A pretty interesting podcast from Quanta Magazine suggests that it might not be a simple matter of training set size: https://www.quantamagazine.org/where-we-see-shapes-ai-sees-t...