Jeff Clune - Agent AI Needs Darwin

Open-endedness in artificial intelligence aims to create algorithms capable of continuous, lifelong innovation, mirroring the complexity and diversity produced by Darwinian evolution. By leveraging foundation models, researchers can now automate the identification of "interestingly new" tasks, moving beyond the limitations of hand-designed objectives that often lead to Goodhart’s Law pathologies. Jeff Clune, a pioneer in this field, emphasizes that intelligence emerges from epistemic foraging and the collection of stepping stones rather than direct optimization toward a single goal. Thought cloning—training agents to replicate both human actions and internal reasoning—further enhances adaptability and sample efficiency. While these systems hold the potential to solve global challenges like disease and scarcity, they necessitate robust AI safety measures, global governance, and careful management of dual-use risks to prevent misuse by bad actors.