Tim Worstall linked to an article about Moore’s Law and I ended up writing a somewhat rambling comment. The original article pointed out that it wasn’t so much technical limitations that would prove the end of Moore’s Law, but the high cost of manufacturing improved chips. Tim mentioned that an economic limit to Moore’s law would be if the cost of more powerful chips outweighed the benefits.
Now, I probably read too much science fiction, and too much Ray Kurzweil, but I’ve got some strong ideas about what the benefits of ever increasing computing power can be. For one thing, I think artificial intelligence is possible. I think the human brain is just a mechanism and that if we can learn enough about how it works we can replicate it. With enough computing power, we can write a program that works like a human brain and is intelligent and self-aware. Now that’s got immense benefits and will surely be worth immense cost.
Just imagine what you can do. For one thing, you can set your thinking machine to the task of inventing better thinking machines. Or inventing anything else you might like. Imagine a computer that can simulate 500 engineers working on designing a new aeroplane. Imagine one that can do it 1000 times faster than real-time. In a year you could have the output of 500 engineers working for 1000 years. That’s going to be a pretty good aeroplane. Everyone will be able to have their own personally designed aeroplane. It’s a bit of a silly example, but it shows how far off the limit to desire for more computing power might be.
The ability of AI to invent better AI could lead to an explosion in progress that Kurzweil describes as a the singularity. (I know Vernor Vinge invented the term but he uses it differently, to mean a point beyond which we can make no meaningful predictions of the future. But I think Kurzweil’s singularity is more interesting because I think we can imagine what we might like the outcome to be.)
Here’s the outcome I’d like: if we don’t have the ability already, we set our computer engineers to inventing Drexlerian nanotechnology: molecular manufacturing. Then we have them design tiny robots that live in our bodies and keep us healthy. The next step is to have the tiny robots interface with our brains and create a brain-machine interface so that we can, if we choose, experience virtual reality — a simulated universe of our own design. Ultimately we might not need bodies at all. Freed from physical limitations we can enjoy infinite wealth, each person a creator of universes. We can commune with our friends in Middle Earth casting spells one day and go parachuting on Jupiter the next.
I think that would be worth a lot of money. So let’s just say I don’t forsee any limit to the desire for more computing power or what people would be willing to pay for it.
But I have one worry, which Tim Worstall’s article alerted me to. What if the next generation of computers is too expensive to be worth building? How would we get to the (next+1)th generation? To double computing power, it might cost 100 times as much, and the things you can do with double the computing power might not be worth that. It would be like a flat spot, or minima, in the cost benefit curve that we would be stuck in forever.
Has this ever happened? I can’t quite think of a technological dead-end like it, but then it would by nature be an obscure technology that few would have heard of. Space flight, for example, could conceivably suffer from this problem. It might be possible to make a fortune mining the asteroids but if no-one can make money launching rockets into space it might never happen. In reality this doesn’t seem to happen: advances in space travel slowed down for a while, but then other technologies made it cheaper and now billionaires are building space-ships with only vague ideas about how to make money from it. In other words, if people want it to happen, it happens.
So while I can imagine advances in computing slowing down a lot if the economic conditions aren’t right, I can’t imagine them stopping for good. New technology in other areas will constantly change the game. And the fact that better computing makes the invention of better computing easier must be a pretty strong impetus to progress.
Hmmm. I have no doubt that thinking machines can be built (nature builds them all the time). But I doubt they will be computers.
I think a great deal about this, one of my intellectual conceits is an attempt to construct a practical mechanical theory of mind. And I think the computer is ill-suited to the task of doing what the brain does.
Now we know for a fact that in theory a universal turing machine (computer) can simulate any algorithm; thus it can simulate any machine. But computer simulations can be inordinately inefficent, depending on the task. Think of raytracing. It takes enormous amounts of computer power to simulate the paths of light through a scene to render a picture. Nature does it in the blink of an eye. Because nature is doing it in parellel (each photon is looking after itself) whereas computers have to do it serially- running through what each photon does one after another. It would take phenomenal amounts of computer power to simulate the activities of all the molecules in my big toe. Again, because nature is parallel, the task is trivial in the real world.
The human brain is composed of some tens of billions of neurons. Each of them has an axon with many dendrites. Some neurons have thousands of connections. A computer would have to run through each one. THen you’ve got the chemical signalling…
You might be able to write a computer programme that does all this, but why bother? Why not build an actual brain instead of computing a simulation of one? The human brain runs on less than 100W. I doubt a computer simulatiing it in real time would be so efficient
I think science went a bit ashtray on this; because computers were the first very complex machines we built, and because they could simulate to a degree a minor part of brain function (the algorithmic part that most brains on the planet, i.e. non human ones, cannot even do) people naturally thought that computers were brains. But they aren’t. The human brain can undertake simple alogorithmic tasks- counting, simple planning, simple equation solving- but generally needs external aids (scratch paper, computers) for all but the most trivial. It’s very bad at logical, algorithmic tasks. It’s good enough to make us what we are compared to the animals. But the interesting part of AI is consciousness- and that is what most of the brain is doing, and it is doing it in parallel by routing connections. In other words, it’s easy to write a programme that can play draughts- but nobody has come close to writing one that can care about whether it wins or loses.
So what we are after is thinking machines, not super draughts players, because it’s thinking machines, conscious machines, that will have true artificial intelligence (and of course, since intelligence is a characteristic, there is no such thing as “artificial”; it is either intelligent or it isn’t). What gives rise to consciousness (I have some more detailed ideas, but too much for a comment box) is the interaction of trillions of connections in real time, connected by a nervous system to the rest of the universe. It’s theoretically possible to simulate that on a computer, but if you knew how to simulate it, you could just build a real machine that works that way, which one may suspect would be very cheap (nature does it with not much more than a kilo of meat).
We may build brains one day; I hope we will. But more computing power isn’t the way there.
Very quick comment before bed: I agree with most of that; I was actually going to discuss specialised hardware but the article was too long; after reading your comment I’m starting to think more seriously about Asimov’s “positronic brains” (I am only now reading I Robot). More on this later…