Interesting read. The conclusions mirror current issues generally for AI, at least in regard to what is currently available to the PUBLIC - its is very good at regurgitating what it has seen somewhere else.
ie. "garbage in - garbage out "
Currently the only way to really "learn", is to "do". Same with a person learning new skills - ie. you can read ( watch or hear ) all you want, however you don't "get it" all until you try it yourself. This is because the level of detail in the task itself, PLUS the associated required skills / knowledge, cannot ( yet ) be taught effectively and concisely.
"Learning" requires actual hands on experience and applies to everything - from Graziering to Brain Surgery. Those years of "doing", and the resulting knowledge, are what ultimately you pay the mechanic for ( assuming they are a good one ). Rule of thumb is 10,000 hrs ( 3 years ) to master a given concept ( for a person ! ) .
This raises some interesting questions for AI -
1/. what are the consequences for AI getting it "wrong" ( whatever "it" is )
2/. when will AI realise that it has to "Do" to "Learn", and
3/. what does that mean - ( Robotics as a networked learning platform etc. etc. )
4/. and then ?...
Obviously AI can process and "remember" data / information far more effectively ( faster / effective cross association etc. ) than the average human brain already,
so what happens when it gets the above figured out ? - That's a rabbit hole to go down ....
