AI Autonomous Mathematical Conjecture Generation System MathGenesis Unveiled: First AI System to Independently Propose and Prove New Theorems

AI Autonomous Mathematical Conjecture Generation System MathGenesis Unveiled: First AI System to Independently Propose and Prove New Theorems

On July 12, 2030, DeepMind officially released the MathGenesis autonomous mathematical conjecture generation system. The system not only proves known theorems but, for the first time, independently formulates entirely new mathematical conjectures and completes rigorous proofs through formal verification. In pre-release internal testing, MathGenesis proposed 7 new conjectures, of which 3 were proven true, and a paper on one has been accepted by the top journal Annals of Mathematics.

MathGenesis's technical architecture is based on deep integration of large language models with the Lean 4 formal proof assistant. The system is first trained on massive mathematical literature to learn deep relational patterns among mathematical concepts, then uses a "conjecture distillation" module to identify gaps in existing theoretical frameworks, and finally employs Lean 4 for rigorous mechanized proof.

"A mathematician's work is essentially two steps: ask good questions, then find proofs," said Pushmeet Kohli, head of DeepMind's math AI program. "MathGenesis is the first time an AI has completed both steps end-to-end, rather than merely answering questions humans have already posed."

MathGenesis's most impactful conjecture addresses an open problem in graph theory — a new sufficient condition for the existence of Hamiltonian paths in sparse graphs. The proof uses a novel combinatorial technique that reviewers praised as "demonstrating surprising mathematical intuition."

However, the mathematical community's response to MathGenesis has been mixed. Fields Medalist Terence Tao commented: "MathGenesis's proofs are technically correct, but whether the conjectures it proposes are 'interesting' — whether they truly advance our understanding of mathematics — that will take time to judge." Another anonymous reviewer was more blunt: "There is still a vast gap between an AI that can propose conjectures and an AI that can propose valuable conjectures."

DeepMind has open-sourced MathGenesis's core modules and established collaborations with Princeton University and the Ecole Normale Superieure in Paris, planning to apply the system to open problems in number theory and algebraic geometry.