Fair point, though I’d argue that there’s inherent selection bias for improvements that could fit a scaling law curve in the small model regime here.