In the critically claimed novel Neuromancer, William Gibson depicts a futuristic dystopia that he intended as a warning towards our ever increasing dependence on technological advancements. Instead though, to his horror, many of those who read the book were inspired and sought to create a world similar to one he depicts. The physical surgical “improvements” that are prevalent throughout the book can be seen through our development of advanced prosthetics or neural implants. The three dimensional internet he depicts is currently under development, and three dimensional interaction can already be seen in the technology in Microsoft’s Kinect. Even the space city of Freeside already inspired orbital living as shown by the International Space Station. However, one of the advancements that we most likely do not have to fear, or, as some view this future, embrace, is that of a fully independent artificial intelligence.
The driving plot of the story is a direct result of the involvement of the two AI’s, Wintermute and Neuromancer, and yet it may actually be impossible to create a synthetic entity that has this degree of aptitude. Yes, “impossible” tasks have been achieved before, such as human flight or space travel, but the very nature of computers prevents them from achieving this level of intelligence. There are two main arguments that back up my claim, the first being P vs. NP. Without diverging into a tangent about the intricacies of this problem, I’ll try to give you a brief overview about why this is fundamental in AI.
P vs. NP stands for algorithms that run in polynomial time versus those that run in non-deterministic polynomial time. An algorithm is runs in polynomial time can provide all the solutions for a given problem, while one that runs in non-deterministic polynomial time is one that can be verified quickly. As an example, take the problem of finding a set of integers that, when added up, equal 0. This problem is NP, as a solution can either very quickly be shown to be true or false, as with the set of {-1,0,1}. It is impossible however to get every single possible combination of all integers that add up to 0, as they are infinite. The question is can we then find solutions and efficient algorithms for all NP problems, showing that P = NP? This is deemed the most important unsolved problem in computer science, and the Clay Mathematics Institute will even pay anyone that proves or disproves this theory one million dollars.
How does this affect AI? Well, the necessary code that would allow computers to be truly independent is code that runs is NP time. Proving P = NP is then required to even show that there are any algorithms that would allow computers to develop what we view as true AI. This has not been achieved to date, and many computer scientists seriously doubt that P = NP. Until this is proven, then AI can never truly exist.
My second argument against AI is one that has arisen through my own research into artificial intelligence here at UR: computers are incredible stupid. These machines are often viewed as all-knowing entities, but in reality they are incapable of understanding anything other than what we tell them to. Computers are incredible adept and speedy at one thing: pattern matching. When you input 2 + 3 into any kind of machines with a calculator function, that machine simply references a list incredibly fast and returns the pre-programmed pattern that 2 + 3 = 5. The machine is incapable of determining why this is true, just as it is incapable of realizing that 3 + 2 is an equivalent statement. Because they are limited to their original programming, it is very improbable that machines’ intelligence will result in anything more that more and more complex pattern matching. This can be seen even in current programs that appear to be intelligent, namely Cleverbot.
Now, some of you reading this may disagree with me. The argument could be posed that just like with all other developments, the technology just has not been developed yet. This may very well be the case, just as it was for so many other advancements. But these two great hurdles of P vs. NP and the very nature of current computing must be cleared, and unlike other unsolved problems where there lies promise, I do not find AI in the foreseeable future.
Your command of CS theory is beyond my limited skills, but I wonder how the human mind solves a simple problem such as 2 x 3.
ReplyDeleteI cannot see a set of answers in my head, but through memorization of multiplication tables in grade school I put every possible combination of single-digit multiples into my long-term memory.
I doubt that every human who can run through every combination from 1x1 to 9x9 did it that way.
You'll enjoy the Atlantic article about the Turing Test we'll read later in the term. While passing the Test seems likely, the resulting "AI" will be nowhere close to Gibson's, since those machines could improvise and learn from mistakes. I'm not sure how human brains do that either, and certainly some brains do better than others.
That's a job for neuroscientists to investigate, I suppose.