Same group, not the same?
An AI preprint claims a counterexample to the Borel conjecture in dimension four
Covers result 320 in OpenAI’s catalog: its one paper. See its connections to other results and the work it cites.
3:43. English captions available. The video loads from YouTube when you press play.
Chapters
The claim, in plain words
In 1953 Armand Borel asked Jean-Pierre Serre a question in a letter, which in modern terms reads: if two closed manifolds are both aspherical with the same fundamental group, must they be homeomorphic? This became the Borel conjecture: a closed aspherical manifold is determined, up to homeomorphism, by its fundamental group, and so are its symmetries: every symmetry of the group comes from a homeomorphism.
It holds in dimensions 1 and 2, and for orientable manifolds in dimension 3, through geometrization. In dimensions 5 and up it is known for hyperbolic and nonpositively curved groups, by Farrell and Jones and by Bartels and Lück. In dimension 4, the surgery tools are known to work only for groups Freedman calls good, and in June 2026 Davis and Hillman wrote that the conjecture is not known in dimension 4 for any group that is not good. Hyperbolic groups, apart from the simplest ones, are not known to be good. The smooth version had just failed in dimension 4, in a 2026 Annals paper.
Result 320 is one OpenAI paper, 60 pages, dated October 4 and made public October 6, 2026. It claims closed aspherical topological 4-manifolds M and N, with the same word-hyperbolic fundamental group, that are homotopy equivalent but not homeomorphic. It also claims a self-homotopy equivalence of M that is not homotopic to any homeomorphism. Both formulations of the conjecture would fail in dimension 4.
What it doesn’t claim
Not claimed: anything smooth, any other dimension, or an explicit manifold, and no invariant of M or N, such as an Euler characteristic, is computed. And it stands or falls with two companion papers of result 305, whose machinery it imports.
Where it stands
- Posted as a preprint
- OpenAI, 60 pages, dated October 4, 2026
- Peer reviewed
- Not yet
- Checked in Lean
- No formalization for this item
Last checked October 19, 2026. The paper is unrefereed and unchanged since release, with no Lean formalization, and it depends on two unreviewed papers of result 305. Combined with a 2012 theorem of Khan, M and N would also be s-cobordant but not homeomorphic; that is our reading, not the paper’s claim. In the first 12 days we found no specialist assessment, no reported error and no competing claim. OSVerify has not reviewed it.
How the proof goes
- A parity in the chamber. The paper builds a 4-dimensional chamber whose boundary can be filled in its double cyclic cover, but not in any odd cover of degree 3 or more. The obstruction is a parity: a filling of an odd cover would force each red and blue edge around a cycle onto one of its corners, with no corner receiving 2 red and 2 blue, and our computer finds that with 5 of each per pair, such an assignment exists exactly when the cycle has even length. Davis’s reflection trick turns the chamber into a hyperbolic aspherical complex.
- Two ways to glue. Its double cover has two manifold models: glue the fillings straight, giving M, or with a twist, giving N. A homeomorphism realizing the half-turn symmetry on M would build a forbidden odd-cover manifold, so none exists. M and N are told apart not by an invariant like the Euler characteristic, but by which symmetries homeomorphisms can realize, using a Sylow argument in a finite group. Our computer illustrates the counting idea on a small example: 63 special elements on one side, at most 52 on the other.
The numbers in the video
- Ours: with five red and five blue edges per pair around a cycle, an assignment with no corner receiving two red and two blue exists exactly when the cycle has even length (checked up to length 15).
- Ours: in a small example of the counting idea, 63 special elements on one side and at most 52 on the other.
Read the source
- Nonhomeomorphic Closed Aspherical Four-Manifolds with the Same Homotopy Type (PDF), 60 pages.
- The collection’s catalog, where this is result 320.
The paper
OpenAI groups related papers into one result; this one has a single paper.
Transcript
Read the full script
0:00The question
In 1953 Armand Borel asked Jean-Pierre Serre a question in a letter, which in modern terms reads: if two closed manifolds are both aspherical with the same fundamental group, must they be homeomorphic? This became the Borel conjecture: a closed aspherical manifold is determined, up to homeomorphism, by its fundamental group, and so are its symmetries: every symmetry of the group comes from a homeomorphism. It holds in dimensions 1 and 2, and for orientable manifolds in dimension 3, through geometrization. In dimensions 5 and up it is known for hyperbolic and nonpositively curved groups, by Farrell and Jones and by Bartels and Lück. In dimension 4, the surgery tools are known to work only for groups Freedman calls good, and in June 2026 Davis and Hillman wrote that the conjecture is not known in dimension 4 for any group that is not good. Hyperbolic groups, apart from the simplest ones, are not known to be good. The smooth version had just failed in dimension 4, in a 2026 Annals paper.
1:06The claim
Result 320 is one OpenAI paper, 60 pages, dated October 4 and made public October 6, 2026. It claims closed aspherical topological 4-manifolds M and N, with the same word-hyperbolic fundamental group, that are homotopy equivalent but not homeomorphic. It also claims a self-homotopy equivalence of M that is not homotopic to any homeomorphism. Both formulations of the conjecture would fail in dimension 4. Not claimed: anything smooth, any other dimension, or an explicit manifold, and no invariant of M or N, such as an Euler characteristic, is computed. And it stands or falls with two companion papers of result 305, whose machinery it imports.
1:56How the proof goes
The paper builds a 4-dimensional chamber whose boundary can be filled in its double cyclic cover, but not in any odd cover of degree 3 or more. The obstruction is a parity: a filling of an odd cover would force each red and blue edge around a cycle onto one of its corners, with no corner receiving 2 red and 2 blue, and our computer finds that with 5 of each per pair, such an assignment exists exactly when the cycle has even length. Davis’s reflection trick turns the chamber into a hyperbolic aspherical complex. Its double cover has two manifold models: glue the fillings straight, giving M, or with a twist, giving N. A homeomorphism realizing the half-turn symmetry on M would build a forbidden odd-cover manifold, so none exists. M and N are told apart not by an invariant like the Euler characteristic, but by which symmetries homeomorphisms can realize, using a Sylow argument in a finite group. Our computer illustrates the counting idea on a small example: 63 special elements on one side, at most 52 on the other.
3:04Where it stands
The paper is unrefereed and unchanged since release, with no Lean formalization, and it depends on two unreviewed papers of result 305. Combined with a 2012 theorem of Khan, M and N would also be s-cobordant but not homeomorphic; that is our reading, not the paper’s claim. In the first 12 days we found no specialist assessment, no reported error and no competing claim. If it holds, in dimension 4 the fundamental group does not determine a closed aspherical manifold up to homeomorphism, even for a word-hyperbolic group.
Spot a mistake?
This episode simplifies on purpose, but it shouldn’t be wrong. If something is, open an issue on GitHub or leave a comment on YouTube. Corrections are noted here with a date.