h2g2 The Hitchhiker's Guide to the Galaxy: Earth Edition

One very convincing argument that artificial intelligence (read "good" artificial intelligence) will never exist, is because of unreachable mathematical requirements.

The argument starts something like this:

The best way to artificially simulate intelligence is to create a simulation of the human brain, as detailed as possible.
The most detailed way to simulate a changing environment like a human brain is to record all of the possible permutations of changing parts. That is to record all of the possible relationships all of the parts have with each other.

Now the problem rears its ugly head. There are a lot of parts in the human brain. Just detailing the brain cells alone is not good enough, one must detail the blood flow, the electro-chemical reactions, and a lot of other boring stuff.

As it turns out, there are SO many things going on in a human brain (or any organic brain for that matter) that the total number of permutations that describes the relationships are extremely large. So large in fact, that if you were to build a computer out of hydrogen atoms, where every atom is a bit, you wouldn't have enough bits (according to the current assumption of how many atoms exist in the universe). Now that is a pretty complex system.

Now, if you know something about the way computers work, you might be saying to yourself that this is a pretty inefficient way to build a computer program to simulate a human brain. After all, we don't really need ALL of the permutations, we just need the really important ones. There are many many shortcuts that we can take to simulate a human brain when we are writing this sort of program. Surely these shortcuts are good enough to make this problem easier to solve.

Now mathematics has to come into play. The problem that we have is that simulating a human brain is a very complex mathematical problem. It is what is called an exponential problem. The more detailed the problem becomes, the harder it gets. And it doesn't just get a little harder. It gets A LOT harder. It get exponentially harder. So in order to get really stupid AI, the kind of AI you see in computer games and weapons systems and the such today, the problem is easily solved by computers. But to get better AI, you need better computers. A LOT better computers. Maybe even computers that require more atoms than the number that scientists think exist in the universe!

The complexity of this problem in mathematical terms is Order X to the Nth power. Where X is a constant greater than one (usually 2 because most computers use binary mathematics), N is the level of detail you need and X to the Nth power is how hard it is to do the problem. So lets say that the type of brain we want to simulate only has 20 things going on in it (20 synapses or whatever) that interact with each other. Then to simulate one change in this type of brain we would need 2 to 20th power computations. Well, that doesn't really seem so bad considering how powerful computers are now days. But when the detail of the type of brain you want to simualte increases to say.. a million or more, then you know you have problems. Because 2 to the millionth power is a really really big number.

Oh, but there are shortcuts. Some of those parts of the brain might not nessisarily need to interact with all of the other parts of the brain all the time. If we can program that sort of shortcut into the program we can reduce the complexity by a certain amount. Unfortunately because our problem is geometric in complexity, the shortcuts may work very well when N is very low, but as N gets larger, the shortcut is less and less useful. That is the complexity of the problem grows exponentially (very very fast) while the usefulness of the shortcuts only grows linearly (which is only very fast, not very very fast). This means that for an extremely hard problem like "good" artificial intelligence, the amount of help that shortcuts would give us is negligable, and the problem is still too hard.

What does this mean? It means that computers will probably never be able to simulate human intelligence well. It would either take a breakthrough in way we understand mathematical problems (finding a way to simplify this particular exponential problems exponentially instead of linearly... which is not likely to happen), or a breakthrough in the way computers compute (something that uses a much more complex mathematics than binary computations... also not likely to happen).

Or there is a third option... and that is to not use a mathematical simulation at all. If we actually build a living thing with all the organic parts that are in a human brain, and if we can make it work, then we have solved the problem.

Remember when I said a more complex computer might help? Well, how about an organic computer that uses a real brain to do it's computation, instead of mathematical computer chips. It is entirely possible that Artificial Intelligence will not be possible until we have mastered Artificial Life.

It's true that there are a silly number of possible brain states, but why do you need to enumerate them? Also, as you pointed out, there is clearly going to be a problem if you try to run an amazingly accurate (cellular? molecular?) simulation.

My expectation is that it is not necessary to create a perfectly accurate model of the brain to get intelligence, any more than you need to create a perfect model of a bird to get flight. There are principles and mechanisms that (neuro)biology instantiates to produces the phenomenon you recognise as intelligence. Once we fully understand the 'essence' of this type of complex adaptive system we can instantiate it in some artificial system which, for the reasons you provided, will probably NOT look like a computer.

The question is Who, and When?

Of course, once we have something that is one percent more intelligence than we are then we simply set it working on it's own successor (one percent smarter than itself!) and that exponential law makes Big Things happen..

Interesting stuff, AI. ALife comes first though.

During my brief lifetime on this somewhat limited planet I have found that when something appears on television it very soon materialises in real life (whatever real is).Also, the realms of technology are advancing, even if it seems to take a long time to achieve progress (whatever progress is). Well, I just watched my first dvd film tonight,the Matrix. Be careful what you wish for, it might come true...

For example:
Cars can talk now, and the are, in fact, even more annoying than they are on TV.

Indeed, be careful what you wish for!

If the question can be interpreted as 'will humans ever produce a good artificial intelligence?', and if the word artificial can be taken to include genetically engineered organisms, I would suggest that the answer is very near indeed.