AI that improves itself: the Darwin-Godel algorithm opens the era of self-programming machines

In recent years, artificial intelligence has shown more and more impressive achievements. But a new stage, which the research community has been striving for for decades, is approaching faster than expected. We are talking about creating AI systems that can recursively improve their own code — that is, program themselves. The new algorithm, dubbed Darwin-Gödel Machines (DGM), represents a significant step in this direction.

The idea is to complete the cycle of self-improvement: AI doesn't just execute commands, but analyzes its own structure, generates improvements, and implements them. This brings the technology closer to the dream of truly autonomous systems that can evolve without constant human intervention.

Darwin-Godel machines combine two principles: empirical adaptation and structural reflection. The architecture of such agents begins with the LLM (large language model), a model trained on a large array of code that can generate, read, and edit programs. This "base agent" undergoes a directed evolution: at each iteration, a new generation of agents is created, each of which receives one modification change in the code proposed by the language model.

The key innovation is the balance between random mutation and controlled progress. Instead of mindlessly selecting only the best solutions, as in classical evolutionary algorithms, DGS preserve the entire population of agents-including those changes that did not show results immediately. This allows you to accumulate non-standard ideas that can become breakthrough later.

During the experiment, the researchers ran the algorithm on two test sets-SWE-bench and Polyglot. After 80 iterations, the performance of agents on SWE-bench increased from 20% to 50%, and on Polyglot-from 14% to 31%. This means that the agents created by the algorithm itself became more and more capable of complex programming-from creating new files to building complex architectures.

It is noteworthy that the algorithm showed better results than alternative methods, including systems where the external model modified agents, and approaches that do not use populations. Especially important was the cumulative progress effect: agents were getting better at getting better.

The limitations and risks of this technology were also not ignored. The researchers implemented security measures by restricting agents ' access to the operating system and network, as well as monitoring all changes in the sandbox. According to lead author Jenny Zhang, future versions may include meta-rewards-rewarding agents for their interpretability and compliance with human guidance.

Although the best automatically generated agent has not yet surpassed the level achieved by manually designed systems (50% vs. 70%), the potential is obvious. With a sufficient number of iterations and resource-intensive calculations, machines capable of self-programming can theoretically go beyond the current human capabilities.

This approach clearly demonstrates a new paradigm in the development of artificial intelligence: not just automation, but the evolutionary design of the intelligence mechanisms themselves.