Terence Tao: “The Age of Heroes” in Math Is Over

https://www.youtube.com/watch?v=7U_fgxzgSrY

无作者

Intro

Terence Tao is standing in front of a blackboard.He’s probably the most famous mathematician alive.He won the Fields Medal at 31.Brady Haran from Numberphile, asks him about something that happened earlier this month.OpenAI dropped 722 math papers on the GitHub in one go.All of them written by an AI.Brady wants to know what we’re supposed to call that.Tao doesn’t have a name for it.He has a verdict though.>> We are seeing the end of the romantic era of mathematics, the age of heroes.>> The age of heroes is one person alone on one problem for years.Andrew Wiles spent about seven years, mostly in secret,on Fermat’s Last Theorem.Tao says that kind of math is ending.Okay, but a famous mathematician being gloomy about AI isn’t news.Except Tao isn’t against it.He’s used these tools for years and he’s said plenty of good things about them.So, what happened?He explained it with three pictures.Some food, some fish,and a marathon.And at the end,

he named the one AI result in that pile he thinks is genuinely clever.

Picture 1: the food

For thousands of years, a proof was hard to get.You needed years of training.You needed to know what everyone before you had tried.And you needed some luck.So, every proof came with something extra.Whoever found it understood why it was true, and they passed that on to everyone else.Tao calls that proof scarcity and he says we’ve left it behind.His picture is food.In the old days, there wasn’t much of it,but every bite was good for you.Now there’s a mountain of it and it’s synthetic.You can eat it, it’s just not the same.>> And now we’re drowning in proofs,but we’re not getting all the understanding um that we used to get.>> That’s the problem with a pile of 700 papers.A proof tells you that something is true.It doesn’t have to tell you why.Tao says he’s only had time to look at about 10 of them.In math, he says,maybe that’s survivable.If an AI settles your favorite problem and nobody understands how,mathematicians get annoyed and the world carries on.

But the same kind of tools are heading for weather models and the power grid.>> [music] >> If they keep handing us answers nobody understands,at some point that’s a real problem.

Picture 2: the fish

The second picture is about students.Math has an unwritten rule.Some problems are medium hard, a senior mathematician could probably crack one,but they leave it alone.It’s kept for a student, because a first paper on a problem like that is how a young mathematician, gets started.That only worked because there were far more problems than students.Tao compares it to fishing back when the sea was full.Nobody needed to split up the water.There were enough fish for everyone.Then the boats changed.>> Now we have this sort of strip mining of mathematics,

where we have these huge machines that can sort of uh, you know,like like dynamite fish, like like hundreds of problems at a time.>> Hundreds at a time.Tao says people are getting reluctant, to share their favorite problems now, because someone with an AI prompt will just grab them.Paul Erdős spent his whole life doing the opposite.He handed problems out to anyone who tried.When OpenAI posted its pile,

nobody checked with the people already working on those problems.>> There’s a lot of juniors, uh research projects that just got wiped out overnight, by this.>> Normally, he says,“If you know a colleague is on the same problem,you quietly check first so you don’t collide.Nobody did that here.” So how does Tao describe it?>> It was extremely rude and and discourteous.>> Mathematicians are still going through that pile and it’ll take months.If you want to hear what they find, subscribe and I’ll cover it when it lands.

Picture 3: the marathon

The third picture explains why this is happening now.Until about 2025, Tao says,“The AI companies needed mathematicians.The tools weren’t strong enough to do math on their own,so the two sides worked together.More or less, everybody was happy.Now the tools are strong enough and the companies want something else.More theorems, more press releases.” So here’s the marathon.Nobody writes a rule that says you can’t drive a car along the course,you could.You’d win.Everyone just understands that you don’t.Someone drove the car.And after the race >> And no one from Open AI is helping us.Um they just They took it as far as they concerned they’re done.>> At a lecture at Caltech on October 9th, his slides were even blunter.

slide

[C] a small piece of film held in a frame that can be shown on a screen when you shine a light through it 幻灯片 SYN transparency
a talk with colour slides

blunt → blunter

characterized by directness in manner or speech; without subtlety or evasion
blunt talking and straight shooting
a blunt New England farmer

Further blind optimization of problem-solving, alone is now actively harmful to the long-term health of mathematics.

The result Tao rates

But Tao is fair about it.Brady asked him if the Riemann hypothesis is about to fall.It’s the most famous open problem in math with a million dollars on it.Here’s the picture.The Riemann hypothesis, says the zeta function’s important zeros all sit on one line at 1/2.For over a century,the best anyone could prove was that they stay away from the line at one.And that safe zone got thinner the higher you went.One of Open AI’s papers says no zero goes past 7/8.A straight wall all the way up.Tao’s verdict on that one?>> Biggest advance on the subject in 50 years, I’d say.>> 7/8 is a quarter of the way from one to a half.Tao says it took three or four old ideas, put together in a way he doesn’t think any single mathematician, alive would have found.He also says the method can’t get past 3/4, without another very [snorts] clever idea.>> It’s It’s been dented um quite a bit.>> So that same pile holds a result Tao calls the biggest in 50 years.It also had three papers withdrawn in the first two days.

and hundreds that nobody has properly read.He expects good stuff and rubbish in there and nobody knows the percentages yet.

What to keep

So what does Tao want?Not a ban.He thinks math will end up like the other sciences, with big teams working on hundreds of problems at once.So it matters less who gets scooped.

scoop

~ sth(informal) to win sth, especially a large sum of money or a prize 获取,赢得(一大笔钱或丰厚的奖品)
He scooped £10 000 on the lottery.
他中彩得了 1 万英镑。

And he wants students to learn the old way first.The way kids still do arithmetic by hand before they get a calculator.On October 15th,the Simons Foundation announced $15 million to help mathematicians, check and use AI results.Two of the three backers are Anthropic and OpenAI.Dow says math is walking into this new era rather unwillingly.>> Exciting new things are possible, but some things are being lost.>> He thinks the age of heroes ends either way.What he’s asking is which parts are worth keeping.If a prompt could hand you the proof,what would you still want to work out yourself?

Outro

On this channel, I read the papers and reports behind the AI math news.If that’s useful to you, subscribe.There’s also a free weekly email, about what AI did in math that, week and what still needs checking.The link’s in the description.Thanks for watching.Bye.