• TWEAK
• Software

Google AI systems achieve breakthroughs in mathematics

Translate from : Google AI systemer opnår gennembrud i matematik

Google has unveiled two cutting-edge artificial intelligence systems that have made remarkable progress in tackling complex mathematical problems, marking a significant leap forward in the field of generative AI. Traditionally, AI models that rely on statistical predictions to determine the next word have struggled with abstract mathematics because of its requirement for advanced reasoning skills reminiscent of human cognition.

DeepMind, Google's AI research arm, revealed that its latest AI models, AlphaProof and AlphaGeometry 2, successfully solved four out of six problems at the 2024 International Math Olympiad, a prestigious competition for high school students. This milestone highlights a significant breakthrough in AI's ability to handle complex mathematical tasks. Google noted in a blog post that while one issue was resolved in minutes, others took up to three days, exceeding the competition's time limit.

Nevertheless, this performance represents the highest scores ever achieved by an AI system in this competition. AlphaProof, which excels in reasoning, was developed by combining Gemini, the language model behind Google's chatbot, with AlphaZero, an AI system famous for outperforming human masters in board games like chess and Go. AlphaProof managed to solve three of the Olympiad's problems, including the most formidable, which only five out of more than 600 human contestants were able to crack.

In addition, AlphaGeometry 2 solved another problem, which testifies to the impressive capabilities of the AI. The new breakthroughs for AlphaProof and AlphaGeometry 2 illustrate the potential for AI to tackle problems that require deep reasoning and understanding, moving beyond just pattern recognition. This progress could have far-reaching implications for various fields that depend on sophisticated problem solving, such as physics, engineering, and economics.

Integrating different AI systems, such as Gemini and AlphaZero, into a unified entity with advanced reasoning suggests a promising future where AI can be even more useful in highly specific scientific research and discovery. However, the extended time required by the AI to solve certain problems indicates that there is still considerable room for improvement. Future advances will likely focus on increasing the speed and efficiency of these models, making them more practical for real-time applications."