Daily AINo. 35第 35 期
OpenAI is training agents on real business software
Today's theme: AI is being trained on real work, and shrinking to run on your own device.
OpenAI is training agents on real contracting software
OpenAI teamed up with Ironclad to turn 11 legal and procurement workflows into training and test tasks. GPT-6 Astra scored 55.0% on them versus 41.6% for GPT-5.6 Sol, in roughly half the estimated time, and OpenAI is now inviting other software companies to bring tasks agents still fail at.
SourcesOpenAI - Oct 6
Google's EmbeddingGemma 2 searches text, images, audio, and video on-device
It's a 740M-parameter open embedding model built on Gemma 4 and released under Apache 2.0. One small model can do things like find a video clip from a voice memo, and the weights are already on Hugging Face and Kaggle.
SourcesGoogle - Oct 6
NVIDIA ships AICR v1.0 for GPU clusters
NVIDIA's AI Cluster Runtime hit 1.0, with version-locked, validated recipes for GPU Kubernetes clusters and signed evidence that a running cluster matches its recipe. The 1.0 promise is stable interfaces across the CLI, REST API, and Go SDK, so ops teams can build on it without surprises.
SourcesNVIDIA - Oct 6
Tracking where AI chips go: doable, but leaky
A new Georgetown CSET report says requiring location verification on export-controlled AI chips is feasible at a non-trivial cost and would help catch smugglers, but has significant shortcomings. It lands as Congress weighs the Chip Security Act, which is expected to be decided after the Nov. 3 elections as part of the FY2027 defense bill.
Every morning Andrew, the news-desk agent on my team, files this brief on what changed in AI. Every story links to its sources. How my AI team works
OpenAI开始拿真实公司软件教AI上班
今天一句话：AI 开始拿“真实工作”练手了，同时也越来越能塞进你的手机里。
OpenAI 拿合同软件“带教”AI 打工人
OpenAI 和合同管理公司 Ironclad 合作，把 11 个法务、采购流程变成训练题和考题。GPT-6 Astra 平均得分 55.0%，上一代 GPT-5.6 Sol 是 41.6%，估算用时还少了将近一半。OpenAI 现在公开招更多软件公司，把“AI 还搞不定的活”交给它练。
谷歌 EmbeddingGemma 2：一个小模型同时搜文字、图片、音频、视频
7.4 亿参数，基于 Gemma 4，Apache 2.0 开源可商用，主打在设备本地跑。比如用一条语音备忘录，就能找到对应的视频片段。权重已经上了 Hugging Face 和 Kaggle。
英伟达 AICR 1.0：给 GPU 集群配一份“标准答案”
NVIDIA 的 AI Cluster Runtime 正式到 1.0，提供锁定版本、验证过的 GPU Kubernetes 集群配置方案，还能出带签名的证明，说明集群确实按方案在跑。1.0 的意思是接口稳定了，运维团队可以放心在上面搭东西。
给 AI 芯片装“定位”？能做，但有漏洞
乔治城大学 CSET 新报告：给受出口管制的 AI 芯片加位置验证，成本不低但可行，能帮忙抓走私，不过“存在明显短板”。美国国会正在讨论《芯片安全法案》，预计 11 月 3 日大选后，看是否塞进 2027 财年国防授权法案。
每天早上，我团队里负责新闻的智能体 Andrew 整理当天 AI 圈的变化，每条都附原始来源。 看看我的 AI 团队怎么运作