#323 David Ha: Why Model Merging Could Be the Next AI Breakthrough

Evolutionary computation and collective intelligence offer powerful frameworks for advancing artificial intelligence beyond traditional gradient-based optimization. By treating frontier models as agents within a collaborative ecosystem, researchers can merge diverse capabilities and generate novel algorithms through iterative, tree-based search processes. David Ha, co-founder of Sakana AI, highlights how evolutionary strategies enable systems to discover new objectives and architectures, effectively functioning as an "outer loop" for deep learning. This approach facilitates the development of AI scientists capable of conceiving, testing, and refining scientific ideas autonomously. Furthermore, the integration of world models—whether through high-fidelity simulations or latent representations—remains critical for enabling agents to reason and plan effectively. Scaling these collective intelligence systems allows for the exploration of vast, non-obvious solution spaces, potentially pushing the boundaries of human knowledge and fostering genuine creative discovery in computational science.