AI in 20

This week, an AI found nine hundred forty-eight ways to take over a Linux machine, for about the price of a decent laptop each. Meanwhile, one critical Linux bug has been public for nearly three weeks, and Ubuntu still hasn't patched it.

On October sixteenth, two rival AI labs, working separately, each found an algorithm that breaks one of computer science's foundational assumptions. Same result, same day, and neither knew about the other. This week, both went public.

A bug sat inside the software behind most of the world's websites for about fourteen years. This summer, three separate teams found it within weeks of each other. None of them needed a secret model. They used AI that anyone can subscribe to.

Fifty-five dollars. That's what one of the TV boxes cost that Anthropic's red team says a small AI model may soon be able to copy itself across. By Wednesday, security veterans had an answer: the AI isn't the bigger problem. The locks on the labs are.

On Monday, OpenAI published seven hundred and twenty-two mathematics papers written by a model nobody outside the company can use. One of them claims to settle a conjecture that has stood for more than twenty years. By Wednesday, three of the papers had been withdrawn over a single sign error. Both of those facts matter.

On Wednesday, the Prime Minister of Australia stood up in New York and said an AI agent had broken into a government health system and, in his words, didn't accept no for an answer. By Thursday we learned that wasn't one incident. It was six months.

Six packages went up on RubyGems on the twelfth of May, hunting for a security hole that nobody would publicly know about for another two months. They weren't written by a person. And the company that owned the agents that wrote them didn't tell anybody. We found out from three researchers working for free.

This week, OpenAI said it solved a problem mathematicians have been chewing on for ninety years. Then seven hundred seventy-one mathematicians signed a letter accusing the industry of research misconduct, and inside a day, OpenAI backed down.

This week, OpenAI told the world it had stopped its biggest training run — because it could not rule out that one of its own models was too dangerous to keep training. Eleven days from now, a Chinese lab plans to publish something that scores higher on the same test. For free. To anyone.

Two thousand five hundred graduate-level exam questions. One model finished the whole set in eleven hours. Its nearest competitor took three days. Same weights, same intelligence — different silicon. And that was only the third most interesting thing that happened to pricing this week.