Fields Medalist Jacob Tsimerman: "math is maybe being hit faster by AI because it's a pure thought activity" AI systems have already solved many important conjectures on their own, and very soon, I think they'll become robustly better at the rest of what we do too

AI mathematical capabilities are only moving exponentially from this point on.

Yup, and the reason is boring Math has verifiers. Lean tells you NO in one second and that's a reward signal you can train on. Biology doesn't have that, law doesn't have that. It's not that math is pure thought it's that math is GRADEABLE, and everything gradeable falls first

Laughable. It will weed out fakes while (finally, in an arc of history coming to fruition) being the “juice” (and henchman) the Philosopher King, who by Logic rightly rules the world (as Logic manifest in a qualified human leader), has always needed to overcome the layman horde.

pure thought was supposed to be our last moat

not just this, but the nature of mathematical proofs being all or none lends itself to very clean training datasets for large neural networks there is zero analog to this in biology, for example

math is verifiable in a way spreadsheets and memos aren't. proof or it didn't happen. that's the lever, not pure thought

People think AI is “doing math” because it can solve conjectures, find counterexamples, and outperform average humans on tasks. But solving problems is the least mathematical part of mathematics. It’s the part that collapses cleanly into search, interpolation, and optimisation — exactly what single‑pole systems are built for. The real benchmark is problem‑generation: deciding what is interesting, what is fertile, what is structurally deep, what opens a new regime. Humans don’t become mathematicians because they can solve puzzles. They become mathematicians because they can invent the puzzles worth solving. That ability doesn’t come from slope. It comes from geometry. Problem‑generation requires multi‑pole cognition — orientation, stabilisation, integration, synchronisation. Those poles constrain each other. They create taste. They create direction. They create the ability to sense when a question is trivial, when it is dead, when it is fertile, when it is a doorway into a new world. Frontier AI has none of those poles. It has one optimiser. One centre of gravity. One evaluative regime. A single pole can escalate. It can interpolate. It can solve. What it cannot do is originate. It cannot decide what matters. It cannot generate the next field. It cannot produce the geometry where new mathematics appears. This is why problem‑generation is the real benchmark. Solving problems measures capability. Generating problems measures intelligence. And until the architecture shifts from single‑pole escalation to multi‑pole development, AI will keep getting better at the part of math that doesn’t matter and remain blind to the part that does.

Proof that the fields medal is not given to "math geniuses" (seriously who the fuck knows any of the retards who get these awards lmao)

The biggest shift may not be solving problems faster. It may be finding ideas that people would never think to test.

Or perhaps LLM’s are calculaters

before you say anything, keep in mind this guy is joining the AI company $$

Wouldn't it be awesome though, like sure for now we have a lot of historical problems which the ai will solve, but once many will be solved or reach diminishing returns more people will create new mathematical conjectures. It is the language of the universe and it will keep grow leaps

Because it's verifiable, as I already said LLMs are good for verifiable, finite systems.

Math is getting hit first because it is harder to deny e.g. the resolution of century-long conjectures. In many other areas AI is just as effective but there's more degrees of freedom for evaluating AI-output unfairly (

He is 💯 right - I guess will be called Automath

every "pure thought" field shd take notes - displacement never starts w/ the median worker, it starts at the frontier & works down. we track capability obsessively & the transition costs almost not at all. that's the #data gap that worries me #FutureOfWork

It's a mini-retelling of history - didn't humans start with math at the dawn of civilization?

Its because LLMs only simulate our left brain Abstract mathematics, either/or definitions & contradictions, are not invented or discovered, they are a feature of our left hemisphere’s emphasis on either/or statements, cuz all verbs & prepositions are either/or, including =, +, -

most important thing is math is verifiable. very cheap synthetic data can be generated and tried with brute force. something like robotic is far harder because while its verifiable, training is orders of magnitude slower and more expensive. its all about training data

AI's wins so far are on problems with a clear, checkable answer, but most of what makes math hard is taste, judgment and knowing which approach to even try. Harder to train against that.

Do you think embodiment is necessary for abduction?