DeepMind’s latest: An AI for handling mathematical proofs

Computers are extremely good with numbers, but they haven’t gotten many human mathematicians fired. Until recently, they could barely hold their own in high school-level math competitions.

But now Google’s DeepMind team has built AlphaProof, an AI system that matched silver medalists’ performance at the 2024 International Mathematical Olympiad, scoring just one point short of gold at the most prestigious undergrad math competition in the world. And that’s kind of a big deal.

True understanding

The reason computers fared poorly in math competitions is that, while they far surpass humanity’s ability to perform calculations, they are not really that good at the logic and reasoning that is needed for advanced math. Put differently, they are good at performing calculations really quickly, but they usually suck at understanding why they’re doing them. While something like addition seems simple, humans can do semi-formal proofs based on definitions of addition or go for fully formal Peano arithmetic that defines the properties of natural numbers and operations like addition through axioms.

To perform a proof, humans have to understand the very structure of mathematics. The way mathematicians build proofs, how many steps they need to arrive at the conclusion, and how cleverly they design those steps are a testament to their brilliance, ingenuity, and mathematical elegance. “You know, Bertrand Russel published a 500-page book to prove that one plus one equals two,” says Thomas Hubert, a DeepMind researcher and lead author of the AlphaProof study.

DeepMind’s team wanted to develop an AI that understood math at this level. The work started with solving the usual AI problem: the lack of training data.

Math problems translator

Large language models that power AI systems like Chat GPT learn from billions upon billions of pages of text. Because there are texts on mathematics in their training databases—all the handbooks and works of famous mathematicians—they show some level of success in proving mathematical statements. But they are limited by how they operate: They rely on using huge neural nets to predict the next word or token in sequences generated in response to user prompts. Their reasoning is statistical by design, which means they simply return answers that “sound” right.