TLDR: Researchers affiliated with KAUST have made significant strides in self-rewriting artificial intelligence, introducing the ‘Darwin Gödel Machine’ (DGM). This AI system, building on Jürgen Schmidhuber’s decades-old concept, can autonomously modify its own code to enhance performance. Unlike the theoretical Gödel Machine, which required mathematical proofs for improvements, the DGM uses evolutionary algorithms to empirically discover beneficial self-modifications. Early experiments show substantial performance gains on coding benchmarks, marking a crucial step towards truly self-improving AI.
A groundbreaking development in artificial intelligence, spearheaded by researchers affiliated with the King Abdullah University of Science and Technology (KAUST), is bringing Jürgen Schmidhuber’s long-standing vision of a ‘Gödel Machine’ closer to reality. The newly introduced ‘Darwin Gödel Machine’ (DGM) represents a significant leap in the field of self-rewriting AI, demonstrating the ability of an AI system to autonomously modify and improve its own source code.
The concept of a Gödel Machine, originally proposed by AI pioneer Jürgen Schmidhuber decades ago, describes a hypothetical self-improving AI capable of optimally solving problems by recursively rewriting its own code. The theoretical challenge, however, lay in the impractical assumption that such an AI could mathematically prove a proposed code change would yield a net improvement before implementation. The DGM addresses this by adopting a more pragmatic approach.
Developed in collaboration with Jeff Clune’s lab at the University of British Columbia (UBC), the Darwin Gödel Machine harnesses the principles of open-ended algorithms, akin to Darwinian evolution, to empirically search for and implement performance enhancements. This method allows the DGM to propose code improvements using foundation models and then leverage innovations in open-ended algorithms to discover a growing library of diverse, high-quality AI agents.
Initial experiments have yielded impressive results, showcasing the DGM’s continuous self-improvement capabilities across various programming tasks. On SWE-bench, a benchmark for resolving real-world GitHub issues, the DGM automatically improved its performance from an initial 20.0% to a remarkable 50.0%. Similarly, on Polyglot, a multi-language coding benchmark, the DGM’s performance jumped from 14.2% to 30.7%, significantly surpassing representative hand-designed agents like Aider.
Among the self-improvements discovered by the DGM are a patch validation step, enhanced file viewing, improved editing tools, the ability to generate and rank multiple solutions to select the best one, and the integration of a history of past attempts and failures to inform future modifications. These capabilities highlight the DGM’s potential to soon outperform AI systems designed solely by human engineers.
This research builds upon Schmidhuber’s decades-long pursuit of self-improving AI. As the co-chair of the Center of Excellence for Generative AI (GenAI) and a professor in the Computer Science Program at KAUST, Schmidhuber has consistently advocated for machines that can ‘meta-learn’ – learning to learn. Earlier work from KAUST and the Swiss Artificial Intelligence Lab IDSIA (where Schmidhuber previously served as Director) introduced Variable Shared Meta Learning (VSML), a method where small neural networks replace conventional node weights to discover their own weight change algorithms, moving beyond human-invented learning algorithms like backpropagation.
Also Read:
- The Evolving Landscape of Computation: From Algorithms to Emergent Intelligence
- DeepAgent: Pioneering AI with Unified Reasoning, Autonomous Tool Discovery, and Memory Folding
The development of the Darwin Gödel Machine marks a pivotal moment in AI research, offering a tangible pathway toward creating truly autonomous and continuously self-improving artificial intelligence systems.
