Google DeepMind announces "breakthrough" scientific AI.

Google DeepMind has just used chatbot models to provide solutions to major problems in mathematics and computer science. The system, called AlphaEvolve, combines the creativity of large language models (LLMs) with algorithms that can examine the model's suggestions to filter and improve solutions. The research is described in a technical report published by the company on May 14th.

"This paper is quite impressive," commented Mario Krenn, head of the Artificial Scientist Laboratory at the Max Planck Institute for Light Sciences in Erlangen, Germany. "I think AlphaEvolve is the first successful demonstration of new discoveries based on multi-purpose LLMs."

According to Pushmeet Kohli, Chief Scientist at DeepMind, in addition to using the system to find solutions to open problems, DeepMind has applied this artificial intelligence (AI) technique to its own real-world challenges. AlphaEvolve has helped improve the design of the next generation of tensor processors—computer chips specifically developed for AI—and has found a way to more efficiently leverage Google's global computing power, saving 0.7% of total resources.

Multipurpose AI

According to Krenn, most successful AI applications in science to date—including the AlphaFold protein design tool—have involved manually designed learning algorithms for specific tasks. But AlphaEvolve is versatile, leveraging the capabilities of LLM to generate code that solves problems in a wide range of fields.

DeepMind describes AlphaEvolve as an 'agent,' as it involves the use of interactive AI models. However, it targets a different point in the scientific process than many other 'agent' AI scientific systems, which are used to review literature and propose hypotheses.

AlphaEvolve is based on the company's Gemini LLM line. Each task begins with the user inputting the question, evaluation criteria, and a suggested solution, from which the LLM suggests hundreds or thousands of revisions. An 'evaluation' algorithm then assesses the revisions based on the criteria for a good solution.

Matej Balog, an AI scientist at DeepMind and co-lead researcher, said that based on the best-performing solutions, LLM proposes new ideas, and over time the system develops a more powerful set of algorithms. He said, "We explore a diverse set of problem-solving capabilities."

Narrow application

In mathematics, AlphaEvolve appears to allow for a significant acceleration in solving certain problems, according to Simon Frieder, a mathematician and AI researcher at the University of Oxford, UK. But it will likely only be applicable to a "narrow portion" of tasks that can be presented as problems to be solved through code, he says.

Other researchers are being cautious about assessing the tool's usefulness until it's tested outside of DeepMind. "Until the systems are tested by a larger community, I would remain skeptical and view the reported results with caution," said Huan Sun, an AI researcher at Ohio State University in Columbus.

According to Kohli, although AlphaEvolve requires less computing power to run than AlphaTensor, it is still too resource-intensive to be offered for free on DeepMind's servers. However, the company hopes that the announcement of the system will encourage researchers to propose scientific fields where AlphaEvolve can be applied. Kohli affirmed, "We are absolutely committed to ensuring that most people in the scientific community can access it."

VNA

Source: https://nhandan.vn/google-deepmind-cong-bo-ai-khoa-hoc-dot-pha-post879748.html