Pierce Freeman and Richard Diehl Martinez discuss Levent Alpöge's announcement of a counterexample found with Claude Fable 5. They use navigation directions to introduce the difference between being able to reverse a function locally and having a global inverse, while acknowledging that their analogy is imperfect.
Pierce Freeman and Richard Diehl Martinez describe a polynomial map in three complex dimensions that meets the nonzero constant-Jacobian condition but sends different inputs to the same output. They explain why adding unused dimensions carries a counterexample into higher dimensions, while distinguishing the still-open two-dimensional question from the simpler one-dimensional case.
Pierce Freeman and Richard Diehl Martinez ask whether the result changes neural-network design. They specifically note that normalizing flows are constructed to be invertible, so the counterexample does not invalidate those architectures. Broader possibilities for new layers or research methods are speculative.
Pierce Freeman and Richard Diehl Martinez distinguish checking one counterexample from proving a universal assertion. They contrast the apparently open-ended search with the tightly guided AI collaboration they associate with researchers such as Terence Tao. Because the original chat transcript was not available to them, their account of how much guidance the model received is an inference rather than a verified autonomous-discovery claim.
Watch on YouTube
