The Japanese startup Sakana AI has opened a Recursive Self-Improvement (RSI) research lab. The main goal of the new lab is to create neural networks that can iteratively rewrite, test, and optimize their own code.
Algorithmic Evolution Instead of Scaling
The company hopes that this approach to algorithmic evolution will allow it to move away from the currently dominant paradigm of compute scaling. Instead of simply increasing the number of parameters and the volume of data, the focus shifts to qualitative, independent improvement of algorithms.
The startup's portfolio already includes successful projects in this direction:
- LLM-Squared: A system where some large language models (LLMs) create training algorithms for other models.
- AI Scientist: A platform that automates the scientific research process.
The Future of Autonomous Agents
The next stage of Sakana AI's roadmap involves the development of advanced autonomous agents that will be able to independently improve their architecture without any human intervention.
Sources: