Import AI 463: Self-improving robots; a 10k Chinese GPU cluster; and an elegiac essay for the human era

NVIDIA researchers have developed ENPIRE, a software framework that applies agentic experimentation loops to physical robotics. According to the research, ENPIRE comprises four modules: an Environment module for automatic reset and verification, a Policy Improvement module for policy refinement, a Rollout module for evaluating policies with one or multiple physical robots in parallel, and an Evolution module in which coding agents analyze logs, consult literature, and improve training code to address failure modes.

The authors describe it as a closed-loop system that turns real-world robot learning into a controllable optimization procedure agents can manage, minimizing human effort. Key ingredients are automatic evaluation, scoring each trial without human judgement, and automatic reset, returning the scene to a fresh initial state.

Hardware details: each station has two YAM arms from I2RT in a fixed bimanual configuration, cameras, and a workstation running a FastAPI server, policy inference, and the station's agent. Each workstation uses an NVIDIA RTX 5090. The authors write that frontier coding agents can autonomously develop policies reaching a 99% success rate on dexterous manipulation tasks such as PushT, organizing pins into a pin box, and cutting a zip tie; they also test inserting GPUs into a motherboard. GPT-5.5 within Codex and Opus 4.7 within Claude Code trade off for best performance, while Kimi-2.6 lags. Larger numbers of agents, such as 8, arrive at higher-scoring solutions sooner, and multi-agent setups sometimes yield a higher absolute score than a single agent.

Challenges remain for fleet instrumentation: coding agents do not fully utilize robot resources while reading logs, writing code, debugging, or waiting for the language-model backbone, and as the number of robots scales, MRU decreases while GPU active utilization increases.

Separately, Tencent released details on ARGUS, which it describes as a low-overhead, fine-grained, always-on tracing and real-time analysis system for large-scale training workloads, with Python, framework, and GPU runtime layers. Tencent says it deployed ARGUS on a production cluster of over 10,000 GPUs for more than six months, diagnosing compute stragglers, communication link degradation, pipeline bubble amplification, JIT compilation blocking, and compute stragglers masked by communication symptoms. Mentioned training runs include a 4,096-GPU video language model job, a 512-GPU audio-model job, and a 12,960-GPU MoE job.

UC Berkeley researchers assembled the Local Ordinance Corpus for the United States (LOCUS), described as a comprehensive corpus and county-harmonized access layer for U.S. municipal and county ordinance codes, containing roughly 2.2 million rows with coverage metadata.