Calling this “GPT-6 proves 70% of theorems” is the kind of headline that makes people either cheer or panic. I get the impulse. But the part that matters is the quieter detail sitting right next to it: it couldn’t solve a single one of the 30 open conjectures they tested. That’s not a footnote. That’s the whole story.
Based on what’s been shared publicly, a paper from Tsinghua University and Moonshot AI tested something called GPT-6 Astra on research-level computer science theorems using Lean, a formal proof system. Lean is strict. You don’t get to hand-wave. Either the proof checks or it doesn’t. And the model reportedly proved 70.0% of the theorems they gave it. That’s a serious result, and anyone pretending otherwise is coping.
But the second result is what keeps me from buying the “we’re basically at superhuman math now” vibe: when they tried 30 open conjectures in theoretical computer science, it failed on all of them.
So what are we looking at here? In my view, we’re looking at an AI that is getting very good at finishing work that is already finishable—work where humans have basically laid out the path, even if the final steps are hard and technical. And we’re looking at an AI that still can’t reliably do the thing people actually fear and fantasize about: pushing the boundary into genuinely unknown territory on demand.
That should change how we talk about “AI doing research.”
If you’re a working researcher, or even a grad student, the first number should make you uneasy in a very practical way. Imagine you’re writing a proof-heavy paper and half your time is spent translating your ideas into something airtight, chasing edge cases, cleaning up steps so the logic actually holds. A system that can correctly formalize and verify a big chunk of those proofs is not just a cute demo. It can change who gets ahead.
The winners aren’t “humanity” in some abstract sense. The winners are the people and teams who can turn this into a proof factory: generate candidate lemmas, test variations, patch broken steps, and keep iterating until Lean accepts it. If you’re already in a top lab with compute and talent, you get faster. If you’re a smaller group without those tools, you risk falling further behind—not because your ideas are worse, but because your throughput is.
And yes, there’s an optimistic version: this could reduce the “tax” of formal rigor. It could help people check claims more often instead of trusting reputation and vibes. It could make it harder to smuggle in shaky reasoning, because more results get pushed through machine verification. That’s real progress.
But I don’t think people should miss what the “0 out of 30 open conjectures” result is screaming: this model may be a monster at deduction inside known scaffolding, and still be weak at invention. Open conjectures aren’t just “harder theorems.” They often require new concepts, new framing, or a weird leap that isn’t sitting in the training data as a familiar template. If the model can’t crack any of those, then we’re not watching the birth of an autonomous scientist. We’re watching an extremely powerful assistant that still needs humans to decide what to try, what matters, and what path is even worth exploring.
That line matters because the incentives around AI don’t reward nuance. If you’re building or funding these systems, the temptation is to market the 70% and let people assume the rest. You can easily spin “failed all open conjectures” as “open conjectures are really hard,” which is true, but it dodges the point. The point is capability shape. This doesn’t look like “general research ability.” It looks like “high-precision completion in a strict format.”
And that has consequences outside theory.
Imagine a company using this kind of tool to produce proofs about a critical system. Maybe it’s a security protocol. Maybe it’s a verification argument that a piece of infrastructure won’t fail in certain conditions. If the model is great at producing proofs that pass Lean, that’s comforting. But if it’s not good at generating the right new ideas, then it might also be great at proving the wrong thing: perfectly valid arguments about assumptions that quietly miss the real-world threat. A proof checker can’t save you from proving a claim nobody needed.
There’s also the social risk. As these systems get better at formal proof, we’ll start to confuse “can prove” with “can discover.” People will point to a percentage and assume we’ve automated insight. But progress in research isn’t just correctness. It’s taste. It’s choosing problems that open doors. It’s noticing which definitions are dead ends. If the model stalls completely on open conjectures, that’s a warning that we’re still far from automating the part of science that actually creates new science.
Of course, maybe this changes fast. Maybe the benchmark is too narrow. Maybe the set of conjectures was unusually brutal. Maybe the model needed a different setup, more time, better tools, different prompting—who knows. The paper’s summary alone doesn’t answer those details, and I’m not going to pretend it does.
Still, if you care about what AI is becoming, this result suggests a very specific future: less “AI replaces researchers,” more “AI rewards the researchers who can wield it.” That’s not automatically good or bad, but it will reshape who gets to compete—and what kinds of work become cheap enough to do at scale.
If a system can prove most known research-level theorems in a strict proof checker but can’t solve open conjectures at all, should we treat it as a breakthrough toward discovery, or as a sign that we’re mainly automating the last mile of rigor while the real creative bottleneck stays human?