The Sequence Radar - Issue 910: Last Week in AI: Google Rewires Its Brain and Meta Hires a Coding Swarm

This was a week about rearranging the people who steer AI, not just the models themselves. Google let Jeff Dean leave after 27 years to launch Discovery Loop, a public-benefit company aimed at automating machine learning, science, and engineering. Google stays in as a founding investor and cloud partner, so this looks less like a clean break than a new structure bolted onto the old one.

Demis Hassabis is making a similarly sharp shift. He is handing Google DeepMind’s day-to-day work to Koray Kavukcuoglu, then moving into the chair role at DeepMind and chief scientist of Alphabet. Hassabis says he will focus more on AGI, scientific discovery, global strategy, and Isomorphic Labs. That is a very different job from running the shop.

The pattern is hard to miss: Google seems to be separating the product machine from the longer-horizon research machine. One clock is Gemini features, developer adoption, and model releases. The other is AGI safety, scientific breakthroughs, and questions that do not fit into a quarterly roadmap. Running both from the same seat may simply have become too much.

Meta, meanwhile, showed what the next contest in coding agents may look like. Muse Code is a terminal-based agent built on Muse Spark 1.2 that can plan changes, write code, validate results, and work across large repositories. For bigger tasks, it can split work among multiple sub-agents in isolated worktrees, and it keeps a record of its actions so it can recover after a crash. That makes it more than autocomplete. It is trying to behave like a small, durable team.

That shift matters because Meta is walking into a field already defined by Claude Code and Codex. But the real competition is starting to move up a level: not which tool suggests the best line of code, but which system can coordinate many agents over long jobs without falling apart. The castle is turning into a network, and everyone is trying to own the gates.