Terence Tao delivered a public lecture titled “Math 2.0” at the California Institute of Technology on October 9, 2026, as part of the AIM Public Lecture Series. Although the event was announced under the heading “Machine Assisted Proof,” Tao has indicated that he prefers the label Math 2.0 for the talk going forward. The lecture appears in Part V of his online AI views summary, grouped with the earlier “Math 1.0 and Math 2.0” section, and the accompanying slides are linked there, with a video and transcript expected to follow.
In the accompanying summary of his views on AI, Tao describes contemporary AI tools as probability kernels: for a given input they generate a random output that clusters around the correct answer but also produces subtle, plausible‑looking errors. He characterizes them as unreliable yet powerful, noting that their competence is spiky—excelling in some areas while making basic mistakes in others. He situates this development within a long tradition of automating mathematical labor, comparing today’s AI to the human “computers” who once produced logarithm and trigonometric tables, to Hendrik Lorentz’s team modelling a Dutch dam, to computer‑algebra systems, and to SAT solvers, each of which delegated routine layers while humans kept overall direction. Tao also historicizes the anxiety, observing that a hundred and fifty years ago a mathematician’s chief usefulness was solving differential equations, and six hundred years ago it was building tables of sines and cosines for navigation—both now done by computer in seconds, yet neither marked the end of the field.
Tao argues that mathematics is a particularly suitable domain for AI because, unlike most applications, the output of a proof can be automatically verified, which curtails the risk of undetectable mistakes. He therefore advises researchers to employ AI only where its output can be red‑team‑ed, matching the tool’s use to tasks where the user retains the ability to check and interpret results. This perspective leads him to weigh comparative advantage and acceptable failure rate, suggesting that AI is best suited for middle‑band tasks where he has some competence but little practice, while high‑stakes, low‑error‑tolerance work remains the domain of human experts. He further distinguishes between artificial friction—tedious computation that can be offloaded if the output can be spot‑checked—and natural friction—genuine conceptual difficulty that should not be smoothed away, as doing so would deprive learners of essential thinking.