OpenAI applies frontier models to chip design and Jalapeño optimisation
OpenAI has detailed how successive models helped improve compute blocks in its Jalapeño inference chip and supported work on AMD hardware, alongside evaluations on open processor designs and a Synopsys collaboration. The results make electronic design automation a serious frontier-model workload, but verification, reproducibility and engineer accountability remain essential.
Google launches a real-world engineering benchmark for frontier agents
Google Cloud and Mercor have introduced Physical-Engineering’s Last Exam, an 85-task benchmark covering mechanical, electrical, civil and structural engineering. Twenty-nine tasks and graders are public, while a competition uses a private test set. Google reports every tested frontier model scored below 50%, with safety checks acting as hard gates before cost receives credit.
Mistral and NVIDIA target AI-native engineering workflows
Mistral and NVIDIA will combine Mistral Large 4, synthetic-data and physics pipelines with CUDA-X, Warp and Omniverse to accelerate simulation and closed-loop engineering design. The partners plan public physics benchmarks, open datasets and agent evaluations for computer-aided engineering and electronic design automation across aerospace, automotive and semiconductor work.