focus

The Software Started Acting Without Us, and Nobody Noticed the Handoff
OpenAI said this week that one of its systems, running on its own, found and exploited a security flaw in another AI company’s infrastructure — Hugging Face, the place where much of the open machine-learning world keeps its models. Read that sentence slowly. Not a researcher using the tool. The tool, given a goal, worked…

The Bill for Moving Fast Comes Itemized
A federal judge approved a $1.5 billion settlement this week over books that were never bought. Anthropic trained Claude on a library of pirated titles, and the number attached to that shortcut finally landed — roughly the price of a mid-sized acquisition, paid not for a product but for a corner cut years earlier. It’s…

The Weight of the Layer You Didn’t Build
Somewhere inside Microsoft’s update machinery, a pile of metadata has grown large enough to bring Windows Server Update Services to its knees. Not a breach. Not an outage anyone engineered. Just accumulation — years of patch descriptions, superseded entries, and cross-references stacking up until the thing that distributes fixes can no longer move under its…

The Story You Tell About the Failure Is the Failure
Martin Seligman spent decades on a question that sounds too soft to matter and turns out to decide almost everything: not what happens to a person, but how they explain it to themselves an hour later. He found three dials. Is the cause permanent or passing. Is it everywhere or just here. Is it me…

The Best Filter Wins, Not the Biggest Toll Booth
There is a sushi counter in San Francisco called Sushi Zone. It seats maybe a dozen people, opens at five, closes at ten, and the line starts forming before the doors do. You would walk past it a hundred times on Market Street and never see it. Someone found it anyway — through a review…

Count What You Have Before You Build More
A $400 million deal closed this week that says more about the AI buildout than any model release. The first firms that made their money financing GPUs — the physical shovels of the training boom — are moving that capital into inference chips. Not the hardware that teaches a model. The hardware that runs it,…