The Drum · Weekly AI news, freeSubscribe →
Share
A groundbreaking new approach called Darwin Gödel Machine combines evolution with self-referential computing, enabling AI to rewrite and improve itself without needing absolute proof of each upgrade's success.
In the ongoing quest for artificial intelligence (AI) systems that can learn indefinitely, one of the most intriguing concepts is the Gödel Machine. Proposed by Jürgen Schmidhuber decades ago, a Gödel Machine is a hypothetical self-improving AI that optimally solves problems by recursively rewriting its own code when it can mathematically prove a better strategy. However, the practical realization of this idea has been hindered by an impractical assumption: the AI must mathematically prove that any proposed change in its code will yield a net improvement before adopting it.
To address this challenge, researchers from Sakana AI and Jeff Clune’s lab at UBC have introduced the Darwin Gödel Machine (DGM). This system leverages principles of open-ended algorithms like Darwinian evolution to empirically search for improvements that enhance performance. Unlike the theoretical Gödel Machine, DGMs are designed to be more feasible and practical.
The core idea behind DGMs is to create a self-improving coding agent that can rewrite its own code to enhance performance on programming tasks. Here’s how it works:
Foundation Models for Code Proposals: DGMs use foundation models (large pre-trained language models) to propose potential code improvements. These proposals are generated based on the current state of the AI and the specific task at hand.
Open-Ended Algorithms for Search: Recent advancements in open-ended algorithms, such as those inspired by evolutionary processes, are used to search for a growing library of diverse, high-quality AI agents. This approach allows DGMs to explore a wide range of possible improvements without being constrained by the need for mathematical proofs.
DGMs introduce several key features that enhance their self-improvement capabilities:
Experiments with DGMs have shown promising results. As they are provided with more computational resources, DGMs continue to improve their performance. This aligns with the trend observed in AI research that systems relying on learning ultimately outperform those designed by hand. The potential for DGMs to surpass hand-designed AI systems is significant.
The Darwin Gödel Machine represents a significant step towards creating AI systems that can learn and self-improve indefinitely. By combining the principles of open-ended algorithms with advanced foundation models, DGMs offer a practical approach to achieving this goal. As more compute resources become available, the performance gains from DGMs are expected to increase, potentially leading to breakthroughs in various AI applications.
For those interested in delving deeper into the technical details and exploring the implementation, the full technical report and released code are available:
Tags
Original Sources
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
13 June 2025
88 articles
Related Articles
Related Articles
More Stories
