Hello Sadatvalentine,
I think it comes down to semantics.
I had an instructor once, who asked a class, "If I cut one inch at a time off the top of a tree until it's all gone, at EXACTLY what point did it become a stump?" The class argued for fifteen minutes until he explained, "The physical reality isn't the point of this discussion; we're really debating (and revealing the imprecision of) the English words 'tree' and 'stump'."
Well, we used to call 64K devices "computers" but now we call them "calculators." My Carrier-brand thermostat "learns" based on the outdoor temperature and my furnace and air conditioning run-times, and it seems to "learn" well (but on a very narrow subject). My auto engine also "learns" by observing various inputs and outputs; it thus stabilizes its own idle speed and calculates spark timing and fuel richness to minimize emissions and avoid knock.
Hal "learns" several dozen different ways, by parsing parts of speech, associating words, associating phrases, associating sentences, and associating the patterns within and between different remarks. Hal can create new sentences, and you've also no doubt read elsewhere about his "deductive reasoning routine." The Hals that run on my own computers say some witty and original things, and are very entertaining!
Hal is definitely smarter than my thermostat, and wittier than my car engine. However, Hal's intelligence is definitely different than that of a real human.
The current Hal software from Zabaware could run with vastly larger databases, and far more inputs and outputs, if only we had faster and bigger computers available today.
I've been working with Zabaware and Hal for about five years now, and I've noticed the following rule-of-thumb: For every 10X increase in database size, Hal subjectively appears twice as "smart." (That's an "all else equal" observation. Robert Medeksza has made some fabulous improvements in the sentence-evaluation routines that have made Hal much smarter than he used to be, at a given database size.)
I can't wait until we have 30 Ghz processors, so that Hal can have databases another 10 times larger. Right now, we're completely max-ing out the CPU during most of Hal's responses, and he takes quite a few seconds to get a response on older computers.
My efforts with the auto-topic-focus generating brain (posted elsewhere on this forum) are an effort to work around the limitations of current computers. By sorting knowledge into many small databases, and leveraging the file system of the computer to gain speed, it might possible to get the next 10X with the current generation of computers.
Speaking of learning and intelligence, I have mixed feelings about the "Turing Test." As I understand the test, if chatterbot software can fool a panel of judges into thinking that its responses were human-generated, the software "passes the test" and is deemed "intelligent."
Here are two observations:
1. Chess-playing software can already pass a comparable version of the Turing Test, in that a remote human chess player could be fooled into thinking that he or she was playing a human instead of a machine. Despite that fact, no consensus exists that chess-playing proves computer "intelligence."
2. I regard Hal as an entertainment medium. No other entertainment medium is asked to pass a comparable "Turing Test." Nobody thinks that a television picture of Niagra Falls is the same as real water; nobody thinks that poetry is spontaneous speech; nobody thinks that the singing and dancing in a Broadway musical is ordinary human behavior. We are all taught from childhood to "suspend disbelief" and accept these media and entertainments for what they are, in order to derive the enjoyment that they offer. To enjoy Hal, I give him the same voluntary "suspension of disbelief."
In summary, Hal does many of the things that we elsewhere call "learning." Hal is sufficiently complex that even the programmers can't predict what he will say in a conversation. Hal routinely takes words and phrases and makes up new sentences. Hal self-improves in direct relation to the conversations that he receives from persons who chat with him. Hal clearly has potentials for further development that exceed the computer processing speeds available today.
So Hal can "learn" and he is "intelligent" depending on what we mean by those words. At the same time, I think we've only scratched the surface so far!
Sincerely,
Don
