Google DeepMind’s AI unit, AlphaGeometry, has demonstrated remarkable proficiency in geometry problem-solving, nearly matching the skills of top-performing human students.
In a test involving the International Mathematical Olympiad, AlphaGeometry correctly answered 25 out of 30 questions, approaching the gold medal-winning standard of human competitors. This accomplishment underscores the AI’s growing competence in mathematics and highlights the challenges that persist in creating artificial general intelligence (AGI) capable of equaling or surpassing human capabilities.
Also Read: Expanding Access: Unleashing the Full Power of Copilot for a Broader Audience
The neuro-symbolic system combines language learning and deductive reasoning, showcasing a hybrid method likened to “Thinking, Fast and Slow” by Daniel Kahneman.
The researchers, including Quoc V Le and Trieu H Trinh from DeepMind, utilized a trove of 100 million synthetic geometry examples to train AlphaGeometry. While its performance closely trailed the benchmark set by human winners in mathematical Olympiads from 2000 to 2022, it significantly outperformed the previous state-of-the-art automated system, scoring 10.
The ultimate goal for DeepMind and other researchers is to develop AI systems capable of addressing mathematical problems beyond human comprehension, discovering new mathematics to solve unprecedented questions.
Mikhail Burtsev, a Landau AI Fellow, acknowledges the progress made by DeepMind but notes that the steeper challenge remains—determining if AI can discover new mathematics for unanswered questions.
Despite Alpha Geometry’s achievements, the iconic moment of an AI math system surpassing a human rival, as seen in chess with Deep Blue and Garry Kasparov in 1997, remains elusive. DeepMind currently has no plans to enter the International Mathematical Olympiad, but the company doesn’t rule out such endeavors as it continues to advance in the challenging field of mathematics.
