New AI Model Advances the “Kissing Problem” and More

AlphaEvolve is a new AI system from Google DeepMind that shows AI can make real scientific discoveries, not just repeat what it has learned. It works by taking a problem written as code and then generating many different solution programs using a large language model. Each solution is checked by another program called a verifier. AlphaEvolve then uses a genetic algorithm similar to natural selection to evolve better and better versions of the code. This simple but powerful approach allowed it to improve the lower bound of the kissing number in 11 dimensions from 592 to 593, which is a new mathematical result. It also created a more general version of an improved matrix multiplication algorithm and solved many math problems at or above known optimal levels. Beyond pure math, AlphaEvolve helped Google optimize data-center orchestration, improve TPU chip designs, and reduce Gemini training time by 1 percent. The DeepMind team believes this method can eventually be applied to natural sciences as well, especially when simulations can act as verifiers. They also aim to combine it with their AI coscientist to create systems capable of producing higher-level scientific ideas.

Read more-https://spectrum.ieee.org/deepmind-alphaevolve