2026-09-04
AI 快讯日报.

Instagram AI 内容标签再陷混乱，The Verge 实测发现只有 Meta 自家信号能触发标注 ↗

Meta has been tagging real photos and original content on Instagram with AI Content labels again, while letting undeniably AI images slip through the cracks.

GPT-6 Astra 基准表现分歧，ARC-AGI-3 效率超人类令 Chollet 提前 AGI 预测 ↗

OpenAI's GPT-6 Astra is drawing contradictory benchmark verdicts. Epoch AI puts it out in front with 169 points, while Artificial Analysis rates it no better than its predecessor and behind Claude Fable 5.1. The biggest surprise comes from ARC-AGI-3, where Astra works more efficiently than the average human for the first time. ARC Prize chief François Chollet doesn't call this proof of AGI, but he does see the progress running "twice as fast" as he expected, and he's moving up his AGI forecast.

为什么 AI 生成的食物图片看起来那么诡异 ↗

Worms, holes, and cracks reveal the technical weaknesses of image generators.

Microsoft 推出面向开发者的 Project Zenith 设备，首批搭载 AMD Ryzen AI Halo 芯片 ↗

Microsoft is launching new developer-focused devices that have a customized Windows install. It’s designed to create a distraction-free Windows 11 for devs.

Sam Altman 就 GPT-6 Astra 混乱发布致歉，付费用户仍被拒之门外 ↗

OpenAI hasn’t provided a clear timeline for when subscribers will get access to the long-hyped Astra.

广汽集团与火山引擎推动一线员工用 TRAE 等工具开发 Agent，落地研产供销全流程 ↗

M&T Bank 向超 1.5 万名员工部署 AI copilots，扩展企业级 AI 应用 ↗

Checking the site connection security

Ugreen 在 IFA 发布 HomeAgent 本地智能家居平台，主打端侧 AI 助手 Uliya ↗

Ugreen’s HomeAgent combines a NAS, a security camera NVR, and an on-device AI assistant in one box, maxing out with a $9,999 Nvidia Jetson Thor version.

群联 CEO 潘健成：明年 NAND Flash 供应仍吃紧，AI 半导体需求才刚开始 ↗

潘健成指出，AI 带动的半导体需求才刚开始，若以台积电持续扩充 AI 相关产能来看，更多 AI 芯片出货势必同步推升 DRAM 与 Flash 用量。

Nvidia 发布开源工具 PAIR，把家庭网络变成本地 AI 迷你数据中心 ↗

Nvidia's PAIR (Personal AI Router) automatically spreads local AI requests across all available devices on a home network, cutting wait times for parallel agent tasks.

消息称三星电子联合 Arm 研发下一代端侧 AI 芯片，采用 2 纳米工艺量产 ↗

据半导体行业今日消息，Arm 已于 8 月底批准拨付了用于开发下一代端侧 AI SoC 芯片的一次性工程费用，并与三星电子正式启动该项目。

工信部印发人工智能中小企业创业支持计划，支持国家级AI开源社区AtomGit发展壮大 ↗

工信部发布《人工智能中小企业创业支持计划（2026-2028 年）》，目标培育 1 万家以上 AI 创业企业，并支持国家级开源社区 AtomGit 发展。计划涵盖算力、数据、开源生态等多方面举措，鼓励创业企业积极参与开源建设。##AI创业##

谷歌助理移动端 9 月 4 日起逐步停用，Gemini 全面接棒 ↗

停用工作已启动，符合条件安卓设备将自动切换，迁移分批完成，后续还将逐步扩展到更多谷歌设备。Gemini 能力更强，配置不达标或未开放地区暂时保留助手。

星尘智能发布 SmoothRL，让在线强化学习跟上大模型的异步推理 ↗

As VLAs and World-Action Models grow larger and inference times stretch longer, robots in real-world deployments have long since switched to asynchronous inference — executing task A while the model computes task B in the background. But this "deployment reality" poses a fresh challenge for online reinforcement learning: the actions the model "planned" and the actions the robot "actually performed" no longer match. Who gets credit for success, and who takes the blame for failure?

Nolan Lawson 谈 AI 冲击下的前端网页开发与前端的未来 ↗

A lot of the educators I admire in the frontend web space seem to be either bowing out or dialing back their efforts: Axel Rauschmayer, Salma Alam-Naylor, Josh W. Comeau, to name a few. Other well-known luminaries like Kent C. Dodds, Addy Osmani, Rachel Nabors, and Lydia Hallie have pivoted from talking about frontend development…

TechCrunch 解析 AI 生成菜单为何雷同难看：模型收敛与恐怖谷效应 ↗

While restaurant owners might look to generative AI as a shortcut to sprucing up their menu, customers can viscerally sense that something is wrong with the food.

后摩智能与国科亿道发布 FN008 AI 笔电，搭载飞腾 D3000M 与 160TOPS 后摩漫界 M50 ↗

后摩漫界 M50 拥有 160 TOPS INT8 算力、100 TFLOPS bFP16 算力，配备 24GB 内存，可本地运行 35B 模型。

联想在 IFA 2026 发布搭载 AMD 锐龙 AI Max+ PRO 495 的 1.6L 迷你主机 ThinkCentre X Ultra ↗

该设备支持至高 128GB LPDDR5X-8533 内存，可将 96GB 分配至 GPU；除 1 个 10GbE RJ-45 网还提供 1 个多功能灵活扩展口。

努比亚 NaviX Ultra 预热：搭载长鑫 10667Mbps LPDDR5X 内存，定位全球首款 AI 智能体手机，9 月上市 ↗

该机由中兴与字节跳动联合研发，搭载国产长鑫 10667Mbps LPDDR5X 内存，配置骁龙 8 Elite Gen5，大模型运算全部本地完成，本月正式上市。

Armature 分析 16，893 次会话：Claude Code、Codex 和 Cursor 如何选择第三方工具 ↗

We watched almost 17k sessions across different types of personas, with 1,163 prompt variations, 75 repositories and 3 coding agents (Claude Code, Codex, Cursor) actually implementing the solutions instead of just recommending one.