An AI handed in 700+ math papers. Solved?

The full narration of the video.

You might think if a computer says it solved a hard math problem, then it's solved. It's not that simple. On October sixth OpenAI put a huge stack of math papers online. 722 papers, all written by an AI that hasn't been released.

They gave this AI about four thousand problems all ones mathematicians had never solved. On average, each result took it about three hours of thinking. Some of these questions are more than a hundred years old. Take this one. A needle has to turn all the way around on a patch of ground. How small can that patch be?

Someone asked this in nineteen seventeen. On a flat surface, the answer is a surprise the patch can have almost no area at all. But in three dimensions, a new question came up however small the space gets, can it ever be as flat as a sheet of paper? Human mathematicians only proved it last year it can't. One paper in this stack says it has proved the same thing with one more dimension.

So, is it solved? Not so fast. Think of an exam. Handing it in is quick, grading is the hard part. Now more than seven hundred answer sheets have landed on the desk at once. Luckily, there's a helper. A computer program built to check math proofs. Every step must be spelled out, skip one, and it won't pass.

It's called Lean. Of the 700 or so papers, 300 have their main result through its check. But there's something it can't catch it checks every step, not whether the question was copied right. If the question is a little off, every step can be right and still answer a different question. Whether it's new, whether it matters, people still have to judge.

The next day, October seventh, OpenAI posted a correction. In one paper, a plus or minus sign was flipped. That one mistake pulled three papers back. Fourteen others had their proofs repaired. Mathematicians say it will take months to tell which of these are truly new.

Some say this is good for math. Others say show us the receipts. How the AI did it, and the exact questions it got, all out in the open. So when a computer says it solved it, does that count? Right now the answer is the paper is handed in, and grading has just begun. Math used to be hard because proofs were hard to write. Soon the hard part may be finding enough people to read them.

Which hard problem would you want an AI to take on? Tell us in the comments.