Israeli AI System "Hal" And The Turing Test

Conspiracy_Of_Doves writes: "Hal, the AI creation of Dr. Anat Treister-Goren of Israel, has fooled child language language experts into believing that it is a 18-month old child. Dr. Treister-Goren says that Hal will probably attain adult-level language skills in 10 years. CNN.com article is here. Yes, it's named after what you think it's named after, and yes, the article mentions why naming it Hal might not be such a hot idea."

Comments

Incredible!

[FortKnox]

It's just like chatting with an 18 month old child! Doesn't know how to type, read, or write at all!

Truely an incredible step in toddler AI!

That's not bad

[cnkeller]

Dr. Treister-Goren says that Hal will probably attain adult-level language skills in 10 years.

I know people I work with who still haven't achieved adult-level language skills...

The Conversation

[PoitNarf]

"Hi, how are you today?"
"Poop!"
"Poop? I don't quite understand what you are trying to say."
"Pee-pee!"
"Indeed."

Baby Hal?

[Dolly_Llama]

Hal, the AI creation of Dr. Anat Treister-Goren of Israel, has fooled child language language experts into believing that it is a 18-month old child.

Dave...I have a load in my diaper...Dave...

Reward -vs- Punishment

[smack_attack]

When Hal was "born," he was hardwired with nothing more than the letters of the alphabet and a preference for rewards -- a positive outcome -- over punishments -- a negative one.

[...] Treister-Goren corrects Hal's mistakes in her typewritten conversations with him, an action Hal is programmed to recognise as a punishment and avoids repeating.

How long until Hal figures out that sending high voltage through the typewriter stops the punishment?

"2001"

[YIAAL]

Funny how all the cultural fears of technology come from books and movies like Frankenstein, Brave New World, Colossus, (remember that one?) and 2001. All of which are fiction, and written the way they are to make an interesting story (who would read a story about a man who created a "monster" that was happy, friendly, and harmless, or a computer that worked perfectly and caused no trouble?) Yet in popular discussion, people treat them as real, and embodying actual dangers with which we have real experience.

We need more Artificial Intelligence -- the natural kind is in too short a supply.

creating computers in man's image, exponentials

[Sebastopol]

neural nets are designed to simulate how the brain works, so it makes sense that they be trained the same way. consider this: perhaps they can absorb information faster than a human brain, but who could deliver interactive teaching at that speed?

now consider:

today (2001): human trains AI, limited by wetware bandwidth

...20 years from now: AI trains AI, limited by neural net bandwidth.

result: all 20 years of training one AI will be compressed into to a fraction of a second training time for the next generation

this is the manifestation of Raymond Kurzweil and James Gleick's observations: the acceleration of everything, the exponential growth of compute power.

hang on for the ride, kids. it's gonna get weird. i bet we see AI legistlation in the next 10 years.

we will be the 'gods' (as in creators) of the new race that will inhabit the earth.

Re:creating computers in man's image, exponentials

[ryants]

neural nets are designed to simulate how the brain works, so it makes sense that they be trained the same way

Actually, neural nets don't simulate, they mimic at some crude level.

But just like mimicking what a bird does (ie tape feathers to your arms and flap) isn't going to get you off the ground, mimicking the human brain will probably only get us so far.

I believe the real breakthroughs will come more or less like it did in aeordynamics: when we understood the principles of flight and stopped mimicking birds, we could fly. When we understand the principles of intelligence and stop mimicking brains, we might be on to something.

Turing tests

[ch-chuck]

here's a funny one...

Turing test is pretty crappy...

[DeadVulcan]

The fact that the Turing Test is probably still the only widely recognized test for artificial intelligence says more about our pathetic understanding of the nature of intelligence than the validity and usefulness of the test.

After all, as any con-artist and magician will tell you, it's really not that hard to fool people. Also, remember that on some occasions, some human beings will actually fail the Turing test! That must be so humiliating...

I freely admit I don't have anything better to offer, but I just wanted to point out that the Turing test is a pretty awful measurement, when you think about it.

If you hate poorly defined software projects... can you imagine being handed the Turing test as a feature spec?

Eliza and the turing test

[z4ce]

I have _personally_ seen Eliza pass the turing test. I set up Eliza on my ICQ uin, one of my friends in crisis messaged me and had 45minute conversation with Eliza (not such a good thing). By the end of the conversation, my friend was convinced that he was talking to a hacker who broke into my account. Oh what a mess that was. He had called his ex-girlfriends's parents and told him her new boyfriend broke into my account. I didn't have any idea a bot could be so convincing. It had some flat out amazing responses to his questions and comments. If I had never seen an Eliza conversation before I would have probably thought it was a person too. But like I said.. setting up such a bot on your ICQ account is not recommended. They will pass the turing test and that's not such a good thing necessarily.. :)

To see many such logs go to www.google.com and do a searh for "aoliza" or even "eliza chat" you'll find all sorts of hillarious conversations.

Ha. And bah.

[Lemmy Caution]

The exaggerated claims are classic for private-sector AI research. Language acquisition is a process of highly structured connectionism that will probably require some hardware isomorphism (if we're talking about something that humans can talk to, anyway) and, ultimately, ontology. I see no reference to any sort of ontology engine - the sorts of successes they've had indicate absolutely no ontological grounding, and there's no way that simply training a network without some way of generating ontologies that allow things like binding pronouns accurately to precedent references could occur.

When I see an AI claim, I check its source - if its a business, I suspect exaggeration; if it's a real research center (public or private, MIT or Bell Labs) then I'm more likely to take the claims on face value. This is hyperbolic investor-porn, no more.

Re:How it works...

[DanMcS]

These researchers are getting back from the program exactly what they put in. At one point in your link, I read:

(person) where is daddy?
(ai) daddy gone
(person) daddy is at work
(ai) daddy gone bye bye
...
(person) wait hal, we need to take something to eat
(ai) banana
(person) ok we can take some bananas
(ai) monkeys eat bananas
(person) you're so smart hal. do you want to take a book?
(ai) juice mommmy
(person) sure we'll take some juice as well

The researcher elsewhere claims that the AI's words "relate to its well-being". This is utter projection- the only reason the AI is stuck on concepts of mommy, daddy, monkey, and juice is because this is the inane crap they insist on talking to it about!

Notice also that they claim the AI is tracking almost exactly with a child its same age. Seem strange? Wouldn't you expect a little deflection over 15 months? Shouldn't the thing be a little smarter or a little dumber than a normal child- just statistically speaking, how likely is it they happened to program one that advances /exactly/ as quickly as a normal human infant?

The paper talks a lot about feedback loops. I've got a huge one for them, but it isn't the AI caught in it, it's the researchers. By expecting the thing to react at a child-level, they're talking to it that way, rewarding it that way, and making it that way. If they started talking to it about quantum mechanics tomorrow, it would bd confused as hell for about a month, but I bet it would pick up real fast after it absorbed the new vocabulary. They claim it cares about monkies and juice?! Those are just words to it, you could just as easily raise it on gluons and dark matter, and I don't think it would notice a difference.