Breakthroughs
Significant results in mathematics, computer science and physics. The bar is deliberately high, so quiet weeks produce nothing at all rather than filler.
For the earlier hand-written round-ups, see 2026 so far and 2025.
Latest digest
Breakthroughs Weekly — 15 November 2026
Experts reproduce AI-found algorithms that refute the Exponential Time Hypothesis
Computational Complexity blog · 28 Oct 2026 · Complexity theory
OpenAI posted a preprint, with a Lean proof, giving an algorithm that decides satisfiability of k-CNF formulas for every fixed k, and even counts solutions, in polynomial time times 2^O(n log log n / √log n); the Lean version proves a slightly weaker bound. That is subexponential, so it refutes the Exponential Time Hypothesis and the stronger SETH, and Anthropic later released its own paper on the same question. Ryan O’Donnell and Shiva Kintali have written simpler follow-up proofs, and Lance Fortnow, who has not checked the argument himself, says it is safe to assume the hypothesis is dead; Impagliazzo, Paturi and Williams are running an open examination and say their role is not a certification. The algorithm gives nothing new for general NP problems such as Subset Sum, and all of these papers are preprints rather than reviewed publications.
Review group says OpenAI’s AI-built proof constructs four-dimensional Yang–Mills theory
Yang–Mills Group (GitHub) · 10 Nov 2026 · Mathematical physics
An independent group that reviews pure Yang–Mills submissions, reportedly chaired by Nobel laureate David Gross, has published its assessment of OpenAI’s 28 October Lean-checked submission. In the group’s opinion it proves that a quantum field theory obeying the Wightman axioms can be built as a scaling limit of four-dimensional lattice gauge theory for any simple compact gauge group. It does not show that the infinite-volume limit is unique or that a mass gap exists, so the Clay Millennium problem, which requires a mass gap, remains open. This is the group’s judgment rather than journal peer review, and press coverage reports that the reviewers say they do not yet fully understand the proof.
Earlier digests
2026
Breakthroughs Weekly — 9 November 2026
An AI-produced, Lean-checked proof of the Aaronson-Ambainis conjecture on quantum speedups gains the conjecture's namesake as a believer; peer review is still pending.
Breakthroughs Weekly — 1 November 2026
Independent experts have now re-derived OpenAI's claimed bound of 9/4 on the matrix multiplication exponent, a large step for algorithms.
Breakthroughs Weekly — 25 October 2026
An AI-produced proof that the Riemann zeta function has no zeros near the edge of the critical strip is drawing serious expert endorsement.
Breakthroughs Weekly — 12 October 2026
A human-written proof that 4-to-1 games are NP-hard lands in complexity theory, and the Nobel Prize in Physics goes to IceCube neutrino pioneer Francis Halzen.
Breakthroughs Weekly — 4 October 2026
A 1971 conjecture of Ronald Graham on reordering residues so all partial sums differ is resolved for large primes, now drawing wide expert attention.
Breakthroughs Weekly — 20 September 2026
A 2004 conjecture on random regular graphs, unnoticed for nearly a year, draws expert praise once Quanta profiles the completed proof.
Breakthroughs Weekly — 14 September 2026
A near-linear-time algorithm delivers a new, more efficient proof of the four-color theorem, cutting coloring time from quadratic to O(n log n).
Breakthroughs Weekly — 8 September 2026
A finite-time blowup proof for forced 3D Euler flows earns Terence Tao's praise as a Navier-Stokes stepping stone, and a five-decade percolation conjecture falls.
Breakthroughs Weekly — 31 August 2026
Claude pushes a Riemann zeta bound to 67%, a 30-year discrepancy-theory bound falls, and detectors catch solar neutrinos scattering coherently.
Breakthroughs Weekly — 27 August 2026
A 1939 conjecture on polynomial maps falls above two dimensions, and reactors show a faint antineutrino afterglow long after shutdown.