Google DeepMind Announces "Breakthrough" Scientific AI

Google DeepMind has used chatbot models to come up with solutions to big problems in mathematics and computer science. The system, called AlphaEvolve, combines the creativity of a large language model (LLM) with algorithms that can test the model’s suggestions to refine and improve solutions. The research is described in a technical paper published by the company on May 14.

“This paper is quite impressive,” said Mario Krenn, head of the Artificial Light Scientist Laboratory at the Max Planck Institute for the Science of Light in Erlangen, Germany. “I think AlphaEvolve is the first successful demonstration of new discoveries based on versatile LLMs.”

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

Multi-purpose AI

Most successful applications of AI in science to date—including the AlphaFold protein design tool—have involved learning algorithms hand-crafted for a specific task, Krenn says. But AlphaEvolve is general-purpose, leveraging LLM’s ability to generate code that solves problems in a variety of domains.

DeepMind describes AlphaEvolve as an 'agent', as it involves using 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 starts with the user entering a question, evaluation criteria, and a proposed solution, from which the LLM suggests hundreds or thousands of revisions. An 'evaluation' algorithm then evaluates the revisions based on the criteria for a good solution.

Based on the solutions that are judged to be the best, LLM suggests new ideas and over time the system develops a more powerful algorithmic ensemble. "We explore a diverse set of problem-solving possibilities," said Matej Balog, an AI scientist at DeepMind and co-lead of the research.

Narrow application

In mathematics, AlphaEvolve appears to offer significant speedups in solving some problems, according to Simon Frieder, a mathematician and AI researcher at the University of Oxford, UK. But it will probably only be applicable to a “narrow subset” of tasks that can be formulated as problems to be solved through code, he said.

Other researchers are cautious about the tool’s usefulness until it’s tested outside of DeepMind. “Until the systems are tested by the broader community, I would remain skeptical and take the reported results with a grain of salt,” said Huan Sun, an AI researcher at Ohio State University in Columbus.

Although AlphaEvolve requires less computing power to run than AlphaTensor, it is still too resource-intensive to be made available for free on DeepMind's servers, Kohli said. However, the company hopes that releasing the system will encourage researchers to propose scientific areas in which to apply AlphaEvolve. "We are absolutely committed to making sure that it is accessible to the broadest possible audience in the scientific community," Kohli said.

VNA

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