Models still can't see determinants, but no-thinking time-horizons double every ~9 months

scaling01 · x · 2026-09-04

Scaling observer notes that current models remain unable to reliably read determinants or eigenvalues from matrices — outputs are very noisy, with models simply favoring low numbers like 1, -1, 2, -2 — though he expects scale to fix this eventually.

Key data point: no-thinking time-horizons are doubling every 373 days (about 9.1 months), depending on whose measurement you trust — a notable trend line for raw model capability growth without reasoning tokens.

Original post →

More from Models

Models channel →