Re: Turing Test is overrated?


ilya2@xxxxxxx wrote:

No dolphin could ever pass Turing Test. Yet no one who spent some time
around dolphins will deny they are intelligent. For that matter a lot
of *humans* will not pass Turing Test, especially if they are from a
different cultural background than the person testing them. I do not
know what Alan Turing was thinking, but his test is enormously
restrictive.

The question he was trying to answer was can machines think.

http://cogprints.org/499/00/turing.html

All he needed was an existance proof of a singular machine acting as if
it could do one sort of thinking (or be one sort of intelligence). He
didn't need to capture the essence of what it was to think and say only
the things that pass the turing test think.

The intepretation I give it is that he was aiming for something that
would convince a sceptical man on the street that a machine could
think.

Think of a robot (and I am fully aware that such robot is beyond
current capabilities) designed for mountain rescue. It has to navigate
very hazardous terrain, locate injured and possibly delirious people,
evaluate the nature of injuries, supply first aid, secure a portable
shelter on very non-flat ground and possibly in gale-force wind, and do
many other things I had not thought of. Such robot would have to have
cognitive functions which would make it intelligent by any reasonable
definition.

But many people use unreasonable definitions.Turings test is designed
with them in mind.

I'll agree that it has not helped make AI and trying to make
chatterbots to pass it is not helpful either. But it will make a good
headline once someone making an AI based on sound principles, passes
the turing test. Although most workers in AI and believers will think
they have created thinking machine a long time before that.

I don't think we have created a better test that will be accepted by
more people.

Yet it has absolutely no need to *talk*, except to reassure
victims - and that can be done easily with ELIZA-type canned phrases.
(All communications with other rescue units/personnel should be done by
high speed electronic file transfer, not through slow and imprecise
human speech.) Such robot will fare worse at Turing Test than your
average parrot.

One practical application which has both the most use for AI and the
most dollars behind it is military. Yet how much intelligence does a
fully autonomous combat machine (call it Ralph) really need? Ralph has
to react to external stimuli -- something almost any animal can do. It
has to make tactical decisions -- something almost any predator can do.
And it has to learn from experience -- something almost any mammal can
do. Realistically, Ralph needs not be more intelligent than a cat.

I really hope we have more intelligent machines on the battlefield than
cats. I know of no way of telling a cat that grey mice should not be
preyed upon, whereas I will want to tell the machine to not attack my
allies (which may have been my enemies previously).

Ralph should have language-recognition capabilities, because humans
convey information best by talking, but any speech on its part will be
superfluous. Given the level of technology required, all friendly
humans are likely to have head-up displays or some other visual
interface, and Ralph will convey tactical information that way.

While I am very interested in this sort of adaptive system, I also see
the value in getting it it to speak or write on the HUD at least
occaisionally. Especially when it needs to acknowledge or repeat orders
it has been given. Remember the imitation game is played out on
teletype so has no need to actually verbalise the words.

Humans
can absorb visual data MUCH faster than audial, which is one of the
reasons talking computers never really took off. And that's another
problem with Turing Test -- computer which talk like people are not
very *useful*. I do not want an AI which can discuss implications of
overfishing in the Gulf of Maine -- I want one which can put Gulf of
Maine onto a multimedia graph, and give ecology time-projections based
on various fishing policies. Something no human can do, at least not in
real time.

I once heard definition of AI as "getting computers to do things that
people do better [than people]". Even if it is achievable (which I am
not at all sure), I'd be much more interested in artificial
intelligence which is fundamentally *different* from human intelligence
-- as different as a dolphin's, or more. An AI which could perceive
problems the way no human ever could *would* be a breakthrough of
unimaginable (literally!) magnitude -- even if it remained clueless
about some things obvious to humans.

But then you will get humans claiming it is not intelligent. Just as
some people claim monkeys/dolphins etc aren't intelligent.

Will Pearson

.