AI is surfacing progress in mathematics

Mathematics has long occupied a strange place in the landscape of machine learning. While AI systems excel at language, vision, and game playing, formal mathematics represents a different kind of challenge: one where a single logical error invalidates an entire line of reasoning, and where creativity must operate within ironclad constraints. Understanding how AI engages with mathematics requires stepping back to examine what mathematical reasoning actually is, how humans have historically approached it, and what it would mean for machines to participate in this uniquely demanding human activity. What mathematical reasoning actually demands Mathematics is not primarily about computation. A human mathematician can spend months on a single problem without arriving at a numerical answer. Instead, mathematical reasoning involves constructing logical arguments, making strategic leaps about which paths to explore, recognizing patterns within abstract structures, and determining which existing theorems might prove useful in new contexts. Proof writing requires both rigid adherence to logical rules and intuition about which strategies are worth pursuing. This combination of constraint and creativity is part of what has made mathematics seem resistant to automation. The role of formalization in mathematics For most of history, mathematical proofs were written in natural language with varying levels of rigor. Euclid wrote proofs that seemed airtight but contained logical gaps. Newton's calculus worked beautifully but rested on shaky foundations. Not until the 19th and 20th centuries did mathematicians develop formal languages and logical systems precise enough to express theorems and proofs with complete precision. These formal systems break mathematics down into symbolic rules that can be checked mechanically. A formal proof is one that could, in principle, be verified by a computer following explicit rules, with no ambiguity or interpretation required. This development was not primarily driven by computers, but it created the conditions under which computers could participate in mathematical reasoning. Previous approaches to computational mathematics Mathematicians and computer scientists have been trying to automate aspects of mathematical reasoning for decades. Computer algebra systems can manipulate symbolic expressions, solve equations, and perform differentiation and integration. Automated theorem provers can search through logical deductions to verify or discover proofs, though they typically work within narrow, carefully defined domains. These systems are powerful within their scope but have historically struggled with the kind of open-ended reasoning required for unsolved problems in pure mathematics. They often rely on brute force search rather than the intuitive leaps that human mathematicians make. The bottleneck has always been the enormous space of possible approaches and the difficulty of encoding mathematical insight into explicit rul