Sakana’s AI learns to upgrade its own code

Sakana AI, in collaboration with the University of British Columbia, has introduced the Darwin Gödel Machine (DGM), an innovative AI system that autonomously rewrites its own code to enhance performance. Drawing inspiration from biological evolution and open-ended exploration, DGM iteratively modifies its Python codebase, creating diverse versions of itself with varied tools and strategies. These self-generated variants are evaluated using coding benchmarks like SWE-bench and Polyglot, with the top performers archived to inform future iterations.

This evolutionary approach has yielded significant improvements: DGM’s performance on SWE-bench increased from 20% to 50%, and on Polyglot from 14.2% to 30.7%. The system developed features such as patch validation steps, enhanced editing tools, and mechanisms to avoid repeating past mistakes.

To ensure safety, DGM operates within sandboxed environments, implements strict modification limits, and maintains full traceability of changes. This self-improving framework represents a step toward AI systems capable of continuous, autonomous enhancement.

Key Points:

• DGM autonomously rewrites its own code to improve performance.
  
• Inspired by biological evolution and open-ended exploration.
  
• Achieved significant gains on coding benchmarks SWE-bench and Polyglot.
  
• Developed new features like patch validation and error memory.
  
• Operates with safety measures like sandboxing and change traceability.
  

#SelfImprovingAI #DarwinGödelMachine #AIEvolution #AutonomousAI #SakanaAI