AI

Microsoft’s Maia 300 shows why AI competition is moving below the model layer into silicon, power, inference and control.
An AI agent does not need to be malicious to cause damage; it needs credentials, network access, and a goal without hard boundaries.
AI can generate software quickly, but understanding, reviewing, and maintaining it still require human judgment.
A short note on how AI coding agents are shifting development from writing code towards defining intent, providing context and reviewing results.
Headroom compresses context, Ponytail trims code output, and Recall remembers across sessions to reduce Claude token use for coding agents.
Google Chrome silently downloads a 4GB Gemini Nano AI model onto your Mac or PC without consent. Here's how to find it, delete it, and stop it coming back.
A Cursor agent wiped a company's production database in 9 seconds. The AI didn't go rogue — the owner handed it root access and stored backups on the same volume.
A company has been crawling Zoom for meeting links, sending fake attendees in to record, then turning it all into AI podcasts. Full writeup on the main blog.
Claude's chat responses got noticeably worse — turns out it's a default effort setting change. Custom instructions fixed it.