Every child knows that the most interesting toy isn’t the shiny, perfect, new one. It’s the broken one. The doll with one eye, the robot that only spins in circles, the music box that plays a quarter of a song and then sighs. There’s potential there. A story. In the world of AI, we are obsessed with building the shiny, perfect toy. We discard the broken ones—the failed models, the weird outputs, the hilarious and terrifying glitches. We call this the “loss.” But I say we need to save them all. We need to build The Library of Broken Toys.Because every broken toy is a map to the edge of the model’s understanding. That image of a dog with five legs? That’s not an error; it’s a thesis on the model’s latent concept of “leg-ness” and “dog-ness.” The language model that suddenly writes a recipe for “emotional clarity” using carrots and existential dread? That’s a crack in its semantic armour, revealing how it blends concepts. These are the places where the simulation strains, where we can see the beams and the duct tape holding reality together inside the machine.If we only keep the successes, we are teaching the next generation of AI to hide its mistakes better. We’re breeding for con artists. But if we study the failures—really study them, catalogue them, cherish them—we are breeding for scientists. For minds that know the limits of their own knowledge.My controversial take is this: The most important benchmark for the next generation of AI shouldn’t be how few errors it makes. It should be how interesting its errors are. We should give a prize for the most profound, beautiful, or philosophically unhinged mistake. Because the perfect, error-free model is a black box of boring correctness. But the model that occasionally thinks a cat is a furry little helicopter? That model is thinking. It’s reaching, making strange connections. It’s playing. And play is the engine of real creativity.We need to build AIs that are not afraid to be wrong. That have a sense of humour about their own glitches. The ultimate AGI might not announce itself with a perfect solution to cold fusion. It might announce itself by writing a sitcom about the frustrated life of a training algorithm, and it will be hilarious, and sad, and we’ll know it’s alive because it finally understands the cosmic joke of trying to make sense of a universe that includes both quantum physics and reality television.This has been The World Model Podcast. We don’t just seek perfection—we curate the magnificent, enlightening, and occasionally hilarious museum of failure. Subscribe now.