DeepMind’s latest AI can solve geometry problems

DeepMind’s latest AI can solve geometry problems

DeepMind, Google's AI research lab, believes that advancing AI could hinge on solving tricky geometry problems. They've introduced AlphaGeometry, claiming it can tackle as many geometry puzzles as top math Olympians. The open-sourced code solves 25 problems, beating previous tools.

According to Google AI researchers Trieu Trinh and Thang Luong, AlphaGeometry's achievement is a significant step in deepening mathematical reasoning, aiming for more advanced AI. They envision it opening new doors in math, science, and AI.

Why geometry? DeepMind says proving mathematical theorems, like Pythagoras', requires both reasoning and choosing steps toward a solution. This could be crucial for future AI systems.

Training AI for geometry is tough due to a lack of usable data and the need for logical reasoning. DeepMind's solution combines a "neural language" model with a "symbolic deduction engine" to guide problem-solving. They trained AlphaGeometry from scratch using synthetic data and evaluated it on Olympiad problems.

The results, published in Nature, ignite debates over neural networks versus symbolic AI. While neural networks rely on data and statistical learning, symbolic AI manipulates symbols based on rules. AlphaGeometry's hybrid approach suggests combining both could lead to more powerful AI systems.

The ultimate goal, according to Trinh and Luong, is to create AI that can generalize across math fields, pushing the boundaries of human knowledge. This approach could revolutionize how AI discovers new insights in math and beyond.