Jul 21, 2026

Today in brief.

Washington moves to ban Chinese open models

The US response to Chinese open weights winning isn't a better model, it's a proposed ban on downloading theirs.

A security engineer got refused by two frontier models on his own codebase, ran the open Chinese one, and it patched fifteen real bugs.

A sixty-seven-year-old open problem in algebraic geometry got a counterexample, and a model produced it.

The largest copyright settlement in AI now has a judge's signature, and it sets the price of training data.

The compute buildout has reached the stage where it uses eminent domain, and voters have started noticing at the ballot box.

Local inference stopped being the compromise option this week - one consumer GPU is now doing 543 tokens a second at 65K context.

Two good pieces landed on the same question: what does it actually cost to run many agents, and what should they read before they act.

A team tried to measure AI writing across arXiv and published the part most people skip - where their own measurement stops working.

Investors dumping AI stocks, tech workers losing financial footing, and a new $1M job title - the money story got three data points in one window.

Two AI leadership exits in one day, at exactly the moment Washington is deciding what to do about Chinese open models.

The efficiency race moved to silicon - Google's new chip is aimed at inference cost, not training records.