AI Improves at Improving Itself Using an Evolutionary Trick

Researchers have developed a new system, dubbed Darwin Gödel Machines (DGMs), that demonstrates an AI's ability to improve its own coding capabilities. These DGMs utilize large language models to read, write, and modify code, then employ an evolutionary algorithm to generate and test variations. The system allows for guided evolution, where changes are made based on the LLM's understanding of programming and then evaluated on their effectiveness in solving coding tasks. Unlike some evolutionary approaches, DGMs retain all generated agents, even those that initially underperform, to allow for potential future breakthroughs. This method represents a significant step towards creating AI systems that can recursively enhance themselves.

This advancement in self-improving AI could dramatically accelerate software development and potentially reshape the landscape of technological innovation.