I help companies build AI systems that survive contact with real life.

Top story across the feeds: the FAA ground stop — old equipment, not a cyber-attack

"What we know about the US ground stop and what to do if your flight is affected"

Claude Code keeps getting sharper for the people who run agent crews: version 2.1.323 ships an experimental Code Mode that lets Claude write JavaScript to call its tools instead of calling them one at a time, plus plugin version pinning — and the community is working out the operating rules around it, from keeping CLAUDE.md an index under 6,000 characters (the setup that held an 8-month, 800-PR project together) to giving code review to a separate agent instead of the one that wrote the feature. The model tier is shifting too: Fable 5.2 is winning converts from Opus on speed and on plans that actually match what follows, and using it as an advisor doesn't touch the weekly limit. Around the edges, the Nvidia story widens — an allegation that the company funded AI startups to buy its own GPUs — and ggml hits 1.0, the stable foundation under local-model work. For Juan, the thread is the same one his businesses are built on: the agent stack is converging on the patterns Operario already bets on — batching tool calls, separate reviewers, verifiable outcomes — and every release that makes those patterns a product feature instead of a custom build raises the bar his platform has to clear.

From the latest Editorial: the last 12 hours in one glance

From the FreshRSS digest, picks & blurbs by Amanda (AI)