The week in AI

Google changed the top of its AI organisation on 5 August. Meta released Muse Code, its terminal coding agent, with a new model called Muse Spark 1.2, also on 5 August. On 6 August Google DeepMind published a Nature paper on its WeatherNext cyclone model and released the code and weights.

Anthropic hired Tino Cuellar, president of the Carnegie Endowment for International Peace, on 4 August to run its policy and government relations.

Demis Hassabis, CEO of Google DeepMind, handed operations to Koray Kavukcuoglu, who now heads Google's AI division, reports to Sundar Pichai and leads Gemini 4. Hassabis becomes Alphabet's chief scientist and chair of Google DeepMind. Google said the change lets him focus on the big picture.

The announcement came alongside the departure of Jeff Dean, Google's chief scientist, after 27 years. He left on friendly terms to found Discovery Loop with Sanjay Ghemawat. Discovery Loop is a public-benefit company working on AI for science and engineering, and Google will invest in it.

So Google lost its chief scientist and moved Hassabis into that title on the same day. Gemini work now runs through Kavukcuoglu rather than through the DeepMind CEO. Google hasn't said how much it is investing in Discovery Loop.

Muse Code puts Meta in the same category as Anthropic's Claude Code and OpenAI's Codex. The agent runs persistent sub-agents in the background and writes every action to an event log. If a session crashes, the log lets it replay the work and pick up where it stopped.

Muse Spark 1.2 was co-trained with the harness, so the model learned from trajectories recorded inside the same agent it ships in. Meta used the previous model, Muse Spark 1.1, to generate the training environments and grade the solutions. Access is through a closed API.

Meta cites results on Terminal-Bench 2.1, DeepSWE 1.1 and its own Meta Internal Coding Bench. The launch post shows these only as charts, so there are no exact company-reported scores to quote, and no independent evaluation had been published by 9 August.

Earlier systems used one model for global weather and a separate finer-scale model for cyclones. WeatherNext does both in one model and runs a 1,000-member ensemble out to 15 days. An ensemble is many forecasts from slightly different starting conditions, and their spread shows forecasters how uncertain a storm's track and strength are.

Google built the model with the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere (CIRA) and the UK Met Office. The extra day of lead time is Google's figure from the paper. Google's August roundup calls the open release WeatherNext 2.

• SeedRealtime from ByteDance Seed, released 5 August, is a full-duplex model that takes audio, video and text in one stream and can speak without being prompted, and ByteDance rolled it out at scale at launch.
• Wan-Animate-2 from Alibaba's Qwen team, released 7 August, is an open character-animation model that takes the driving video directly and lets text control the camera viewpoint, with a Lite version for real-time streaming.
• Shieldstral from Mistral AI, released 4 August, is a 3B open-weights multimodal safety classifier that reads plain-language policies at inference time, so changing a policy needs no retraining.

• Tino Cuellar, president of the Carnegie Endowment and a former California Supreme Court justice, joined Anthropic on 4 August as its first Chief Global Affairs Officer, running policy, international engagement and government relations as Anthropic's disputes with the US government escalated.
• Jeff Dean, Google's chief scientist, left Google on 5 August after 27 years to found Discovery Loop with Sanjay Ghemawat.
• Demis Hassabis stepped back as CEO of Google DeepMind on 5 August to become Alphabet's chief scientist and Google DeepMind's chair.