Machine intelligence will surpass human intelligence, within a few decades.

Very good point, but computing devices don't have scaling problems that the other technologies you listed have. Nuclear would have had a similar generality if, for example, you could safely run a fission power device on your phone, for example -- but you can't; it just didn't scale down at the required safety level. Similarly, air travel doesn't scale down as people thought it would (still no flying cars) even though it works very well on large scale (airlines, combined, have near monopoly on long-distance travel). Computing scales both up (supercomputers) and down (cell phones). It may be that current silicon chips won't scale down to cellular level, but it also may be that (a) such scaling is not required for human-level AI (I believe it isn't), or that (b) other technologies such as graphene may replace silicon.

> There is still much debate as to what drives the brain

There is no debate. The short answer is: sugar.

> whether there is more to human intelligence than just the physical brain

No.

> Even if we could completely duplicate a human brain biologically, would it just start working automatically

Not unless you teach it. Humans who grow up in the wild on their own are quite ape-like -- clearly, human brain doesn't start working automatically on its own (at least working well). Also, as demonstrated by machine learning, you can teach a non-biological system just as efficiently as you can teach a biological one, which means that the fist AI will reside (or already resides) on a run-of-the-mill silicon chip.

> Would it be able to learn if we could attach sensory inputs?

It would have to, otherwise it would not be human-level AI by definition.

> It's quite possible that there is much more to human intelligence than massive computing power.

Yes, and that is years of learning and good enough (just good enough) learning algorithms.

> See for example http://en.wikipedia.org/wiki/Mind-body_problem

My foot tells me it doesn't have that problem.

Machine intelligence will soon diverge from Kurzweil's hyper-optimistic predictions and timetables so dramatically that he won't be able to gloss over it any longer.

As for where AI stands today, I'd say we're not even at the level of simulating a jellyfish, or even a single cell. Simulated protein folding is still a pipe-dream for our fastest supercomputers, nevermind simulating the intelligence of a sentient being.

On top of that, algorithmic systems as we understand them may be a shallow model of what it means to have 'human intelligence'. How do you design a circuit that feels pain, or know when you've succeeded?

Your rhetorical questions of course have rhetorical answers: yes, if it is programmed that way. "AI" isn't a singular mind design. The rest of your comment is weird, I'll refrain from guessing what you're implying (if anything).

> Are there some articles/papers that I can look in to?

There was little real science done because "real" science is when you do experiments, and a study of this kind would involve the so-called forbidden experiment, which would create a huge ethical stir. The most often-cited report is this one: http://academic.research.microsoft.com/Publication/2053957/g...

Here is a brief Powerpoint with an overview of interesting cases: http://www.tulane.edu/~howard/LING411/ppt/d39-FeralChildren....

While it is often said that it is hard to draw conclusions from the few reports that exist -- I think that the one conclusion that can be drawn is that nearly everything that makes us "human-like" is a product of learning and social interaction. No confirmed "feral" child has ever been reported to have acquired language in the full sense of the term.

> Is there something you can refer me to that compares the efficiency of learning in humans vs machines?

Just look up any introduction on machine learning. Nearly everything we do every day (recognizing objects, thinking in terms of goals, etc.) involves a classification problem, and every classification problem can be reduced to a regression on data. So there is nothing more fundamental to learning than fitting a model to data. Humans are pretty efficient learners, but they have their limits (your Bayesian spam filter is probably more efficient than you on recognizing spam, even though it may fail in a few cases because it doesn't quite "live" in the world you live in, meaning that its world model is not as deep/rich as yours). To acquire a deep/rich world model, you would need more training (provided you have a correctly balanced algorithm), which would ultimately be as expensive as raising a child. There is no reason though to think that it won't be done one day.