起售价 74.99 美元

    Amazon 旗下个人 AI 硬件品牌 Bee 发布被收购后的首批新品:可戴在手腕或别在衣服上的 Bee Frontier 售价 74.99 美元,信用卡大小、磁吸在手机背面的 Bee Eclipse 售价 99.99 美元。两款设备延续记录日常对话并生成摘要、待办和提醒的思路,免费提供对话摘要与 14 天记忆,11.99 美元每月的 Bee Pro 解锁无限摘要以及与 Amazon.com、Alexa+ 和第三方应用的集成。设备采用私有计算安全架构,不存储原始音频,转写内容由设备生成并控制的密钥加密。

    Bee users can make purchases from Amazon or ask Alexa for their Bee insights, summaries, or to-dos hands-free. Bee also works with your existing productivity software, like Gmail, Google Calendar, and Microsoft Outlook, so it can draft emails, set up meetings, surface action items, and more. Plus, you can upload photos to Bee or connect it with Apple’s HealthKit to add a picture of your overall wellness to be integrated into your provided insights. The Bee Frontier and Bee Eclipse both have just

    匹配字段正文

  5. LMSYS Blog

    SGLang 与 Miles 为 Muse Spark Open 提供 Day-0 支持

    LMSYS 发布博客称,SGLang 与 Miles 为 Meta 开源多模态 MoE 模型 Muse Spark Open 提供 Day-0 支持。该模型采用 816B 参数文本 Transformer 主干、每 token 约 42B 激活参数,支持文本、图像、视频和音频输入以及 1M token 上下文窗口,checkpoint 提供 NVFP4 与 BF16 版本。在 8 张 NVIDIA B300 GPU 上以 NVFP4 checkpoint 跑 OpenHands 多轮智能体编码回放,并发 1 时单 GPU 吞吐 14.8K tokens/sec、单用户 543 tokens/sec(p50),并发 128 时为 80.3K tokens/sec、单用户 61 tokens/sec(p50)。支持平台包括 NVIDIA Blackwell(GB300、B300、GB200、B200)、H200、AMD MI350X 与 MI355X,以及 Intel Xeon 6 P-core,并首次通过 SGLang-TorchTPU 为开源权重模型提供 Day-0 TPU 支持。

    Muse Spark Open maps text, image/video and audio inputs into a shared decoder with GQA and sliding-window/global attention; its MoE block routes full-width inputs while pooling expert features, normalizing each expert output before the weighted sum and restoring the model width with an unpooler.

    o enable cost-efficient, production-ready serving for agentic workloads. TPU recipes for Muse Spark Open and access to SGLang-TorchTPU will be available soon. Sign up through this Google Form to receive the latest updates. Reinforcement Learning Miles provides distributed training support for Muse Spark Open through Megatron-Core, covering the model's compressed MoE, normalization, and residual gating. Each expert output is normalized independently, then combined as a routing-weighted sum in FP3

    匹配字段正文

  6. Unite.AI

    NVIDIA 发布 cuPotential,半导体厂商采用 CUDA-X

    NVIDIA 发布 CUDA-X 新库 cuPotential,供开发者构建模拟原子与分子相互作用的 AI 智能体,并公布 Synopsys、Cadence、Siemens、Lam Research、KLA、Applied Materials 等半导体厂商对 CUDA-X 库的采用情况。cuPotential 目前处于早期访问阶段,NVIDIA 称早期测试中它让已 GPU 优化的流程再加速 2-3x,Synopsys 已用它加速材料模拟。NanoVDB 在 OpenVDB 中开源,NPP 通过 NVIDIA CUDA Toolkit 提供,Lam Research 的 NanoVDB 加速版 SEMulator3D 计划于 2026 年底发布。

    hnical accuracy, clarity, and responsible coverage of the rapidly evolving AI compute landscape. Discover More Microsoft Research Publishes AI Weather Forecasting Study in Nature; Google Rolls Out AI Scam Defenses Across Android Messages and Calls; OpenAI Alleges Employee Who Rejoined Apple Took Unreleased Device Files; Baseten Adds Runway’s Gen-4.5 and Aleph 2.0 to Dedicated Inference; Illumina Debuts BioInsight Platform Unifying AI, Software, and Data; Kyndryl Applies Agentic AI Framework to T

    匹配字段正文

星期三

  1. Publickey

    微软正式发布 Windows 智能体隔离环境 MXC

    微软宣布 Windows 11 上用于安全隔离 AI 智能体的 Microsoft Execution Containers(MXC)正式发布,可限制智能体可访问的文件与网络,并在运行时执行策略。MXC 提供多种隔离后端,包括基于 AppContainer 的进程隔离、会话容器隔离、WSL 容器(WSLc)以及硬件级 MicroVM,用户可按需设置文件与网络访问策略,并计划支持通过 Windows Intune 做组织级策略配置。微软表示 OpenClaw、Anthropic Claude Code 等合作伙伴已宣布支持 MXC。

    に注力; さくらインターネット、GPU専有によりトークン消費量を気にせず定額で利用できる「さくらのAI Engineプライベートエディション」提供開始を発表; AWS、DuckDBをAurora PostgreSQLに統合。ETL不要で直接データレイクを読み取り可能にしたと発表; マイクロソフト、VS CodeにAIモデルのオーケストレーションで高品質低コストなAIを実現する「HydraFusion」をプレビュー実装; FDE(フォワードデプロイドエンジニアリング)によって開発されたAIエージェントの7割は使われず、FDEの名前で単なるコンサルティングが販売されるようになる危険が高まるとガートナーが予測; Cloudflare ContainersがAIエージェント向けに刷新。起動が6倍速、ファイルシステムのスナップショット機能など; Google、サンドボックス型コンテナランタイムの「gVisor」をCNCFに寄贈したと発表 Publickeyについて/運営者について; お問い合わせ; 広告について; プライバシーポリシー; 読者が安心して記事を読めるようにするための取り組みについて

    匹配字段正文

  2. Factory 研究 / 产品

    Factory 推出自托管智能体沙箱编排器 OpenCompute,进入私有预览

    Factory 发布 OpenCompute,这是其自托管的智能体沙箱编排器,可运行在客户自选的运行时上,目前处于私有预览阶段。它通过出站 TLS WebSocket 连接 Factory,无需开放入站端口,控制平面用 Go 编写、状态存于客户环境内的 Postgres,并以声明式生命周期模型做状态协调。每个 provider 集成以独立 gRPC driver 实现,分为必需的 Core 以及可选的 Suspend、Snapshot、Fork 一致性档位,Factory 计划将其开源。

    Introducing OpenCompute

    s no shortage of work happening on sandbox environments for agents. Third-party cloud sandbox platforms include E2B, Daytona, and Modal. Projects such as Kubernetes Agent Sandbox, Google's Agent Substrate, and AWS Lambda MicroVMs also show growing demand for bring-your-own-cloud and bring-your-own-infrastructure solutions. Each platform makes different tradeoffs across isolation, persistence, startup speed, operational simplicity, and deployment flexibility. A backend that is right for short-liv

    匹配字段正文

  3. The Next Platform

    微软公布 Clea Maia 300 推理 XPU 规格

    微软在 Open Compute Project Summit 2026 上公布第三代自研推理加速器 Clea Maia 300 的规格,采用台积电 2 纳米 N2P 工艺和 chiplet 架构,封装内含 7 颗芯片与 8 个 HBM4 堆栈,配备 288 GB HBM4 和 18 TB/sec 内存带宽。微软称一机架 96 颗 Maia 300 的 token 吞吐量最高可达一机架 48 颗 Maia 200 的 7 倍,每瓦性能提升 2.5 倍,单机架功耗约 167 千瓦。该芯片面向更大规模的智能体推理负载,ATL 2.0 网络可将 128 个机架、12,288 颗加速器连入同一一致性域。

    oherent fashion across their rackscale systems, or even the 576 or 1,152 multi-tiered NVSwitch coherent fabrics that Nvidia is promising for production environments in the future. Google’s “Sunfish” TPU 8t training chip (which is not yet shipping) tops out at 9,600 TPUs in a coherency domain, while the “Zebrafish” TPU 8i inference chip has a “boardfly” topology that peaks at 32 in a boardfly (a dragonfly topology of eight dragonfly nodes with four TPUs per node) and 1,152 in a group of 36 boardf

    匹配字段正文

  4. The Guardian Technology

    苹果发布 HomeView 智能家居中枢,搭载升级版 Siri AI

    苹果发布智能家居新品 HomeView,这是一款可挂墙或置于桌面的类 iPad 设备,起售价 299 美元,配备 7 英寸屏幕和面部识别,可控制门铃、恒温器和智能锁,今日开启预售,预计 11 月 13 日发货。该设备搭载升级版 Siri AI,苹果称其更具对话能力,可调用 Messages、Calendar 等个人应用信息作答。苹果同时更新了 HomePod 和 Apple TV,并联合 LG、施耐德电气推出门铃、摄像头、恒温器和智能锁等配件。苹果表示,由于 Siri AI 受欧盟法规限制无法在当地使用,HomeView 不会在欧盟销售。

    , Apple said. Apple first released its Home app to control smart gadgets in 2016, ⁠with HomePod smart speakers following in 2018, but those offerings are competing with Alphabet’s Google Home speakers and Amazon’s Echo line. Analysts said two major hurdles for Apple were making its devices work with a variety of third-party gadgets and ‌the limited ability of Siri to give useful answers. The HomeView device includes a revamped version of Siri AI, the company announced. An Apple press release des

    匹配字段正文

星期二

  1. The New Stack

    Workday 用 TypeSafe 的 Jev 实时生成界面,Sana 成为默认首页

    Workday 将去年以 11 亿美元收购的 AI 助手 Sana 设为 HR、财务和 IT 软件的新首页,界面由 TypeSafe 的决策模型 Jev 在运行时从设计系统组件中挑选拼装,而非由模型生成完整界面。Workday 首席 AI 官 Joel Hellermark 称,用分类模型从现成组件中选择可以做到即时响应,TypeSafe 称 Jev 的响应时间为 70 至 500 毫秒。Jev 只返回最多 255 个选项中的一个、分数或是否答案并附带概率,不生成文本;TypeSafe 于 9 月 15 日发布 Jev,并在上周宣布由 Andreessen Horowitz 领投的 8.7 亿美元 A 轮融资,估值 75 亿美元。

    zmaier, Workday’s president of product and technology, said in the briefing, and users lose the Jev-driven interface and some context along the way. Workday’s May partnership with Google Cloud already covers A2UI , Google’s spec for agents to describe interfaces that host apps then render. MCP Apps , a Model Context Protocol extension, brings interactive UI into Claude and ChatGPT. Neither shows up in Workday’s announcements this week, so for now, it seems, employees who reach Sana through Claud

    匹配字段正文

  2. Computerworld

    微软发布第三代 AI 芯片 Maia 300,计划 2027 年部署

    微软发布第三代 AI 芯片 Maia 300,称其针对最新推理模型和智能体优化,可降低运营成本。微软云与 AI 执行副总裁 Scott Guthrie 表示,Maia 300 每兆瓦 token 数预计比主流 GPU 系统高 40%,公司将从 2027 年起把 Azure Maia 系列扩展到多吉瓦部署,使其成为微软产品与服务的主要推理系统。上一代 Maia 200 于今年 1 月推出,目前支持 Microsoft Copilot 负载,GitHub Copilot 和 Microsoft Foundry 将在 2026 年底前跟进。

    26 1 min; newsAnthropic’s new framework will let AI agents control hardware Aug 28, 2026 1 min; newsMicrosoft moves to limit AI use by its employees Aug 5, 2026 1 min Show me more Google positions Gemini 4 Argon for long-horizon agentic work China clears Alibaba’s Qwen AI just in time for iPhone Duo Microsoft knows it robbed publishers to train AI. It just doesn’t care Can your tech stack actually handle AI? Microsoft Copilot Growth, ClaudeBleed Risk, LinkedIn GDPR Complaint | Ep. 84 AI agents a

    匹配字段正文

  3. Computerworld

    Apple 推出 299 美元 HomeView 智能屏及多款家居新品

    Apple 发布 HomeView 智能显示屏,售价 299 美元,搭载 A18 芯片和 Apple Intelligence 驱动的 Siri AI,支持最多 6 名用户通过人脸识别切换各自内容,因 Digital Markets Act 限制暂不在欧盟上市。同期还推出搭载 A19 芯片、售价 299 美元的 Apple TV 4K,支持 Siri AI 的新款 HomePod mini 2(149 美元),以及与 LG、施耐德电气合作、支持 Matter 和 Thread 的 Apple Home Collection 智能家居配件。

    the company experimenting with ways to provide ambient AI experiences that actually benefit consumers. It puts the company up against products in the ilk of Amazon’s Echo Show or Google’s Nest. It also echoes Samsung’s SmartThings ecosystem, but platform integration and privacy may give it the edge. Apple Apple Home Collection partnerships The Apple Home Collection is a rare excursion into partnership, with Apple working with both LG Electronics and Schneider Electric to introduce a new range o

    匹配字段正文

  4. Computerworld

    Workday 发布基于 Sana 的默认首页与四个自动执行智能体

    Workday 在 Workday Rising 大会上发布基于 Sana 的默认首页 Workday Home,把任务、决策、会议和智能体活动集中到一处,并可按角色实时生成表单、日历和审批卡片。同时推出 Sana Concierge 员工服务入口,以及招聘、薪酬、会计、合规四个自动执行智能体,可在用户离线时按心跳运行。Workday Home 已可用,Sana Concierge 面向早期采用客户开放,预计 2027 年第一季度全面可用。

    tion at every stage of functions such as financial closing or preparation of workforce plans by HR. Sana also works within other AI tools and platforms including Anthropic Claude, Google Cloud’s Gemini Enterprise, and Microsoft Copilot, as well as within Microsoft Teams and Slack. But, Workday said in its announcement, “third-party AI tools and platforms never get raw access to Workday data — they can now access this data through Sana, and Sana does the reasoning, applies each company’s own perm

    匹配字段正文

  5. CNET

    苹果发布 299 美元 HomeView 智能音箱,7 英寸方形屏配 A18 芯片

    苹果在纽约 Welcome Home 活动上发布 HomeView 智能音箱,采用 7 英寸方形 Liquid Retina 屏、12MP 支持 Center Stage 的摄像头、两个全频驱动单元,搭载 A18 芯片,可桌面摆放或壁挂。设备支持家庭成员人脸识别,可查看 Messages、FaceTime、Calendar、Reminders 和 Find My 内容,并与支持 Apple Intelligence 的设备配合提供个性化回答。美国、澳大利亚、加拿大、印度、新西兰和英国用户现已可预购,售价 299 美元,11 月 13 日发货,提供银色、星光色、玫瑰粉和深空灰四种配色。

    ctional interface or something in between. Customers in the US, Australia, Canada, India, New Zealand and the UK can preorder the HomeView right now for $299, with devices available from Nov. 13. It comes in silver, starlight, rose pink and space gray, so there’s bound to be at least one option that will fit your home decor scheme. The Latest from Katie Google’s Demis Hassabis Puts a Hopeful Spin on AI’s Environmental Impact I Put the Oppo Find X10 Pro’s Camera to the Test. These Are the Results

    匹配字段正文