Terence Tao: neural networks have grown into the role of junior co-author in mathematics
One of the strongest mathematicians of our time, Terence Tao, increasingly works together with neural networks. In an interview with The Atlantic, he acknowledges that chatbots have reached a level where they can collaborate with living mathematicians, and this opens up “another way to do mathematics”.
The conversation was prompted by reports that generative models have solved several open problems from Erdős’ list. This list contains over a thousand questions of varying difficulty, compiled by the Hungarian mathematician Paul Erdős. Some solutions created with AI, including ChatGPT, did pass verification, but they mostly concerned the easiest problems.
Tao himself is a judge of these proofs. His assessment is restrained: the solutions “impress but do not overwhelm”, so far they are mostly “cheap victories”. According to him, AI systematically goes through a long tail of little-known problems and picks the easiest ones — those that an expert would solve in half a day using standard techniques.
At the same time, the mathematician notes the benefit even of such successes. Some proofs use techniques from old works that he himself did not know about, and this expands the arsenal. But there is also a fundamental difference: when a person solves a problem, they leave “trails” – intermediate ideas that colleagues can use. AI, on the other hand, takes you straight to the goal, like a helicopter, and skips the journey itself.
Tao expects hybrid human-machine contributions to appear in the coming months. Back in 2023, he predicted that by 2026 AI would become a reliable co-author of technical papers. According to him, the timeline is almost on track: neural networks are currently used at the level of a junior co-author who willingly takes on routine calculations and analysis of tedious cases.
For collaboration to become full-fledged, changes are needed from developers. Tao points to the lack of honest self-assessment of confidence in models: AI always claims to be completely confident in its answer, which reduces its usefulness. He also criticizes companies’ drive toward fully autonomous systems—mathematicians need a dialogue with the machine, not a button that launches a ready-made solution. According to him, the community will have to develop standards for responsible AI use in a much shorter time than was the case with computer proofs.
Primary source: www.theatlantic.com ↗