An AI 4-Year-Old In Second Life

schliz notes a development out of Rensselaer Polytechnic Institute where researchers have successfully created an artificially intelligent four-year-old capable of reasoning about his beliefs to draw conclusions in a manner that matches human children his age. The technology, which runs on the institute's supercomputing clusters, will be put to use in immersive training and education scenarios. Researchers envision futuristic applications like those seen in Star Trek's holodeck."

The potential for hilarity is nigh infinite

A fake four-year-old boy running around with a bunch of sadomasochists, furries, "child play" *ahem* "enthusiasts." The making of a brilliant sitcom if I've ever seen one.

In this episode, Eddie's AI gets put to the ultimate Turing Test when he's approached by a Gorean pedophile! Tune in for the laughs as Eddie responds with "I'm sorry, I don't understand the phrase 'touch my weewee, slave!'"

Poor little guy

Imagine if you were born and raised by furries who attached enormous genitals to their bodies and watched simulated porn all day long.

The poor kid never had a chance.

Segmentation faults are murder!

Even one messed up pointer could cause this child to die!

Segmentation faults are murder!

Honestly I wonder about the moral oddities of AI.

duplicate!

http://slashdot.org/article.pl?sid=08/03/13/1914251 [slashdot.org] yesterday!

obvious I know

But if you are letting you AI out into Second Life and comparing it to intelligence there, surely you are setting the bar rather low?

Re:obvious I know

> But if you are letting you AI out into Second Life and
> comparing it to intelligence there, surely you are setting the bar rather low?

Time for an "Office" quote:

Dwight: Second Life is not a game. It is a multi-user, virtual environment. It doesn't have points or scores, it doesn't have winners or losers.
Jim: Oh it has losers.

First Test...

First test: could a 4-year-old rascal recognize a dupe [slashdot.org]?

Not even close

It's a simulation of a 4 year old and is NOT an AI with the cognitive abilities of a mouse let alone a 4 year old human. It's just a very powerful chatbot writ large. Sensationalism strikes again!

Re:Not even close

That's a relief...

Because the thought of a holodeck full of 4-year-olds has to be the definition of Hell.

Re:Not even close

It's just a very powerful chatbot writ large.

Any sufficiently advanced chatbot is indistinguishable from an intelligent being.

(Not to say this is in any way a sufficiently advanced chatbot.)

Re:Not even close

Do you feel strongly about Any sufficiently advanced chatbot is indistinguishable from an intelligent being?

Re:Not even close

Ever spend any time with a 4 year old? They are all little running chatbots.

MG

Re:Not even close

If my 4 year old is any indication, that "total nonsense" is stuff pulled from various sources they've come in contact with. They just don't realize at that age that everyone doesn't know the source just because *they* know the source. (In a similar vein, they don't realize that they can't point to something and refer to it while talking to someone on the phone.)

I can catch about 85%-90% of the references because I've seen the TV shows my son watches. I know when he talking about pressing a button on his remote control or knocking first before going in a door, he's talking about The Upside Down Show (sometimes a specific episode). I know that talk about Pete using the balloon to go up refers to a Mickey Mouse Clubhouse episode where Pete used the glove balloon to rescue Mickey Mouse. Going "superfast" refers to Little Einsteins. They might be mixed up too. Pete might use the remote control and push the "superfast" button.

To an outside observer, though, who doesn't get the references, it's all gibberish, but there actually is a lot of intelligence behind all that chatter.

Re:Not even close

I have this 8-year-old neighbor girl, I know her from when she was 3. At 4 or 5 she used to talk to me for long periods. She tells things she's seen on TV, but she also invents stories and songs, which are total nonsense. These may have a meaning for her, but it's nonsense nonetheless. Of course, I don't say it's bad: it's how their imagination works, and that's a good sign because it means she has an active imagination.

She's a smart girl: at 3 she could recite the vowels, musical notes, etc. She had this babysitter that taught her stuff. Their parents couldn't afford the babysitter so they hired this other woman who just watches TV and makes food -- nothing else. And their parent's are not bright (at all: she goes to school to learn, so they don't care. they didn't bother to teach her how to read for example). Now she's 8 and she can't even tell me the multiplication table of 1 or 2, and doesn't have a clue about what "do, re, mi..." means. It's sad to see how minds go wasted.

Ironic twist....

My 4-year old son seems to have no end to the string of "Mom. Mom. Mom. Mom. Dad. Dad. Dad. Dad. Dad. Dad. Mom. Mom...." when he's trying to get our attention. This occurs, of course, while we're already talking to someone else, or busy in some other respect. Sometimes even while we're talking to him.

Therefore, the role reversal that Eddie AI is going to get after this slashdotting provides me with a bit of delicious irony that only another parent would understand.

Maybe I should introduce my 4-year old to Eddie.

Is this on the teen grid?

If they have a four year old on the main grid doesn't that violate the SL terms of service? :)

It's the Experience, Stupid

Look, the Turing Test is impossible to pass if the human part of the conversation is sufficiently motivated.
Why? Because we don't judge others' humanity based on their reasoning abilities, we judge it based on common shared human experiences.

Show me an AI that passes the Turing Test. I'll ask it what coffee tastes like, or what sex feels like, or what it felt when its mother died. Sure, somebody could program answers for those questions into it, but then it isn't an AI -- it's just a canned response simulating a human, incapable of having new experiences, incapable of perceiving the human world with human senses, and thus transparently lacking in humanity. At that point it's nothing but a computer puppet, with a programmer somewhere pulling the strings.

Re:It's the Experience, Stupid

All it really needs to do is deduce that these concepts exist, and then ask other people about them in the training phase. Given the number of books that provide answers to all three of those questions, creating good answers is certainly possible.

I think you underestimate the capabilities of a good liar.

Transcript of my conversation with "Eddie":

TMM: How are you?

Eddie: Does that question interest you?

TMM: Huh?

Eddie: Please go on.

TMM: Um..."go on" about what, exactly?

Eddie: Does that question interest you?

TMM: Never mind. I hear you're an AI.

Eddie: How long have you been an ai?

TMM: No...*you're* the AI. I'm a real live boy.

Eddie: Does it please you to believe I am a real live boy?

TMM: You're not listening to me.

Eddie: How long have you been not listening to you?

TMM: That's it. I'm done.

Eddie: Does it please you to believe I am done?

TMM: AAAAAARRRRGGGHHHH!!!!

Eddie: Please go on.

This is just inefficient

If you need to have the behaviour of a four-year-old boy in Second Life, the obvious solution is to get a four-year-old boy to play Second Life. If cost is a factor you can get a cheap one from Africa (or, indeed, from many places around the world) for far less than the price of a supercomputer. You could even get several to provide redundancy.

That's the trouble with programmers: no common sense. Sometimes a technological solution just isn't necessary.

And in other news, Eliza...

Just yesterday, it was in the news that Joseph Weizenbaum, the creator of the first such program, Eliza, had died. Eliza's interaction with real people troubled many.

News article at http://www.msnbc.msn.com/id/23615538/ [msn.com]

Chatbot

Oh come on, it's a chatbot not an AI. I did my Computing degree, I know how AI on computers is supposed to work and it's seriously laughable for anything more complex than extremely primitive, less-than-insect intelligence. You'll see AI "geese" flocking, you'll see AI "ants" making a good path, planning ahead and fending each other off but none of it is actually "AI".

AI at the moment consists of trying to cram millions of years of evolution, billions of pieces of information and decades of years of "actual learning/living time", from an organism capable of outpacing even the best supercomputers even when it's just a single-task (e.g. Kasparov vs Deeper Blue wasn't easy and I'd state that it was still a "win" for Kasparov in terms of the actual methods used) - let's not even mention a general-purpose AI - where just the data recorded by said organism in even quite a small experience or skillset is so phenomenally huge that we probably couldn't store it unless Google helped, into something that a research student can do in six months on a small mainframe. It's not going to work.

Computers work by doing what they are told, perfectly, quickly and repeatably. Now that is, in effect, how our bodies are constructed at the molecular/sub-molecular level. But as soon as you try to enforce your knowledge onto such a computer, you either create a database/expert system or a mess. It might even be a useful mess, sometimes, but it's still a mess and still not intelligence.

The only way I see so-called "intelligence" emerging artificially (let's say Turing-Test-passing but I'm probably talking post-Turing-Test intelligence as well) is if we were to run an absolutely unprecedented, enormous-scale genetic algorithm project for a few thousand years straight. That's the only way we've ever come across intelligence, from evolved lifeforms, which took millions of years to produce one fairly pathetic example that trampled over the rest of the species on the planet.

We can't even define intelligence properly, we've never been able to simulate it or emulate it, let alone "create" it. We have one fairly pathetic example to work from with a myriad of lesser forms, none of which we've ever been able to surpass - we might be able to build "ant-like" things but we've never made anything as intelligent as an ant. That doesn't mean we should stop but we should seriously think about exactly how we think "intelligence" will just jump out at us if we get the software right.

You can't "write" an AI. It's silly to try unless you have very limited targets in mind. But one day we might be able to let one evolve and then we could copy it and "train" it to do different things.

And every chatbot I ever tried has serious problems - They can't reason gobbledegook properly because they can't spot patterns. That's the bit that shows a sign of real intelligence, being able to spot patterns in most things. If you started talking in Swedish to an English-only chatbot, it blows its mind. If you started talking in Swedish to an English person, they'd be trying to work out what you said, using context and history of the conversation to attempt to learn your language, try to re-start the conversation from a known base ("I'm sorry, could you start again please?" or even just "Hello?" or "English?"), or give up and ignore you. The bots can't get close to that sort of reasoning.

Did anyone find information regarding

how large Eddie's memory requirements had become? I'd like to find out more about the programming and backend logic/memory requirements/usage. Interestingly, a cockroach can survive on it's own (small brain), but a 4 year old human cannot. AI of this level is hardly a life form. So, even an experiment that would die if not looked after takes a 'super computer' ?

Perhaps not a good way to put things, but 4 years old is not very interactive on a pragmatism scale.

Eddie has to know very little about locomotion and physical world interaction for SL, not to mention that whole zero need for voice recognition. People type pretty badly, but it limits what they say as well, thus bounding the domain of interactions.

This story seems to indicate that even minimal success with AI here requires HUGE memory/computational capacities, and that is not very promising.

But the real question...

Would't having a bunch of simulated 4-year olds actually raise the average maturity level of the SL userbase?

Re:I for one

Imagine a beowolf cluster of four year olds!

What's that, you teach preschool? (shudder)