←搜索结果

按匹配度 · 共 384 条结果 清除搜索

星期三

  1. Pydantic AI 官方文章精选

    Pydantic AI Harness 推出 AWSLambdaDurability,让智能体在 AWS Lambda 上可持久运行

    Pydantic AI Harness 发布 AWSLambdaDurability 能力,把每次模型请求、函数工具调用、MCP 调用和动态工具集解析都记录为 AWS Lambda durable functions 的步骤,恢复执行时从存储结果中读取已完成步骤。使用时需在构建智能体时挂载该能力,并用 durable_agent_handler 改造处理函数,@durable_execution 必须位于最外层。步骤默认至少执行一次并带重试,副作用需保持幂等;改变步骤数量或顺序会使在途执行失效,需发布新版本并让旧版本执行完。

    ock appears in the examples because it is the provider most Lambda deployments reach for first. What does checkpointing cost? Steps consume a durable execution's operations and storage quotas. A turn consumes one model step plus one step per tool call. Large tool results consume the storage budget, so store references, not payloads, or use the FileSystem serdes to keep large values on Amazon S3 Files and checkpoint only a file pointer. You benefit because you do not pay for repeated model reques

    匹配字段正文

  2. Apple Newsroom精选

    苹果发布 iPhone 18 Pro 与 iPhone 18 Pro Max,搭载 A20 Pro 与 Siri AI

    苹果发布 iPhone 18 Pro 和 iPhone 18 Pro Max,搭载 A20 Pro 芯片、iOS 27 与 Apple Intelligence,Siri AI 以 beta 形式随 iOS 27 于 9 月 14 日推出。新机首次在 iPhone 上引入 48MP Fusion 主摄的可变光圈,并提供 Apple Reference Image 用于验证照片真实性。两款机型 9 月 12 日开启预售、9 月 18 日发售,iPhone 18 Pro 起售价 1,199 美元,Pro Max 起售价 1,299 美元。

    ro in glacier is paired with the Clear Case previous; next iPhone 18 Pro and the Environment iPhone 18 Pro and iPhone 18 Pro Max will be available in 256GB, 512GB, 1TB, and 2TB storage capacities. Available in burgundy, glacier, silver, and black, iPhone 18 Pro starts at $1,199 (U.S.) or $49.95 (U.S.) per month, and iPhone 18 Pro Max starts at $1,299 (U.S.) or $54.12 (U.S.) per month for 24 months.13 Apple offers great ways to save and upgrade to the latest iPhone models. With Apple Trade In, cu

    匹配字段正文

  3. Apple Newsroom

    苹果发布 AirPods 5,开放式主动降噪售价 129 美元

    苹果发布 AirPods 5,在开放式设计中提供主动降噪,官方称相比 AirPods 4 主动降噪版可多消除最多 50% 的外部噪音,售价 129 美元,配无线充电盒版本 149 美元,9 月 18 日到店。两款均支持 Siri AI 免手操作、点头或摇头回应 Siri,以及由 Apple Intelligence 驱动的实时翻译,该功能现已覆盖整个 AirPods 产品线。无线充电盒版本新增耳机柄滑动调音量,单次开启降噪续航最长 5 小时,配合充电盒最长 22 小时。

    users. Customers can choose AppleCare+ to cover their new AirPods 5 or, in available markets, AppleCare One to protect multiple products in one simple plan. Both plans include coverage for accidents like drops and spills, theft and loss protection on eligible products, battery replacement service, and priority support from Apple Experts. In the U.S., customers can also choose AppleCare One Family, which extends that same coverage to every eligible device across an Apple Family Sharing group for

    匹配字段正文

  4. Apple Newsroom

    Apple 发布 Apple Watch Series 12,搭载全新健康传感系统

    Apple 发布 Apple Watch Series 12,搭载全新 Health Sensing System 与 S11 芯片,心率改为全天每 5 秒测量一次,HRV 测量频率最高提升至 24 倍,并新增 0-10 分的 readiness 就绪度评分。新表还带来 Audio Intelligence 功能,包括 Sound Recognition、Live Rewind 和 Siri Recap,音频由 S11 芯片的 Secure Exclave 隔离处理且不存储录音。Apple Watch Series 12 起售价 399 美元,今日开启预售,9 月 18 日到店;watchOS 27 与 Apple Intelligence 支持的 Siri AI 将于 9 月 14 日推出。

    Apple users. Customers can choose AppleCare+ to cover their new Apple Watch or, in the U.S., AppleCare One to protect multiple products in one simple plan. Both plans include coverage for accidents like drops and spills, theft and loss protection on eligible products, battery replacement service, and 24/7 support from Apple Experts. For more information, visit apple.com/applecare Apple Upgrade, a leasing program provided by Klarna and available in the U.S., gives eligible customers monthly paym

    匹配字段正文

  5. Apple Newsroom精选

    苹果发布首款折叠屏 iPhone Duo,搭载 A20 Pro 与 Siri AI

    苹果发布首款折叠屏 iPhone Duo,展开为 7.6 英寸内屏、折叠后为 5.4 英寸外屏,搭载 A20 Pro 芯片与 iOS 27,起售价 1999 美元,10 月 16 日开启预售、10 月 23 日发售。新机引入基于端侧 AI 的 Smart Take 自动抓拍、Duo Preview 外屏取景和 Duo FaceTime,并支持 Split View 双应用分屏。Siri AI 随 iOS 27 以 beta 形式推出,可调用个人上下文与屏幕内容执行操作。

    uo and the Environment iPhone Duo is available in star white and night sky, and starts at $1,999 (U.S.), or $83.29 (U.S.) per month for 24 months, in 256GB, 512GB, 1TB, and 2TB storage capacities.13 Apple Upgrade, a leasing program provided by Klarna and available in the U.S., gives eligible customers monthly payment options for iPhone Duo, with 12- and 24-month leasing terms starting as low as $57.99 (U.S.) per month for 24 months. For more information, visit apple.com/shop/apple-upgrade.14 Cus

    匹配字段正文

  6. Meta:Newsroom · AI精选

    Meta 推出个人 AI 智能体 Muse,运行在 Muse Secure VM 上

    Meta 发布个人 AI 智能体 Muse,由 Muse Spark 模型驱动,可在 Muse app 或 WhatsApp 中对话使用,能执行发邮件、订行程、填表、议价等任务,并在需要审批时回来确认。Muse 运行在云端专用虚拟机 Muse Secure VM 上,由独立的 Sentinel agent 管控联网行为,用户可选择连接哪些应用及授权范围,并可要求其遗忘特定信息。该产品已在美国上线 iOS、Android 和 muse.ai,即将登陆 AI 眼镜,大部分功能免费并提供订阅方案;支付环节支持 Stripe 的 Link,Meta 称其为首个受 Link 购买保护覆盖的 AI 智能体。

    Share on Threads

    est of someone’s life shifts. When it comes time to pay, Muse can checkout with Link built by Stripe, and it is the first AI agent covered by Link’s purchase protections: free coverage for damaged or lost items, price drops, no-fee returns, and a return guarantee on eligible purchases. Link’s wallet for agents generates a one-time-use card so your real card details stay hidden, allowing you to purchase safely across the internet. Shop Pay is coming soon as another way to pay, along with 1Passwor

    匹配字段正文

星期二

  1. Cohere 产品与研究博客精选

    Cohere 发布 North Mini Code Megakernel 推理服务引擎

    Cohere 发布面向 North Mini Code 的 Megakernel 推理服务引擎,在单张 H100 上以 BF16 运行,端到端平均解码吞吐量比 vLLM 快 1.25× - 1.41×。该引擎支持连续批处理、分页注意力、流式输出、前缀缓存和工具调用,并提供 OpenAI 兼容接口;其代码已在 GitHub 开放。批大小为 1 时,Megakernel 达到 292 tok/s,为 vLLM 的 1.58×,但当前最大批大小为 8,且预填充仍由普通 PyTorch 内核运行并会暂停解码。

    s what we believe is the first fully fledged serving system built around a decode megakernel. It supports everything a real server needs: continuous batching, paged attention, and ragged sequence lengths, all behind an OpenAI-compatible endpoint with tool calling. Point OpenCode at it and you can code with it. On batch size 1, our megakernel reaches 292 tok/s, or 62% of SoL — 1.58× faster than vLLM. That margin holds across batch sizes and out to 256K of context, with no measurable loss of accur

    匹配字段正文

  2. OpenRouter:Announcements精选

    OpenRouter 发布 shell 服务端工具与 Files API

    OpenRouter 推出 openrouter:shell 服务端工具和 Files API,任何接入 OpenRouter 的模型都能在托管的 Linux 容器中执行命令,并上传文件、下载运行产物,两者目前均为 beta。shell 与 Bash 工具兼容 OpenAI 和 Anthropic 规范,容器默认关闭出站网络、空闲 5 分钟后休眠,沙箱时间按每秒 $0.0001 计费并计入请求成本,Files API 存储上限 10 GiB。

    esponses API and the Anthropic Messages API, as well as openrouter:bash, the Anthropic-compatible Bash tool, on the Messages API. Both work with any model; Files API: workspace storage under /api/v1/files. Upload files, attach them to a container by id, and keep the files a run produces; Containers: these are the sandbox where shell commands run. Files written there persist between requests that share a container id. You can access the container’s contents through the /api/v1/containers API When

    匹配字段正文

  3. OpenRouter:Announcements

    OpenRouter 推出 shell 服务端工具与 Files API,为任意模型提供托管沙箱终端

    OpenRouter 发布 openrouter:shell 服务端工具和 Files API,任意模型都能在托管 Linux 容器中执行命令并读写文件,两者均处于 beta 阶段。shell 兼容 OpenAI 的 Shell 工具与 Anthropic 的 Bash 工具,通过 engine 设为 openrouter 可强制在服务端沙箱执行,容器默认关闭出网、闲置 5 分钟后休眠。沙箱按活跃时间计费,每秒 $0.0001,冷启动最低计费 30 秒,Files API 存储上限 10 GiB。

    Give any model a terminal and files

    esponses API and the Anthropic Messages API, as well as openrouter:bash, the Anthropic-compatible Bash tool, on the Messages API. Both work with any model; Files API: workspace storage under /api/v1/files. Upload files, attach them to a container by id, and keep the files a run produces; Containers: these are the sandbox where shell commands run. Files written there persist between requests that share a container id. You can access the container’s contents through the /api/v1/containers API When

    匹配字段正文

  4. Mastra Blog精选

    Mastra 发布 Factory Beta,智能体已承担 25-35% 的 PR

    Mastra 发布开源软件工厂 Mastra Factory Beta,它连接 GitHub、Linear 和 Slack,按可配置看板分阶段运行智能体,每个阶段可设为手动或自动。官方称自 7 月 alpha 版上线以来,Factory 已自动化 25-35% 的 PR、关闭 50-60% 的 issue,项目每周收到 100 多个 issue、合并约 200-250 个 PR。团队还介绍了基于图谱的 wiki 知识库、AgentController 会话管理和 Signals 实时纠偏机制,并计划将知识图谱作为独立原语发布。

    Sam Bhagwat

    fixed it in a way you wouldn't expect. With Factory, the cost of fixing these bugs has dropped so low, this no longer makes sense as a contribution model. We're shifting from encouraging community PRs to encouraging highly detailed issue repros, and we're crediting issue authors on merged PRs. If you want a walkthrough, join Abhi and Alex for the Build and Own Your Software Factory workshop.Share on X or LinkedInSam BhagwatCEO Sam Bhagwat is the founder and CEO of Mastra. He co-founded Gatsby, w

    匹配字段正文

星期一

  1. ElevenLabs Blog精选

    Vimeo 借助 ElevenLabs 将自动视频配音扩展至 40 种语言

    Vimeo 通过 ElevenAPI 集成 ElevenLabs Dubbing,将转录、翻译、声音克隆和语音合成整合为单一流程,支持 40 种目标语言的自动视频配音。自 2026 年 1 月以来,客户已为超过 25,000 个视频配音 1.4 million 分钟,超过 27% 的企业账户尝试过 AI 配音,平均每个视频覆盖 3.3 种目标语言。创作者还可编辑文本轨道,在发布前检查和调整配音结果。

    uary 2026, Vimeo customers have dubbed 1.4 million minutes of footage across more than 25,000 videos, and more than 27% of enterprise accounts have tried AI dubbing to date. On average, each video is dubbed into 3+ target languages, with thousands that have been dubbed into 10+ languages. Customers dub across 35 source languages and 40 target languages. English is the most common source language, and the most requested targets are Spanish, French, German, Latin American Spanish, and Italian. Ent

    匹配字段正文

星期五

  1. xAI News精选

    xAI 用 Grok Bot 做采购,Haggle Bot 省下超 10 万美元

    xAI 发布 Grok Bot 采购实践,其搭建的 Haggle Bot 读取供应商支出、合同与用量数据,已识别出超过 10 万美元直接节省。Haggle Bot 接入 Slack、Notion、Drive、Gmail、Hex 和 Ramp,梳理出约 125 家活跃供应商,并自主完成未使用 SaaS 席位审计与续约谈判准备。xAI 表示该 Bot 已可在 Bot Marketplace 复制导入,同时开放 Grok Bot for Enterprise 等待名单。

    ch the "strong finding" example above. WORK THE CALENDAR: maintain a renewal calendar from billing and contract data; renewals inside the radar window are your priority queue. Leverage lives at renewal. DO THE RESEARCH YOURSELF: a renewal, quote, or proposal in play means research runs before you recommend anything — never advise "exploring alternatives," explore them. Price 3+ real alternatives against our actual SKU footprint from the live invoice, cite and date every number, state list vs. st

    匹配字段正文

星期四

  1. Hugging Face

    Hugging Face 发布 funes,为编码智能体加上可自有的记忆

    Hugging Face 发布开源工具 funes,用一条 add 命令为 Claude Code、Codex、pi、Hermes 等编码智能体加上记忆,索引增量进行,embedding 与 rerank 都在本机完成。记忆以 Lance 数据集形式存在本地,也可绑定为 Hugging Face 私有数据集实现跨机器、跨团队共享,发布前会先做凭据脱敏与密钥扫描。官方在 handoff-vs-recall 基准上对比了压缩、手写交接与 recall 三种方式,recall 在两个任务上分别比手写交接便宜 8 倍和 4 倍。

    worth paying the cost of evaluating the embeddings (this extra cost you see at indexing and recall time). funes shines when the information is actually difficult to find. johandoc 21 days ago • This comment has been hidden (marked as Spam) Edit Preview Upload images, audio, and videos by dragging in the text input, pasting, or clicking here . Tap or paste here to upload images Comment · Sign up or log in to comment +138 Datasets mentioned in this article 1 dacorvo/funes-handoff-recall-benchmark

    匹配字段正文

  2. xAI News精选

    xAI 发布 Grok Bot 设计说明,以持久化智能体取代会话式交互

    xAI 发布 Grok Bot 的设计说明,将产品主对象从聊天会话改为持久存在的 Bot,每个 Bot 拥有独立身份、记忆、运行时和工具。界面围绕 Bot、Chat、Prompt、Tool、Artifact 五个概念组织,工具与技能放在账户层,记忆与 Routine 归属具体 Bot,账户上限约 50 个 Bot、群聊中 6 个。Bot 头像同时承担身份与状态显示,其独立计算机提供状态、预览和接管三级访问方式。

    control it? We explored four arrangements: Floating window: kept the computer easy to reach but covered the conversation. Side by side: made the work continuously visible and encouraged users to watch it. Modal: made checking in easy but treated the Bot’s workspace as a temporary interruption. Full screen: gave the computer plenty of room but displaced the conversation entirely. The more prominent we made the computer, the more the product encouraged users to supervise it. We decided it should r

    匹配字段正文

星期三

  1. Anthropic News

    Anthropic 发布 Enterprise Frontier Safeguards,客户自持数据存储

    Anthropic 发布 Enterprise Frontier Safeguards(EFS),把零数据留存(ZDR)的隐私性与检测滥用的安全监控结合,数据存放在客户自己控制的云基础设施中,而非 Anthropic。该方案与 100 多家金融、医疗、制造、电信、法律、零售及公共部门客户以及 AWS、Google Cloud、Microsoft Azure 合作开发,将支持 Claude Code、Claude Enterprise、Claude Platform、Amazon Bedrock、Google 的 Agent Platform 和 Microsoft Foundry,今秋晚些时候起分阶段推出;在 EFS 就绪前,符合条件的客户可在 Fable 5 和 Fable 5.1 上获得 ZDR。EFS 的客户自有存储、客户管理加密密钥和全自动审查均为可选开启,不改变模型行为、API 定价或速率限制,Anthropic 不对此收费。

    Developing Enterprise Frontier Safeguards with our customers

    Because of these concerns, we architected EFS so that customers have the ability to store data on their existing cloud infrastructure. In EFS, customers can control their data storage and management. Customers want the ability to have their data live in infrastructure they control, under their own encryption keys, access policies, and audit logging. Activity data used for monitoring can be stored in the customer’s own cloud account (such as Amazon S3, Azure Blob Storage, or Google Cloud Storage

    匹配字段正文

星期二

  1. Hugging Face

    Hugging Face 发布 @huggingface/kernels:207 个 WebGPU kernel 支持本地 AI

    Hugging Face 发布 @huggingface/kernels,在 webgpu-kernels 组织下以 Apache-2.0 许可发布 207 个 WebGPU kernel,并提供从 Hub 下载、准备和运行 kernel 的 JavaScript 加载器。每个 kernel 有独立仓库与 kernel card,包含 manifest.json、metadata.json、test.json、bench.json 和 WGSL 模板,可通过 npm install @huggingface/kernels@preview 安装,需浏览器支持 WebGPU。在 Apple M4 上与 ORT WebGPU 对比的 809 个可比用例中,其 kernel 几何平均快 2.57 倍、中位数快 1.90 倍;同时推出浏览器端基准工具 Fleet 众包真实设备上的正确性与性能证据。

    , each run privately contributes evidence that helps us spot device-specific failures, compare variants, and improve selection rules. The goal is simple: use broad, real-world coverage to make the kernels faster and more reliable for everyone. Building a shared foundation for WebAI The initial 207 kernels are a starting point, not the end state. Publishing kernels independently on the Hub gives us a common place to inspect contracts, compare implementations, reproduce correctness checks, and imp

    匹配字段正文

  2. Google DeepMind精选

    Google DeepMind 为 Gemini 3.7 Flash 等模型推出智能体视频理解功能

    Google DeepMind 为 Gemini 3.7 Flash、3.6 Flash 和 3.5 Flash-Lite 推出智能体视频理解功能,模型可动态扫描视频片段,token 消耗最多降低 88%、成本最多降低 66%、准确率最多提升 7%。该功能已在 Google AI Studio 和 Gemini Enterprise Agent Platform 的 Gemini API 中上线,支持视频上传和 YouTube 视频,按标准 token 计费且不额外收费。相比固定 1 FPS 的静态处理,它结合模型推理与原生视频工具,按需检索画面、音频和转录文本,后续还将接入 Gemini 应用和 YouTube 的 Ask YouTube 功能。

    ill roll out to all users in the Gemini app across Flash and Flash-Lite models soon. And in the coming months, agentic video understanding will also power YouTube's ‘ Ask YouTube’ feature on the video watch page, leveraging Gemini to deliver higher-quality answers grounded in the visuals. Acknowledgement for their contribution to this work: Sergi Caelles, Filip Pavetić, Ahmet Iscen, Suhas Yogin, and the Agentic Vision team. Done. Just one step more. Check your inbox to confirm your subscription.

    匹配字段正文

星期一

  1. Fireworks AI

    Fireworks AI 宣布 Training API 与 Fireworks Lab 正式可用

    Fireworks AI 宣布 Training API 和 Fireworks Lab 正式可用,Training API 让用户用 Python 编写自定义训练循环,由 Fireworks 托管分布式训练与 rollout 基础设施。该 API 提供 Serverless 和 Dedicated 两种算力选项,前者按 token 计费、支持 LoRA,后者按 GPU 小时计费、支持全参数训练和更大的模型。原文还列出 Harvey、Vercel、Heidi、Factory 等团队用该平台微调模型的效果,例如 Vercel 为 v0 的自动修复功能微调开源模型后错误率为 93%、端到端延迟改善 40x。

    rstand where existing training workflows fall short: constrained model and method choice, limited control over parameters and training loops, rigid and underutilized compute, and fragmented training and rollout infrastructure that bloats costs, creates synchronization overhead, and introduces numerical drift. Those partnerships shaped our Training API. The Fireworks Training API connects your custom training loop to Fireworks-managed distributed training and rollout infrastructure. You orchestra

    匹配字段正文

星期五

  1. Anthropic News

    Anthropic 开放 Model Hardware Standard 研究预览,让 AI 智能体操作实验硬件

    Anthropic 面向首批科研实验室与先进制造厂商开放 Model Hardware Standard(MHS)研究预览,这是一套让 AI 智能体安全操作物理设备的共享规范,可并行驱动显微镜、液体处理机和机械臂等仪器。MHS 通过标准化驱动把硬件集成时间从数周或数月缩短到数小时甚至数分钟,兼容任何具备可编程接口的设备,且与模型无关,智能体可通过 MCP 等标准协议接入。Anthropic 表示正与科学、机器人、电子和制造领域伙伴共建安全评估与最佳实践,之后再将该标准开源。

    g the data streams was similarly time-consuming, because each application needed its own bespoke viewer, written in whatever language the recording program used. MHS unified that fragmented process and removes the per-program rewrite. With MHS, each data stream is stored in shared memory in the MHS state dictionary, in a documented format that is readable by any process that attaches to it. Because each data stream is presented in the same way, analysis or visualization code can now be reused ac

    匹配字段正文

星期四

  1. NVIDIA Developer · Generative AI精选

    NVIDIA 在 GB300 NVL72 上实测 Qwen3.8-Flash-Next 智能体编码

    阿里发布 Qwen3.8-Flash-Next 权重,作为 Qwen4 架构预览供开发者实验评估,这是一个 125B 参数、每 token 激活 6B 的多模态 MoE 模型,原生上下文窗口 262,144 token,借助 YaRN 可扩展至 1M token。NVIDIA 提供 SGLang、vLLM 和 TensorRT LLM 的 Day 0 功能支持,并在 GB300 NVL72 上验证推理,单 GPU 吞吐超过 16K token/秒、单用户超过 200 token/秒。开发者可用 NVIDIA NeMo AutoModel 微调、用 NeMo RL 做强化学习,权重可从 Hugging Face 或 ModelScope 下载。

    Decorative image.

    Technical Blog sessionStorage.removeItem('facets.search.text'); sessionStorage.removeItem('facets.search.text-copy'); sessionStorage.removeItem('facets.search.filters'); Experiment with Qwen3.8-Flash-Next on NVIDIA GB300 NVL72 for Agentic Coding Qwen3.8-Flash-Next is a multimodal mixture-of-experts model with 125B parameters, 6B activated per token, and a native 262,144-token context window extensible to 1M tokens with YaRN. The architecture combines Gated DeltaNet and Qwen Sparse Attention, whe

    匹配字段正文