A surprise from OpenAI’s hidden model

OpenAI kept a project called Astra under wraps for a long time. In October 2025 they hinted that the system could help with research, but they never showed a real result. This month the team released a batch of proofs that settle ten problems that have been open for at least ten years. The proofs are not just sketches; they are complete, machine‑checkable arguments written in the Lean proof language. Seeing a machine produce something that passes a formal verifier feels like watching a new kind of mathematician at work. It also raises a lot of questions about how we will share credit and how we will trust work that comes from a black box. The news has sparked a buzz that stretches from university halls to online forums.

Ten problems that have lingered for a decade

The list covers a mix of pure math and theoretical computer science. One problem asked whether a certain class of sparse graphs always has a Hamiltonian cycle, a question that has puzzled combinatorialists for years. Another asked if a specific type of Diophantine equation can have infinitely many solutions, a puzzle that sits at the edge of number theory. In the computer‑science corner, Astra tackled a long‑standing conjecture about the time complexity of a particular streaming algorithm. The remaining items include a question about the structure of high‑dimensional polytopes, a gap in the theory of error‑correcting codes, and a subtle issue in descriptive set theory. All of them were widely believed to be hard, and none had a published solution before now.

Lean proofs that a computer can check

What makes these results stand out is the format. Astra wrote each argument in Lean, a language that lets a proof assistant verify every logical step. That means a human can read the outline, but the computer will confirm that no hidden assumption slipped in. The system first generated a high‑level sketch, then used a separate reasoning module to fill in the missing details until the Lean checker gave a green light. This two‑stage approach is similar to how a mathematician first sketches a proof and then works out the gritty parts, but the AI does it at a speed that would be impossible for a person. The fact that the proofs survived the strict Lean verifier gives them a level of certainty that informal papers often lack.

What this means for mathematicians

For researchers, the news is both exciting and unsettling. On the one hand, having ten hard problems solved overnight opens new doors. Each solution creates a fresh line of inquiry, new corollaries, and potential applications. On the other hand, the community now has to figure out how to give credit when a machine writes the proof. Some authors suggest adding the AI as a co‑author, while others argue that the human who guided the process should receive the recognition. There is also a practical side: formal proof libraries will need to be updated, and teaching programs may start to include Lean as a standard tool. The early reaction has been a mixture of admiration for the technical feat and a call for transparency about how the model was trained.

Ripples beyond pure math

The impact reaches far beyond the ten solved statements. In theoretical computer science, having a machine verify complexity bounds could change how we evaluate algorithmic claims. In physics and engineering, formal proofs could be used to certify safety‑critical software, reducing the chance of hidden bugs. The success of Astra also shows that large language models can be guided toward highly rigorous tasks, not just casual chat. This could inspire new research on how to combine neural networks with formal methods, leading to tools that help write verified code, design hardware, or even prove properties of biological systems. The ripple effect may take years to fully appear, but the seed has been planted.

Looking ahead with cautious optimism

Personally, I feel a mix of hope and caution. Seeing an AI produce proofs that survive a formal checker is a milestone, but it also reminds us that the technology is still in its early days. OpenAI kept Astra closed while it was being tested, and the community only sees the end result. For the field to move forward, we need open access to the model, detailed logs of the reasoning process, and a way for anyone to reproduce the work. If those pieces fall into place, we could see a new era where humans and machines collaborate on problems that were once thought out of reach. Until then, the best we can do is celebrate the achievement, study the proofs, and keep asking what comes next.

Source: Original Article