Fields Medalist Terence Tao explains why AI can’t be trusted on its own, and why that doesn’t make it useless. He shares how mathematicians are already using large language models to surface forgotten approaches, why AI’s real power lies in tackling millions of medium-difficulty problems rather than the famous ones, and how education must shift toward teaching students to verify and critique what they see. Plus, Terry and I debate whether math is truly a language and what it would take for AI to police itself.
🎯 Support Into the Impossible on Patreon — get my weekly M.A.G.I.C. Message, unfiltered bonus content, and live monthly Office Hours with me: https://www.patreon.com/drbriankeating
The discussion explores the intersection of AI, mathematics, and education. The speaker emphasizes that while AI is currently prone to hallucinations and lacks grounded reasoning, it serves as a powerful complement to human experts when paired with rigorous verification. The conversation also touches upon the nature of mathematics as a language of efficiency and the need to evolve educational methods to prioritize critical thinking over passive knowledge acquisition.
I think as long as you pair these AIs
with good verification, you only use the AIs to
the extent that you can verify the outputs and no further.
I see them more as complementing human scientists and mathematicians.
Has AI actually enabled new discoveries in mathematics
or new proofs that otherwise would
not have existed?
Slowly, it's beginning to.
I mean, by itself, the big weakness
of these AIs right now is that they can begin to produce
output that looks like, say,
a human mathematician reasoning their way through a problem, but
it's not grounded.
they often make mistakes and much like you know
if i were to get a student to solve a problem on
a blackboard and they're nervous
and they just say the first thing comes to mind they might get it
right or they might get it wrong
but if it's a weak student and they don't have fundamental
knowledge of of what they're actually doing once they go off
the rails they can go really off the
rails and this is something which is a fundamental problem
with the current large language models but
if you use them as a component of rigorous
and grounded reasoning systems so if you if you
converse with a large language model for them
to make suggestions but you've you understand
with the output and you can verify it.
So then people have had some success
talking about their math problems
over a large-language model.
One will produce some suggestions,
some of which the human expert
can dismiss as not viable.
And some of us would be thinking,
oh, I thought of that already.
But one or two is,
oh, that's actually something
which I should have known,
I should have come up with myself,
but I just didn't realize.
So one thing where AI models
are already beginning to be useful
is in literature review type tasks
where there is a classic problem
and in the literature,
there are maybe, say,
a dozen ways already
to attack this problem.
And if you're even working on the problem,
maybe you can remember six of them,
but you forgot the other six
and it don't come to your mind.
And you can use these models to prompt you,
remind you of the missing six.
They may also hallucinate for you more
that they don't exist.
So you do need, you can't trust them.
Supervise them.
Yeah, you have to verify.
There is hope in the future.
You know, so there's this separate math technology
to software that can automatically verify
certain types of proofs.
We mentioned, you know, inevitably,
when we talk about LLMs,
you know, the middle L is language.
We'll get to my friend Galileo.
I brought a book I want you to get your impression on,
one of his math books and treatises. But he
said, you know, that the book of knowledge of nature,
the universe, is written in the language
of mathematics. And it kind of, you know,
was echoed later, you know, by Wigner and so forth,
as we already discussed. But is it really a language?
Yes, it has a vocabulary and it has
a syntax, you know, but in the same text,
you know, Shakespeare and math, if they're truly,
you know, at root some proto-language
or something like that, then they should have
more combinations or similarities, I would think.
But again, I want your opinion. Do you think of
math as a language or is it much more than that?
Well, certainly when mathematicians talk to each
other or to other scientists, I mean,
they have to use math as a language. I think the difference
between mathematical language
and natural language is that language sort of has evolved over time
to describe the underlying mathematics
as efficiently as possible.
Language, natural language,
is not always about efficiency.
I mean, you also want to convey nuance
and emotion and art
or just express frustration or whatever.
So it isn't driven purely by efficiency.
But mathematics pretty much is
partly because over time
we try to do more and more
ambitious mathematical tasks.
And if we didn't use our math language
in this fashion,
we would not be able to do
these more complicated tasks.
And the same is true in the sciences.
We keep updating our laws of nature
so that we can make more complex predictions.
When you optimize a language for efficiency,
you're basically just trying to compress
a description of the universe
into as minimal and elegant a form as possible.
And so when you're doing that,
you are somehow getting to the essence
of how the universe actually works.
So presumably the universe does operate
by some laws of nature,
which maybe you don't know yet,
but we'd like to believe
that these are simple, predictable laws
and it isn't just some big chaotic,
there isn't some agent
that's just making things up
as they go along.
And the whole history of science
has been sort of validating
that belief,
that naturalistic philosophy.
And mathematics has been trying
to do the same thing
to mathematical theories,
trying to find the most elegant,
minimal inputs
that would explain
lots and lots of mathematical phenomena.
So maybe that's why
they sort of converge over time
and why Wigner also observed
that the types of mathematical language
and formalism
that is good for mathematics,
for example,
the language of curved space to describe all kinds
of geometries happens to coincide quite
well with the language that describes the universe,
like the use of that same language
to describe space-time.
The figure of space, yeah, exactly.
When we look at the future of education,
you're not only a Fields medalist, a mathematician,
a father, and everything else that you do,
but you're a teacher and you're an educator.
Talk to me about your vision for the future.
What's your philosophy of teaching?
Yeah, so it needs to evolve quite a bit for many reasons.
yeah so the world has become infinitely more complex
and unstable and unpredictable and now
you know and and now with AI you know humans used
to be sort of have a monopoly on cognitive tasks
like you know and and now AI so one of the problems
of AI actually i mean the way the subject develops
it's not so much that they overtake human like
research level mathematics or any other discipline
in the near future but already undergraduate level mathematics
for instance many of the
homework assignments that we teach we assign right now they can
be done by by ai yeah so we
have to reinvent the way we we teach
so one thing that will become more important is students will
need to have much more training in how
to validate information that they see you know so in the past
we had like a small number of authoritative source information
or textbooks and your teacher or
something and and you know we didn't have social media
and and and the internet and all kinds of
information of now and now ai was a really variable quality
on the other hand the um in the past when
had information that was low quality in content it
was also low quality in presentation so you
could tell that like a really well-produced
textbook would likely have more accurate content
than then you know something written in crayon
or something but but now our ability to produce
high quality presentation has far outpaced
our ability to produce high quality content
so you cannot have have you know youtube videos
or textbooks that look flawless okay and now ai
generated output but have got lots of fundamental mistakes
so yeah we we need to to encourage
critical thinking you know i already see teachers experimenting
with things like you know here is a
question that i would have assigned but i've given it to to chat
and this is this is the answer that
they give it's wrong please critique it i correct it interesting
and these are i think more of the
skills more interactive you know
so not treating knowledge as a passive thing to be acquired by
authority but something that you always have to question
and struggle with interesting you know
That kind of reminds me of John Preskill at Caltech,
talking about quantum computing and quantum supremacy and so forth.
And one of the ways to overcome some
of the issues with error correction
and quantum computing is to throw more qubits at the problem.
And I wonder, will we throw more AIs at the problem?
You know, this kind of flipped it through natural human brains
at an AI to prove where it's wrong.
But will we be at a place where AI could police itself?
And so what would it take to trust them?
It's good to make them more reliable.
But I think, well, maybe if we use
a very different architecture from the current AI.
so by nature
they are inherently
unreliable
but we have ways
to use unreliable tools
random number generators
are the most
unreliable
device technology
we have
but they're extremely
useful for all kinds
of things
I think as long
as you pair
these AIs
with good verification
and you only
use the AIs
to the extent
that you can
verify the outputs
and no further
then they can be
a great tool
I see them
more as complementing
human scientists
and mathematicians
so because there are
so few human scientists
and we don't only have
so much time
to work on research,
we tend to focus
on sort of high value,
high priority,
isolated problems.
But in mathematics
and in the sciences,
there are millions
and there's a long tail
of lots and lots
of less well-known problems
which should require
some attention
and they're not
the most difficult
or important,
but it would be good
to have someone
or something look at them.
And so I think AI actually,
their best use case
is not to target them
on the most
high-profile problems,
but actually on the millions
of medium-difficulty problems.
And they may fail and they may only solve 10%
of these million problems, but that's 100,000 problems solved.
So scale is the big advantage.
You cannot scale a graduate student this way.
Not legally.
No, not legally or ethically.
But AI, I think that's where the real value lies.
What's your highest priority task right now?
Well, research-wise, what I'm interested in most nowadays
is new workflows to modernize mathematics
and make it more collaborative, more accessible to the public,
and to integrate in these new tools like AI.
The way we've done mathematics has
not changed fundamentally in centuries.
We have blackboards in my office.
We still work with pen and paper.
We use computers a little bit, but not so much.
And our collaborations are still very small.
We work with two, three people.
In the sciences, of course, we're thousands.
In large part because we don't know how
to incorporate contributions to the general public.
I mean, there's a barrier to entry, first of all.
A lot of what we do is very technical.
But we need to synthesize proofs where
every single step has to be verified.
So if we had thousands of people with a thousand little components,
it wasn't feasible until very recently.
Also, because of all these factors,
we don't collaborate as much with the other sciences as we ought to.
Especially the new sciences, which are so data-driven and connected
in new ways, you know, like social network analysis or whatever.
So that is, I think, the direction in which my research is going.
to it. It's almost more the sociology
of mathematics actually than the technique. And
more recently, I've been interested in trying
to secure funding for my research that has become
very unstable in recent years.
Terry and I went much deeper on this, including whether AI can
actually do original mathematics.
Full conversation is linked right here. I'm Professor Brian Keating.
I don't dumb things down. I'll be talking
to Terry again soon, so subscribe for more.