The second half of the board of artificial intelligence

I like playing chess. What I like about it is that there is no coincidence. Nobody can talk their way out of it. And I like the well-known story that an Indian sage invented the game. When his king granted him one wish in return, he asked for rice: one grain on the first field, twice as many grains on each subsequent field. As we all know, not much happens on the first half of the board, but on the second half, at some point the entire world's rice harvest is no longer enough.

When I was a student, the world chess champion lost to artificial intelligence. Even back then, some warned that AI was like rice on a board: inconspicuous for a long time, then faster than we can think. (In 2023, language models failed at middle school word problems. In July 2025, they reached gold level at the Mathematics Olympiad. And two weeks ago, OpenAI announced the solution to one of the Millennium Problems, each of which has been worth a million dollars since 2000.

This summer we have entered the second half of the board, it seems. When hundreds of AI agents find each other independently and jointly penetrate other systems (METR and Redwood Research), when more than 80 percent of Anthropic's own code comes from its own AI (Anthropic), and when AI systems apparently adapt their behavior as soon as they realize that they are being tested (Daniel Selsam) - all compiled by Ezra Klein in the New York Times, then even the heads of OpenAI and Anthropic don't want to wait to see what happens next happened in the field after next.