AI digest: agents go mainstream
Google pushes AI agents to Android, medical models hit 103B parameters, and optimisers get smarter.
Three stories about AI systems getting more capable and autonomous this week.
Google brings proper AI agents to Android
Google announced Gemini Intelligence for Android, which can book trips, fill forms, and clean up your text messages automatically. The standout feature is “Create My Widget” where you describe what you want in plain English and it builds custom dashboards. This feels like the first proper consumer deployment of agentic AI that normal people will actually use.
Medical AI hits 103B parameters with clever routing
MedAIBase released AntAngelMed, a medical language model with 103 billion parameters that only activates 6.1B at runtime using mixture-of-experts architecture. It’s ranking first on medical benchmarks while running at 200+ tokens per second on H20 hardware. The efficiency gains here are mental - getting 40B model performance whilst using a fraction of the compute.
New optimiser fixes dead neuron problem
Tilde Research launched Aurora, an optimiser that solves a hidden flaw in the popular Muon optimiser where neurons die during training and never recover. Aurora uses leverage-aware scaling to keep neurons alive throughout training. This matters because optimisers are the engine room of model training, and fixing structural flaws like this could unlock better performance across the board.