> constraining the models enormous capacity for creativity without stifling its ability to learn If I recall my AI history correctly all the 60s/70s research

> constraining the models enormous capacity for creativity without stifling its ability to learn If I recall my AI history correctly all the 60s/70s research focused around logic based systems and inference producing things like Prolog. It was thought that one just simply needed to come up with an appropriate rule set to generate powerful AI. The issue of course is writing enough rules to give powerful AI outside some very constrained problem domains (see https://en.wikipedia.org/wiki/SHRDLU https://en.wikipedia.org/wiki/SHRDLU) just isn't feasible. The problem of current machine learning models being too clever for their own good and needing appropriate constraints feels similar. If only you had the correct constraints you could do all kinds of things. History repeating itself?