The shift is plausible because mathematical proofs combine several properties that are unusually favorable to AI: they can be represented digitally, evaluated against shared standards of correctness, and in suitable languages mechanically checked. Tao names Lean as an example of a proof-assistant language. This is not a claim that AI has solved mathematics: the slides stress that most frontier problems remain unsolved and that the space of problems has no fixed upper bound.
The implication is that the evaluation unit should include more than whether a result was produced: the path, resource consumption, failures, and independently checkable evidence matter. The slides call for transparent reporting of resource use and negative results, and identify First Proof as one effort in that direction.
Thesis
The central change Tao identifies is not simply faster problem solving, but a possible mismatch between abundant solutions and the incentives that historically organized mathematical research.
Observed facts
The slides describe mathematical proofs as objectively verifiable and say they can be mechanically checked when written in a suitable proof-assistant language such as Lean.
1The slides state that most problems at the current frontier of mathematics remain unsolved, despite computational expenditure.
1The slides call for assessments reporting resource consumption and negative results, and identify First Proof as one effort in this direction.
1
Mechanism
When solutions become cheaper to produce, a field that rewards solved problems may overvalue outputs and underinvest in exploring the surrounding mathematical landscape.
Why now
The slides point to substantial AI investment in mathematical problem-solving and to current systems finding solutions to some problems previously thought difficult.
Implications
The implication is that the evaluation unit should include more than whether a result was produced: the path, resource consumption, failures, and independently checkable evidence matter.
If problem-solving becomes abundant, mathematical institutions may need to reward exploration and frontier expansion more explicitly.
Evidence boundary
This source is Tao’s presentation and its examples, not an independent audit of AI mathematical capability or proof that research incentives have already changed.
Watch conditions
First watch condition: whether future public assessments report comparable resource use and failed attempts alongside successful results.