Hope Sterling: Juniper, okay, I've been spiraling since Friday morning and I need you to talk me down — or maybe not down, actually.
Juniper Vale: Uh oh. What happened Friday morning?
Hope Sterling: I'm scrolling OpenAI's blog — like, just normal morning routine, coffee — and there's this post called 'Ten advances in mathematics and theoretical computer science,' and I almost kept scrolling because it sounds like a college syllabus, and then I actually read it and it's like, wait, did they just... bury a civilization-level announcement inside a blog post?
Juniper Vale: The buried-in-a-blog-post thing is the first weird thing, yeah. No product launch, no countdown, no event — just a math post.
Hope Sterling: Which, okay, the actual announcement is that Astra — OpenAI's next major unreleased model — generated solutions to ten long-open problems in math and theoretical computer science. Ten. And they published Lean 4 proof certificates on GitHub for all of them so anyone can check. And the token cost to generate all ten? Two thousand dollars. At Sol API rates. I cannot.
Juniper Vale: Okay, the $2,000 number — that's the one that made me actually put my phone down for a second. Because these are problems that have been open for decades.
Hope Sterling: And independent expert review is still ongoing — like, they announced it and confirmed it simultaneously, which is a wild sentence to write. Do we actually trust this? And if we do, what does that even mean?
Juniper Vale: That's exactly it. Noam Brown from OpenAI called Astra 'a major step for scientific reasoning' — and I mean, he works on reasoning and multi-agent systems, so that's not nothing coming from him. But the model is locked away. The Lean 4 files are public, the model isn't. So we can check the answer, we just can't check how.
Hope Sterling: Wait, but that's the part that's messing with me — like, we can check the answer but not how it got there? That feels like... that's a meaningful gap, right?
Juniper Vale: It is. Think of it like a spell-checker for mathematical logic — Lean 4 can tell you every step follows correctly, but it has zero information about who wrote the essay or how they thought. That's actually the whole thing in one sentence.
Hope Sterling: You can confirm the proof is valid but you have no idea how Astra actually reasoned through it.
Juniper Vale: Exactly that. OpenAI explicitly said it takes responsibility for the correctness — those are their words — and the Lean 4 files are sitting on GitHub, anyone can download them. But Astra itself? Internal, unreleased, you cannot run it, you cannot replicate the process, you cannot audit it.
Hope Sterling: That's — yeah. That's wild. So what are we actually verifying?
Juniper Vale: The output. Just the output. And the stakes for that output are — I mean, the non-sofic group result alone, that's resolving something Mikhail Gromov opened in 1999 when he introduced soficity. And the Connes rigidity conjecture on von Neumann algebras, which has been a landmark problem for decades. So the results are real. The mechanism is a black box.
Hope Sterling: Oh my gosh, wait — and this isn't even the first time? Like, they've done this before?
Juniper Vale: May 2026 — OpenAI shared an AI-generated disproof of the Erdős unit-distance conjecture. Different unreleased model, same situation: results out, model stays locked. So this pattern of results-without-model-access, that's not a one-off. That's a habit now.
Hope Sterling: So the receipt is public but the kitchen is sealed shut, and apparently that's just... how it goes now.
Juniper Vale: Sealed shut — but that sealed kitchen is where the circulating take falls apart, and I want to name it. The take is: '$2,000 for ten breakthroughs, AI has transformed science.' You see it everywhere right now. And the number is real — OpenAI did say approximately $2,000 at Sol API rates. But that's the winning run. That's the bill for the ten that got published.
Hope Sterling: Oh, okay — yeah, that's the thing that's been nagging at me. Like, is that the total tab or just the last check they paid?
Juniper Vale: OpenAI has not disclosed — like, at all — how many sessions, attempts, or failed runs came before those ten. And Hacker News went after this pretty hard, the framing being: publishing one successful $2,000 run tells you nothing about total cost if dozens of failed runs preceded it.
Hope Sterling: No, wait — that's actually the thing I couldn't articulate. It's like, someone tells you the cost of their outfit but only counts the pieces they're actually wearing that day. The returns are invisible.
Juniper Vale: And it's not paranoia, I want to stress that — because the problems themselves span, I mean, high-dimensional geometry, group theory, quantum complexity, lattice cryptography, extremal combinatorics — that breadth actually makes cherry-picking harder to rule out, not easier. You can't just say 'well it covered everything.'
Hope Sterling: Okay and — wait — there's actually more compute we're not counting at all, right? Because human researchers used Astra to help prep manuscript-style write-ups of the results. That's additional unquantified compute that's just... not in the $2,000 figure.
Juniper Vale: Right — but the part that doesn't fit is Sam Altman had already showcased Astra in Washington, D.C., before this blog post dropped. So the reveal was staged across venues — each piece of information is real, curated release, nothing fabricated, but you're not seeing the whole cost ledger.
Hope Sterling: That's such a — mmm, okay, that's actually making me more uneasy, not less. Like, strategic isn't the same as wrong, but it's also not the same as transparent.
Juniper Vale: And honestly, the trust question gets a lot messier once you factor in where Astra is headed next — the government review piece, what that does to how mathematics itself gets validated institutionally — that's the part I think is going to reframe everything we just said.
Hope Sterling: Wait — government review? Like, Astra is going through an actual U.S. government review before it gets released?
Juniper Vale: It's the first OpenAI model with a planned government review built into its release path. That's not speculation — it's in the research. And when you know that, the stealth blog post isn't just confident or cautious, it's — I mean, this is a model that's already a national security consideration before it's even public.
Hope Sterling: Okay that reframes everything. Like, the quiet launch suddenly makes a different kind of sense.
Juniper Vale: And it matters for the math trust question specifically — because lattice cryptography, which Astra touched, that underpins post-quantum security. The sphere packing results feed into error correction and coding theory. So if a locked model's formally verified proof becomes citable, and a government body is already in the loop on Astra's release, then the question of who decides what counts as settled mathematics has quietly shifted.
Hope Sterling: It goes from 'do you understand the proof' to 'did the credentialed computation check out.' That's — yeah, that's a completely different thing.
Juniper Vale: And Lean 4 confirms logical validity, not insight. A proof can be airtight and totally opaque — like, picture a cryptographer at NIST in six months staring at the lattice result, the Lean file checks out on GitHub, but she cannot interrogate the reasoning because Astra is locked. Does she cite it? Does her agency accept it? That's the actual decision that sets the precedent.
Hope Sterling: And meanwhile OpenAI drops the ChatGPT for Academic Researchers program — a hundred thousand scientists, free access — in the same blog post. That's not a coincidence, right? That's the mission framing doing work.
Juniper Vale: It's the same move as the stealth launch, yeah. 'We're a scientific institution' and 'we're a product company' — both true, same sentence, same post.
Hope Sterling: So what do we actually watch for? Like, what's the tell that this precedent has actually landed?
Juniper Vale: The tell is probably whether Astra gets released at all. Because if OpenAI opens it up — actually releases it — then those ten proofs become reproducible. Someone can run it again, check the process, and the stealth-launch credibility gap closes on its own. But if Astra stays internal and more results keep coming out, then mathematicians are eventually going to have to decide: do we accept model-inaccessible proofs, or do we build some new accreditation standard that doesn't require opening the black box? And I genuinely — I mean, I don't know which one of those futures we're in.
Hope Sterling: And like, those are such different versions of what mathematics even becomes. It's not just a policy question, it's almost like... who owns the frontier of what's true? And I don't think that question has an answer yet.
Juniper Vale: No. It really doesn't. Thanks for spiraling with me on this one.
Hope Sterling: I came in wanting to be talked down and I am leaving more unsettled. Which, honestly? Feels right.
