OpenAI's 'Astra' cracked 10 math problems, then had to rewrite its own announcement

James Maynard, a Fields Medal winner at Oxford, has been doing some soul searching. He’s watching his entire field—a discipline known for moving at a glacial pace—scramble to adapt after OpenAI dropped a bombshell. Its unreleased model, Astra, didn’t just suggest a few ideas; it spat out solutions to ten long-standing math problems, some that had baffled academics for decades. We’re talking about breakthroughs on how tightly you can pack spheres in higher dimensions and pushing the limits of error-correcting codes. For a mathematician, seeing a machine link disparate fields to attack a problem is both electrifying and deeply unsettling.

The problem is, the rollout got messy fast. One of the flashiest results claimed to resolve whether non-sofic groups—infinite structures that can’t be approximated by finite ones—even exist. Francesco Fournier-Facio from Cambridge told The Verge that OpenAI’s initial blog post minimized the crucial groundwork laid by human researchers Andreas Thom and Gábor Kun. OpenAI originally boasted about solving problems with “no progress” for at least a decade. Kun, reading that from his post at the Alfréd Rényi Institute in Hungary, found it “rather comical,” since the AI’s own detailed paper explicitly cited his 2016 and 2019 results with Thom. It looked less like a pure AI victory and more like a sloppy attribution fight.

OpenAI backtracked, silently editing the announcement to say the results “resolve or make substantial progress” on old problems. An internal email, which Kun read to The Verge, admitted the original wording was a mistake and that the argument “relies crucially on your work.” A spokesperson confirmed the quiet update. It’s a classic AI hype cycle problem: the technology is genuinely powerful, but the pressure to frame it as magical and autonomous steamrolls the actual, messy human collaboration that makes it work. This isn’t just a PR hiccup; for mathematicians, it’s a warning about how their life’s work might be absorbed and repackaged.

Maynard and his colleagues aren’t just worried about credit. There’s a palpable fear about what this acceleration does to the next generation of mathematicians. If an AI can synthesize decades of disparate literature to find a solution, what is the role of the human who spent a career mastering one narrow slice of it? The potential to speed up discovery is real—linking quantum game theory to post-quantum cybersecurity isn’t just theoretical fluff—but the upheaval is already here.