[enord]

I never mentioned limits, and the reasoning is purely inductive, based on previous events.

We are far away from understanding intelligence in all directions. Top down, bottom up, neurologically, psychologically, logically, mathematically and last but not least philosophically.

There's optimism at the moment, because we are doing more stuff with more annotated data (the annotations providing the semantic grounding, as in "Not hot dog" vs. "Not in category 339492-883764-399274"). The key difference this time being access to (and processing power for-) sufficiently large "training sets" (read "samples") for deep-learning algorithms (read "statistical models"). From an AGI point of view, this is nothing but an expensive parlor trick, because the "intelligent" part is the annotation, not the categorization after the fact.

[YeGoblynQueenne]

The annotation is not even particularly smart. In supervised classification, labels are basically scalars, standing for... whatever the researcher means them to stand for. The class represented by the label can be as broad or as narrow as the researcher wants it to be. Even the relation of the class with the data it is supposed to represent is arbitrary and its choice entirely unprincipled and based on instinct alone.

Which goes to show that a) our machine learning models are dumb as bricks and b) they are as far from AGI as worms are from building a rocket to go to the moon, where their god lives (see all those holes up there?).