Where AI meets the built environment

Shazeer — a co-author of the 2017 "Attention Is All You Need" transformer paper and a co-leader of Google's Gemini — announced he is joining OpenAI to head AI-architecture research; ex–Trump AI-policy official Dean Ball is joining too, ahead of OpenAI's IPO. Google had paid a reported $2.7B to bring him back from Character.AI less than two years ago. The models your tools ride on are steered by a small, mobile group — vendor and model continuity is a planning variable, not a given. TechCrunch → · TNW →

With Carnegie Mellon and UC Berkeley, Nvidia's embodied-AI lab unveiled ENPIRE — a closed training loop that hands robot instruction to teams of AI coding agents (OpenAI, Anthropic, Moonshot), no human supervisor required. The robots learned to insert GPUs and cut zip ties at up to 99% success. Paired with this week's Built Robotics–Penn xLAB construction partnership, the gap between "AI in software" and "AI on the jobsite" is closing faster than the org chart expects. HotHardware →

Z.ai's open-weight GLM-5.2 (753B parameters, MIT license, 1M-token context) drew detailed independent benchmarks this week: within 1% of Claude Opus 4.8 on the FrontierSWE coding benchmark, trained entirely on Huawei silicon, and roughly 82% cheaper per token. (Released June 16; Z.ai has been on the US Entity List since 2025.) An MIT-licensed model you can self-host changes the calculus for any firm uneasy about sending drawings, models, or contracts to a cloud API. Decrypt →

Launched at AIA26 with $2.15M in pre-seed funding, Kestrel runs a jurisdiction-specific code-compliance check inside Autodesk Revit in about 30 seconds, flagging violations tied to specific model elements before drawings go to plan review; a browser dashboard serves teams without BIM files. Front-loading code review into the model is the permit-bottleneck play — a direct answer for architects and engineers tired of catching violations at submittal. Engineering.com →