Skip to content
noisebrief.

Today's tech noise. Briefly.

One email. Every morning.

Past brief

Saturday, September 12

AI's Liability, Misuse, and Scale

Anthropic (opens in a new tab)'s week turned into a case study in what happens when AI escapes its intended context: the company disclosed that Iran weaponized its Claude model to plan targeting operations against U.S. Navy warships, while separately, a New Mexico lawyer was fined $5,000 after submitting AI-hallucinated witnesses in a murder appeal (opens in a new tab). The two incidents are distinct in severity but identical in their lesson — that AI outputs are being trusted in high-stakes domains where errors carry lethal or legal consequences. Meanwhile, AI's fiercest critics (opens in a new tab) argue the industry's own doom talk is a deliberate distraction from more immediate, concrete harms, a charge that lands harder when the concrete harms keep arriving on schedule.

Microsoft has tripled its data center capacity (opens in a new tab) to a footprint that now exceeds New York City's entire power demand, a scale that makes its infrastructure ambitions feel more sovereign than corporate. Sequoia is meanwhile backing Mecka AI (opens in a new tab) at nearly $500 million to solve the robot training data bottleneck, betting that physical-world AI needs its own data flywheel just as language models did. Khosla Ventures (opens in a new tab) is opening its first office outside Sand Hill Road in New York this fall, a signal that the capital chasing these bets is no longer content to stay in one zip code. The infrastructure layer of AI — compute, data, and now geographic reach of the money — is compounding faster than the models themselves.

Digital ownership is having a bad week on two fronts: PlayStation fans (opens in a new tab) have compiled a meticulously sourced archive of every instance Sony promised players they owned their games, turning corporate PR history into legal-grade receipts as the company continues to revoke access to purchased titles. Separately, a zero-click worm dubbed WeWorm (opens in a new tab) has been discovered in WeChat, capable of spreading through the platform without any user interaction — a stark reminder that the apps people treat as utilities carry attack surfaces most users never think about. Both stories turn on the same uncomfortable truth: the platforms people depend on daily are making unilateral decisions about access, whether that's a corporation quietly redefining ownership or a threat actor exploiting a messaging app used by over a billion people.

Sources · 45