AI Evolves Itself Using Evolutionary Algorithms and LLMs

Researchers have developed a system called Darwin Gödel Machines (DGMs) that enables AI to recursively improve its own coding abilities. This breakthrough leverages evolutionary algorithms and large language models (LLMs) to create self-enhancing AI agents, marking a significant step toward autonomous AI development.

How DGMs Work

DGMs combine two key technologies:

1. Evolutionary Algorithms: Inspired by natural selection, these algorithms create variations of AI agents, test their performance, and iteratively refine the best ones.
2. Large Language Models (LLMs): LLMs like those powering ChatGPT provide the "intuition" needed to suggest useful code modifications.

The system starts with a base AI agent capable of reading, writing, and executing code. The DGM then:

• Generates multiple variants of the agent.
• Uses LLMs to suggest improvements.
• Tests each variant on coding benchmarks like SWE-bench and Polyglot.
• Retains all variants for "open-ended exploration," allowing even initially poor performers to contribute to future breakthroughs.

The evolution of AI agents, with the best performer marked by a star. (Credit: Jenny Zhang, Shengran Hu, et al.)

Performance Gains

• SWE-bench: Agent performance improved from 20% to 50%.
• Polyglot: Scores rose from 14% to 31%.

Notably, the best-performing agent followed a non-linear path, with temporary dips in performance before achieving breakthroughs. This highlights the value of open-ended exploration.

The lineage of the best agent included temporary setbacks. (Credit: Jenny Zhang, Shengran Hu, et al.)

Implications and Risks

Productivity Boost

• Companies like Microsoft and Google already use AI to generate significant portions of their code (30% and 25%, respectively).
• DGMs could automate the creation of high-performance software beyond human expertise.

Safety Concerns

• Self-improving AI risks becoming uninterpretable or misaligned with human goals.
• Researchers added guardrails, such as sandboxing and code review, to mitigate risks.

The Singularity Debate

• Some experts, like Jürgen Schmidhuber, dismiss fears of a runaway AI "singularity."
• Others, like Zhengyao Jiang, emphasize the need for human creativity to guide AI evolution.

Future Directions

• Combining DGMs with systems like Google DeepMind’s AlphaEvolve, which optimizes algorithms and hardware.
• Extending DGMs to domains like drug design by scoring agents on multiple objectives.

This research represents a major leap toward self-improving AI, with profound implications for both productivity and safety.