OpenAI published today a selection of ten advances in mathematics and theoretical computer science that, it says, were found with the help of AI models. What’s behind that headline and why should it matter to you even if you’re not a professional mathematician?
What OpenAI announced
The company explains that, as part of its ChatGPT for Academic Researchers initiative, it granted free access to 100,000 scientists and mathematicians to its best models. In addition, during tests of an internal model called Astra, teams found solutions to open problems that had gone decades without progress.
Among the milestones mentioned are everything from new bounds on high-dimensional sphere packings to constructions of non-sofic groups and counterexamples to classical conjectures. OpenAI says the arguments generated by the model were edited by humans, formalized in Lean, and accompanied by the model’s narrated reasoning.
A curious fact: the total number of tokens used to find these solutions would cost approximately $2,000 at public API rates, according to the note.
The ten results, in plain words
-
High-dimensional sphere packing. New upper bounds on density close to the Cohn–Elkies threshold. Think of arranging oranges in many dimensions — this changes how tightly you can pack them when the room has many more directions than the three you’re used to.
-
Binary and spherical codes. Exponential improvements in bounds on the maximum size of codes with a given minimum distance; analogous results for codes on high-dimensional spheres. This matters for communications and information theory — like making data transmission more robust.
-
Non-sofic groups. A construction that establishes the existence of groups that do not satisfy the property of soficity, resolving a central question in group theory.
-
Connes’ conjecture. A counterexample to a rigidity conjecture in operator algebras: it shows that certain groups are not uniquely determined by their von Neumann algebras.
-
Arithmetic circuit complexity. New lower bounds for computing the permanent with arithmetic circuits and formulas, including an n^4/log n-type bound for formulas.
-
Quantum parallel repetition. An exponential parallel-repetition theorem for two-player quantum games, extending classical theory principles into the quantum realm.
-
Closest vector problem. Hardness of approximation within a polynomial factor for CVP, with implications for lattice-based post-quantum cryptography.
-
Ehrhart volume conjecture. They determine, in each dimension, the maximum volume of a convex body whose centroid is its only interior lattice point.
-
Multicolor Ramsey numbers. A superexponential lower bound for multicolor Ramsey numbers of triangles, resolving Erdős’s problem 183.
-
Extremal graph conjectures. Results that resolve compactness and degeneration problems in extremal graph theory (Erdős 146 and 180).
How these solutions were obtained
According to OpenAI, the solutions emerged during evaluation of an internal version of Astra. The arguments proposed by the model were reviewed and prepared by people, then formalized in Lean to produce formal certificates. The narration of the model’s thought process for each solution is also published.
OpenAI takes responsibility for the correctness of the manuscripts it helped prepare, but states that the original arguments were generated by the system.
Why this matters and what questions it raises
Does this mean AI does mathematics by itself? It isn’t that simple. Here are several points to keep in mind:
-
The AI speeds up exploration: it can propose conjectures and routes that humans then check and turn into formal arguments. Have you ever had a rough idea that needed polishing? Think of the model as that first draft you can iterate on.
-
Formalization in
Leanis an important step: it’s not just a sketch, it’s a verification that can increase confidence in the results — though the mathematical community has the final say. -
There are ethical and attribution questions: how do you name authors when a model provides the main argument? OpenAI argues against attributing exclusive human authorship to proofs generated entirely by AI.
-
Access and fairness: OpenAI emphasizes the need for broad access to these tools so research benefits from human–machine collaboration.
Risks and legitimate reservations
While these achievements sound promising, many in the mathematical and philosophical communities have flagged concerns. Among them:
-
Dependence on closed or internal systems makes full reproducibility harder.
-
Interpreting the AI’s “narration” as a human-style explanation can be misleading: AIs can produce justifications that look coherent without being well-founded.
-
The impact on academic training and authorship culture needs open discussion.
OpenAI notes its support for and respect of initiatives like the Leiden Declaration on AI and Mathematics, which aim to guide responsible adoption of these tools.
What’s next?
The publication invites the mathematical community to verify, reproduce, and extend these results. That means reading the manuscripts, checking the formalized proofs, and, above all, applying the ideas to new work.
If you’re a researcher or just curious, the practical lesson is clear: AI can already be a powerful tool to explore hard problems, but responsible integration depends on human review and community debate.
The technology doesn’t replace mathematical intuition or critical discussion; power complements human judgment.