AI is limited more by language than anything else. NLP which I do a lot of work on at http://plexinlp.com goes a long way, but the computer doesn't have the ability to get metaphors, or sarcasm, or have any sort of "BS" detection.

CoreNLP (which isn't as good as our NLP but is the "standard" only gets the parts of speech right 95% of the time, by words, but about 75% of the time by sentence. Imagine you were a waiter and you only understood 75% of sentences. You'd not last too long.

But somehow we expect computers which not only are limited to 75% of sentences being "tagged" correctly, but then not having any idea about the concepts behind them.

AI as we see it today is not so much "AI" as it is Computer Driven Bucketing. Build a formula that gets as many things in to pre-sorted buckets as possible. For me it isn't AI until the computer can make leaps of its own. Create a new bucket and explain why it was needed, or create a new solution that wasn't proposed by the programmer.

I am working hard to bring all of that to fruition, but honestly we are a long ways off.

Here are some milestones I think we will need to hit before we get to AI.

Software powerful enough to convert Wikipedia (or any knowledgebase) to structured Data automatically.

Query system that takes any user input and converts it to something resembling a Database query (This is what I'm working on)

NPC's for Games/MMO's etc that pass a turring test with in the limits of their Virtual world. (Basically they don't have to pass for real, just as real as the DnD character a human would play)

These are really "simple" milestones that are the step before we see true AI start to work.