Recursive Self-Improvement

Sakana AI achieves recursive self-improvement of models through collective intelligence, building a population of models that search for better weights through a simulated evolutionary process similar to a genetic algorithm.

The concept of recursive self-improvement was first proposed by AI safety researchers including Nick Bostrom in discussions of intelligence explosion.

Unconstrained recursive self-improvement can lead to performance drift, runaway costs, and even contamination of training data through leakage.

Recursive Self-Improvement refers to an AI system using its own capabilities to optimize its next version or assist in designing better training processes, forming a positive feedback loop.

Recursive self-improvement has a known pain point in the code generation domain of overfitting to a small number of test cases, where models make tests pass rather than genuinely solving the problem.

RSIGame is a framework for autonomous game development that employs a recursive self-improvement approach to enable large models to autonomously develop games.

Recursive Self-Improvement refers to the possibility that when an AI system participates in improving its own architecture or weights, a positive feedback loop may form, causing the pace of capability improvement to exceed the rhythm of human oversight, and it is directly related to the intelligence explosion hypothesis.

In recursive self-improvement (RSI), exploration strategy optimization is the most critical and most resource-intensive stage, because each strategy iteration requires placing the agent in a real environment to run complete online rollouts.

OpenAI published an internal research report defining the day AI coding assistants took off within the research team as 'RSI Day' (Recursive Self-Improvement).