Briefly, Moore's law is the observation, valid for the last four decades or so, that the number of transistors you can pack on a chip has been doubling every two years. This translates to computing power mostly because you can flip a smaller switch faster.
Eventually this must come to an end when the circuits reach the atomic level, around 2020. Some futurists (Vernor Vinge, Ray Kurzweil) think some other unspecified technology will allow the law to continue beyond this. It's fun to speculate about a technological singularity, where an AI becomes smarter than humans, and can advance the field so fast that humans can no longer participate meaningfully. Vinge predicts the end of the human era before 2030.
But lets be a little more sober for a sec. First of all, Moore's law is very nearly at a standstill already. The problem is, clock speeds reach a hard limit around 4GHz. After that they run too hot. Combine that with the fact that they can process a number of instructions on the order of 1, and you have a problem. That's why in recent years they have been making multi-core processors, and developing architectures that can execute two or even three instructions per clock tick. That will help for a while, but what happens when you have as many cores as processes? (note: someone recently put 1000 cores on one wafer)
Second, there is much more to AI than computing power. A computer can win at chess, by brute force computation. But the best go programs in the world cannot compete with even casual club players. And it's not for lack of trying. It's just that computers are not good at judgement. Correct me if I'm wrong, but no robot can run on two legs, because they can't balance well enough. Surely they already have the necessary computing power for that task. And visual perception is primitive. And is it really all that hard to figure out what letters are scribbled in those little boxes to foil spammers? Even the tiny brain of a mosquito exceeds our best AI efforts.
So, sorry Vernor. I don't know what's in store for 2100, but your singularity just ain't happening before 2030.
Micropython on the Air602
-
Lately I have been forgetting to close my garage door. Not fun to find it
open in the morning, and wondering if some of my tools might have walked
off. The...
4 years ago
No comments:
Post a Comment