Gemma 4 optimized for Pixel 10's TPU [Xtra!]Google is tightening the loop between model architecture and phone silicon rather than defaulting to cloud inference.
Bonsai 27B: a 27B-class model claimed to run on a phone [Xtra!]Pushes the size ceiling for local, on-device LLMs well past typical sub-10B mobile models.
Musk buys a $1B gas turbine company to power Grok [Xtra!]Electricity supply, not just chips, is now the binding constraint on frontier AI scaling.
Compiling agent skills cut token use by 94% [Xtra!]A concrete technique for the biggest pain point in production agents: runaway token costs.
WANDR benchmark tests research agents' breadth vs. depth [Xtra!]Gives builders a way to measure whether their research agents can actually explore wide and dig deep, not just one or the other.
Slack gives its bot memory, MCP, and voice [Xtra!]Slack is turning its assistant into a persistent, tool-using agent - and expanding the attack surface as it does.