Monday, April 19th, 2010

The next Labs Night will be April 29, which is next Thursday, here at the Mountain View Mozilla HQ. More details as we get closer to the event.

Computer scientists have been working on artificial intelligence for over 50 years now. The results have been disappointing for anyone who was looking forward to a world of intelligent robots. AI research has produced some useful algorithms, which combined with Moore’s Law have allowed us to brute force certain narrow problems (like winning at chess); not to belittle the cleverness and hard work of AI researchers, but everything we’ve seen is just cleverer ways of solving problems by rote calculation. We have yet to see a computer program do anything remotely like what we’d think of as “true” Artifical Intelligence: independent reasoning, original thought, self-awareness, comprehending human language, etc.

So the thought I had is that maybe strong AI is just not something that Turing machines are capable of. No matter how fast or powerful our computers are, they’re still Turing machines, and Turing machines have been mathematically proven to be incapable of solving certain types of problems, like the halting problem (given a program, correctly determine whether the program ever halts or whether it loops forever). This is just a hunch, not something I could back up with any solid evidence, but it seems to me like the halting problem is much easier than strong AI. Intuitively, a computer that has what we think of as true intelligence ought to be smart enough and capable of enough introspection to avoid getting caught in self-referential loops of the kind that make the halting problem un-computable. Therefore, a true intelligence could not be a Turing machine.

If this is true, then trying to build strong AI on a Turing machine is like trying to build a web browser out of gears and springs: with extreme cleverness you might solve some very specialized sub-problems, but it’s the wrong tool for solving the main problem. Not just the wrong tool, but the wrong type of tool.

Strong AI might still be possible, but if my hunch is true, then it would have to wait for some future type of hardware which is not simply a faster Turing machine, but a completely different computing paradigm.


Get every new post delivered to your Inbox.

Join 40 other followers