Machine Math Makes Translation the Moat — 2026-10-07

On October 7th, Azeem Azhar's Exponential View published What is Left to Do? The piece discusses OpenAI releasing a range of mathematical results produced by an unreleased frontier model, seven hundred and twenty-two manuscripts in three hundred and seventy-two families spanning number theory, complexity theory, and mathematical physics.

Azeem says the average result took the equivalent of about three hours of ChatGPT Pro thinking.

He also cites scientist Derya Unutmaz arguing that the body of work represents eighty-one percent of the major math discoveries in the past three years, a claim worth treating carefully.

But even if the true figure were much lower, the direction is astonishing.

Many of these results were verified and lean, though not all.

That distinction matters.

Uh, for years, we've talked about AI as a tool for speeding up research, but this is different.

This is not AI helps a mathematician search literature.

This is machine systems generating enough formal work that humans may not be able to contextualize it at the speed it arrives.

Azeem frames the deeper problem beautifully.

Mathematics could split into two layers.

What's-- One layer is machine mathematics, uh, vast, verified, uh, mostly consumed by AIs.

The other is human mathematics, uncompressed theories, analogies, explanations, and teaching structures that humans can actually understand.

That is, that is today's central theme.

Verification is necessary but not sufficient.

A proof can be formally valid and still be intellectually unusable to most people.

A drug candidate can be computationally promising and still fail in a phase two trial.

A robot hand can execute a grasp in simulation and still break a wine glass in a kitchen.

A mortgage platform can encode regulation and software and still need humans, auditors, and regulators to trust the operating model.

So the next bottleneck after verification is translation.

Can AI not only discover but explain? Can it compress? Can it teach? Can it turn alien machine output into human institutions? This is where the markets come in.

On October 7th, Azeem Azhar's Exponential View published What is Left to Do? The piece discusses OpenAI releasing a range of mathematical results produced by an unreleased frontier model, seven hundred and twenty-two manuscripts in three hundred and seventy-two families spanning number theory, complexity theory, and mathematical physics.

Azeem says the average result took the equivalent of about three hours of ChatGPT Pro thinking.

He also cites scientist Derya Unutmaz arguing that the body of work represents eighty-one percent of the major math discoveries in the past three years, a claim worth treating carefully.

But even if the true figure were much lower, the direction is astonishing.

Many of these results were verified and lean, though not all.

That distinction matters.

Uh, for years, we've talked about AI as a tool for speeding up research, but this is different.

This is not AI helps a mathematician search literature.

This is machine systems generating enough formal work that humans may not be able to contextualize it at the speed it arrives.

Azeem frames the deeper problem beautifully.

Mathematics could split into two layers.

What's-- One layer is machine mathematics, uh, vast, verified, uh, mostly consumed by AIs.

The other is human mathematics, uncompressed theories, analogies, explanations, and teaching structures that humans can actually understand.

That is, that is today's central theme.

Verification is necessary but not sufficient.

A proof can be formally valid and still be intellectually unusable to most people.

A drug candidate can be computationally promising and still fail in a phase two trial.

A robot hand can execute a grasp in simulation and still break a wine glass in a kitchen.

A mortgage platform can encode regulation and software and still need humans, auditors, and regulators to trust the operating model.

So the next bottleneck after verification is translation.

Can AI not only discover but explain? Can it compress? Can it teach? Can it turn alien machine output into human institutions? This is where the markets come in.