Solving Sudoku reasoning via Energy Geometric models

Name: Solving Sudoku reasoning via Energy Geometric models
Availability: InStock
Author: epokh

by epokh·Feb 12, 2026·18 points·4 comments

Visit Project View on HN

AI Analysis

●●SolidWizardryBig BrainBold Bet

Differential geometry applied to constraint satisfaction—elegant math, but Sudoku is a solved novelty.

Strengths

•Novel application of Riemannian curvature to CSP scheduling; unifies classical heuristics as degenerate cases.
•Staggering performance: 270K puzzles/sec vs. LeCun's EBM at ~3 puzzles/sec; real GPU engineering.
•Generalizes beyond Sudoku to finite-domain CSPs like scheduling and logistics—genuine algorithmic contribution.

Weaknesses

•Sudoku is a solved problem with diminishing practical returns; the 'hardest puzzle' claim is marketing.
•Heavy mathematical framework (Regge calculus, holonomy groups) may not scale or matter for real CSP instances outside toy domains.

Post Description

We were genuinely impressed by Yann LeCun’s recent announcement about his Sudoku solver based on Energy-Based Models here: https://logicalintelligence.com/yann-lecun

So we (a software engineer and prodigy mathematician) worked in the weekend to see if we can beat it and we did!

If Kona solves puzzles in 313 milliseconds, we currently solve 270,000 puzzles per second!

We do not know the internal mechanics behind Logical Intelligence’s system, but we are happy to share ours in the webpage there is a full description, code and also a paper published.

Why people are obsessed with Sudoku: we think is because being a CSP, it can solve some serious business problems such as work scheduling, resource allocations and logistic optimization.

For any questions reach me out to: [email protected].