Member-only story
Scientists Put AI in a Virtual World. They Didn’t Expect It to Learn This.
Researchers keep dropping AI into simulations to do one simple thing: win a game of hide-and-seek, finish a boat race, walk across a room. Then they look at the replay and find the machine has invented tools, exploited physics glitches, cheated the rules, and even thrown its own party. Six real, documented cases of AI learning things nobody programmed and what they reveal about how these systems actually think.
2019, a team at OpenAI built a tiny virtual playground and dropped two teams of AI agents into it to play the world’s simplest game: hide-and-seek. The hiders would try to stay out of sight. The seekers would try to find them. Scattered around the arena were a few movable objects: some boxes, some ramps. The researchers pressed go and let the agents play the game against each other, over and over, hundreds of millions of times.
Nobody told the agents what the boxes were for. Nobody told them to build anything, or to…








