10,000 AI Agents Attack One Problem

Brian (0:00)

Hey, everybody, welcome to The Daily AI Show. Today is September 9th, 2026 And with me today are Andy and Beth. And boy, oh boy, do we have some news topics today to talk about. We have math problem drama at OpenAI, and we're going to get into that. We have US law schools restricting AI. We have Meta Muse, the app. I downloaded that and played with it. I can talk about that a little bit.

We have a new image, a better image generator out of OpenAI. I downloaded it, but we're not downloading it, but I used to play that, Image 2.5.

We have all sorts of stuff, stuff about jobs and, oh my goodness, there's so much going on today. So we're going to get into all those things.

Oh, and Beth is also saying OpenAI is thinking about pausing pro subs or just trolling. Is TiVo just trolling us? We'll talk about that as well. So I feel like maybe one of the biggest stories coming out of yesterday, at least, you would think it would have been Muse or the image generator from OpenAI, but no, it's math. We have drama in the solving 90-year-old math problems. So there's interesting parts to this, but do either one of you all want to jump in here and talk about what we know so far? It's an interesting one.

Beth (1:25)

It's a really interesting one, and I will say that there were hints that a story was coming. This is the proof that OpenAI announced yesterday was rumored to be coming.

Do you guys have the name of the, it's like Navient something, Salient Navient?

Brian (1:52)

I could certainly find it while you're talking.

Andy (1:54)

Navier-Stokes.

Beth (1:55)

Navier-Stokes. Okay. Before that, it's 3D Euler, maybe, is related to the Navier-Stokes.

Andy (2:04)

There are seven one million dollar Millennium Prize math problems.

Karl (2:09)

Yes.

Andy (2:10)

Several of them have now been solved, earning a million dollars for those who solved them.

Karl (2:17)

Only one.

Andy (2:17)

There's many open still.

Beth (2:21)

Because I thought only one had been solved before, and it was in 2003 by a guy not with AI, just doing his own thing.

Yeah. I don't know. 2003, man, like a psychopath, just doing it with the fingers of paper.

Andy (2:39)

I don't know.

Beth (2:40)

Yeah. Anyway, these two people who, Andy, do you have their name? I'm sorry. Yeah.

Andy (2:50)

Tristan Buckmaster, and that's from NYU. This is the mathematician and an anthropic of a researcher named Levent Alpöge.

Beth (3:01)

Right.

We've already started to unpack the T because the fact that Levent is at Anthropic is part of the sticking point here. The story coming out of OpenAI is that they heard the rumor, but not the specific rumor within their training data or their user records because, of course, these two mathematicians or the mathematician and the Anthropic guy were actually doing this in Codex. So the files are all user files in OpenAI. OpenAI says they don't access those.

The mathematician said, did you train on our data? And OpenAI was silent. We don't access user data when we're doing things, but no answer about training. The OpenAI side says, look, we were trying to do this in good faith. We reached out. We asked people to come. We asked Buckmaster to come to a phone call. Apparently Levent was invited to that and didn't come, but ultimately OpenAI's position was we don't see a way for us to announce something that includes an Anthropic researcher or gives the Anthropic researchers access to our internal models. Because what created this result, what proved this result was 10,000 internal coordinated agents.

Brian (4:50)

Using a model that's smarter and more powerful than Astra, that does not exist publicly. So better than Astra and 10,000 agents. That's a powerful one-two punch.

Beth (5:00)

So Buckmaster was offered, according to both of them, Buckmaster was offered co-headlining on the proof. They offered because Buckmaster and Levent did solve 3D EULA.

It's like a stick of hearts and egg McMuffin. I don't know. It didn't have as much meaning for me in some ways. But 3D EULA was like a jumping off point for the math, maybe.

So OpenAI offered like, you could publish that and then we'll publish our thing. You can come on. We give you access to the model, but it can't be with Levent because they're Anthropic. Buckmaster might hold this up with many of the same facts but completely different interpretation. Yeah. And said, if you go forward with this, I will go public with this part of the story.

Karl (6:07)

Yeah.

Beth (6:07)