alphaevolve, a coding agent from google deepmind, uses gemini to design and improve algorithms. it has moved beyond research into practical use across many fields. in genomics, it reduced variant detection errors by 30% in deepconsensus, helping pacbio analyze genetic data more accurately and cheaply. for electricity grids, it boosted a graph neural network's ability to find feasible solutions for optimal power flow from 14% to over 88%, cutting the need for costly post-processing. in earth sciences, it improved natural disaster risk prediction accuracy by 5% across 20 categories like wildfires and floods.
the system is also advancing scientific research. in quantum physics, it suggested quantum circuits with 10 times lower error for molecular simulations on google's willow processor, enabling key experiments. mathematicians, including terence tao, used it to solve erdős problems and improve bounds for the traveling salesman problem and ramsey numbers. alphaevolve also aids in neuroscience, microeconomics, neural network design, cryptography, and ai safety. these results show how automated algorithm discovery can speed up progress in both theoretical and applied science.
alphaevolve is now part of google's infrastructure, optimizing next-generation tpu chip designs and improving cache replacement policies in days instead of months. it reduced write amplification in google spanner by 20% and cut software storage footprint by nearly 9%. through google cloud, commercial partners are using it: klarna doubled transformer training speed, substrate sped up semiconductor simulations, fm logistic improved routing efficiency by 10.4%, wpp gained 10% accuracy in ad models, and schrödinger achieved a 4x speedup in molecular force field training. this matters for ai and data science because it demonstrates how ai-driven optimization can directly enhance hardware, cloud services, and enterprise machine learning pipelines, making systems faster, cheaper, and more capable.
Source: Google DeepMind: AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields