Specialized in Problem-Solving Algorithms
Applied to the Design of Google's AI Chips
DeepMind, Google's artificial intelligence (AI) division, has unveiled a new AI agent specialized in algorithms (problem-solving procedures) called 'AlphaEvolve'.
On May 14 (local time), Google DeepMind introduced AlphaEvolve as an 'evolutionary coding agent' via its blog, describing it as "an AI that evolves algorithms necessary for real-world applications in mathematics and computer science."
AlphaEvolve operates by discovering algorithms that can be applied across various fields and autonomously improving its own code. The underlying model is based on Google's family of AI models, Gemini. In particular, it combines 'Gemini Flash', which is fast and efficient, with 'Gemini Pro', which has enhanced reasoning capabilities.
AlphaEvolve uses automated evaluation metrics to verify, execute, and assess the generated programs. Google explained that these evaluation metrics objectively quantify accuracy and quality, making the agent especially useful in systematically measurable domains such as mathematics and computer science.
Over the past year, Google has internally utilized AlphaEvolve to apply new algorithms across data centers, hardware, software, and more. As a result, Google DeepMind explained that the company was able to save 0.7% of its total computing resources.
AlphaEvolve was also used in the development of Google's AI chip, the Tensor Processing Unit (TPU). It generated design code that eliminated unnecessary computations and increased the speed of multiplication operations required for AI training by 23%, thereby reducing Gemini's training time by 1%.
When AlphaEvolve was applied to more than 50 unsolved problems in areas such as mathematical analysis, geometry, combinatorics, and number theory, it identified new solutions for 75% of the problems. For 20% of the problems, it proposed solutions that were better than existing ones.
Google DeepMind stated, "AlphaEvolve is evolving from discovering algorithms in specific domains to developing algorithms that solve more complex real-world problems," adding, "We expect this model to evolve further and drive significant change across materials science, drug discovery, sustainability, technology, and business."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
