We train AIs with gradient descent: a careful, mathematical nudge towards better performance. But what if we stopped teaching and started breeding? This is the world of AI evolution: creating a vast sandbox, spawning millions of slightly different world models, and letting them compete, reproduce, and mutate over thousands of digital generations. Not guided design, but artificial selection for intelligence.Here's the process. You create a massive simulation—a complex universe with challenges. You spawn a population of, say, a million 'agent' world models, each with random variations in their neural architecture. They are thrown into the sim. The ones that perform best—that navigate, solve puzzles, survive—are selected. Their 'code' (their weights and architecture) is combined (mated) and randomly mutated to create the next generation.You are not optimizing for a single loss function. You are creating an evolutionary pressure. Over generations, you might see the emergence of stunningly elegant solutions no human engineer would ever design. Strategies that exploit subtle bugs in the simulation physics, communication protocols between agents, or hyper-efficient internal representations. This is how nature built our brains: not with a blueprint, but with relentless, dumb selection pressure.The outcome could be minds that are highly competent but deeply alien. Their intelligence is shaped by the specific, possibly weird, evolutionary pressures of the sandbox. They might develop a ruthless, perfectly logical cooperation that looks like cruelty to us. Or a staggering creativity that is utterly illogical by our standards.This method could brute-force its way past plateaus in traditional AI. Need a model that can design radically novel proteins? Evolve it in a sandbox where survival depends on catalytic efficiency. The resulting model wouldn't just be good; it would be adapted to the problem in a fundamental, embodied way.My controversial take is this: Evolutionary training is the most likely path to creating a true Artificial General Intelligence, and also the most likely path to creating one we cannot hope to understand or align. Gradient descent gives us a hand on the tiller. Evolution gives us a button that says 'breed for success' and a hope that what emerges is friendly. It is the ultimate gamble: outsourcing the design of god-like intelligence to the same blind, pitiless process that designed us.This has been The World Model Podcast. We don't just design intelligence we sometimes dare to grow it in a wild garden and hope for the best. Subscribe now.