This "efficiency" is likely why these models feel dumb, they lose reading of nuance when overtrained for specific tasks. It's like autists understanding of normal speech.