On July 16, the AI story is shifting from raw model size to efficiency and infrastructure: Google tuned Gemma 4 for the Pixel 10's TPU, Prism ML claims a 27B-parameter model that runs on a phone, and Elon Musk bought a $1B gas turbine company just to keep Grok's compute fed. At the same time, agent tooling is maturing fast - one engineer cut token use 94% by compiling agent skills, Perplexity published a new benchmark for research agents, and Slack shipped memory, MCP, and voice for its bot - while security researchers are simultaneously documenting how those same integrations (context bombs, a Claude-Slack exploit) can be hijacked into unauthorized actions. Layer on OpenAI's hardware ambitions colliding with Jony Ive's Apple pedigree, enterprise vendors rebuilding around AI-native workflows, and a widely-read essay asking what engineering work survives automation, and the throughline is clear: the frontier has moved from 'bigger models' to 'who controls the agent, the compute, and the workflow around it.'

• Gemma 4 tuned specifically for Pixel 10's on-device TPU.
• Bonsai 27B claims to be the first 27B-class model running on a phone.
• Musk buys a $1B gas turbine company to power Grok's compute.
• Compiling agent 'skills' cut one engineer's token use by 94%.
• Slack adds memory/MCP/voice to its bot as researchers flag exploit risk.