Re: How Robots Will Steal Your Job

[Calum]

Well, they can, but someone has to make the program that programs
them.

Those programs could be considered part of the initial program. And
they wouldn't be solving any _new_ problems, just ones the original
programmer anticipated.

I am sure you can make a self-modifying program that is sufficiently
complex that you cannot predict how it is going to end up. In this
case, will it be correct to say that it was _you_ who created the end
product?

I suppose you could have genetic programming - whereby you have a pool
of programs each recombining, getting randomly mutated, trying to solve
a problem. The problem is that you need some kind of fitness function,
and being one bit away a solution could render a program worthless. The
chances of getting a working solution out of such a system are quite small.

You may as well just start at the number 0 and test every possible
program you can generate to see if it solves your problem. But there
isn't enough time to do that.

I think in the future, computers may be able to exceed their original
programming? But I can't imagine how! The question is as much _what_
to solve as how to solve it. With an animal it's easy: how to
locate/catch/disable/open nutrition, how to escape cage, how to woo
mate, how to kill rival, how to locate/build home. There is no logical
reason to solve those problems, but without those urges the animal would
die out. What problems would a computer decide to solve???

Calum

 

[Richard Heathfield]

Here's [yet another attempt at a working definition of intelligence]:

The ability to obey the Three Laws of Robotics.

This would necessitate some rewriting if we were discussing, say, dolphin
intelligence ... "nor by inaction allow a dolphin to come to harm", etc,
but its original form would be good enough to cover both humans and
robots/computers built by humans.

Note that I am associating intelligence with the *ability* to obey the
three laws; that doesn't necessarily imply a *willingness* to obey them,
nor the *necessity* of obeying them. For example, I am (reluctantly) ready
to concede that the UK Prime Minister, Mr Blair, qualifies as
"intelligent", even though he has undoubtedly broken at least two of
Asimov's three laws during the course of his political career.

Can anyone obey the three laws of robotics?

Probably not, in which case we must all be dense.

It appears to me that, even
if we could agree on the definitions of "harm" and "human" (and we
can't) situations would arise in which any action or inaction would
change which individuals were harmed. In which case some individual
could always claim to be harmed by the action or inaction.

Positronic robots dealt with that by going for "least harm" - which, you
could argue with plenty of justification, is not the same as "no harm".

In at least one story, Asimov played games with the definition of "human"
too. (Was it "The Tricentennial Man"?)

 

[Les Cargill]

Les, I couldn't parse the above sentence (on
which the next sentence seems to depend).

Srorry. I keep thinking eveyrbody knows vi .

So substituted, it would then read.
Intelligence:
The ability to choose in every commonly encountered
type of situation an executable action which produces
an optimal consequence.

All it does is use a slightly more measurable
verison of the word "desirable".

 

[Les Cargill]

If you're looking to maximize satisfaction over the long term, then choosing
the optimal action in every state won't always be your best net...there's
tonnes of stuff about this in the reinforcement learning literature...

Fred.

That's one thing I like about my version - it doesn't specify the
domain over which the choice is omptimized

 

[Les Cargill]

Here's mine:

The ability to obey the Three Laws of Robotics.

This would necessitate some rewriting if we were discussing, say, dolphin
intelligence ... "nor by inaction allow a dolphin to come to harm", etc,
but its original form would be good enough to cover both humans and
robots/computers built by humans.

Note that I am associating intelligence with the *ability* to obey the three
laws; that doesn't necessarily imply a *willingness* to obey them, nor the
*necessity* of obeying them. For example, I am (reluctantly) ready to
concede that the UK Prime Minister, Mr Blair, qualifies as "intelligent",
even though he has undoubtedly broken at least two of Asimov's three laws
during the course of his political career.

Dunno. As I read Asimov, the Three Laws were intended as constraints on
already intelligent "beings". They're an ethical construct, not a
description of "how to" or "is it".

 

[Roedy Green]

(I would appreciate criticisms and improvements of
this definition.)

You need something to measure the degree of intelligence in your
definition. By your definition an electric eye to open a door might
be considered intelligent since it decides whether to open the door or
not.

 

[Roedy Green]

The ability to choose in every commonly encountered
type of situation an executable action which produces
an optimal consequence.

that's a bit strong. We consider humans intelligent and we humans
rarely hit on the optimal solution to a problem.

In a Darwinian sense, it means finding solutions that are good enough
for survival within the given time constraint.

 

[Roedy Green]

By
this definition, computers are not intelligent because they cannot
program themselves.

In genetic algorithms, computers change their own programming. Neural
nets constantly tinker with their own programming.

What we are reserving for humans is the task of framing new problems
that would be interesting to solve.

Even that is not strictly human any more. Some theorem proving program
go looking for interesting theorems to prove, then prove them.

We used to think we were special because we were alive. DNA
sequencing squashed that bit of pomposity. We still have pockets of
pride left based on the fact we are human. Yet this superiority is a
temporary accident, not some fundamental feature of the universe.

We are probably one of the stupidest creatures ever to walk the earth,
in a Darwinian sense. No other creature managed to destroy its
environment or create so many way to make itself go extinct in so
short a time. We likely will not last another 100 years. We have
simply created too many avenues for our demise. You get a multiplying
odds effect.

I see our best hope as artificial intelligence that is capable of
overpowering man's primitive wetware, designed for tribal warfare and
intra-tribal competition. If we are lucky, it will force us to
survive. On the other hand, it may decide man has to go to save life
on planet earth. On the other hand, it may have no sentimentality
about carbon-based life, and set about eliminating it all and
replacing it with designed life forms based on silicon, germanium etc.

 

[Roedy Green]

I am sure you can make a self-modifying program that is sufficiently
complex that you cannot predict how it is going to end up. In this
case, will it be correct to say that it was _you_ who created the end
product?

I remember back in the early 70s when my high voltage transmission
line program started doing things, "developing a personality" that I
DID NOT CODE INTO IT.

It made my hair stand on end. The "personality" emerged from the
thousands of trigonometric equations and decision rules I put into it.

This has lead me to speculate that many of the higher abstract traits
we so highly value in humans are emergent properties of our neural
wiring. Therefore we can expect similar wooly things to magically
appear when we start creating electronic analogs.

 

[Roedy Green]

Those programs could be considered part of the initial program. And
they wouldn't be solving any _new_ problems, just ones the original
programmer anticipated.

To be fair, if you put a human in an isolation cell, he would not
likely come up with thousands of interesting problems to solve.

People need input too.

 

[Les Cargill]

that's a bit strong. We consider humans intelligent and we humans
rarely hit on the optimal solution to a problem.

You have to consider "optimal with resepect to what". If thing
A hits on a better solution than thing B, I'd say A is more
intelligent than B.

Intelligence means going beyond the "good enough" level of
solution, in other words. If you don't then the definition
doesn't dovetail with Markov models of evolution.

In a Darwinian sense, it means finding solutions that are good enough
for survival within the given time constraint.

Actually, not so much survival as the ability to procreate.

 

[Alf P. Steinbach]

Yep.

Nope.

(OK, you can tell I'm bored out of my mind since I'm replying to
cosmology questions on comp.programming!)

For an infinite universe you consider just the size of any particular region
of space, defined wrt. an inertial point. At Biggy Bangy time that region
is zero size, which is not very meaningful, and which is why in that view the
Biggy Bangy is considered a singularity. Conceptually I think it's much more
useful to take the opposite view of shrinking matter; in that view the
region stays fixed size while matter shrinks; looking back in time matter
expands, and at Biggy Bangy time (the singularity, which cannot be achieved)
any little speck of matter is infinite size wrt. the region size. Note: it
doesn't matter for that particular problem whether the universe is infinite
or not. The little speck will be infinite at Biggy Bangy time no matter what.

One reason this viewpoint is, to me, so much more useful is that it's then
easy to visualize a sub-region where the rate of matter shrinking is larger;
corresponding to a bubble universe. Then we can think of such bubbles
colliding. Or what happens when stuff escapes from such a sub-region: is it
then still smaller than stuff in the surrounding space, or not?

Of course there are problems no matter which conceptual view is used...

The main problem being: what is the "measurement rod", what is all this
expansion or shrinking relative to? But that's not a problem of needing
infinite time. It's just a problem of scientists evidently putting off that
little question for later consideration, and using absolutely mindblowingly
silly explanations (it's relative to galaxies!) in popular accounts...

Which is why a large body of cosmologists refuse to use the word
'infinite' in relation to the universe. Most of them (a vast majority?)
seem to agree that the universe is bounded.

I don't believe they can be that ignorant. If you have a bounded universe
then you can put two great circles across it, in parallell, with some relative
speed (all in the abstract). Then you can either ditch special relativity or
you can go back in time as much as you want.


Cheers!

 

[George W. Cherry]

Srorry. I keep thinking eveyrbody knows vi .

So substituted, it would then read.
Intelligence:
The ability to choose in every commonly encountered
type of situation an executable action which produces
an optimal consequence.

All it does is use a slightly more measurable
verison of the word "desirable".

Intelligent behavior--given an agent's bounded rationality
and imperfect information--aims at "good enough" solu-
tions, not optimum solutions. (Although, sometimes, op-
timum solutions are both possible and essential, as in the
Apollo Lunar Module's guidance and control system. I
designed the ascent and descent algorithms and the LM's
digital autopilot, and they were on the money optimum;
but such solutions are often impractical and/or unnecessary.)
I could not conceive or design an optimum investment strat-
egy. (I wish I could!)

George W. Cherry

 

[George W. Cherry]

that's a bit strong. We consider humans intelligent and we humans
rarely hit on the optimal solution to a problem.

Right. As Nobelist Herbert Simon famously said: we aim to
"satisfice", to define a "good enough" solution.

In a Darwinian sense, it means finding solutions that are good enough
for survival within the given time constraint.

Exactly. One must act "in time": the "optimum" action--
executed too late--is worthless.

George W. Cherryy

 

[George W. Cherry]

You have to consider "optimal with resepect to what". If thing
A hits on a better solution than thing B, I'd say A is more
intelligent than B.

It's not a "one time" matter. Intelligence is demonstrated
over an extended period in a number of situation types.

Intelligence means going beyond the "good enough" level of
solution, in other words. If you don't then the definition
doesn't dovetail with Markov models of evolution.

"Markov models of evolution" ????

Actually, not so much survival as the ability to procreate.

Well, Les, survival is a prerequisite for procreation, n'est pas?

 

[George W. Cherry]

You need something to measure the degree of intelligence in your
definition. By your definition an electric eye to open a door might
be considered intelligent since it decides whether to open the door or
not.

In its limited domain, the automatic door may operate
intelligently. For example, the engineers must brain-storm
the situations that can occur: what to do in the case of
a power failure, the approach of a very small creature
like a squirrel, whether to allow an over-ride by a button
push, how to handle a very windy day, what to do when
there is a fire in the building, and so on.

if (situationA) {actionA;}
else if (situationB) {actionB;}
else if (situationC) {actionC;}
else if (situationD) {actionD;}
else if (situationE) {actionE;}

George W. Cherry
http://sdm.book.home.comcast.net

 

[Richard Heathfield]

I remember back in the early 70s when my high voltage transmission
line program started doing things, "developing a personality" that I
DID NOT CODE INTO IT.

Write Langton's Ant. It has a distinct emergent personality that you don't
code into it. Give it an environment, and watch it adapt!

It made my hair stand on end. The "personality" emerged from the
thousands of trigonometric equations and decision rules I put into it.

The emergence is not surprising. What /is/ surprising is the behaviour of
your hair. Were you not expecting emergent behaviour from a complex system?

 

[Sir Charles W. Shults III]

Well, they can, but someone has to make the program that programs
them.

I am sure you can make a self-modifying program that is sufficiently
complex that you cannot predict how it is going to end up. In this
case, will it be correct to say that it was _you_ who created the end
product?

Making such a program is nearly trivial. Conway's Game of Life is proven to
be unpredictable in the sense that you can only see what the outcome will be by
running it. In the same vein, you can easily make a program that shares this
property. Therefore, making unpredictable results in self-modifying code can be
almost guaranteed, given a little unsupervised running time.
As for the question, "who created the final outcome", that is for the
philosophers. It is true that you made the efforts that produced the program
that wrote the program, and if it is done on a computer using pseudo-random
number generators, then it was wholly and completely deterministic- so, you are
directly responsible. But that does not mean that you envisioned the end
product!
Sort of like dropping a jar of jam from 10 stories up and knowing what the
splat would look like.

Cheers!

Chip Shults
My robotics, space and CGI web page - http://home.cfl.rr.com/aichip

 

[Gerry Quinn]

Gerry Quinn wrote:

Positronic robots dealt with that by going for "least harm" - which, you
could argue with plenty of justification, is not the same as "no harm".

Certainly has more scope for debate, given that such things tend to be
unpredictable!

In at least one story, Asimov played games with the definition of "human"
too. (Was it "The Tricentennial Man"?)

Don't know - I've never actually been a big fan of Asimov.

- Gerry Quinn

 

[Les Cargill]

It's not a "one time" matter. Intelligence is demonstrated
over an extended period in a number of situation types.

Right, but intelligence would be a measure from
results of tests iterated over all such "comparisons".

If "good enough" is the only standard, then the state space
of the thing being measured has a bound at one end. I suspect
that's a problem.

Given person A and person B, if A consistently chooses a
closer to optimum performance than B, it seems B is
less intelligent than A. And we're kind of conflating
intelligence and evolutionary fitness, which I'm not sure
is correct.

I thought this was a Stephen Jay Gould hypothesis? Can't
remember where from.

If intelligence is modelable as a continuous distribution,
I believe that the measures it's distributing have to be
well ordered. Intelligence isn't the same thing
as evolutionary fitness.

"Markov models of evolution" ????

I don't mean strictly Markov, but each mutation can be analogous
to a "game move", moving the animal closer to or father from
optimum. I'm using the term "optimum" because things like
sharks haven't changed in millenia - looks like a
stable species to me.

Well, Les, survival is a prerequisite for procreation, n'est pas?

Not always. There's hysteresis there - sometimes the
procreation is at the expense of the survival of the
parent(s). This is especially true with bacteria and
very small animals.