kalomaze shares hilariously incoherent early outputs of tiny byte-level code models

kalomaze · x · 2026-10-24

AI developer kalomaze shared early convergence outputs from training tiny non-BPE two-byte models (65,536 vocab) on pure code data: the generated "code" is syntactically chaotic with bizarre variable names, which he says makes you "feel like a stroke victim." The screenshot shows garbled lines like fiskingCallremember = outp(vhi) — a genuinely funny glimpse of early-stage small-model training.

Original post →

More from Fun

Fun channel →