🚀 go-pugleaf

On Sun, 1 Jul 2018 06:47:13 -0400
Peter Flass <peter...@yahoo.com> wrote:

> Ahem A Rivet's Shot <ste...@eircom.net> wrote:
> > On Sat, 30 Jun 2018 05:46:46 -0400
> > Peter Flass <peter...@yahoo.com> wrote:
> > 
> >> maus <ma...@mail.com> wrote:
> > 
> >>> Are there any reallife AI achievements?
> >>> 
> >> 
> >> Getting closer. Watson beat the champ at jeopardy, which requires a
> >> certain amount of AI.
> > 
> > 	It is interesting what people think of as AI which seems for
> > most people to focus around conversational criteria (winning Jeopardy,
> > passing a Turing test) and far less on non-verbal criteria such as
> > creating a picture from a short description or detecting lies in court
> > or playing video games or diagnosing illness or ... All of which (and a
> > good many more) have been reported as achieving or exceeding human
> > capability in the last year or so.
> > 
> 
> I don't think creating a picture or playing video games qualify as
> intelligence - maybe the picture part.

	Imagine trying to design algorithms to do either of those - there
are neural nets that play first person shooter games better than people
(one was a PHD thesis project!) which is not an easy thing to contemplate
writing code to do. It might not be an AI but I'd hate to meet a descendant
of that code running in one of those robot guns mounted on a spider robot.

	The pictures I have seen included a lot of detail and background
that wasn't in the description putting what was asked for into an
appropriate context. Not something I'd care to attempt to write code for.

	There's an old saying in the AI field - "Once we know how to do it
it's not AI" therefore AI will never be achieved as long as there's
anything people can do that AIs can't - not sure how long that will be,
perhaps forever or perhaps way sooner than anyone thinks.

	Some people reserve AI for either general artificial intelligence
or artificial sentience - neither of which is happening yet - and use ML
(Machine Learning) for everything else. It avoids the poorly defined
concept 'intelligence'.

	What interests me about what people will and will not call AI is
that it really digs into the question of what do we mean by 'intelligence'.
It's going to get even more fun if/when there's something around that
might be artificially sentient - the arguments are going to be wonderful
(but perhaps not from the point of view of the sentient AI whose life and
rights are at stake).

	The way I see it is that if the machine learns how to do it from
examples or practice with or without guidance then it's AI (or ML if you
prefer) if the machine has to be programmed in detail then it's not.

> Computers are getting really good at diagnosing illness. This isn't so
> much AI as an expert system.

	The pure expert system approach was used as far back as the early
1980s, I recall reading a report on a system that was claimed to be better
at diagnosing stomach pains than human doctors - I still wonder why it's
not in every GP's surgery. AIUI the more modern ones are hybrids of expert
systems and neural nets.

> I think they can detect lies, too, but maybe
> not in real time.

	Yes they can viz:

https://futurism.com/new-ai-detects-deception-bring-end-lying-know-it/

	Futurism tends to be somewhat enthusiastic (although the links to
source material usually help with that), but this article is better than
most with quotes from the people who did the work.

> The thing about jeopardy is that (if I got this right)
> is that you have to be able to pick up clues from context which is
> hard(er).

	Yes that's the sort of thing that modern AIs can do.

-- 
Steve O'Hara-Smith                          |   Directable Mirror Arrays
C:\>WIN                                     | A better way to focus the sun
The computer obeys and wins.                |    licences available see
You lose and Bill collects.                 |    http://www.sohara.org/