NVIDIA发布ENPIRE实现机器人自我改进；腾讯公开ARGUS追踪万卡GPU集群；UC Berkeley发布LOCUS地方法规数据集
英文摘要
NVIDIA researchers introduce ENPIRE, a framework that lets coding agents autonomously improve physical robot policies through automatic trial, evaluation, and reset. On dexterous tasks like push-T, pin insertion, and GPU installation, the best AI agents (GPT-5.5, Opus 4.7) reach 99% success, while multi-agent setups yield higher final scores. Tencent describes ARGUS, a low-overhead tracing system deployed on a 10,000+ GPU cluster for over six months to debug large-scale training jobs, including Hunyuan MoE and video models. UC Berkeley releases LOCUS, a corpus of 2.2 million rows of U.S. local ordinances harmonized for legal AI research. The issue also includes a cautionary essay on the historical failure of technology predictions and a long piece on AI-driven human disempowerment.
中文摘要
NVIDIA研究者推出ENPIRE框架，使编程智能体通过自动试验、评估和环境重置在真实世界中自主改进机器人策略。在PushT、插针、安装GPU等灵巧任务上，顶尖AI智能体（GPT-5.5、Opus 4.7）的成功率达99%，且多智能体系统最终得分更高。腾讯公开ARGUS——一款低开销追踪系统，已在超过1万个GPU的生产集群上稳定运行六个月，用于诊断大规模训练（如混元MoE和视频模型）中的故障。加州大学伯克利分校发布LOCUS，一个包含220万行美国地方法规的语料库，以推动法律AI研究。本期还包含一篇提醒科技预测常常出错的文章和一篇关于AI导致人类失权的长文。
关键要点
NVIDIA’s ENPIRE framework enables autonomous real-world robot policy improvement via coding agents, achieving 99% success on dexterous tasks and showing benefits from multi-agent parallelism.
NVIDIA的ENPIRE框架让编程智能体在真实世界中自主改进机器人策略，在灵巧任务上达到99%成功率，并展现出多智能体并行的优势。
Tencent’s ARGUS tracing system has been deployed on a 10,000+ GPU cluster for over six months, providing fine-grained, always-on diagnosis for large model training jobs including Hunyuan MoE.
腾讯的ARGUS追踪系统已在一万多个GPU的集群上部署超过六个月，为混元MoE等大模型训练提供细粒度、不间断的诊断。
UC Berkeley released LOCUS, a corpus of 2.2 million rows of U.S. local ordinances, making machine-readable local law accessible for legal AI research.
加州大学伯克利分校发布LOCUS，一个包含220万行美国地方法规的语料库，使机器可读的地方法律可用于法律AI研究。
A historical review argues that both skeptics and optimists consistently fail to predict technology’s impacts, cautioning against complacency about AI’s future effects.
一篇历史回顾指出，怀疑者和乐观者都一贯预测错技术的影响，提醒不要对AI的未来影响掉以轻心。