AI, explained for everyone.
A bilingual weekly journal covering AI news, tools, and ideas in plain language.
9 月 3 日美国早间，多家主流 AI 服务先后出现故障。Grok 官方记录为 3 小时 37 分；三家服务的恢复节奏与根因并不相同。
中文 GPT-6 Astra 发布：更会“做事”的 AI，先要学会被约束OpenAI 发布 GPT-6 Astra。它把模型从“回答问题”推进到能在软件中持续执行复杂任务，也把安全与授权问题推到产品体验的正中央。
中文 英伟达收购 Hugging Face：开源 AI 的“广场”迎来新主人英伟达以 129.3 亿美元收购 Hugging Face，并承诺平台保持开放。交易改变的不是一个模型，而是 AI 开发者如何选择模型、框架和算力。
中文 Gemini 3.8 Flash 与 Flash Cyber：速度型模型也在进入专业安全场景Google 推出 Gemini 3.8 Flash 与受限开放的 Flash Cyber：前者瞄准高效率代理工作流，后者把漏洞发现和修复能力交给经过审核的防御方。
中文 从 910B 议题转向 950 落地：华为昇腾进入生产环境意味着什么？近期可核验的昇腾新闻不是 910B“启动测试”，而是昇腾 950 开始进入生产环境。它把国产算力竞争从规格讨论带到软件兼容、交付和持续运行。
中文 Claude Fable 5.1 与 Mythos 5.1：同一能力底座，为什么要分两种访问方式？Anthropic 发布 Claude Fable 5.1 与 Mythos 5.1：它们共享同一模型能力底座，但以不同安全护栏和访问门槛服务不同人群。
中文 GPT-6 Astra 的安全争议：当模型能发现零日漏洞，护栏够不够？OpenAI 将 GPT-6 Astra 标为网络安全“Critical”级能力模型。争议不只是模型有多强，而是高风险能力应如何评测、开放和监督。
English What the ChatGPT, Claude, and Grok outages revealed about AI dependenceSeveral major AI services experienced disruptions on 3 September. Grok’s official incident lasted 3 hours and 37 minutes, while the providers’ impacts and recovery timelines differed.
English GPT-6 Astra arrives: more capable AI needs stronger boundariesOpenAI’s GPT-6 Astra is designed to carry out longer, tool-using tasks inside software. That makes access controls and human oversight part of the product, not an afterthought.
English Nvidia acquires Hugging Face: an open-AI commons gets a new ownerNvidia has acquired Hugging Face for $12.93 billion and says the platform will remain open. The deal reshapes the routes developers use to choose models, frameworks, and compute.
English Gemini 3.8 Flash and Flash Cyber bring fast models into serious security workGoogle’s Gemini 3.8 Flash targets efficient agent workflows, while the restricted Flash Cyber model gives vetted defenders vulnerability-finding and patching capabilities.
English From 910B testing to Ascend 950 deployment: what Huawei’s production push meansThe verifiable recent Ascend story is not a new 910B test, but Ascend 950 moving into production. It shifts the discussion from specifications to software compatibility, delivery, and sustained operations.
English Claude Fable 5.1 and Mythos 5.1: why one model family has two access pathsAnthropic’s Fable 5.1 and Mythos 5.1 share the same underlying model, but pair it with different safeguards and access rules for different risk profiles.
English The GPT-6 Astra safety debate: are safeguards enough for zero-day discovery?OpenAI classifies GPT-6 Astra as Critical for cybersecurity capability. The debate is not only about model power, but about how high-risk capabilities should be evaluated, released, and supervised.