News Summary

DeepMind, a subsidiary of Google, has unveiled a groundbreaking AI project called AlphaEvolve, which is capable of designing algorithms that go beyond human expertise. By combining the coding skills of the Gemini AI model with innovative testing and evolutionary methods, AlphaEvolve has developed more efficient algorithms for complex computations. Notably, it has improved upon the Strassen algorithm, which has been a staple for 56 years, by enhancing computational efficiency. AlphaEvolve’s capabilities extend to optimizing real-world problems, such as data center task scheduling, computer chip design, and the algorithms used in building large language models like Gemini. According to Pushmeet Kohli, head of AI for science at DeepMind, the AI agent surpasses known solutions in certain tasks, marking a significant advancement in the AI ecosystem. While some experts, like Princeton University’s Sanjeev Arora, note that the advancements are modest and specific to certain algorithms, the potential for AI to generate novel ideas through experimentation and evaluation is evident. The project showcases AI’s ability to produce innovative solutions that are not present in training data, as confirmed by research lead Matej Balog and supported by Josh Alman from Columbia University.