Chollet: Benchmark Scores Are Meaningless Without Cost

fchollet · x · 2026-07-07

Keras creator François Chollet believes that reporting a single scalar benchmark score (e.g., "75% on XYZ") is now entirely meaningless. He argues that evaluation results must include an efficiency score, such as "75% at $10 per task," otherwise they fail to reflect the model's true value.

Related event: Chollet: AI Benchmarks Must Include Marginal Costs(2 posts)→

Original post →

More from Research

Research channel →