Brian Reed: Hey. Did you see the Mathathon thing?
Eliza Ward: The Caltech one. Yeah, just read it.
Brian Reed: So Sathvik Redrouthu — Caltech math student, former Y Combinator founder — posts on August 30th that he's organizing this hackathon. October 9 through 11. First one, he says, exclusively devoted to research-level mathematics. A hundred teams, forty hours, frontier AI models. And the headline number is a million dollars plus in compute credits.
Eliza Ward: Right — and the participant pool he's describing is IMO gold medalists, PhDs, researchers from OpenAI. Which — wait, actually — that's where I stopped.
Brian Reed: Where you stopped meaning — none of them have confirmed?
Eliza Ward: Zero named individuals from any of those groups had publicly said they're attending. As of the announcement date. The post got 103,000 views, 500-plus likes — so it moved. But the participant pool is, so far, a description of who he wants.
Brian Reed: So the gap between what's claimed and what's confirmed — that's the actual story here.
Eliza Ward: That's exactly where I'd want to start, yeah.
Brian Reed: But hang on — before we get to whether the people are real, what even is this thing? Like, what does a research-level math hackathon actually mean?
Eliza Ward: Okay, here's the plain version. Imagine a PhD student at her kitchen table at midnight. Frontier AI model open on one screen, a list of unsolved mathematical conjectures on the other. Thirty-nine hours left on a clock, a team scattered across three time zones. That's the Mathathon. Not building an app. Not solving a problem with a known answer. Actually trying to move mathematics forward — using the AI as a research tool.
Brian Reed: So it's not a competition in the IMO sense — where there's a correct answer waiting.
Eliza Ward: Right. That's actually the — wait, that's the precise distinction. The International Mathematical Olympiad has fixed problems with known solutions. The Mathathon is targeting open conjectures. Nobody knows the answer. Which is either genuinely exciting or a very good way to make a hackathon unverifiable.
Brian Reed: So what did the social response actually confirm, then? Because 103,000 views is a number, Tony Yue Yu posting is a name — but I want to know what those signals actually mean.
Eliza Ward: Tony Yue Yu is a Caltech math professor — that's a faculty endorsement, not a student hype post. His reply got 70 likes, he mentioned Caltech undergrads already preparing. That's real institutional warmth. But — and this is the part that stuck with me — Jason Rute from Mistral AI also replied, and his reply wasn't enthusiasm. He asked how problems would be recorded and whether transcripts would be made available to the broader community.
Brian Reed: That reads like due diligence, not a yes.
Eliza Ward: Exactly. An engaged researcher — from Mistral AI, so someone who works in this space — wanted to know what would be documented before committing to anything. And the compute offer, the million-plus in credits, the specific labs providing it are still not publicly named. So you've got real interest, a credible faculty endorsement, genuine engagement from someone at Mistral — and underneath all of it, no named attendees and no named funders.
Brian Reed: So the signal is real. The roster — that's still a projection.
Eliza Ward: And the projection part is where the dominant take actually falls apart. The thing already circulating is: a million in compute equals serious people are coming. That's the implicit logic. And I don't think it holds.
Brian Reed: Because the compute offer itself — who's actually putting that up?
Eliza Ward: Unnamed labs. Unnamed startups. The specific providers aren't publicly named in any available source. So you can't verify the number, you can't verify the commitment. It's a budget line, not a signed agreement anyone can read.
Brian Reed: Right — but let me push on the deeper assumption, because even if the million were fully verified, I'm not sure it would do what Redrouthu thinks it does. An IMO gold medalist, an OpenAI researcher — they're not scanning hackathon announcements for a good GPU allocation. That's not how they decide what's worth their time.
Eliza Ward: No — and actually, wait — this is the inversion. Compute follows credibility in that world. It doesn't create it.
Brian Reed: So then the organizer's profile matters a lot here. And Redrouthu's profile is — I mean, it's genuinely unusual. Instachip through Y Combinator, Procyon Photonics, Etched, now a Caltech math graduate student. That's a real trajectory. But he's a grad student, not a faculty director, not a lab head. And that raises a fair question about, like — what's the institutional standing to deliver commitments from unnamed labs?
Eliza Ward: Tony Yue Yu's endorsement helps. A Caltech math professor publicly backing this is not nothing. But endorsing the vision isn't the same as underwriting the participant list.
Brian Reed: And the 'first math hackathon' label — that's his framing, not a verified claim. Which is completely consistent with how you'd launch a startup, and completely inconsistent with how mathematics culture actually works. There's a version of this where what actually happened in October tells a very different story than the August announcement — and that's what I want to get to.
Eliza Ward: And that's actually what makes the AI capability piece worth separating out — because the premise isn't invented. OpenAI o1 hit gold-medal-level performance on IMO problems. July 2025. That's confirmed.
Brian Reed: Wait — gold medal level. On actual IMO problems.
Eliza Ward: Reported by OpenAI, yes. And AlphaGeometry — the neuro-symbolic model — solved 25 out of 30 IMO geometry problems, which outperformed every prior automated theorem-proving baseline. So when Redrouthu says frontier AI can work on serious mathematics, that's not wishful thinking. The capability is there.
Brian Reed: Okay but — problems have answers. Someone already knows whether the solution is right. Open conjectures don't work that way. So I'm not sure the o1 result actually gets you to 'a 40-hour sprint can advance research math.'
Eliza Ward: No, it doesn't. That's — wait, that's exactly the line. Competition math and research math are structurally different problems. One has a finish line. The other, you don't even know what the finish line looks like.
Brian Reed: Which is where Jason Rute's question actually becomes the concrete test, right? He asked about transcripts, documentation, community access. If the Mathathon produces recorded problem-solving sessions — actual outputs the community can examine — that's a path to legitimacy. If it doesn't, I mean, what did you actually produce in 40 hours?
Eliza Ward: A demonstration. A well-funded one, maybe. But peer review doesn't run on a 40-hour clock — it runs on months. So the question for October 9th isn't 'did teams make progress?' It's whether any output clears the bar Rute was pointing at: documented, accessible, something a mathematician sitting outside Caltech can actually evaluate.
Brian Reed: So the thing to actually watch — it's not the participant list at this point. It's whether there are transcripts on November 1st.
Eliza Ward: That's the signal. Named participants, compute delivery, and documented outputs by the time October 11th closes. Any one of those missing and the gap between the August announcement and the actual event tells you exactly what kind of event this was.
Brian Reed: So where I land — and I'm genuinely uncertain about this part — is that the question isn't really whether Redrouthu can bootstrap credibility through momentum. It's whether named participants confirm before October 9th. Because that's the one thing that separates 'announcement that became real' from 'announcement that stayed one.'
Eliza Ward: Yeah. That's the cleanest version of it.
Brian Reed: And I keep wondering — I mean, math culture doesn't work the way Silicon Valley does. You don't get credit for the conjecture you're about to prove. You get credit when the proof exists and someone else has checked it. So can you even bootstrap into that world using announcement logic? Or does the community just wait and see what October 11th actually produces?
Eliza Ward: October 9th answers it more cleanly than anything we say right now. Either IMO gold medalists and OpenAI researchers show up — publicly, by name — or the gap between the August 30th post and the actual event is the whole story. We don't have to speculate. The date's right there.
Brian Reed: Then that's what we're watching for. Named confirmations before the event starts.
