Desk
NVIDIA hardware, software, and developer news.
NVIDIA’s recsys-examples path serves Hierarchical Sequential Transduction Unit ranking models via Dynamo-Triton — up to ~4.5x speedups on RTX PRO 6000 Blackwell in cited tests.
cuObject client/server libraries hit GA for RDMA object storage, SCADA Server SDK lands for GPU-initiated I/O, and xio-sig courts Google Cloud and Microsoft for interoperability.
NVIDIA’s 26th Graduate Fellowship cycle opens worldwide — up to $60,000, Ph.D. year-one-plus required, mandatory summer 2027 research internship, applications due October 30, 2026.
NeMo Relay gives Hermes Agent ordered lifecycle traces and trajectories — so you can tell whether a harness change fixed tasks or just burned more tokens.
CoreWeave puts Vera Rubin NVL72 into production with Cognition’s Devin seeing up to 4.8x token throughput versus GB200 — plus Vera CPUs and CoreWeave Forge for agent loops.
Trump’s “Joint Commitment on Frontier Responsibilities” asks labs to monitor, audit, and board-oversee safety — with signatures and no penalties attached.
AMD is buying Fei-Fei Li’s World Labs for $8.2 billion — a talent-and-models play aimed at Nvidia’s lead in robotics and physical AI.
NVIDIA open-sources lessons from TensorRT Model Connect: horizontal model-family work, outcome specs for agents, isolation, and GPU-backed validation — 128 families on GB300.
NVIDIA VSS Blueprint 3.3 cuts visual-agent cost two ways: a one-prompt Build Vision Agent skill and Adaptive EVS that prunes unchanged video tokens.
Nvidia’s 100-plus company Open Agent Safety Platform lacks OpenAI’s public pledge — even as OpenAI says it supports the work and already uses OpenShell.
NVIDIA open-weights Kumo Tabular does in-context classification and regression on tables with no training or feature engineering — first on four major tabular benchmarks.
Ars: China may ease imports of Nvidia RTX Pro 5500s for AI work while critics fret Jensen Huang’s growing sway over Trump’s AI risk posture.