On Sun, 14 Sep 2003, Nadav Har'El wrote:
> On Sat, Sep 13, 2003, Orna Agmon wrote about "Re: [hackers-il] Book Review: The Age of Spiritual Machines":
> > > Here's my review of Ray Kurzweil's 1999 book
> > >
> > > "The Age of Spiritual Machines -
> > > When Computers Exceed Human Intelligence"
> > > accelerates in its current pace, is that around 2020, a $1000 computer will
> > > have the computing power of a human brain. Very quickly afterwards the
> > > computer "intelligence" will surpass those of humans. In the following decades
> > Computers are indeed getting faster and with more memory, but is this
> > really comparable to the human mind? Ever since the 60's (at least) they
> > have been trying for AI, but I do not think the development in this field
> > can be compared to the physical development of computers.
> > Does he say anything about how he thinks this will happen?
> People who plan to read this book very soon might not want to read on, because
> I'll be spoiling some of your surprises ;)
> Yes. Kurzweil concedes that having a computer as strong computationally as
> a human brain doesn't necessarily mean that we'll have the software as good
> as the "software" that "runs" on the human brain.
> His response to this attack is two-pronged:
> First, he claims that together with the improvement of computing power, we're
> going to see improvement in other relvant technologies. One of these
> technologies is scanning human brains using MRI-like technologies with ever-
> increasing resolution (what he calls "bandwidth"). By 2020 Kurzweil predicts
> we could destructively scan the entire structure of a (dead) human brain, and
> by 2050 we will be able to non-destructively scan a live human brain (I'm
> writing this from memory, so perhaps I got the dates wrong). Scanning the
> complete structure of a brain will supposedly allow us to replicate exactly
> the "mind" of a person on a computer, which is one way of making a computer do
> what a human can do. Such scanning will also allow scientists to better
> understand how the brain is built, copying "algorithms" from it (like how to
> do face recognition, how to read, how to understand language, etc.) into
> artificial neural networks on a computer.
> One really interesting observersion Kurzweil makes is that the "algorithms"
> in the human brain are much "smaller" than appear in first glance. He estimates
> the amount of DNA that specifies the human brain to contain only 10 MB of data.
> Yes, something 1/10th the size of Open Office ;) How can 10 MB of code specify
> something like our brain that contains thousands of times more information?
> The "trick", Kurzweil says, is that a fetus brain starts out with a lot of
> random neural connections (this, of course, doesn't need any "data" to specify)
> and appropriate algorithms to build correct connections based on input data
> the baby gets in the very first years of its life. Similarly, we might
> theoretically build a computer that has a few-megabyte-large program and then
> goes on to listen, see and read, like a child normally would, until it built
> the knowledge of an adult human. But how do we write this 10MB program?
> Understanding the human brain might give us some ideas. Evolution has worked
> on it for a lot more years than our human programmers can spare ;)
> Kurzweil's second reponse to your "attack" is that "AI" is a moving target:
> Whenever a computer can do something it couldn't do before, we suddenly say
> this is not "real" intelligence. For example, a computer can now beat the
> world chess champion and it couldn't do so in the 60s. Did we conclude that
> computers have become smarter than humans? No, we concluded that playing chess
> doesn't require intelligence :) Similarly, computers can now read written
> text (OCR), understand words spoken to it ("continuous speech recognition"),
> translate texts (as varying degrees of accuracy), create music and paintings
> of certain complexity, and other stuff they weren't able to do in the 60s.
> Right, computers still don't pass the "Turing Test". But the Turing Test
> basically requires a single computer to have all the faculties of the human
> brain - understanding and producing language, recognizing patterns, memory
> and knowledge of the world, the concept of "self", emotions, sense of humor,
> and so on. Would a computer that knows how to do just one of those things,
> or just a few of them be "intelligent" or not?
> Another thing to remember that in 2020 computers will be (according to
> Kurzweil's predictions) as strong (computationally) as the human brain,
> but in 2030 they will be 1000 stronger. Given such huge margins, it is
> conceivable that even lousily-designed software we put on these computers
> will appear to be as intelligent as a human.
> You might like to read Kurzweil's original arguments, rather than my "broken
> telephone" (how do you call that in English?) version of his arguments.
I think, out of all those tasks that a computer needs to be able to do in
order to pass the Turing test, the humor part is the most difficult. Humor
changes between places and times. I wonder, what defines humor?
It ranges from sharp tongues (??) such as word games to slapstick, and we
recognize it all as humor.
Can anyone draw a plan as to how to teach a computer to laugh? Say we
define laugh as print "LOL", and define smile as print ":)". How would a
computer know when to print any of those, and when to operate an Eliza
I do actually classify OCR and speach recognition under AI.