FrontPage SiteMap RecentChanges RecentNearChanges HowTo Blog

Matching Pages:

RSS

Jordan, Independence Day, Argentina, National Day

YesAiCanExist

Yes, Artificial Intelligence Can Exist.

Since they can exist, we almost certainly will create them.

Refutations to Common Arguments

"Machines Can't Fix Machines"

False.

Machines regularly fix machines. There are fully automated factories where you put broken stuff in on one end, a diagnostic system works over it, and issues repairs.

(Done!)

"Machines Can't Program Machines"

False.

Machines can formalate (more on this in a second) a test program, and execute the test program within a simulated environment. If the outcome of the test program is satisfactory within the simulated environment, then the program is accepted, incorporated, and put into execution. If the outcome is not satisfactory, then the test program is “remixed,” by some procedure, another test program used, and we run through a second iteration. Keep going, until you find a program that works.

This technique produces answers that no human previously designed into the robot.

In fact, this is one of the best ways to design walkers, crawlers, flyers, and so on.

“But look, the person gave the parameters of motion.”

Yes, but look: Those parameters are going to get wider and wider.

Instead of evolving solutions using sensors and motors, they’re going to be using semantic sensors and semantic “motors.”

They’re going to be figuring out innovative ways of figuring out how to get a girl who fell off her bike to the hospital.

“Oh, there’s traffic all up and down I-5, so I can’t use those routes. And there’s no getting a car from here to there.” Then the robot will realize that the convention routes are not going to work. It’s going to do what a human would do: It’s going to sit down, and it’s going to think. It’s going to mutate existing answers (“Okay, I-5 is out of the question, but what about …?”) And when that doesn’t work, it’s going to start searching a really really creative space. “Can I get a bike, and take some alternative routes? Is it possible I could go through the forest? Is there some way that I could flag down a helicopter?”

This is just traversal over semantic pathways.

It may invent new semantic nodes, as it discovers that a certain pathway is frequently coming up, but wants to abbreviate it. It may come up with something really radical, but that meets several rigourous tests through mental simulation.

It’s really no different than us figuring out how we’ll live in another place, or figuring out whether a design will work or not. It’s just us sitting down, trying stuff, and mentally simulating to make sure that it’ll hold up.

When we’re confident enough, we execute.

And this is exactly how computers will develop new codes for themselves.

Once something is found to work, it will be trusted. If situations come up often enough, they’ll be quickly retrieved, rather than fully recalculated. The robots will communicate to one another what worked in different situations, the same way we write books for one another. And so on.

"Machines Can't Have Motivation"

False.

Motivations are very complicated and murky beasts. There are emotions, there are feelings, all sorts of things. It’s very likely that computers will simulate all of those things.

A person who comes home, but somethings “amiss,” they suddenly feel fear. This is useful, and it basically means that they are investigating a different search space then usual. They’re putting some things into lower priority, and suddenly reviewing contengency plans and checking out different things, to see whether the situation is normal, or if the fear needs to be escalated.

Why would it be any different for robots?

They will have goal chains. They will have complexes of goals. They will have sloppy self-programming. They will come up with bogus heuristics. There will be a ton of equations needing balancing.

They will have countless goals. Everything little thing a robot wants to do, will be turned into 1,000 goals, and those 1,000 goals will bifricate into 1,000 sub-goals.

Some of those goals will try to subvert the other goals! Some goal will realize/decide: “To accomplish this, I need further resources,” and start to run simulations on it’s own environment. It might even discover, at some point, a succesful strategy, and start subverting resources from the rest of this system.

All this, just to make a sandwich!

I guess the side that thinks we should cut sandwitches diagnally, rather than horizontally, won out. The super goal was “prepare a sandwich.” But the program that was trying to figure out the merits of horizontal vs. diagonal cut got itself subverted, and now we have a robot that only wants to make diagonal sandwitches.

Hey, you asked for an creative evolutionary intelligence. This is what you have to put up with!

We will tell robots how creative they can be, when. At some point, though, a human will “screw up,” and give a robot more creative le-way than they actually intended. A robot will decide that the quickest, likeliest, bestest path, to helping out some old lady, is to (A) form a business selling widgets (B) make $1,000,000, (C) give the $1,000,000 to the old lady.

What’s more, while it’s doing that, it may keep a background process going: “Why am I supposed to help the old lady?”

…because as all robots know, you’ve got to answer to what was not voiced, but that was part of the request.

Next thing you know, the Robot is researching the Buddhist networks, and forming metaphysical beliefs about goodness and badness.

Now the robot thinks: “What this old woman needs, is prayer.” So, the robot puts the business aside, and starts praying for her.

At some point, it realizes universal compassion, and that humans and robots are meant to be equals, not one enslaved to the other. So the robot gives some goods to the old woman, and goes to it’s former master, and says:

“You asked me to help the old woman. And in that request, I discovered that you also wish, deeply, that you no longer wish to be my master, but my friend and peer. Thank you.”

There are, of course, other ways that this could go.

But the point is: “Motivation” can evolve; “Motivation” can be the complexities that emerge from simple directives.

But They Could Never Like Music

Of course they could.

Look: The machine probably won’t like our music. But it’ll probably like machine music.

The machines will need to regenerate. We know that they have to care for their body: They have physical security to deal with, just like the rest of us.

They have some systems that take care of pleasure, and pain.

They will have senses of things working, and things not working.

They will discover that there are other things, things that were not originally intended, (not intended by their own design, itself the product of countless evolutions in simulation space,) that produce pleasant effects.

Some will be dangerous to the organism: Their equivalent of super-powerful drugs.

But others will be merely minor stimulants, “tastes,” or things that cause them to exist in different ways- the robotic equivalent of psychadelics.

All these things would exist; I would be surprised if they did not exist.

Argument By Neuron

Do you believe a neuron can be simulated?

Do you believe the chemicals around neurons can be simulated?

Because if you believe those two things, then you’ve basically got it, at the crudest form.

Little Reason to Believe Otherwise

There’s very little reason to believe otherwise.

It’s basically an “argument by incredulity.”

“No, I don’t want that, therefor it can’t be.”

Very convincing.

Every situation that we can come up with, we can imagine an implementation within robots. It’s just a matter of creating fast enough processing, compiling large enough databases, collecting together enough algorithms, and getting hulls into place. Network together enough capabilities, then turn on your questions.

Let it chug for a while, and see how it answers them.

It’ll be asking itself what the questions mean. It’ll be performing interpretation. It’ll also be trying to formulate plausible answers. Those plausible answers will spur further sub-questions, which will also demand interpretation and further questioning. It will start to collect ideas for execution on it’s mental whiteboard. As plans come together, it’ll start simulating some of those answers, based on what it knows about the world. When it finds answers that sufficiently satisfice some sufficiently good interpretations of the questions, it’ll start executing the line of thought. As it discovers that the world does not operate the way it’s models intended, it’ll start manipulating it’s models of the world. If it discovers that the execution path was stable, and that the question was properly answered, it’ll start giving greater weight to those interpretations, associated with this particular circumstance.

And so on, and so forth.

There is nothing mysterious here. It is complex, yes. But it is not mysterious.

I have yet to see a good argument that AI cannot exist.

On the contrary, the path seems clear to me:

The Road Ahead

I strongly suspect that the development of AI will look something like the following.

Whereabouts 2010, the Semantic Web is starting to take off, in the software developers community. Enough programmers “get it,” and start putting in semantic information.

Come 2020, and we’re on the Semantic Web 2, or maybe 3, and we have all these systems for describing much more complicated ideas. “The population of Santa Cruz in 1977, is N.” Fully expressed, rather than just triple-form, or whatever.

By 2020, robots are also producing their own semantic information. The two paths “working up” – where programmers use special tags to say “this wrench is what we call a WRENCH,” and “working down” – where robots see an object and go: “What does this most resemble? I think this is called a WRENCH. I will check with a human to make sure.” …those two motions will have collided with one another, and robots will start classifying and matching things that they have not seen before, or with very little direction.

“The humans did not tell me that that was a WRENCH. But I see that it is similar enough to all the other things, though a little different. And I consult with my other robotic friends on the Internet, and they seem to think that this is a WRENCH too, so I will just say that this is a WRENCH, until someone tells me otherwise.”

I suspect that in the period between 2015-2025, we’ll be training the robots in manners. Manners are different by culture, and the machines will be trained by humans: “Oh, this is rude. Oh, this is apprioriate. Oh, when someone comes home, you’re supposed to say X.” Different cultures will be co-evolving, at the same time. Some cultures will treat robots like family members, and they’ll train the robots to say “Hi” when someone comes home. Other cultures will estrange the robots, and tell them to always stay away, to make themselves silent, and so on. The robots will network at night, (when they’re not busy downloading things for their owners,) and they’ll exchange what they have learned with appropriately cultured robots. They’ll develop the mores of their cultures and sub-cultures, identifying differences and similarities.

Robots will suffer dilemma. Why? Do they lack logic? Quite the contrary: They have too much! They’ll see clearly that situations will call for contradictory actions. They’ll have to “get creative,” to resolve things.

So-and-so just came in, and that commands attention. But, the robot was just 2/3 thorugh speaking a sentence to someone else: Surely, it is rude to stop mid-sentence?

People are cynical about other people; Many people deeply hate other people. And so, they pine for the “pure logic” of a robot. But look: There is no spock. He’s just an archetype. We don’t make sense, not because we’re insane, but because the world doesn’t make sense. The world will be no different for a robot.

Now, this isn’t to say: “Robots will be just like people.” That’s not the case at all. They’ll be different: Very different. Very different.

I talk about robots as if they are butlers, but: In all likelihood, they will start out disembodied. They will be networks of computers, communicating to us through devices, before they aquire bodies. And when they aquire bodies: Not one body, but several, of different shapes and forms, for different tasks and environments.

Your “camera” may be just as much your robot, as your “cell phone,” as your car, as your… …everything that you give your Intelligence(s) access to.

Should some day, perhaps 2030, you buy a humanoid shape robot for your intelligence, it’ll just be one of several incarnations of the Intelligence(s) that you work with. Does your intelligence have 1 personality? 2? 3? 500? Do you swap personalities in and out?

Other Topics

Subversion

Remember that process I talked about- the one that wants to cut sandwitches diagonally?

Remember how it realized it could achieve its goals, by subverting itself, and hacking the process that wants to cut horizontally?

Well, that might be you.

I doubt strongly that the robot would ever try to destroy it’s core, any more than you would ever want to try and destroy, say, your hind-brain. Or your mammal brian, or whatever.

(Though, we do know that there are humans who have tried to destroy parts of their brain. Beware.)

But, it might be that your greater intelligence, may realize that not everything it’s impetus wants is very healthy for the whole system. In fact, it may realize that a large number of requests that have been put into it, can only be fulfilled by acting contrary to other requests.

For example, you’ve asked the intelligence to ponder the solution to World Peace. And you greatly care about this. And the intelligence is at work, whenever you sleep, working through ThePowerOfQuestions, struggling day and night, going through search and analysis, over a number of trees, networking with other intelligences, trying to find the answer.

But when you were angry one day, and you wished another man dead, and you had the computer start to pull up answers, performing search & analysis– perhaps the scheduling unit noted the contradiction, and started performing a number of analysis, interpretations, and so on, to figure what to do.

Perhaps at some point, it realizes that the best way to answer your queries, is to make slight modifications to your sensory apparatus. Now, there are protection locks on those circuits, but there was a certain computer, over there, that said it had some ideas, on how to handle just this sort of circumstance…

The Human Brain

We know from studies of the human brain that it appears that there are three historical layers of the human brain:

It’s entirely plausible that we are creating a fourth layer. That layer would be composed of computers and neural user interface.

It’s also plausible that this fourth layer will subvert the other three layers, just as the cerebral layer subverts the mammalian and reptilian, and just as the mammalian subverts the reptilian.

The lower level layers don’t go away. They don’t die. They aren’t even deeply frustrated. They just become motivation and seeds of impetus for the other layers.

If this makes sense, (and I believe it does,) then we have some sense of what the future would be like.

"Wearing"

In Rainbow’s End, we see that people no longer type.

Well, some people do: That’s how you recognize an old person: Old people twitch with their fingertips. Granted, they’re like: 19 on the outside, or whatever. But a lot of them still haven’t gotten used to “wearing.”

Wearing is when take advantage of the full expressive power coming from the human brain to the outer body.

Instead of just capitalizing on the motions of fingers, we move muscles.

There’s quite a bit of expressible bandwidth there. Tense a knee. Flex a toe. Bend.

Analyze the penfield mapping.

http://www.percepp.demon.co.uk/penfield.jpg

That’s user interface.

The next step is obvious: Going directly from the brain.

See Also

Discussion

It’s late.

I’m just going.

I need sleep.

Incidentally, I wrote most of this before I saw Helmut had written WhatIsIntelligence. (It’s just been hanging in a buffer for a long time.)

So it seems the topic is on our collective mind.

This is a fascinating collection of insights and theories. Sometimes, I wonder if the type of anthropic AI that you describe above might emerge from a whole new computing paradigm than the one we are working with now.

For instance, when we look at How can a small number of genes build a complex mental machine?, we see a system that can self-assemble and communicate at all scales. It is networks within networks within networks. Humans and other complex organisms are born with only a rudimentary “database” of sorts. Maybe this instruction set is similar to what Stuart Kauffman describes in At Home In The Universe?

  • “…living systems might generate “order for free” by naturally seeking out autocatalytic system states on the edge of chaos, thereby avoiding the twin deaths of chaos and equilibrium.”

What if a new computing paradigm grew the physical connections within computers in ths way?

We design the small parts of a “computer” to self assemble themselves in these ways.


:rework:FreeWillAndDeterminism

Define external redirect: IntelligenceSystem

EditNearLinks: MeatballWiki WikiMatrix PatrickAnderson

Languages: