Odyssey 发布下一代多智能体世界模型 Agora-2，支持最多 20 个人类与智能体在同一共享环境中实时交互。其研究预览中，最多 4 名人类可组队与 16 个智能体对战，整个环境由世界模型实时模拟，底层没有游戏引擎。Odyssey 认为多智能体世界模型将越来越多地支撑 AI 训练、AI 安全、机器人、自动驾驶、国防、能源、网络安全和游戏等应用。
This research preview of Agora-2 enables up to 4 humans to join together to battle 16 agents, in an environment all simulated in real time by the world model. There's no game engine under the hood. https://t.co/PLpeZOdBqH
匹配字段正文
Gemini 3.8 TTS 支持复刻用户自己的声音，也能通过提示词设计完全自定义的音色。流程分三步：录制 20 秒本人说话音频加一句同意声明，通过 API 调用创建声音，之后在任意请求中使用，风格参数放在 speech_metadata 里。作者还给出了一段可直接粘贴给智能体的提示词，让智能体先检查 GEMINI_API_KEY、ffmpeg、gemini-skills 等环境，再协助录制两段音频、创建声音并生成一条可试听的测试语句。
can replicate your voice or design a completely custom one from a prompt 1. Record 20s of you talking + the consent sentence 2. Create your Voice via API call 3. Use it in any request, style goes in speech metadata Past this into your agent "Read https://t.co/8DjBBAfAhv and walk me through creating my own voice for Gemini 3.8 TTS. Check my setup first (GEMINI API KEY, ffmpeg, gemini-skills), help me record the two clips, create the voice, and generate a test line I can listen to." or read below.
匹配字段正文
LangChain 发布 LangSmith Engine v2，新增红队测试功能，可在问题影响生产前主动发现模型幻觉和系统提示词违规等问题。Engine v2 还能识别低效的智能体路径以及错误率、延迟和成本的变化趋势，并在人工审核前自动复现故障、测试和验证修复方案，用户可在 LangSmith 中一键提交 PR 部署。红队测试和修复验证目前面向现有 LangSmith Deployment 用户开放 Private Beta，Engine v2 已在 LangSmith Plus 和 Enterprise 计划的 SaaS 部署中可用。
Key Takeaways Red team your agents: Engine proactively finds issues in your agent before they emerge in production; Detect more issue types: Engine detects inefficient agent paths and trends in error rate, latency, and cost; Automatically test proposed fixes: Engine now tests and validates prompt and code fixes before showing them to you We built LangSmith Engine to make fixing agents faster, so developers can spend more time building new capabilities and less time fighting bugs. Engine is an in
匹配字段正文
LangChain 推出 LangSmith Fine-Tuning 和 smithtune CLI，可从 LangSmith 轨迹生成数据集、在 Fireworks 或 Baseten 上训练并评估微调模型。smithtune 目前支持监督微调（SFT），训练走 LoRA，无需自行管理 GPU 或训练基础设施。官方实测中，Qwen-3.8-27B 经 SFT 后 F1 从 48.9% 升至 53.7%，模型调用减少 29.8%、工具请求减少 29.4%。
Key Takeaways Use the LangSmith Fine-Tuning CLI and skill - smithtune - to leverage agent trajectories stored in LangSmith to a fine-tuned model in one end-to-end workflow; In addition to driving training, the smithtune CLI handles evaluating the fine-tuned model and uploads eval results to LangSmith for easy analysis; We partnered with Fireworks & Baseten to create a seamless link between LangSmith trajectories and creation of a dataset for fine-tuning both the platforms Today we're launching L
匹配字段正文
Anthropic 发布 Claude Opus 5.5，官方称其在多数任务上达到 Claude Fable 5.1 的水平，运行成本比 Opus 5 低 40%，Claude Code 的五小时额度提升 20% 并新增额度重置机制。OpenAI 同期推出 GPT-6 Luna 和 GPT-6 Sol，仅先在 Codex 和 ChatGPT Work 提供，价格比上一代便宜 50%，Luna 为每百万 input/output token 0.10/0.50 美元，Sol 为 2/10 美元，作者实测 Sol 跑约 4 小时的长任务消耗了 100 美元 Codex 周额度的 22%。
detailed research” detroitbench - “run the three new models and add them to the results” Every says it’s pulling their Codex converts back to Claude and I feel the same. My go-to agent has changed to Claude Code with this release. GPT-6 Luna and Sol from OpenAI . There’s no Terra this time. Don’t think anyone was even using it. Both GPT-6 Luna and GPT-6-Sol are only available in Codex and ChatGPT Work (not regular chat) for now. Key part: both of them are 50% cheaper than their previous version
匹配字段正文
Meta 工程博客介绍把 Private Processing 机密计算基础设施扩展到 Meta AI 眼镜，让流式转录、上下文搜索和长期记忆等云端 AI 负载在机密虚拟机（CVM）内执行，Meta 自身也无法读取用户数据。该方案基于 TEE 硬件隔离，要求远程认证、非定向路由、加密存储和公开可验证的透明账本，并扩展漏洞赏金计划覆盖 AI 眼镜上的 Private Processing。此前 Meta 已在 2025 年为 WhatsApp 和 Meta AI 应用推出 Private Processing。
chers with the tools, CVM binaries, and documentation needed to audit our implementation, test our attestation chains, and hold our platform accountable. Private Processing for an Agentic Future To date, Private Processing has focused on discrete tasks, like summarizing a message. But the future of AI is agentic and multimodal. In the future, your glasses will have the capacity to take actions on your behalf across different sessions, in a range of real-world contexts. As we work towards launchi
匹配字段正文
Google Cloud API Gateway 进入公开预览，可充当远程 MCP 服务器：在已有的 OpenAPI 3.0.x 或 3.1.x 规范上添加 x-google-api-management.mcp 注解并部署后，现有 REST 操作即可作为 MCP 工具供智能体调用，无需另建服务器。网关在 /mcp 路径接收 MCP JSON-RPC 请求，把 tools/call 转码为对应 REST 请求并沿用原有 JWT 或 API key 认证、配额和日志策略。该预览版覆盖 REST 与 OpenAPI 3.x 后端，MCP resources、prompts、响应流式传输和 Model Armor 载荷检查仍在路线图中。
Facebook; Twitter; LinkedIn; Mail Most enterprise capability sits behind REST APIs that agents cannot see. To make one callable by an agent today, teams typically stand up and operate a separate MCP server that re-implements the routing, authentication, and quota logic their gateway already handles. The Model Context Protocol (MCP) has become the standard way for agents to discover and invoke tools, and frameworks like the Agent Development Kit (ADK) and Gemini Enterprise speak it natively. Goog
匹配字段正文
Google 开始测试名为 Call for Me 的新功能，让 Gemini 代替用户给商家打电话。该功能首批面向美国付费订阅 Gemini 的 Pixel 11 用户，并需使用 Google Phone 应用的 beta 版本。Gemini 可在通话中分享用户批准的个人信息，处理查询库存、餐厅订位、改约、预留商品等任务，能自我介绍、应对自动语音菜单、等待接通并完成对话，用户可实时查看通话转写并随时接管，通话直接使用用户本人的手机号码拨出。
With AI agents like Meta’s Muse and Instinct now able to make calls on users’ behalf, Google is now letting Gemini call businesses with a new feature called “Call for Me.” The feature will initially roll out to Pixel 11 owners who pay for a Gemini subscription in the U.S. It will also require the use of the beta version of Google’s Phone app for Android, as it’s still supposed to be an experiment. While Google has long offered calling features guided by AI, prior versions were limited in what th
匹配字段正文
Baseten 在 Model Library 上线 NVIDIA Nemotron 3 Diarization，提供预录、流式和实时转写三种预设，单次推理即可标注最多 8 位说话人，间隔最低 320ms。该模型约 100M 参数，基于 NVIDIA Streaming Sortformer，同一 checkpoint 支持 0.32 至 30.4 秒四档算法延迟；在 RTX PRO 6000 上可维持 500 路以上并发分离流，加入转写后为 190 路。AISHELL-4 低延迟档 DER 为 9.8%，对比 Streaming Sortformer v2.1 的 27.2%，模型已在 Hugging Face 和 Baseten Model Library 开放。
n over 500 concurrent hour-long diarization streams (190 with transcription added). The same per-speaker activity can be used to inform VAD, endpointing, and turn-taking for voice agents. Speaker diarization segments an audio stream by when each distinct speaker is talking, outputs timings for each distinct voice, and assigns every span a consistent speaker label. In conjunction with transcription, diarization can attribute transcribed words to who said them in a conversation. The NVIDIA Nemotro
匹配字段正文
NVIDIA Nemotron 3 Diarization 已在 Hugging Face 和 Baseten Model Library 上线，提供批量、流式和实时说话人标注转写三种预设。该模型基于 NVIDIA Streaming Sortformer，约 100M 参数，单次前向即可输出说话人活动，最多标注 8 位说话人，同一 checkpoint 支持 0.32 至 30.4 秒四档算法延迟。Baseten 在 RTX PRO 6000 上优化后，单卡可维持超过 500 路并发小时级分离流，加入转写后为 190 路，批量吞吐比 NVIDIA 参考部署高约 1.35 倍。
n over 500 concurrent hour-long diarization streams (190 with transcription added). The same per-speaker activity can be used to inform VAD, endpointing, and turn-taking for voice agents. An end-to-end streaming architecture Speaker diarization segments an audio stream by when each distinct speaker is talking, outputs timings for each distinct voice, and assigns every span a consistent speaker label. In conjunction with transcription, diarization can attribute transcribed words to who said them
匹配字段正文
Meta 发布 Meta VR Glasses，整机约 100 克，比 Meta Quest 3 轻五倍，采用眼镜加外接 puck 的两段式设计，puck 负责算力、电池与存储。设备搭载 5K Infinite Display 与 micro-OLED 面板、37 PPD，支持 Dolby Vision 和 Dolby Atmos，由高通 Snapdragon Reality Elite 处理器驱动，puck 电池可支持约三小时高分辨率媒体播放和 45W 快充。Meta AI 智能体已集成进操作系统，可通过语音、眼动和手势操作，无需手柄；产品将于 2027 年春季开售，售价 1299.99 美元。
y scene come alive with ultra-vivid color, sharp contrast, and lifelike detail, and Dolby Atmos spatial audio is integrated directly into the frames. We’ve integrated your Meta AI agent directly into the operating system. You can just talk to Meta VR Glasses or use your eyes and natural hand gestures to do things like play a movie, open an app, adjust your workspace, or search for whatever you’re looking for. No controllers needed. Meta VR Glasses are powered by Qualcomm’s new Snapdragon Reality
匹配字段正文
Meta 在 Connect 上发布首款音频眼镜 Ray-Ban Meta Audio，重 43 克，单次充电可听音乐、通话并与 AI 助手对话最长 12 小时，配合充电盒可再延长 48 小时，起售价 349 美元，10 月 13 日发货。同时 Meta 将个人 AI 智能体 Muse 引入 AI 眼镜，用于管理日程和长期目标，并发布 Ray-Ban Meta（Gen 3），续航提升至 9 小时，配备 6 麦克风阵列可削减 90% 以上背景噪音，起售价 449 美元。Meta 还宣布与 LISA 合作推出 Meta Glasses 系列，并计划到年底在 Ray-Ban、Oakley 和 Meta Glasses 下提供超过 100 种眼镜选择。
Takeaways Today at Connect, we announced we’re bringing Muse, a personal AI agent, to our AI glasses to help you stay on top of things, take tasks off your plate, and turn long-term goals into action plans, all hands free; We also introduced Ray-Ban Meta Audio, our first-ever audio glasses, and our biggest expansion of AI glasses; Our expanded glasses lineup includes an all new Meta Glasses collection designed with international icon, LISA; improved battery life and controls for Ray-Ban Meta (Ge
匹配字段正文
Prime Intellect 宣布 Prime Sandboxes 正式可用，每个沙箱是一台完整 Linux 虚拟机，支持 Docker Compose、后台任务和依赖内核的工作负载，可通过 CLI/SDK 独立使用或接入其 RL 套件。官方称已创建约 3000 万个沙箱，可运行数万个并发沙箱，并接入 verifiers、prime-rl 与 Prime Tunnels，提供超过 365,000 个预置环境。定价按用量计费、无订阅和最低消费，官方称截至 12 月 22 日价格为其他大型沙箱供应商的三分之一。
In agentic RL, the sandbox is the foundation of a training system. It creates the state that the agent acts on, enforces isolation boundaries and produces trajectories that feed directly into the reward signal. However, model training requires running thousands of concurrent sandboxes, which leads to complex configuration and can be very costly. Prime Sandboxes were built to solve this, responding to the needs of research teams like ours as we were unable to find adequate options on the market.
匹配字段正文
NVIDIA 在 ROSCon 发布 Isaac ROS 5.0，为基于 ROS 的机器人开发引入智能体工作流与新的平台支持，并新增用于环境搭建和操作的 Isaac 技能。该版本支持 ROS Lyrical 和 Ubuntu 24.04，FoundationPose 的智能体推理库让机器人感知和跟踪物体位姿的速度最高提升 5.5 倍，抓取放置流程也作为独立技能开放。Isaac ROS 5.0 支持从 Jetson Orin Nano 到 Jetson Thor 的可扩展算力，Mentee Robotics、Universal Robots、ROBOTIS 等公司已将其用于各自机器人平台。
humans build robots. NVIDIA Isaac ROS 5.0 — a collection of GPU-accelerated packages built on ROS, released today at the ROSCon conference in Toronto, Canada — helps humans and AI agents build robots together. The release introduces new agentic workflows and platform support to help developers build, customize and deploy robotics applications faster. ROS provides the open source foundation for much of modern robotics development, giving developers common tools, libraries and standards for buildi
匹配字段正文
OpenAI 的 GPT-6 Sol 和 GPT-6 Luna 现已在 Amazon Bedrock 正式可用，两者 API 定价均明显低于前代 GPT-5.6。GPT-6 Sol 面向开发与运维中反复出现的复杂任务，可写功能、调试、重构与审查代码、分析数据并完成跨工具的多步流程；OpenAI 内部事实性评测显示其事实性错误约为 GPT-5.6 Sol 的一半。GPT-6 Luna 面向高频重复任务，如从大规模文档中抽取信息、摘要、分类和定向问答，并支持按请求调整推理投入。两者均支持显式提示词缓存，推理数据不用于模型训练。
drock can support your team? Connect with us to start the conversation. Tanvi is a Product Marketing Manager for Amazon Bedrock at Amazon Web Services (AWS), where she helps customers adopt and scale AI applications and agents with Amazon Bedrock. Chris is a Member of Product Staff at OpenAI focused on the OpenAI APIs. His work includes collaboration with AWS on Amazon Bedrock to make OpenAI’s frontier models widely accessible to developers. Manish is a Senior Product Manager for Amazon Bedrock.
匹配字段正文
Fireworks AI 发布 Specialized Intelligence Index（SII），汇集开放、闭源与专用模型在医疗、法律、网络安全、金融、客服、生产力、软件七个领域的真实任务基准。该索引不做跨领域加权汇总，只保留基准与领域层面的结果，并提供分数对成本、分数对耗时的视图；每个基准标注可复现范围，新前沿模型需通过生产可部署性筛选后才会进入。Fireworks 同时开放合作提交，伙伴按 Harbor 兼容格式提供任务与数据，评测由 Fireworks 免费运行并在伙伴同意后公布分数。
easoning, recall, or both. Does the test represent the work we care about? Representativeness concerns the coverage and composition of the task set. Wang and colleagues studied 43 agent benchmarks and found a concentration in computer and mathematical work, a category accounting for 7.6% of U.S. employment in their analysis. Management and legal work were underrepresented, as were interpersonal skills common across occupations. Real-work benchmarks must represent their intended domain. Evaluatin
匹配字段正文
Google 发布新版 Gemini 托管智能体，缓存命中率最高提升 22%，成本最高降低 30%。新增 Files API 支持在沙箱中上传、列出和下载文件，Credentials API 为 MCP 服务器、OAuth2 和第三方 API 提供安全出口代理。Google AI Studio 和 Gemini API 提供免费层供试用。
Today we are releasing a new Gemini managed agents version with better caching, lower costs, a Files API, and a Credentials API. 📉 Up to 30% lower costs with up to 22% higher cache hits 📁 Files API: Upload, list and download files from sandbox easily 🔐 Credentials API for MCP servers, OAuth2, and 3rd-party APIs as secure egress proxy 🆓 Free tier to experiment in @GoogleAIStudio & Gemini API
匹配字段正文
Google 与联合国系统发布 UN System Data Commons，这是一个基于 Google Data Commons 构建的开源平台，把联合国各机构的全球统计数据整合为可搜索的 AI-ready 知识图谱。平台支持自然语言查询和交互式可视化，并基于 MCP 等开放标准让 AI 智能体直接获取权威数据、生成图表或报告草稿。联合国系统计划未来一年继续接入更多数据集，目标到 2027 年覆盖 80% 的联合国系统统计数据集。
. Most importantly, every dataset is validated with UN system statisticians and technical experts, so every answer stays grounded in trusted, official facts. Putting AI to work as agentic research assistants Today’s launch also brings AI assistant capabilities directly to the research workflow. Instead of spending hours manually searching for numbers and assembling spreadsheets, you can prompt an AI assistant to do the heavy lifting. Built on open standards like the Model Context Protocol (MCP),
匹配字段正文
Baseten 发布 Hosted Tools，首个工具为网页搜索 Baseten Grounded Inference，首批接入 Exa、Keenable、Parallel 和 You.com 四家搜索服务商。开发者只需在 Messages、Chat Completions 或 Responses 请求中加一项配置即可启用，搜索在 Baseten 推理栈内服务端执行，模型按原生工具调用能力自行决定是否检索并迭代。Baseten 称其基准测试中端到端延迟比客户端工具方案降低 15%，该功能目前以 playground 预览形式开放，限速 25 RPM 并赠送 2 美元额度。
and lowers latency. Baseten Grounded Inference; Launching with Exa, Keenable, Parallel and You.com; Baseten Hosted Tools; Server-side tool execution = fewer round trips and faster agents; One platform for models and agents Today we’re introducing Baseten Hosted Tools, a new ecosystem for capabilities like web search, code execution in sandboxes, browser use and automation, and more to enhance models and agents running on Baseten. The first category of tools we are introducing is Web Search, and
匹配字段正文
Baseten 发布 Hosted Tools，首个类别为网页搜索，联合 Exa、Keenable、Parallel 和 You.com 四家搜索提供商，开发者只需在 Messages、Chat Completions 或 Responses 请求中加一行配置即可启用。Baseten 称托管工具把工具调用放进服务端智能体循环，相比客户端工具端到端延迟降低 15%，并支持流式输出工具调用与结果。该功能目前处于 playground 预览阶段，限速 25 RPM，附赠 2 美元免费额度。
y we’re introducing Baseten Hosted Tools, a new ecosystem for capabilities like web search, code execution in sandboxes, browser use and automation, and more to enhance models and agents running on Baseten. The first category of tools we are introducing is Web Search, and we are excited to launch it today with Exa, Keenable, Parallel, and You.com, bringing their search technologies into the Baseten platform with a consistent developer experience. Until now, developers building apps and agents wi
匹配字段正文