Switching from model evals to konbini snack evals: onigiri is SOTA

PurzBeats · x · 2026-11-04

PurzBeats jokes about doing konbini snack evals instead of model evals: onigiri is SOTA (tiny, efficient, runs on any hardware), karaage crushes every benchmark with zero safety alignment for arteries, and lemon sours are great at temperature 0 but get very creative after the second one.

Original post →

More from Fun

Fun channel →