The Path to Mathematical Superintelligence | Tudor Achim | TED

Let's take a look at this clay tablet.

It might not look like much,

but it's actually some of the oldest mathematics we have.

It’s a 4,000-year-old message in a bottle from ancient Babylon --

a precursor to the quadratic equation.

And for four millennia,

people have been doing math basically the same way.

Someone will have a brilliant idea,

they’ll write it down, and their peers will discuss and check it.

It’s a process built on creativity, communication and, most importantly,

trust between people.

And what might seem like a humble or simple process is anything but.

It's not just been successful.

It's been, as the physicist Eugene Wigner famously put it,

unreasonably effective.

Wigner was pondering and trying to unravel a deep mystery.

Why should the abstract, creative,

and often bizarre ideas that spring from a mathematician's imagination

so often be the perfect language with which we understand the universe?

Why should the strange laws of non-Euclidean geometry,

which were originally conceived of as a thought experiment

in the 19th century,

turn out to be the exact mathematics

that Einstein needed for general relativity?

Why should the esoteric math of group theory,

which was originally designed to study the abstract nature of symmetry,

be fundamental to understanding everything from particle physics

to the patterns in crystals?

Well, there's no logical reason it has to be this way.

This strange connection

between pure mathematical thought and the real world

has actually been the invisible engine driving human progress.

Every piece of technology that defines our lives

was ignited with a mathematical spark.

If you take the device in your phone,

its brain is based on the quantum mechanics of semiconductors.

And that's a theory built on linear algebra

and complex numbers.

The wireless signals that get data to it,

they're just a concrete manifestation of Maxwell's equations.

And finally, the security that protects your data online

is based on number theory,

which for a long time was truly considered the most pure

and least applicable possible branch of mathematics.

And now it safeguards trillions of dollars in the global economy.

And now we come to AI.

Modern AI is not just built with math, it's forged from it.

A neural network is just a monumental structure of applied mathematics.

And when AIs learn,

they're using the tools of calculus

to navigate vast landscapes of possibilities

with billions of dimensions.

So AI is, in its soul,

a mathematical idea that's given life through computation.

So we agree that math is the foundation that modern civilization is based on.

But that foundation is starting to show some signs of strain.

The very process of human-led discovery

that's gotten us to this point is nearing a breaking point,

buckling under the weight of its own success.

And now AI, which is one of mathematics’ greatest creations,

is accelerating us towards that breaking point

faster than the world's ready for.

So let's just look at some evidence.

Consider the Poincaré conjecture.

This is a legendary problem.

It's a fundamental question

about the nature of three-dimensional shapes

originally posed in 1904.

And for nearly a century,

it stood as an unconquered Everest of mathematics.

Until in 2002,

a Russian mathematician working in isolation

named Grigori Perelman

posted a series of three short, cryptic papers online.

He didn’t bother submitting them to a journal --

he just put them on the internet and walked away.

His fellow mathematicians had to stop what they were doing

and try to decipher it.

And several teams working independently of the best of colleges in the world,

took the next four years to try to unpack the arguments,

fill in the logical gaps

and eventually, at the end, after they really reviewed it,

declare that yes, he did it.

He proved the Poincaré conjecture.

But that's interesting

because it took one person to write a proof

and a global, multi-year intellectual mobilization to check it.

And that's in the best case, when the proof is correct.

Consider Andrew Wiles's proof of Fermat's Last Theorem.

With the electrifying announcement in 1993 in Cambridge, the world celebrated.

But during the peer-review process, deep in it,

a single thread was found out of place

in that magnificent tapestry of a proof,

and when we started to pull on it,

the proof started to unravel.

And this wasn't a small mistake.

Andrew Wiles and his collaborator Richard Taylor took two years of heroic,

secret effort to try to fix it.

And that effort included some insights that Andrew Wiles said

were among the most important in his life.

And that's before we throw AI into the mix.

Two short years ago,

AI could barely solve entry-level high school math-contest problems.

They were very clever, but brittle.

Now, in 2025,

they can compete with the best of us

at the International Math Olympiad,

which is the premier precollege math competition.

But the interesting bit is the following.

The AI might work for four hours and produce a purported solution,

which takes an expert human mathematician maybe up to an hour to check.

And we all know the exponential trend that AI is on.

So we can expect it’s not going to be one proof in an afternoon --

it’s going to be a thousand pretty soon.

And they're not going to be attempts to solve math-contest problems.

They're going to be attacks on the most fundamental

and important questions of the day,

whether it's the Riemann hypothesis,

Navier-Stokes or P versus NP,

just to pick a few.

We simply don't have the human bandwidth

to review all these proofs.

There's only a couple thousand mathematicians

that are qualified to do it, and they already have day jobs.

And it's not just a verification bottleneck.

The very process by which we train these AIs

is taking the data off the internet,

which is from humans,

post-training them with human feedback,

and so we're essentially baking in the cognitive biases

and the flawed reasoning of humans into these future engines of discovery.

So the conclusion is in some sense obvious.

Humans are becoming the bottleneck of verification for AI.

And now the question is, where does that leave us?

Is this the end of the road for reliable mathematical discovery?

Are we resigned to drowning in a sea of unverified claims

where we can't really tell truth from fiction?

And are we about to squander the opportunity

for AI to revolutionize math?

Well, the good news is no.

But it does mean it's time

to upgrade the 4,000-year-old operating system of math,

and move away from the imprecise and ambiguous nature of human language,

and towards a language that computers can understand.

The solution is formal mathematics.

But before I tell you how this futuristic idea works,

we should first recognize that it has a deep and fascinating history

dating back to the 17th century,

where a mathematician actually laid out the road map with stunning foresight.

Four hundred years ago,

in a Europe torn by religious and political conflict,

a polymath named Gottfried Wilhelm Leibniz

had a vision of breathtaking ambition.

He was a contemporary of Newton and a cocreator of calculus,

but his dreams went far beyond that.

He dreamed of something called a universal characteristic,

which was a system for perfectly encoding all scientific

and philosophical thought.

And the system had three parts.

First, you need a perfect logical language.

Second, you need a grand encyclopedia written in language

that contains all verified human thought.

And third,

and this is the masterstroke,

you need a so-called engine of reason,

a system of mechanical rules

by which you can automatically derive new facts from that library

as surely as a calculator performs arithmetic.

Now, Leibniz thought this would revolutionize humanity.

With a system like this,

if two people had an intellectual conflict,

they would resort to logic and not rhetoric to resolve it.

They would simply sit down,

say “calculemus” -- “let us calculate,”

and get to the bottom of it.

In some sense, it was meant to be a universal calculator for truth.

Now, Leibniz was a bit of an optimist.

He thought this would take a small group of people five years to build,

and he was off by several centuries.

But what I think is really remarkable

is that in 2025,

truly for the first time in history,

it's actually possible to realize this philosopher's dream.

So what do we need?

Well, we need a perfect, logical language.

Turns out we've got it.

It's called Lean.

Lean is a programming language,

but it's also what's known as a proof assistant.

You can think of it as a programming environment

for mathematical proofs,

where it doesn't just give you feedback

if you have a syntax error here or there --

it's actually looking at the core of the mathematical argument

and telling you if you have any problems anywhere in it.

Great.

What's the second thing we need?

We need the grand encyclopedia.

Well, the good news is we've got that too.

It's called Mathlib.

Mathlib is an open-source project.

It's about two million lines of code in Lean,

and it covers a lot of the undergraduate and graduate math curriculum.

You can think of it like a Wikipedia for proven truth,

where every edit is computationally certified for correctness.

OK, we've got the language,

we've got the encyclopedia,

what about the engine of reason?

Well, we could try to have humans do it,

but you've got to write a lot of Lean code

and the level of robotic precision you need to write a formal proof

is not something that human creativity is so well suited for.

And that's how we've come full-circle.

It turns out that AI is the key to making this whole thing work.

In the future, AI is not just going to be writing math papers in English

for humans to read.

They're going to be writing math proofs in Lean

for computers to check.

And that is the fundamental key

that makes it possible to use Leibniz's vision

to unlock the full potential of AI in mathematics.

Because when a math AI spits out a proof in Lean of, let's say,

the Riemann hypothesis,

we're not going to need humans to go through every single line of the proof

in painstaking detail,

check every single case,

and understand the possibly strange and alien logic of the proof

just to see if it's correct.

Instead, all we're going to do is we're going to take those files,

we're going to give them to a Lean compiler, and if it builds,

we can know with absolute certainty it's correct.

And this is what fundamentally alters our relationship with AI.

AI can now become a true collaborator,

one whose word we don't have to take on blind faith.

We get to trade in the tedium of checking for the creative joy of discovery.

Humans get to use our intuition and judgment,

we ask the questions,

we chart the course, we propose the brilliant conjectures

and then we delegate to AI to explore the vast oceans of logic,

to find the correct answer,

and then a computer confirms that we've gotten to the destination.

And the amazing thing

is that this isn't just some far-off science-fiction dream.

It turns out that at this year’s International Math Olympiad,

automated systems were able to find solutions to five of the six problems

in a way that computers could check and require no human review whatsoever.

And that’s enough to get a gold-medal-level performance.

So the transition is already happening.

So are humans going to be the bottleneck for math research?

Well, the answer is yes, but only if we refuse to change.

Only if we insist on being the only thinkers

and the only checkers.

But if we're able to realize this 400-year-old vision,

we're not going to replace ourselves,

we're going to elevate ourselves.

We're going to put ourselves in the driver's seat

as the explorers, the architects and the question askers.

And that means that formal mathematics is the key to this new era of discovery,

based on the powerful and essential partnership between human imagination

and mathematical superintelligence.

Thank you.

(Applause)