Humans are adaptable enough to extend or modify their goals in many situations where these goals go “off distribution”. The ACE algorithm I mentioned is capable of the same thing, in a much weaker form. Learning can allow entities to extend their goals intelligently to novel situations (rather than extending them blindly).
Wait, why is this more of a problem for fixed-weight models vs other kinds of models?
Humans are adaptable enough to extend or modify their goals in many situations where these goals go “off distribution”. The ACE algorithm I mentioned is capable of the same thing, in a much weaker form. Learning can allow entities to extend their goals intelligently to novel situations (rather than extending them blindly).