Google introduced Alphaevolve – a new agent based on artificial intelligence, which combines the creative potential of large language models and a strict test of solutions for the search and optimization of algorithms. This system is able not only to generate functions, as previous projects like Alphatensor did, but also to develop entire code bases, evolutionarily improving them with each iteration.
The system has already proved its practical value: the algorithms developed by Alphaevolve are used in Google data centers in the design of chips and in acceleration of AI training. For example, the agent proposed a new heuristic method of loading in the data centers, which allowed to return to the circulation about 0.7% of computing resources around the world. And all this is in the form of a code read by a person, which makes it not only effective, but also convenient for implementation.
In the hardware sector, Alphaevolve helped rewrite the code of the code in Verilog, optimizing the work of matrix multiplication without loss of correctness. This change is already integrated into the future version of TPU – a specialized AIC AIC from Google. And in the field of teaching models, Alphaevolve found a way to break matrix operations into more effective subtasks, accelerating the work of one of the key components of Gemini architecture by 23% and reducing the total training time by 1%.
Even on such low -level tasks as optimizing instructions for GPU, where only compilers work, Alphaevolve has achieved impressive results – up to 32.5% of acceleration in the implementation of the Flashattention nucleus for transformers.
In addition to practical applications in IT infrastructure, Alphaevolve was also useful in scientific tasks. He developed a new gradient optimization procedure and offered improved algorithms for multiplying complex matrices 4 × 4, surpassing the famous Strassen algorithm. Moreover, the system successfully decided or improved about 20% of more than 50 open mathematical tasks in geometry, combinatorics and numbers theory. One of the achievements is the new approach in the Kissing Number task in 11 dimensions: Alphaevolve found a configuration of 593 related spheres.
Google considers Alphaevolve as a step towards universal AI systems that can not only write code, but also help in scientific discoveries. Now the company is preparing an early access program for researchers and is developing a convenient interface of interaction with the system. In the future, Alphaevolve can be applicable not only in IT and mathematics, but also in areas such as materials science, the development of drugs, sustainable development and business analytics.