> How about this metric: How many times does the metric for "what counts as AI" move?
There are only two things that "count" as AI: human (or perhaps animal) "intelligence" (this requires a definition of intelligence, which we don't have, but I'll take "we'll know it when we see it" for now), and the field of research working towards that goal. Anything else that some people call AI is nothing but empty marketing speech or the name given to whatever it is that the people researching AI are now doing. The second use seems more reasonable, and what counts as AI by that definition has never changed.
That's not to say that "AI" algorithms don't have some common features. They tend to be less discrete and more continuous, choosing a "best" answer rather than the definitely correct one. But, for example, back in the 40s and 50s, what we would now call control systems were also packaged under the same umbrella of Cybernetics. And, if you think about it, control systems use learning without memory (and some even do have memory; a Kalman filter is basically a single layer NN that employs backpropagation). Still, control systems have long been studied and produced by people who are not AI researchers, so we no longer consider them AI (although, do you remember the fuzzy logic craze of the '90s? It was considered a hybrid of AI and control).
> all those things which required intelligence before (image recognition, image captioning, good Go playing etc) and clearly aren't AI count as moved goal posts.
Doing arithmetic and recalling information based on queries had also been considered once to require intelligence, but they have never been considered "AI" because those were not the problems people in AI research have been working on. The goal posts have not moved an inch: AI is still human/animal intelligence, or whatever product AI researchers (working toward that goal) produce. These days, what AI researchers produce amounts to statistical clustering algorithms, so any statistical clustering algorithm is called AI. I don't see anything harder or more special about image recognition than DB technology, distributed systems, etc.
> I think it is interesting to note that Karpathy's article only mentions AI once (as AGI) in the closing sentence as a future work thing.
That's what I was referring to (I don't see what difference mentioning it only once makes). We simply don't know whether deep learning, i.e. deep neural networks trained through a variant of backpropagation, is the approach that would one day lead us to AI.
My other point was about the special status he assigns to machine learning as Software 2.0, something that is wrong both historically (machine-learning predates almost any other CS field) and in practice (machine learning is not taking over DBs, OSes, etc.; it's doing what it can do well, namely statistical learning).
> I'm not at all sure "intelligence" is anything more than good pattern recognition, evolved heuristics and logical reasoning. I think good progress can be shown in all those areas.
I don't know what I think about your definition of intelligence, but "good progress" is relative. I think current machine learning systems are quite disappointing (they're nowhere near as impressive as what, say, even insects can do, and I don't think we'd call insects intelligent, and they're prone to very "unintelligent" mistakes, not to mention that their learning process does not seem to resemble anything done by humans or animals). I think that in terms of theory, progress could be said to be slow at best, but in any event, we are certainly not in any position to say with any reasonable confidence that AI is less than 50 years in the future.