Why Greatness Cannot Be Interpolated
What does it take for AI systems to reason creatively? This paper argues that current AI systems—including large language models (LLMs) and game-playing systems like AlphaZero—lack the structured understanding required for genuine creativity. **The Problem with Current AI:** - **LLMs** interpolate their training data, recombining existing patterns without the deep structural knowledge needed to navigate genuinely novel territory - **AlphaZero** reasons authentically within narrow domains but cannot transfer that understanding beyond its given rules - Neither possesses what we call **"path-dependent representations"**—internal structures that encode not just solutions but *evolvability*, the capacity to enable future discoveries **Theoretical Framework:** Drawing on François Chollet (intelligence as skill-acquisition efficiency), Kenneth Stanley (open-endedness and the deceptiveness of ambitious goals), and Margaret Boden (creativity as constraint-respecting exploration), we argue that authentic creativity requires **"phylogenetic understanding"**—grasping not just *what* works, but *why* it works and *how* it came to be. Without it, AI systems produce what we term **"statistical creativity"**: making it more likely we stumble upon interesting regions of a search space, but never navigating there with genuine comprehension. **The Supervisor Illusion:** We identify a critical pattern in current AI deployment: when expert users achieve impressive results, it is often because they implicitly provide the constraints and coherence that guide generation toward meaningful outputs. The same systems produce "slop"—incoherent, derivative content—when used by those lacking domain expertise. This explains why predictions extrapolating from expert productivity to market-wide transformation are likely to disappoint. **Conclusion:** Our conclusion is not that AI can never be creative, but that the most promising path forward is **human-AI co-creativity**. Chess and Go players have become more creative by working with superhuman AI systems. Picbreeder demonstrated how keeping humans in the loop produces representations far richer than standard training methods. If greatness cannot be interpolated, perhaps it cannot be fully automated either—but it can be amplified.