I don’t doubt that - my point though is that maybe 3 can only solve things in its training data and 4 can figure things out.

3 seems to be more rigid. It needs babysitting to solve things. Which means it can only solve things I already know. 4 is more flexible and can solve things by itself.