Close to the gold medal level of high school mathematics competition! Google DeepMind's AI makes a new breakthrough, "a crucial step towards building AGI"
AlphaGeometry was able to answer 25 out of 30 Olympiad questions, almost reaching the highest level of human performance.
AlphaGeometry, a system developed by DeepMind, a subsidiary of DeepMind, has demonstrated mathematical proficiency comparable to top Olympiad athletes in the latest research.
According to reports, a paper published on Wednesday in Nature showed that AlphaGeometry, developed by Alphabet's DeepMind, was able to answer 25 out of 30 challenging international math problems, equivalent to the level of a gold medalist. This achievement signifies another leap forward in the mathematical computing capabilities of artificial intelligence.
Specifically, within the training dataset constructed by the DeepMind team, consisting of 100 million synthetic geometric data, AlphaGeometry achieved a score of 25 (out of 30), close to the average score of Olympiad winners from 2000 to 2022 (25.9), and far surpassing the scores of previous state-of-the-art automated systems (10).
Quoc V Le, a researcher at DeepMind, stated:
"This is a crucial step towards AGI (Artificial General Intelligence)."
"This serves as a great example of how AI can drive technological progress and help us better understand the order of the world."
According to DeepMind, AlphaGeometry possesses "fast and rigorous" logical thinking abilities. Researcher Trieu H Trinh stated that this ability is particularly suitable for solving geometric problems.
However, the research found that AlphaGeometry still struggles with certain specific problems, such as an intersecting circle problem from the 1979 Olympiad.
AlphaGeometry is described as a neural-symbolic AI system that combines language learning and deductive reasoning. DeepMind has set the goal for AlphaGeometry to be an AI system that "goes beyond the scope of human mathematical thinking."
Mikhail Burtsev, a researcher at the London Mathematical Science Research Institute, stated that DeepMind's work is a significant step forward, but "limited to the challenges it has set for itself." He expressed:
"More formidable challenges still exist, namely whether AI can solve unsolved mathematical problems."