A Systems View of Recursive Self Improvement (github.com)

A new initiative, Exo, aims to advance the concept of recursive self-improvement (RSI) in artificial intelligence by enabling models to autonomously modify and enhance themselves. Traditionally, RSI has been broadly defined as using AI to expedite AI development, but Exo proposes a more rigorous interpretation, emphasizing the need for systems to carry their state throughout the improvement process safely. This involves creating a robust framework where agents can access their own code, clone, rewind, and rebuild themselves incrementally without extensive external support. The significance of Exo lies in its potential to revolutionize AI development by allowing models to escape rigid, hand-engineered confines and evolve more organically. By implementing a canonical state—an immutable log that preserves the execution history—Exo ensures that agents can learn from past mistakes and maintain lineage across iterations. This innovative design addresses the challenge of enabling maximum flexibility in self-evolution while safeguarding critical components. Ultimately, Exo suggests a future where agents not only improve themselves but potentially lay the groundwork for broader advancements in AI capabilities.