Google is relocating its roughly 90-person AI responsibility unit, which evaluates Gemini for chemical, biological, radiological and nuclear risks, from DeepMind to report to Kent Walker's global-affairs organization starting September 1. Several team members who sought transfers to stay inside DeepMind had those requests denied.
why it mattersPlacing the team that flags dangerous-capability findings inside the division that manages Google's relationships with regulators raises an obvious independence question.
Quartz
Z.ai launched GLM-5.3-Flash, a natively multimodal 320-billion-parameter mixture-of-experts model with 18 billion active parameters and a 1-million-token context window, publishing the weights on Hugging Face under the MIT license at roughly a tenth of its predecessor's price.
why it mattersDay-one open weights scoring close to Claude Opus 4.8 on coding benchmarks keep pushing the price of capable models toward zero.
Z.ai
OpenAI and Broadcom showed the first benchmarks for Jalapeno, OpenAI's in-house inference chip, claiming 1.5x-1.9x more AI work per watt and up to 3.6x lower latency than Nvidia's Blackwell systems while drawing 700W versus 1,200-1,400W, with results partly verified on-site by SemiAnalysis on its public InferenceX benchmark.
why it mattersThe first credible outside evidence that a hyperscaler's custom silicon can beat Nvidia on inference efficiency, not just cost.
CNBC