Sotaku: neural net teaches itself sudoku rules, hits 98.9% SOTA on hardest dataset
CatAstro_Piyush · x · 2026-09-03
Cheng Lou unveiled Sotaku, a toy research project where a neural net automatically discovered the rules of sudoku and learned to solve them, achieving a new state-of-the-art 98.9% on one of the hardest sudoku datasets — while remaining game-agnostic and beating all sudoku-optimized architectures.
Notable unconventional findings:
- Reverse curriculum consistently beat curriculum learning, counterintuitively
- Emergent reasoning-like capabilities appeared
- The writeup also muses on the future of traditional programming
Full motivations and experiments are documented in the linked post.
More from Research
- Nature review: AI is designing 'weird' physics experiments humans can't intuit — MarioKrenn6240 · 2026-09-03
- ai& and Tenstorrent launch JapanFold: free sovereign open-source drug discovery in Japan — DavidBennett__ · 2026-09-03
- One int8 op makes exact rescoring nearly free in PLAID-style late-interaction search — antoine_chaffin · 2026-09-03
- 10 AI research labs worth following, from AI2's AutoDiscovery to Stanford SAIL — goyalshaliniuk · 2026-09-03
- LoopArena: even the best controller model hits just 24.69% managing coding agents — rohanpaul_ai · 2026-09-03
- SnapBench shows image noise cripples mobile snap-and-ask multimodal retrieval — Zirong Chen · 2026-09-03