Hassabis steps aside, Jeff Dean leaves, and Google’s AI roadmap gets shakier
Alphabet reorganised Google DeepMind on 5 August. Demis Hassabis stepped away from running DeepMind day-to-day to become Alphabet’s chief scientist and DeepMind’s chairman, with chief technology officer Koray Kavukcuoglu taking operational control. The same week, Jeff Dean — Google’s 30th employee, there since 1999 — left the company (his last day was 6 August) to co-found Discovery Loop, a public-benefit startup that wants to use AI to automate scientific and engineering research. He took three senior researchers with him: Sanjay Ghemawat, Google Brain co-founder Quoc Le, and former Gemini technical lead Oriol Vinyals. Alphabet is a founding investor and will supply the compute for at least the first year; Radical Ventures and Khosla Ventures are co-leading the seed round. The backdrop, as the coverage frames it, is a company seen to be trailing OpenAI and Anthropic and racing to ship its next Gemini flagship.
For a company standing up AI, the useful part is not the drama — it is roadmap risk. The people who set Gemini’s direction just changed, and the flagship is late. A workflow hard-wired to one model inherits that vendor’s turbulence; a workflow that keeps models swappable — same API, a different engine behind it — can wait out a slipped release without re-plumbing anything.
Xiaomi open-sources a robot-control model you can download and run
Xiaomi open-sourced Xiaomi-Robotics-1 (XR-1) on 5 August: a vision-language-action model that takes what a robot sees plus a written or spoken instruction and outputs the next physical action. Xiaomi says it was pre-trained on more than 100,000 hours of real-world manipulation data and post-trained on over 10,000 hours across different robot bodies, pairing a Qwen3-VL backbone with a diffusion transformer. The company reports it tops the RoboCasa365 and RoboDojo benchmarks, runs on consumer GPUs, and ships with the full pipeline — post-training, deployment and benchmarking — on Hugging Face.
Robots are not the near-term use for most regulated European buyers. The pattern is what matters: another frontier-scale set of open weights out of a Chinese lab, now reaching into physical AI, that you can download, inspect and run on your own hardware so nothing leaves your network.
xAI ships Grok 4.6 — with no benchmarks to check it against
xAI released Grok 4.6 on 7 August. Elon Musk said on X that it reuses Grok 4.5’s 1.5-trillion-parameter foundation with improved supervised fine-tuning and reinforcement learning, and that a larger 2.1-trillion-parameter Grok 4.7 will follow “a few weeks later.” As of launch, xAI had published no model card and no benchmarks; independent Arena scores are expected within the week.
The recurring discipline here: an announcement is not a benchmark. A point release with no third-party numbers is a claim, not a measurement — and the decision to move production traffic onto a new model should wait for someone other than the vendor to score it.
Quick Hits
- Anthropic starts an in-house chip team — Anthropic confirmed on 5 August that it is hiring engineers to co-design silicon alongside Claude, aiming to cut per-token inference costs by roughly half. It frames the effort as cost optimisation rather than a break from Nvidia, and says it will keep its existing chip supply — Nvidia and AMD GPUs, AWS Trainium and Google TPUs.
- White House finalises its frontier-model cyber framework — The administration met its 1 August deadline to finalise a voluntary framework for up-to-30-day pre-release government testing of frontier models for cyber risk, then briefed the major labs — OpenAI, Anthropic, Google, Microsoft, Meta and Nvidia — on 4 August. Its contents remain undisclosed, and open-weight models are reportedly outside its scope.