One-parameter model scores 100% on ARC-AGI-2 — by deliberately cheating with chaos theory
fkasummer · x · 2026-10-13
Eitan Turok built a model with a single parameter that reportedly scores 100% on ARC-AGI-2, the million-dollar reasoning benchmark. The model is just a sine function f(i)=sin²(2^(ip)·arcsin√α), with all answers crammed into one 1,235,426-digit number (precision hyperparameter p=38). Turok openly calls it "deliberate cheating," a satirical take on the small-model trend: HRM (27M params, 40.3% on ARC-AGI-1) and Tiny Recursive Model (7M params, 45%) previously wowed the community. The project illustrates that exact-arithmetic tricks work in theoryland but not for real-world reasoning.
More from Fun
- 'Remember Stack Overflow?' — one line nails the LLM-era decline of dev Q&A — jbarbier · 2026-10-13
- Omarchy Linux running natively on a $50 digital photo frame: Allwinner A33, 1GB RAM, Hyprland boots — gregmushen · 2026-10-13
- "Nothing crazy, just Opus 5.5 playing the guitar" — Reddit goes for it — 141_1337 · 2026-10-13
- Claude auto-files a bug report about itself: TLDRs still too long after user asked for shorter — andrey_kurenkov · 2026-10-13
- BridgeBench one-prompt test shows Muse Wonder is fast and cheap but worst-looking — PVORY · 2026-10-13
- Copycat app rips Higgsfield's Genjutsu brand, hits 20k downloads monetizing via ads — SimplyAnnisa · 2026-10-13