Marks points out that model collapse means that AI can’t simply keep getting better by training itself until it becomes a superintelligence.