Downloading The Mind [Slashdot]
Kurzweil figures we'll have strong AI by 2029 and be able to copy a human mind about a decade after that.
I've met Ray on several social occasions and discussed his vision for the future. There is a huge flaw (or blind spot) in his vision. All of his massive advances in AI and general computing functionality are based on extrapolating trends like Moore's Law, Metcalf's Law, etc. into the near future. He infers that because they predict that massive CPU power and network bandwidth will be available, that the software to match it will naturally come along, too. Ray's a hardware guy, for sure. Unfortunately, we've already seen a plateau in the demand for CPU cycles and network bandwidth. Without market forces to drive these trends, why assume they'll be sustained?
The problem with depending on hardware and network advances to drive his vision is that software engineering simply cannot keep up with the pace of advances on the hardware front. Anyone who has ever read the "Mythical Man Month" understands this at a basic level. Humans simply cannot organize themselves well enough to tackle software projects of the magnitude that Ray envisions, at least not by 2029.
Ray dismisses this argument by saying we'll have software that writes the software. Well, there's a tautology for you. If you can't write the software you need because it's too complicated, how can you possibly be expected to write the software to write that software? Genetic algorithms are useful for some very specific trial and error sorts of problems. But using them to random walk our way to a billion lines of debugged, functional AI code seems a bit of a stretch.
My money sure isn't on Ray's pony...
8:27:44 AM 
|
