February 16, 2026
Manifold Mania hits Sudoku
Show HN: Solving Sudoku reasoning via Energy Geometric models
HN erupts: GPU Sudoku “curvature” magic—brilliant or buzzword bingo
TLDR: A new GPU Sudoku solver claims wild speed by steering work to the “hard” spots of a puzzle. The crowd’s split between wowed and wary: is it genius math or a SAT solver in fancy clothes? Big if true, but people want fair benchmarks and code they can test.
Bee Rosa Davis dropped a “Show HN” bomb claiming a geometry-guided, GPU-powered Sudoku solver that blitzes puzzles at roughly 270,000 per second. The pitch: use “curvature” (think a heat map of where the puzzle is hardest) to focus computing power, plus an automatic pipeline that picks the right strategy on its own. The flex includes beating the “world’s hardest” Sudoku and dunking on prior work like Yann LeCun’s energy-based model. Cue the crowd split.
On one side, dazzled readers cheer the speed and the “math superhero” vibe; on the other, skeptics smell marketing perfume on old tricks. The spiciest jab called it “AI larping,” with extra side-eye over the résumé math and “patent pending” gloss. Another chorus asked, in plain English, whether this is just a SAT solver (a classic puzzle solver) wearing designer labels. Memes rolled in fast: “Manifold mania,” “wavefront my dishes,” and “geodesic to the fridge.” Benchmark drama bubbled too—GPU vs Python CPU? Apples vs bowling balls? Commenters want apples-to-apples against top CPU solvers, and code they can run. Meanwhile, fans point to the preprint and say the math explains why it’s fast: pick the cells that unlock the most information first. Whether it’s revolution or rebrand, the thread’s pure popcorn energy.
Key Points
- •A preprint proposes a geometry-aware, GPU-accelerated solver for finite-domain CSPs, demonstrated on Sudoku, using curvature on a “Davis manifold” to guide computation.
- •The solver uses a three-phase CUDA pipeline: (I) wavefront propagation with arc consistency, (II) manifold relaxation via curvature-adaptive gradient descent, and (III) curvature-directed iterative-deepening DFS.
- •An automatic trichotomy parameter Γ classifies instance complexity and routes each problem to the optimal phase combination without manual algorithm selection.
- •Mathematical definitions include a local curvature K_loc (from saturation, scarcity, coupling), an information value V(v) for variable ordering, and a Davis energy functional combining path length, curvature-weighted complexity, and holonomy deficit.
- •Reported performance: ~270,000 puzzles/sec, 3.7 µs per-instance latency, 7.8 ms average solve time (v4), solving 11/11 hardest-known Sudoku, with claimed speedups of 1,226× vs Python CPU and 40,128× vs Kona 1.0, and a comparison to Yann LeCun’s EBM.