Generalising from Self-Produced Data: Model Training Beyond Human Constraints How do we train AI systems to go beyond human benchmarks—and generate knowledge autonomously? In this paper we propose a framework in which AI agents self-improve by manipulating real-world variables (like disk space or wallet value) without human-labeled data. Currently LLMs are bottlenecked by two things: reliance on human-generated data and benchmarks and the necessity to operating within a single abstraction layer. We argue that this limits genuine innovation—and propose an escape route. We demonstrate this with an agent that autonomously explores, strategizes, and writes code to annex disk space. Any barrier to expansion becomes a new problem to solve, and thus a new skill to learn. Successes are saved, failures pruned. Over time the models generalise to independent world models based on their own empirical data. We suggest that the path to ASI does not lie in building ever bigger models and hoping that one begins to learn independently, but rather to create a system capable of learning and then giving it an incentive to grow. This is the foundation of Superior Agents. Read the full paper: OS implementation: