Sakana AI Uses Evolutionary Algorithms to Optimize AI Models

The traditional development of AI models involves extensive trial and error and human insight. However, Sakana AI has adopted nature-inspired evolutionary algorithms to streamline this process. Recently, the company used LLMs to act as evolutionary algorithms, creating novel optimization techniques.

Sakana AI’s groundbreaking method, LLM² (LLM-squared), leverages LLMs to propose and refine preference optimization algorithms. This pioneering approach has led to the discovery of the Discovered Preference Optimization (DiscoPOP) algorithm, a game-changer that has consistently outperformed existing methods like Direct Preference Optimization (DPO). DiscoPOP and other new algorithms have not just demonstrated superior performance, but have set a new benchmark across multiple evaluation tasks.

The results, including model checkpoints and the discovery process code, are open-sourced on GitHub and HuggingFace. Collaborations with the University of Oxford and Cambridge University underscore this project's academic rigor.

What does it mean for you?

This breakthrough means more efficient and effective AI models, leading to smarter applications in various fields.

How can you use it?

Developers and researchers can leverage these open-sourced tools to enhance their AI models and optimization techniques.

What does it mean for jobs?

Automating parts of AI research could shift human roles towards more strategic and creative aspects of AI development, potentially reducing the need for extensive manual experimentation.