Zvi: This Benchmark Progress Isn't Suspicious—The Dramatic Drops Are

TheZvi · x · 2026-09-05

AI commentator Zvi weighs in on a recent benchmark movement: this particular progress seems "highly not suspicious" and consistent with all other observations, because reality is consistent. It is precisely that consistency which makes other, more dramatic benchmark drops look even more suspicious. The remark needs its quoted context to fully land, but the core point stands: plausible gradual changes cohere, while isolated crashes signal data-integrity problems.

Related event: Zvi and Gary Marcus debate whether new model gains are real or benchmark-gaming(4 posts)→

Original post →

More from Models

Models channel →