AI Meets Mathematical Creativity
Pushmeet Kohli, Head of Science and Strategic Initiatives at DeepMind, shared the announcement on X (formerly Twitter):
“Google DeepMind has been collaborating with Terence Tao and Javier Gómez-Serrano to use our AI agents AlphaEvolve, AlphaProof, and Gemini Deep Think for advancing math research. AlphaEvolve can help discover new results across a range of problems.”
Kohli pointed to a particularly exciting example: using AlphaEvolve, researchers discovered a new construction related to the finite field Kakeya conjecture. The system Gemini Deep Think then verified the result, and AlphaProof formalized it in the Lean proof assistant.Tao later expanded on these results in a detailed blog post, describing how the collaboration blurred the boundary between human creativity and machine exploration.
What Is the Kakeya Conjecture?
The Kakeya conjecture addresses a seemingly simple geometric puzzle: in a finite mathematical space, one where numbers “wrap around” like in modular arithmetic, what is the smallest possible set of points that contains a line in every direction?Imagine a chessboard where you need to draw arrows in all directions but want to use the fewest possible squares. That’s the essence of the Kakeya problem.Mathematician Zeev Dvir proved a finite-field version of the conjecture in 2008, marking a major milestone. However, what AlphaEvolve achieved is not a contradiction of Dvir’s result but an optimization of it, offering smarter, more compact constructions that improve on known bounds in certain dimensions.
An Evolutionary Approach to AI-Driven Math
At its core, AlphaEvolve operates on a principle inspired by biological evolution. The system begins with a population of small computer programs generated by large language models (LLMs). Each program is evaluated according to how well it performs on a given mathematical objective, for example, how efficiently it constructs a Kakeya set.Programs that perform best are “mutated,” slightly modified to create new variants, and the process repeats over many generations. This evolutionary loop allows the AI to explore vast solution spaces, often discovering structures that human researchers might never consider.As the researchers describe it:
“The evolutionary process consists of two main components: (1) A Generator (the LLM) responsible for introducing variation, and (2) An Evaluator (provided by the user), which acts as the ‘fitness function.’”
In mathematical contexts, this evaluator rewards compactness, precision, and adherence to known constraints, essentially teaching the AI to “think like a mathematician.”
Results Beyond Expectation
The research team tested AlphaEvolve on 67 diverse mathematical problems across geometry, combinatorics, and number theory. In most cases, the system successfully rediscovered known results, a key validation of its reasoning ability. More impressively, in several problems, AlphaEvolve discovered improved or entirely new constructions.In certain cases, the AI went a step further, generalizing specific numerical findings into formulas that apply universally, a feat often associated with deep human insight.These results, though subtle in numerical terms, refine existing mathematical bounds and introduce methods that could influence practical applications in fields such as coding theory and signal processing.
Human–Machine Collaboration in Mathematics
For Terence Tao, one of the most celebrated mathematicians of our time, the experience has been transformative. In his blog, he wrote that AlphaEvolve serves not as a replacement for human intuition but as a partner capable of proposing unconventional yet valid ideas.Gómez-Serrano echoed this view, emphasizing that their role was to guide the AI, setting evaluation criteria, validating its results, and interpreting its discoveries. The synergy between human intuition and machine computation proved crucial to achieving meaningful results.This new mode of collaboration may represent a turning point in mathematical research: rather than competing with humans, AI systems like AlphaEvolve act as co-discoverers that expand the scope of what can be explored.
Wider Implications for Science and Technology
The consequences of this breakthrough extend far beyond pure mathematics. Since mathematical principles underpin many modern technologies, from cryptography and data compression to quantum computing, even small theoretical improvements can yield practical benefits.By improving our understanding of how to construct efficient mathematical structures, systems like AlphaEvolve could lead to faster algorithms, stronger encryption, and better error-correcting codes.However, this development also raises profound philosophical questions: can an AI truly invent, or is it merely optimizing human-generated ideas? The line between creativity and computation is becoming increasingly blurred.
Conclusion
As this news report from Karina Web highlights, AlphaEvolve represents a landmark in the collaboration between humans and artificial intelligence. By combining the evolutionary power of algorithmic search with the creative insight of mathematical reasoning, DeepMind and its collaborators have opened a new era in mathematical discovery.AlphaEvolve may not replace mathematicians, but it undeniably changes what it means to be one. The future of mathematics, and perhaps all of science, is no longer solely human; it is human and machine, thinking together.
