Google launches AI initiative for mathematics

Google DeepMind and Google.org have created an initiative to use artificial intelligence in mathematical research alongside five leading institutions. The goal is not to replace mathematicians, but to help them find important problems, test ideas and accelerate new discoveries.

Five centers for AI research

The AI for Math Initiative initially brings together these institutions:

• Imperial College London
• Institute for Advanced Study
• Institut des Hautes Études Scientifiques
• Simons Institute for the Theory of Computing at the University of California, Berkeley
• Tata Institute of Fundamental Research

Google will provide funding through Google.org and access to several of its tools. These include Gemini Deep Think, a mode specialized in reasoning; AlphaEvolve, an agent that searches for algorithms; and AlphaProof, a system that helps complete formal proofs, meaning mathematical arguments that can be verified step by step.

The joint work will focus on three areas: identifying problems where AI can contribute something new, building the tools needed to investigate them and speeding up the validation of results. The idea is to create a continuous collaboration loop between mathematicians and AI researchers.

What AI has already demonstrated

Google DeepMind enters this initiative with several recent results. In 2024, AlphaGeometry and AlphaProof reached a level equivalent to a silver medal at the International Mathematical Olympiad.

At the 2025 edition, a Gemini model with Deep Think correctly solved five of the six problems and scored 35 points, a gold-medal-level result. That does not mean AI can solve any mathematical problem: olympiad tests are difficult, but they are bounded and have verifiable answers.

Another system, AlphaEvolve, was applied to more than 50 open problems in analysis, geometry, combinatorics and number theory. It improved on the best known solution in 20% of cases.

One of its most concrete results involves a core operation in computer science: matrix multiplication. For 4-by-4 matrices, it found a method that requires 48 scalar multiplications, compared with the previous record of 50, set by Strassen's algorithm in 1969. An improvement like this may seem small, but these operations appear in graphics, simulations and AI models.

AlphaEvolve also helped find mathematical structures showing that some complex problems are even harder for computers than previously believed. These results do more than answer questions: they also help clarify the limits of computation.

What changes for you

There will be no immediate consumer application or new tool for doing math homework. The effect will be more indirect, but it could matter: better algorithms can reduce the cost of computations used in AI, science, engineering and software.

The most relevant part is the collaboration model. AI can explore many possibilities and detect patterns, while mathematicians contribute intuition, context and judgment to decide whether an idea is genuinely useful. The initiative will try to turn that combination into a regular research process.

It remains to be seen how far AI can advance on problems with no known answer and no clear path to finding one. What matters now is whether these systems can produce new ideas that other mathematicians can verify and reuse, rather than simply solve exercises governed by established rules.