I'd say that strong artificial intelligence is far more likely to happen via bottom-up modeling of the existing human brain as a starting point. The computational capacity for reductionist simulations that will exist probably 20 years from now, and as computational capacity increases those simulations start to become emulations. I don't expect to see massive gains over that time in the other way of doing it.
Once you have crude human brain models, a lot of very unethical things will start to happen in the course of further development. (e.g. countless deaths of intelligent beings if you're in the continuity identity boat, and probably a lot of unavoidable pain and suffering regardless of your take on identity). That seems hard to prevent given the enormous advantages that will accrue as a result; there is a strong incentive for people to adopt the pattern identity point of view so as to justify what they are going to do with mind copies in the course of development.
The important point is that once you have human brain models, from there the path to many different forms of strong artificial intelligence is just a matter of iterating those models. This seems far more likely to produce results than constructing new forms of intelligence progressively and de novo.
There is firm evidence he had at least 3 serious views (and one fairly archaic one not many want to talk about) on the cause of physical gravitation. It is only the "modernists" who need for Newton be be not beholden to any of them that elevate and reiterate "Hypotheses non fingo" to some sort of rallying cry as if he had it tattooed on his torso.
There is nothing to say that evolution has arrived at the perfect method.
We can and probably will build reasonably general AI, but you far more likely to see each generation of AI being an ever smaller improvement than any sort of runaway exponential progression. Not to mention hardware progress has slowed to a relative standstill.
The whole "diminishing returns will stop us" just seems like a comforting fairy tale for those who don't want to think about the consequences (which I suspect a bit of thought does show won't be as rosy as Kurzweill imagines).
Edit: Hardware progress is still mostly following "Moore's Law". The only that's not increasing is processor speed. But if we build a "reasonably general AI", how could that box's capacities not be increased by tightly integration with other similarly intelligent boxes?
PS: S curves often look like exponential curves but the real world has real limits so you can't have unlimited exponential progression of any type period end of story. And it looks like we are on the down slope when it comes to transistors. http://www.extremetech.com/computing/123529-nvidia-deeply-un... "Nvidia deeply unhappy with TSMC, claims 20nm essentially worthless" And that's for video cards which are embarrassingly parallel.
It's been over 300 years and we're still trying to figure out exactly how gravity works. But in a way it doesn't matter: his law has given humanity tremendous abilities it didn't have before.
When Deep Blue won the Jeopardy contest a couple of years ago, it was obvious that it wasn't intelligence in the sense that we commonly understood it. Yet it was able to beat the human champions. Deep Blue wasn't a model of a human player, but it didn't matter because for the purposes of its construction it performed just as well as one.
My money says the singularity happens the exact same way -- we are able to "fake" more and more things that look exactly like intelligence until one day we're able to fake intelligence to a degree that it's virtually indistinguishable from our own. We're eventually able to do something that looks like moving our sentience into a computer even though it "won't really" be doing it.
My guess is that we're hundreds of years away from that date, but whether it happens or not, the fact that an expert right now in the complexity of the underlying physical system has an opinion on the computational problem that probably won't be solved until after 2100 doesn't seem to me to be very relevant. Of course it's complicated. Of course we don't understand it. And of course we can't duplicate the structure of things we don't understand. I believe for any layman in the field all of that goes without saying?