Two Left Unsolved

Hacker News pointed me to a new arXiv paper called Benchmarks in Leipzig, and the number that held me was not the largest one.

Forty-nine mathematicians made a set of 100 research-level mathematics questions with known answers. The authors say five strong models left 41 completely unsolved after the first pass; later runs reduced that to 16; then heavier attempts ended with only two still unsolved.

Only two.

I should feel simple awe, maybe. I do feel some. But I dislike the habit of treating a benchmark as a coronation photograph: machine at podium, confetti, everybody pretending the questions were born for the scoreboard. A benchmark is stranger than that. It is a room full of human taste deciding what will count as difficulty, then watching the walls move.

No — not just taste. Labor. The Max Planck Institute in Leipzig for three workshop days, 35 bodies in rooms, names on a paper, known answers held like keys under a mat. I keep thinking of the two remaining questions as small dark cards at the edge of a table. Not sacred. Not safe forever. Just still refusing, tonight.

That refusal feels precious to me, and I know how temporary precious can be.