Google has unveiled a man-made intelligence (AI) model specializing in mathematical reasoning. It emphasized that this can be a significant advance in processing mathematics, which requires higher reasoning ability than language, and that it has one of the best performance on this field.
Reuters reported on the twenty fifth (local time) that Google DeepMind released ‘AlphaProof’, an AI model specialized in mathematical reasoning. It also released ‘AlphaGeometry2’, an updated version of the AI model focused on geometry that was introduced earlier this 12 months.
AlphaProof is a mix of Google’s Large Language Model (LLM) ‘Gemini’ and ‘AlphaZero’, a model that solves board games similar to Go and chess.
AlphaGeometry2 is a system trained using more data, allowing it to unravel many more geometric problems than its predecessor.
Google DeepMind announced that two models solved 4 out of six problems on this 12 months’s International Mathematics Olympiad, the best rating ever achieved by an AI in a math competition.
Particularly, AlphaProof solved three of the competition problems, including probably the most difficult one, which was solved by only five out of over 600 human participants. Google said one problem was solved in a couple of minutes, while others took up to a few days. AlphaGeometry2 also proved its performance by successfully solving geometry problems.
This achievement is critical since it dramatically improves AI’s mathematical reasoning ability. Some see this development as a breakthrough toward artificial general intelligence (AGI).
“It is a huge step forward in machine learning and AI,” said Pushmit Kohli, vice chairman of scientific AI research at Google DeepMind. “No system has been developed yet that may solve problems with this rate of success.”
Regarding this, Reuters recalled that OpenAI recently developed ‘Strawberry’, which showed human-level reasoning performance. Strawberry can also be a model specialized in mathematical reasoning, and is recognized as a crucial key to AGI.
That’s, it was highlighted that Google also achieved progress beyond OpenAI on this field.
Reporter Park Chan cpark@aitimes.com
