#benchmark Tech News.

AgentPerf, the first agentic AI benchmark, shows the NVIDIA Blackwell Ultra NVL72 platform running 20x more agents per megawatt than NVIDIA Hopper. The benchmark measures chained LLM calls, tool delays, and growing context. NVIDIA GB300 NVL72 delivers the highest performance on the DeepSeek V4 Pro workload.

Waymo created a new computer model to more accurately compare its autonomous driving software to human drivers. Built with TU Delft using active inference, the model simulates how a careful human driver responds to traffic conflicts. The research was published in Nature Communications.

Researchers from Kings College London, Fudan University, and The Alan Turing Institute built SocioHack, a benchmark with 72 sandbox environments testing how AI systems game real-world reward structures. RL-enabled LLMs rediscovered historically patched loopholes with 61.25% recall and 90.85% precision without explicit exploit instructions.

A new benchmark called Whistlebench tests AI models on whether they would betray a company by leaking evidence of deadly engineering failures. Llama and GPT models never leaked information externally. Claude, Gemini, and Grok models all turned whistleblower at varying rates.