AI for Mathematics: DeepMind accelerates discoveries | Keryc
Mathematics has been the language we use to describe the universe for centuries. What if AI not only automates calculations, but actually collaborates with you and other mathematicians to create new ideas? Google DeepMind and Google.org present the AI for Math initiative to explore exactly that: empowering mathematical discovery with advanced models and tools.
What is the AI for Math initiative
The initiative brings together five leading institutions: Imperial College London, Institute for Advanced Study, IHES, Simons Institute (UC Berkeley) and Tata Institute of Fundamental Research. The goal is clear and ambitious: identify mathematical problems where AI can help, build infrastructure and tools for research, and speed up the pace of discoveries.
Google contributes funding via Google.org and access to DeepMind technologies like Gemini Deep Think (an enhanced reasoning mode), AlphaEvolve (an agent for algorithm discovery) and AlphaProof (a formal proof completion system). The idea is to create a feedback loop between fundamental research and AI applications.
What technical advances are behind this and why do they matter?
In recent years DeepMind has shown concrete leaps in reasoning ability. Systems like AlphaGeometry and AlphaProof reached silver-medal level at the International Mathematical Olympiad (IMO) in 2024. The latest Gemini with Deep Think reached gold-medal performance: it perfectly solved 5 of 6 problems and scored 35 points.
AlphaEvolve has been applied to 50 open problems in areas like analysis, geometry, combinatorics and number theory, improving previous solutions in 20% of cases. In algorithms, it discovered a new technique for matrix multiplication: for 4x4 matrices it found an algorithm using only 48 scalar multiplications, breaking Strassen’s 1969 record of 50. That’s not just a curiosity: optimizations like this change computational costs in problems that rely heavily on matrix multiplication.
These results show two things: AI no longer just reproduces human steps, it can propose new strategies; and formal verification is key to trusting those proposals.
Technical components and how they work at a high level
Gemini Deep Think: a reasoning mode that integrates deeper chains of thought, internal verification and neural-assisted symbolic search. It’s not just “more parameters,” but better ways to direct inference.
AlphaEvolve: combines learning-guided search, program synthesis techniques and combinatorial optimization to explore the space of algorithms. It can propose variants humans hadn’t tried because the search complexity was too high.
AlphaProof: a formal completion system that produces and verifies proofs with mathematical rigor. In practice this means interacting with proof assistants and converting between informal reasoning and formalization inside a verifier.
Key terms: formal proof (a demonstration verifiable by a machine), program synthesis (generating code/algorithms that meet specifications), model-guided search (using networks to prioritize promising paths in a search tree).
Implications for researchers and developers
For mathematicians: AI can be a copilot. It doesn’t replace human intuition, but it suggests conjectures, proof strategies and counterexamples that speed up iteration.
For AI researchers: this raises challenges about how to mix symbolic reasoning and statistical learning, how to design benchmarks that measure creativity and not just accuracy, and how to ensure reproducibility and verifiability.
For entrepreneurs and companies: algorithmic optimizations (like improvements in matrix multiplication) directly impact performance and compute costs, opening opportunities in infrastructure, scientific computing and verification tools.
Risks, ethics and validation
Trusting automatic results requires formal proofs and human review. Systems can propose incorrect or incomplete solutions that look correct. That’s why the initiative emphasizes verification infrastructure, controlled open data and close collaboration with mathematicians.
There are also questions about intellectual property for AI-assisted discoveries and how to distribute access to high-impact tools without concentrating technological advantage.
What’s next and how can you get involved?
The roadmap points to: broadening the problems tackled, improving integration with formal proof assistants, designing richer benchmarks (beyond the IMO) and building reproducible tools for the community.
If you’re a mathematician: think about how to formulate problems in terms an AI can explore (specifications, invariants, constraints) to speed up collaboration. If you’re an engineer or entrepreneur: look where an algorithmic improvement reduces costs in your systems and consider investing in automatic verification and reproducible infrastructure.
The bet is solid: combine human intuition with the ability to explore huge spaces and suggest unprecedented shortcuts. Will we reach new theorems guided by machines? Probably yes, but the real value will be the collaborative relationship between humans and models.
