DeepMind Releases a “Stunning” General-Purpose Scientific AI

This system has been used to enhance chip design and solve mathematical problems, but it has not yet been opened to researchers outside the company.

DeepMind stated that AlphaEvolve has helped improve the design of AI chips.

The DeepMind team under Google has developed an AI system named AlphaEvolve using large language models (LLMs), successfully solving several important problems in the fields of mathematics and computer science. The uniqueness of this system lies in its combination of the creativity of LLMs and the analytical capabilities of intelligent algorithms, enabling it to automatically screen and optimize the solutions proposed by LLMs. DeepMind detailed this achievement in a white paper released on May 14, 2025.

Mario Krenn, head of the Artificial Science Laboratory at the Max Planck Institute for the Science of Light in Erlangen, Germany, commented, “This paper is quite amazing. I believe AlphaEvolve has for the first time successfully demonstrated the possibility of achieving new scientific discoveries based on general-purpose LLMs.”

Pushmeet Kohli, the scientific director of DeepMind, pointed out that in addition to solving mathematical problems, the company has already applied AlphaEvolve to practical work. For example, it has optimized the design scheme of the new generation of AI chips (TPUs) and improved the utilization efficiency of Google’s global computing resources, saving approximately 0.7% of the total resource consumption. “It has had a significant practical impact,” said Kohli.

General artificial intelligence

Currently, most AI tools that have been successfully applied in the scientific field (such as the well – known protein structure prediction tool AlphaFold [1]) usually require specially designed learning algorithms for specific tasks. In contrast, AlphaEvolve is more general. It can leverage the powerful capabilities of large language models (LLMs) to generate code and thus solve complex problems in multiple fields.

DeepMind describes AlphaEvolve as an “intelligent agent” because it involves interactive collaboration among multiple AI models. However, compared to other “AI research agents” that focus on reviewing literature or generating hypotheses, AlphaEvolve is more focused on solving the problems themselves.

Specifically, each time AlphaEvolve is used, the user first provides a problem, clear evaluation criteria, and an initial solution. The large language model (LLM) then quickly proposes hundreds or even thousands of improved solutions. Next, an “evaluator” algorithm filters and scores these solutions one by one based on the user’s criteria—for example, reducing resource waste when optimizing Google’s computational tasks.

Based on the highest-scoring solutions, the LLM continues to generate new ideas, iteratively evolving better algorithms. “We explore multiple possibilities for solving problems,” said Matej Balog, the DeepMind scientist leading the research.

In fact, the core idea of AlphaEvolve originates from DeepMind’s FunSearch system introduced in 2023. FunSearch successfully used a similar evolutionary approach to outperform humans on previously unsolved mathematical problems. In contrast, AlphaEvolve can handle more complex, larger-scale code and spans multiple scientific fields.

DeepMind states that in the field of matrix multiplication—a mathematical computation widely used in neural network training—AlphaEvolve has discovered a method faster than the fastest algorithm proposed by German mathematician Volker Strassen in 1969. Despite being a general-purpose system, its performance even surpasses DeepMind’s own AlphaTensor system from 2022, which was specifically designed for matrix computation.

Krenn believes that this method can be widely applied to optimization problems in any scientific field where clear evaluation criteria can be defined, such as designing new types of microscopes, telescopes, or even developing new materials.

The applicable fields are still limited

Simon Frieder, a mathematics and AI researcher at the University of Oxford, pointed out that AlphaEvolve can indeed significantly accelerate problem-solving in mathematics. However, he noted that the system may only be applicable to problems that can be expressed through code, meaning its range of applications is relatively limited.

Other researchers have suggested that AlphaEvolve’s current results should be viewed with caution until the system is opened up to a broader research community and undergoes thorough validation. Huan Sun, an AI researcher at Ohio State University, said, “I will remain cautious about the current reported results until it has been tested more widely in the community.” Frieder also expressed a preference for waiting until researchers develop similar open-source versions rather than relying on DeepMind’s proprietary system, which could be modified or withdrawn at any time.

Although AlphaEvolve requires fewer computational resources than the earlier AlphaTensor system, Kohli admitted that it is still not available for free external use at this time. However, he said the company hopes that by publishing these research results, it will inspire other researchers to explore more applications. “We will definitely do our best to ensure that more people in the scientific community can use it,” Kohli said.

Original article link: https://doi.org/10.1038/d41586-025-01523-z