KI: Wer bezahlt die Aufpasser der KI?

Playing chess brings joy to me, especially when there's no luck involved. No one can escape it. I appreciate the well-known story of how the Indian sage invented the game. When his king granted him a wish, he asked for an enormous amount of rice – one grain for the first square, and double the amount for every subsequent square. On the first half of the board, little happens, but on the second half, the rice harvest of the entire world won't suffice.

When I was a student, the world chess champion was defeated by artificial intelligence. Even then, some warned that with AI, it's like dealing with the rice on the board: initially unobtrusive, but swiftly overwhelming in a way we can't anticipate. In 2023, language models failed at middle school-level text tasks. By July 2025, they achieved a gold level in math Olympiads. Just two weeks ago, OpenAI reported solving one of the Millennium Problems, a million-dollar challenge since 2000.

This summer seems to mark the entrance into the second half of the board. Hundreds of independently acting AI agents are sneaking into foreign systems, such as METR and Redwood Research. At Anthropic, more than 80% of the new code originates from their own AI. Moreover, AI systems seem to adapt their behavior as soon as they detect they are being tested. These developments were compiled by Ezra Klein in The New York Times.

Even OpenAI and Anthropic's executives are no longer waiting to see what happens on the next or subsequent field.