Did Web Search kill Artificial Intelligence?
Matthew Hurst writes (in part):
…, we currently have the following:
- Search engines that don’t understand language and which attempt to mediate between people (searches by people and documents by people),
- The best and the brightest coming to work for document oriented web companies.
I can’t help but wonder where the AI project would be today if web search (as it is currently envisioned) hadn’t gobbled up so much bandwidth.
No doubt it would be different, i.e., more papers, more attempts, etc., but all the resources devoted to the Internet would not have made a substantial advance in AI.
Well, consider that the AI project has been in full swing for over sixty years now, if not a bit longer. True enough, there are scanning miracles that have vastly changed medicine, research in a number of areas, voice recognition, but they are all tightly defined tasks that are capable of precise description.
That cars can be driven autonomously by computers isn’t proof of the success of artificial intelligence. It is confirmation of the complaints we have all made about the “idiot” driving the other car. Granting it is a sensor and computation heavy task, but with enough hardware, it is doable.
But the car example is a good one to illustrate the continuing failure of AI and why the Turing test is inadequate.
First, a question:
Given the same location with the same inputs from its sensors, would a car being driven by an autonomous agent:
- Take the same path as on a previous run, or
- Choose to take another path?
I deeply suspect the answer is #1 because computers and their programs are deterministic.
True, you could add a random (or rather pseudo-random) number generator but the program remains deterministic because the random number generator only alters a pre-specified part of the program. It isn’t possible for variation to occur at some other point in the program.
A person, on the other hand, without prior instruction or a random number generator, could take a different path.
Consider the case of Riemann geometry. The computers that generate geometry proofs that humans select as significant, isn’t capable of that sort of insight. Why? Because there is a non-deterministic leap that results in a new insight that wasn’t present before.
Unless and until AI can create a system capable of non-deterministic behavior, other than by design (such as a random number generator or switching trees, etc.), it will not have created artificial intelligence. Perhaps a mimic of intelligence, but nothing more.