That you need world models to sensibly deploy "thinking" machines in the real world. Else they do stupid shit like drive straight into water. You can bruteforce some semblance of thinking by training on literally all knowledge that can be digitized but even that is proving to not be quite enough.