<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>智身居 · NVIDIA</title><description>智身居里标了「NVIDIA」的稿。</description><link>https://notesofone.com/</link><item><title>9 月 30 日：NVIDIA 把视觉智能体和模型连接写成可核对的工程</title><link>https://notesofone.com/notes/daily-2026-09-30/</link><guid isPermaLink="true">https://notesofone.com/notes/daily-2026-09-30/</guid><description>NVIDIA 公开 TensorRT Model Connect 的做法：按模型族隔离、可逆改动、GPU 校验。Metropolis VSS 3.3 把视频检索与摘要接到 Cosmos、Nemotron、RAG 和 MCP。Apple 一篇 NeurIPS 论文用往返协议测链式推理把树状表达式写成自然语言之后还剩多少。成本下降和分数没有独立复现。</description><pubDate>Wed, 30 Sep 2026 00:00:00 GMT</pubDate><category>NVIDIA</category><category>TensorRT</category><category>VSS</category><category>MCP</category><category>Apple</category></item><item><title>9 月 24 日：3D CT 开始「逐步想」，机器人仿真上 GPU</title><link>https://notesofone.com/notes/daily-2026-09-24/</link><guid isPermaLink="true">https://notesofone.com/notes/daily-2026-09-24/</guid><description>英伟达开源 NV-Reason-CT，把放射科式思维链接到三维 CT，并给出 CT-RATE 上的 Macro-F1 与 Macro-AUROC。SWE-Serve 用 53 道推理工程题分开本地测试和线上服务。Hugging Face 上的 Warp / MjWarp 教程把 MuJoCo 场景迁到 GPU。数字是作者自述。</description><pubDate>Thu, 24 Sep 2026 00:00:00 GMT</pubDate><category>NVIDIA</category><category>医学影像</category><category>三维CT</category><category>具身智能</category><category>MuJoCo</category><category>GPU集群</category><category>每日更新</category></item><item><title>9 月 23 日：机密计算上的生产推理，实验室把 ROS 2 搬进 GPU 缓冲</title><link>https://notesofone.com/notes/daily-2026-09-23/</link><guid isPermaLink="true">https://notesofone.com/notes/daily-2026-09-23/</guid><description>英伟达把生产推理放进机密虚拟机与机密 GPU，Isaac ROS 5.0 用 CUDA 缓冲做零拷贝。数字和架构都按厂商博文转述。</description><pubDate>Wed, 23 Sep 2026 00:00:00 GMT</pubDate><category>AI</category><category>机密计算</category><category>具身智能</category><category>ROS 2</category><category>NVIDIA</category><category>Isaac ROS</category><category>vLLM</category><category>GGUF</category></item><item><title>9 月 17 日：Thor 跑完边缘代理基准，仿真场景开始交给智能体</title><link>https://notesofone.com/notes/daily-2026-09-17/</link><guid isPermaLink="true">https://notesofone.com/notes/daily-2026-09-17/</guid><description>NVIDIA 技术博客给出 Jetson AGX Thor 上 MLPerf 边缘代理基准 6.4 倍加速；Omniverse 智能体把 Blender 场景补进机器人仿真。数字是提交方自述。</description><pubDate>Thu, 17 Sep 2026 00:00:00 GMT</pubDate><category>每日更新</category><category>NVIDIA</category><category>Jetson</category><category>TensorRT</category><category>Omniverse</category><category>机器人仿真</category><category>PyTorch</category><category>具身智能</category></item><item><title>9 月 16 日：NVIDIA 用激活参数比较 MoE 吞吐，IBM 把 Agent 一致性单独拿出来</title><link>https://notesofone.com/notes/daily-2026-09-16/</link><guid isPermaLink="true">https://notesofone.com/notes/daily-2026-09-16/</guid><description>NVIDIA 技术博客把 Nemotron 3.5 Lightning 写成每 token 只激活 30B 里的 3B，并和稠密的 Gemma 4 31B 比吞吐。同日前后还有 Vera Rubin 能效、NVLink 6 和 FLARE 的官方摘要。Hugging Face 上 IBM Research 讨论 Agent 一次做对、下次不一定还对。数字都是作者自述。</description><pubDate>Wed, 16 Sep 2026 00:00:00 GMT</pubDate><category>NVIDIA</category><category>MoE</category><category>Agent</category><category>评测</category></item></channel></rss>