Daily AI Roundups
H Company releases 27B Holo4 agent models for GUI, code and API work. NVIDIA announces a layered agent-safety design that pairs broadly available OpenShell with a BlueField-4 watchdog reference system. In one Nscale test, NVIDIA reports 49.2% more aggregate inference throughput under the same provisioned power budget. Cloudflare's founders question how agent traffic reshapes the web, while Ben Thompson argues Meta and Microsoft could control agent distribution. Reporting on China's AI-talent travel curbs and data-center borrowing costs rounds out the edition, alongside Simon Willison's annotated keynote, a Washington Post satellite analysis and Xiaomi's postmortem on repeated agent tool calls.
- agent models
- agent security
- inference economics
- AI infrastructure
- AI analysis