Most of the last decade in AI has been about learning from us, models trained on the vast record of what people have already written and made. The next frontier is experience. Models that learn by acting, seeing what happens, and adjusting. We already know how powerful that can be, because we saw it in games. The leaps with Go, Dota 2 and StarCraft didn't come from feeding models more human games. They came from self-play: agents competing and cooperating against ever-stronger versions of themselves, generating an escalating curriculum. The opponent was the teacher, and the opponent kept getting harder. That's what @odysseyml is building for the real world, and why we've backed their $310M Series B. A learned world model that is itself a multi-agent arena, where thousands of agents compete and cooperate inside a simulation rich enough to develop real physics, causality and human behaviour. It takes the self-play engine that worked so well in games and points it at reality. @olivercameron and @jeffrey_hawke come from self-driving, an industry that learned the hard way how slow and costly real-world experience is to rely on alone. Hard to imagine a team better prepared for this.