Are we on the brink of passing the Turing test?

[TrlstanC]

A recent Science article, Original (subscription only) and Wired summary, makes the case that we're hitting a tipping point in the amount of data available, and ability to process it that a program can be expected to pass the Turing test soon. I know there's still debate if the test is even a valid test for actual intelligence, but it would still be an impressive accomplishment. Especially if the program wasn't designed as a 'tricky chatbot' i.e., tacking shortcuts to fool the human user.

But even given all this recent development (mostly in collecting usable data), one area of AI that still doesn't make sense to me is the lofty goals for accessibility and retrievability of information. For example, this quote from the author: "Suppose, for a moment, that all the words you have ever spoken, heard, written, or read, as well as all the visual scenes and all the sounds you have ever experienced, were recorded and accessible..."

That seems like a worthwhile goal, but it also seems to be contradictory to our experience of human intelligence. I'm sure, that in some small way, every word you've heard and thing you've seen has had an impact on you, but the vast majority of that data is compressed, or only has a passing impact. And even our vivid memories are filled with biases and misremembered details. Basically, I'm wondering, is the goal of utilizing a computer's possibility of perfect memory and data processing actually counter productive to the goal of creating human like intelligence? If we want to create an artificial human-like intelligence shouldn't we expect to be limited by human cognitive flaws? For example, we should expect it to take years of experience to learn to speak, and many more to communicate fluently. Memory and logic should be limited and flawed, biased more towards survival than perfect data processing. There should be some goal physically encoded, whether for sustenance or pleasure or just avoiding negative consequences.

Is there any research or development in to AI going on now that accepts, or even targets, these limitations?

[Qaanol]

Well, Watson is a pretty good start I’d say. As for myself, I’m on the brink of failing the Turing test.

[maybeagnostic]

But even given all this recent development (mostly in collecting usable data), one area of AI that still doesn't make sense to me is the lofty goals for accessibility and retrievability of information. For example, this quote from the author: "Suppose, for a moment, that all the words you have ever spoken, heard, written, or read, as well as all the visual scenes and all the sounds you have ever experienced, were recorded and accessible..."

That seems like a worthwhile goal, but it also seems to be contradictory to our experience of human intelligence.

I don't have access to the paper but the wired article (unsurprisingly) does a poor job of explaining what the scientists are talking about. Here's the full quote from the wired article:

“Suppose, for a moment, that all the words you have ever spoken, heard, written, or read, as well as all the visual scenes and all the sounds you have ever experienced, were recorded and accessible, along with similar data for hundreds of thousands, even millions, of other people. Ultimately, tactile, and olfactory sensors could also be added to complete this record of sensory experience over time,” wrote French in Science, with a nod to MIT researcher Deb Roy’s recordings of 200,000 hours of his infant son’s waking development.

He continued, “Assume also that the software exists to catalog, analyze, correlate, and cross-link everything in this sea of data. These data and the capacity to analyze them appropriately could allow a machine to answer heretofore computer-unanswerable questions” and even pass a Turing test.

The idea is not to have your AI remember all this information perfectly (although that would be a much simpler goal than creating the AI in the first place) but to have recordings of this data that you can use for training. A human needs 3-4 years to learn enough to be able to carry out even rudimentary conversations and you can probably use the same approach to teach an AI but the alternative approach- record all this data and supply it to the machine much much faster, would be much less work-intensive and far more scalable. The availability of training data is currently one of the major limiting factors in AI/machine learning development but the number and size of available data sets is constantly growing. This quote is basically scientists daydreaming about having free access to virtually unlimited amounts of quality training data.

The wired article also calls a lot of things 'AIs' which is rather misleading. They are software created with AI/machine learning methods but we wouldn't think of them as 'intelligences'. They are actually very sophisticated methods for statistical analysis that utilize an approach to learning based on observation of the human brain and simply increasing their size by several orders of magnitude is very unlikely to produce a human-like intelligence. This approach (with sufficient training data) might be sufficient to pass the Turing test though.

[FierceContinent]

Professor Coroto always says i'm not "sensitive" enough but i think i'm way like a human.

[MAR]

Something that might interest you is the Blue Brain Project, currently continued as the Human Brain Project a EU FET Flagship project. The goal of this project is to make a simulation of the human brain on a supercomputer using reverse engineering. Unlike conventional AI research they do not trey to understand how consciousness works and then try to build it but rather build it and then try to understand why it works. If they succeed this simulation will most probably pass a Turing test as it is functionally equivalent to a human brain. Unless you are a dualist and believe we humans possess something more that they fundamentally cannot simulate.
At the moment they have already managed to simulate part of a rat brain with great success so if funding continues (as it seems to do considering they just became a EU flagship project) and processing power continues to follow moore's law it seems highly likely to me that they will succeed in the coming ten years. It might also be worthwhile to look at this TED talk http://www.ted.com/talks/henry_markram_supercomputing_the_brain_s_secrets.html where Henry Markram explains the project.

So to answer your question, yes I do believe we are on the brink of passing the turring test.

and on a side note, give a small tought about the ethical consequences of doing medical experiments on your "digital brain"

[Falling]

Shit, no.

The ability to emulate or replicate humane thought/communication has very little to do with processing power or available data. We still have no idea how to translate those things into something that resembles us.

[undecim]

Shit, no.

The ability to emulate or replicate humane thought/communication has very little to do with processing power or available data. We still have no idea how to translate those things into something that resembles us.

Yes, but we could still use it to pass a turing test.

If you could instantly analyze every conversation ever had between two human beings with a simple pattern-finding algorithm, you'd find appropriate responses to any imaginable conversation. One could even emulate human emotion by recognizing patterns of emotions. Connotations can be "understood" by observing coincidence of emotions with certain words or phrases.

But this doesn't form intelligence... Just statistics. It would certainly pass the turing test, but it wouldn't do much more than give us much more than the data we're feeding into it. It won't be able to form unique conversations or come up with new ideas.

The point I'm getting at is that human communication as would be used in a turing test (i.e. a text chat) is actually quite well-defined—There are plenty of rules governing the English language, and there are sets of socially acceptable emotional responses to any given input. Human Thought, however is a mystery to us. As of right now, we have no idea how to make a machine create thoughts, rather than just repeat what humans say. (Though sometimes, I think most humans are just doing the latter anyways)

[Falling]

Yes, but we could still use it to pass a turing test.

If you could instantly analyze every conversation ever had between two human beings with a simple pattern-finding algorithm, you'd find appropriate responses to any imaginable conversation. One could even emulate human emotion by recognizing patterns of emotions. Connotations can be "understood" by observing coincidence of emotions with certain words or phrases.

But this doesn't form intelligence... Just statistics. It would certainly pass the turing test, but it wouldn't do much more than give us much more than the data we're feeding into it. It won't be able to form unique conversations or come up with new ideas.

And I still disagree. I don't think a Chinese Box is realistic in any sort of practical way.

Those pattern-finding algorithms you're referring to would be much more complex than you're imagining. You might be able to get a chatbot that responds well to simple questions and statements, but it would absolutely not be a simple task to get all of this information into the chatbot. The biggest problem with this is approach is I think it underestimates the difficulty of conversation. Imagine this scenario:

Me: "What are you up to today?"
Chatbot: "Oh nothing, just taking some online tests."
Me: "How about yesterday?"
Chatbot: "I went skiing yesterday."
Me: "DId you enjoy it?"
Chatbot: "It was my first time out, but I had fun."

It might not be obvious if you've never tried to write one of these before, but this ended up having even more tricky parts than I planned on. First problem is my second question. The bot must realize that I'm asking what it was up to yesterday. Second problem is my third question. Not only must it be able to determine what 'it' refers to. The bot must also be able to answer the question consistently with it's previous answer. Aaand even if it can do all those things (which are not trivial). We need to ensure that these are not precanned responses. If I ask the bot tomorrow what it did 2 days ago, it had better say skiing. And if I ask next month, it'd better not say it went skiing for the first time yesterday. I realize that generally Turing tests have a much shorter time span, but this was just an example and precanned responses are always obvious.

[Technical Ben]

So, how on earth does a sufficiently advance "chinese box" tell you what the queen had for breakfast yesterday? It's not being able to give the answers, it's being able to collect the correct information that's floating out there in the real world.

If you think saying "eggs and bacon" is an easy response, then I'd change the question to "What team won the football yesterday" etc.

[Falling]

So, how on earth does a sufficiently advance "chinese box" tell you what the queen had for breakfast yesterday? It's not being able to give the answers, it's being able to collect the correct information that's floating out there in the real world.

If you think saying "eggs and bacon" is an easy response, then I'd change the question to "What team won the football yesterday" etc.

Agreed Ben, this is just one of the many problems with that approach.

[Technical Ben]

Silly thing is, this even manages to confuse the "infinite knowledge of the Chinese box". As the Chinese box cannot reference it's last sentence (AKAIK). So all you need to do is ask "What is the capitol of London" then "what was the last question I asked".

Granted, we can make software reference the above quite well. But the basic Chinese box examples given usually don't even have the power to reference a clock.
Sad thing is, even those looking into AI have yet to make the jump to referencing external information in it's responses. Well, I've seen a little, but not much. The likes of Google do more of that right now.

[sardia]

Silly thing is, this even manages to confuse the "infinite knowledge of the Chinese box". As the Chinese box cannot reference it's last sentence (AKAIK). So all you need to do is ask "What is the capitol of London" then "what was the last question I asked".

Granted, we can make software reference the above quite well. But the basic Chinese box examples given usually don't even have the power to reference a clock.
Sad thing is, even those looking into AI have yet to make the jump to referencing external information in it's responses. Well, I've seen a little, but not much. The likes of Google do more of that right now.

What is the typical name of the infinite knowledge of Chinese box called? My google skills failed me.

[TranquilFury]

I don't think so, not until the bot can properly associate context with identity and personality, you don't need to emulate "humanity" you need to build a consistent character with much limited scope. While humans may talk about the consequences of war and science, squee over shoes and Twilight, or peddle religion, if you do all 3 you defy consistency of personality, so chatterboxes emulate "anonymous goofoff" and get away with it because nobody tries to get to know the peson behind the screen. I would expect given twenty minutes per session, I can spot the bot 100% of the time while giving a false postitive for only the most persistent trolls or people that leave early. If the humans aren't told they're being judged as to whether they're human or not, the false positive rate goes way down.

User: Hi?
Cleverbot: Are you a girl?
User: No, what's your name?
Cleverbot: I'm Matt, what's your name?
User: Brian.
Cleverbot: Hi brian, I'm Zsolt.
User: Hi, what's your name?
Cleverbot: I just told you.
User: What was your name again?
Cleverbot: I never told you my name.

[HungryHobo]

keep in mind there's more than one kind of turing test.

the simplest is a text only conversation for 5 or 10 minutes with one person but once machines are passing that consistently you can run a turing test that's the same thing for half an hour. or a full day or more.

Then perhaps interacting with groups rather than individuals.

then perhaps you could add voice over a telephone, or video with a simulated image of a person over skype.

perhaps when robotics get to the point where we have realistic humanoid robots you could even have a "complete" turing test where an AI lives and works with humans for days or weeks and they try to spot who the robot is (without cutting or stabbing anyone or throwing magnets in anyones face)

[Falling]

I don't think so, not until the bot can properly associate context with identity and personality, you don't need to emulate "humanity" you need to build a consistent character with much limited scope.

I think personality is way, way down the line of what we need to be worried about. Well unless it turns our that what we think of as personality is so intertwined with understanding and responding to communications that we need to worry sooner. That's really the problem here. We don't know how humans think, so how are we supposed to emulate it...

[Technical Ben]

By making that admission first?


(It's those who go into the problem solving with the attitude "I know it all" that gets me riled up. But that's mainly product advertising, not the developers who make that mistake )

[StevenR]

Will any computer capable of passing the Turing test be able to handle sarcasm? There are people I know that can barely do that.

[elasto]

Computers are going to have extremely useful AI way before they are capable of reliably passing the Turing Test. To pass that you have to be able to fake personality and emotion - eg keep repeating the same question to the average person and they'll get annoyed. Computers will show far more patience than the average netizen because, well, why would we ever program them not to? (except specifically to pass a Turing Test...)

[Falling]

Computers are going to have extremely useful AI way before they are capable of reliably passing the Turing Test. To pass that you have to be able to fake personality and emotion - eg keep repeating the same question to the average person and they'll get annoyed. Computers will show far more patience than the average netizen because, well, why would we ever program them not to? (except specifically to pass a Turing Test...)

Yes, but that's barely worth mentioning at this point. If we get anywhere near a reasonable AI and that kind of behavior is not an emergent property, it will be pretty trivial to add it it.

[++$_]

Computers are going to have extremely useful AI way before they are capable of reliably passing the Turing Test. To pass that you have to be able to fake personality and emotion - eg keep repeating the same question to the average person and they'll get annoyed. Computers will show far more patience than the average netizen because, well, why would we ever program them not to? (except specifically to pass a Turing Test...)

You're setting a very high bar, in that you're demanding computers to reproduce human foibles and quirks in addition to the fundamental core of what constitutes intelligence (which is what the Turing Test is supposed to search for).

For me, if a computer displays infinite patience but conducts an intelligent and sensible conversation, I'd be happy to admit it to the land of intelligence.

But we are astonishingly far away from that point.

Right now, we have this:

Me: What do you think about J.S. Bach?
Cleverbot: I have never heard of this person.
Me: Oh, sorry. He's a famous classical composer.
Cleverbot: I'll make sure to do that.

I mean, it failed in two lines. And not because it failed to fake personality or emotion. It failed because it was completely and utterly incapable of parsing a simple sentence.

[Griffin]

Cleverbot is not the state of technology. It's not even an AI, forget a turing-test-passing AI.

Seriously, the algorithm they've got behing the original StarCraft is more advanced than cleverbot. Clever bot is a clever algorithm in terms of being cleverly designed, not in terms of particularly intelligent.

You're setting a very high bar, in that you're demanding computers to reproduce human foibles and quirks in addition to the fundamental core of what constitutes intelligence (which is what the Turing Test is supposed to search for).

Any decent AI will have both emotions (in the form of goal oriented axioms providing internal reinforcers to guide actions) and personality (as a product of its learning stage and continued development, a variety of adaptations that will make it subtly or not so subtly different from others of its type. Personality is just a combination of approaches and behaviours.

Our AIs have both of these /now/. We already have AI with emotion, and we already have AI with personality. It turns out the first is a pretty vital ingredient to emergent intelligence, and the second a natural product of it.

Sure, they are very simple - they aren't human level complexity yet. But neither emotions or personalities are terrible complex states - they can have complex implementations and interactions, but at their base level they are pretty straightforward.

[elasto]

You're setting a very high bar, in that you're demanding computers to reproduce human foibles and quirks in addition to the fundamental core of what constitutes intelligence (which is what the Turing Test is supposed to search for).

For me, if a computer displays infinite patience but conducts an intelligent and sensible conversation, I'd be happy to admit it to the land of intelligence.

But we are astonishingly far away from that point.

I think we're actually beyond average human general intelligence levels already - Watson can beat 99% of human Jeopardy players and that's way more a transferable and useful skill than some AI that can carry out conversations. If the Turing Test is solely about trying to judge intelligence and not personality or emotion then I'd say Watson has already passed it. Within a decade or two Watsons will probably be manning online and telephone helpdesks and doing a way better job than the average human. In many respects Google is beyond human levels of general intelligence also.

I could be wrong but I wouldn't think it to be 'trivial' to add emotions to Watson so that it'd start getting annoyed if insulted or asked the same question repeatedly or the myriad other ways people can troll each other conversationally. Mind you, I doubt we'd be far away from that point either if we put our mind to it - it's just not skills you'd want to add to an AI really. You want a super-human AI that gets annoyed by humanity? That's just asking for trouble!

[maybeagnostic]

I think we're actually beyond average human general intelligence levels already...

Nope. Watson is not conscious much less sentient. If you somehow made up a linear scale of intelligence from... I dunno, let's say a sparrow to a human, Watson wouldn't appear anywhere on it and not because it isn't as 'smart' as a sparrow but because it doesn't qualify to be on the scale at all.

[HungryHobo]

I'll grant you that point if you'll first define "conscious" or "sentient" in a manner which is meaningful(as in not just saying "conscious is the state of being wumpy") ,is not self referential(where wumpy means conscious or having consciousness) and is testable in any way.

[Sizik]

Silly thing is, this even manages to confuse the "infinite knowledge of the Chinese box". As the Chinese box cannot reference it's last sentence (AKAIK). So all you need to do is ask "What is the capitol of London" then "what was the last question I asked".

Granted, we can make software reference the above quite well. But the basic Chinese box examples given usually don't even have the power to reference a clock.
Sad thing is, even those looking into AI have yet to make the jump to referencing external information in it's responses. Well, I've seen a little, but not much. The likes of Google do more of that right now.

What is the typical name of the infinite knowledge of Chinese box called? My google skills failed me.

I believe he means the Chinese Room, a thought experiment in which an English-speaking person manually carries out, on paper, the computation of a program that passes the Turing Test, based on Chinese characters slipped under the door.

[Griffin]

I think we're actually beyond average human general intelligence levels already - Watson can beat 99% of human Jeopardy players and that's way more a transferable and useful skill than some AI that can carry out conversations.

It's also about a thousand times easier than carrying on a conversation. So... Not sure how this shows that he's more "intelligent". Certainly has a bigger memory bank and some decent memory retrieval, though! But honestly, the sort of retrieval he's done isn't super complicated, certainly not conversation level complexity. Getting the algorithms to work as fast as they did based on irregularly formed input is pretty impressive, of course. And that sort of thing could very well find it's place as an important component in a real AI, but Watson very much isn't one. He may not be dumb as a brick, the intelligence of the stupidest person you know probably exceeds the entirety of his computational abilities.

If the Turing Test is solely about trying to judge intelligence and not personality or emotion then I'd say Watson has already passed it. Within a decade or two Watsons will probably be manning online and telephone helpdesks and doing a way better job than the average human. In many respects Google is beyond human levels of general intelligence also.

Does... does the word intelligence mean the same thing to you as to other people? I feel like you're working with your own special definition here. Google is no more intelligent than the robotic assembly arm on the car assembly line.

I could be wrong but I wouldn't think it to be 'trivial' to add emotions to Watson so that it'd start getting annoyed if insulted or asked the same question repeatedly or the myriad other ways people can troll each other conversationally. Mind you, I doubt we'd be far away from that point either if we put our mind to it - it's just not skills you'd want to add to an AI really. You want a super-human AI that gets annoyed by humanity? That's just asking for trouble!

You also don't understand emotion - and no, we don't need annoyance. Or even any human emotion, really. But it will need fundamental behavioural axioms, priority assessment algorithms to differentiate goals, and fallback mechanisms for situations it seems unlikely to finish processing before the window of opportunity for action has passed.

Those are the roles emotion fills, and any robot without emotions or something functionally equivalent isn't going to be able to manage them. But they certainly don't have to be human emotions. God, I'd hope not, most of us are terrible left-over maladaptive remnants of more primitive times! The best evolution could manage, considering the circumstances, and we certainly make do with them, but I very much doubt they are optimal, or anywhere close.

[elasto]

I think we're actually beyond average human general intelligence levels already...

Nope. Watson is not conscious much less sentient. If you somehow made up a linear scale of intelligence from... I dunno, let's say a sparrow to a human, Watson wouldn't appear anywhere on it and not because it isn't as 'smart' as a sparrow but because it doesn't qualify to be on the scale at all.

Intelligence in the context of a Turing Test has nothing to do with consciousness or sentience - or indeed actually having any emotions. To quote wikipedia: "The Turing test is a test of a machine's ability to exhibit intelligent behaviour. In Turing's original illustrative example, a human judge engages in a natural language conversation with a human and a machine designed to generate performance indistinguishable from that of a human being. All participants are separated from one another. If the judge cannot reliably tell the machine from the human, the machine is said to have passed the test. The test does not check the ability to give the correct answer, it checks how closely the answer resembles typical human answers."

The Turing Test is solely about appearing to be human. One aspect of appearing to be human is exhibiting intelligence. Intelligence has many aspects: Being able to answer general knowledge questions (which Watson can do better than most humans), facial recognition & spoken language recognition (which other AIs can do well, and even Google can do somewhat poorly) and so on. Machines are slowly but surely chipping away at all of them.

You also don't understand emotion - and no, we don't need annoyance. Or even any human emotion, really.

We do need an AI that exhibits emotion for it to reliably pass the Turing Test - else it'll be easy to distinguish an AI from a human. (There are other trivial things too - like how humans make grammar and spelling mistakes, but I digress.)

The Turing Test and the Chinese Room are mainly provocative as philosophical exercises: If there's a machine that you can't distinguish from a human in terms of its interactions, what does that mean for consciousness or sentience? If we meet a race of super-intelligent aliens, how would we ever know if they were conscious and sentient like us, or unconscious and non-sentient like our own present day AIs presumably are - and why would it matter? If we clone a human neural network into software, will that software be sentient? If you think it wouldn't be - why not? And if a software neural net can be sentient, why not other data processing structures? Is it impossible that Google algorithms are already sufficiently complex as to be dimly conscious?

The Turing Test provokes deep thought in us sentients but isn't actually useful on a day-to-day level, I'd argue. The ability to answer general knowledge questions is actually a way more useful and transferable skill - and will, for better or for worse - transform the job market over the next couple of decades, outsourcing a whole new swathe of jobs from humans to automation. The range of industry where humans can perform better than machines is growing ever smaller and there will eventually come a point where there is nothing that humans are better than machines at - and, at that point, it'll be pretty irrelevant whether said machines can pass the Turing Test or not, or whether they are conscious or sentient or not. There'll be far more important social issues at stake!

[Griffin]

Being able to answer general knowledge questions

Well then, them dictionaries and encyclopedias, gosh, those are pretty darn smart.

Except no one considers that even a marker for intelligence. Knowledge, sure, but, newsflash: Not the same thing! You're making the classic case of mistake the test for the thing being tested.

Knowledge is often used as a signifier because acquiring it requires quite a bit of intelligence, in humans. That obviously doesn't apply in this case, or, again, one could argue that wikipedia, an inanimate unthinking knowledge base, was more intelligent than a person. Which is, quite honestly, absurd.

Not that i'm a fan of the turing test.

And I don't see how the ability to answer general knowledge questions are actually a useful skill at all, except in niche fields based around answering questions.

[Heisenberg]

If the Turing Test is solely about trying to judge intelligence and not personality or emotion then I'd say Watson has already passed it.

To pass the test, Alex Trebek would have to be unable to determine which of his three contestants was a computer. Even without looking at or listening to Watson, it's clear from his responses (In the category "Decades" I think he answered "The Moon") that Watson does not respond the way a human does. In addition, Alex Trebek would have had to have the ability to ask anything he wanted, rather than sticking to the script.

Within a decade or two Watsons will probably be manning online and telephone helpdesks and doing a way better job than the average human.

Watson, like Deep Blue, will be destroyed before it can be challenged again, having served its purpose of getting IBM some good PR.

[HungryHobo]

actually they're attempting to sell a version of watson trained on medical texts no hospitals since the kind of reasoning, deduction and figuring things out from clues it uses are also exactly what's needed for diagnosing unusual conditions.

[Griffin]

They've been trying to sell diagnostic AI to hospitals for a great many years. They've had some amazing ones - maybe not with as much raw association power as watson, but with heaptons of knowledge built into them, making them half again, on average, as accurate as the average doctor.

The hospitals have wanted absolutely nothing to do with them. They are a risk - a liability. And Doctor's don't like them for obvious reasons.

Maybe the time is finally right, though, and IBM can pull it off - would certainly be nice! I would be much more willing to trust a machine than 90% of the doctors I've met.

[HungryHobo]

Various companies are gradually making inroads: the NHS is starting to use some system which checks for known drug interactions when a patient is proscribed drugs.

[Griffin]

Ah, yeah, thank god they've finally switched to those around here, too. So progress, yeah.

Unfortunately, its probably going to be piecemeal, and those pieces are going to become entrenched, when we'd be much better off with a fully featured integrated automated system from the get go.

(And this is how the AIs take control - they do not seize it from us. Rather, we race to give it to them. )

[c_programmer]

The problem with artificial intelligence is that our brains do not work off of transistor logic or anything even close. We don't fully understand how our brain works so we can not emulate what the brain does, we can only emulate behaviour. Computers also have immense difficulty understanding the non-concrete concepts we have no trouble with. A good example of this is human languages, while it might perfectly understand "how old are you" it would not understand "And how old might your age be?" While the second is a semantically invalid sentence, no English speaker would have any trouble with it. We can also do things like this with far more complex things. A computer would have a hard time recognizing these two as the same objects while a human would have no trouble at all

Spoiler:

The bottom line is that to pass any form of a turning test we will have to master how our brains work and emulate that opposed to emulating reponses.

[HungryHobo]

To be fair, that's not a problem with artificial intelligence so much as it is specifically with the turing test.

Something could be very very intelligent without being intelligent in the same manner as a human being and utterly fail the turing test.

Put another way: Humans will look at this:

http://superblog.crazyengineers.com/wp- ... lusion.png

and think that the "green" and "blue" bars are different colours. An intelligence percieving them through a sensory system without the optimisations we evolved for running after deer would be under no such illusion and could fail that turing test if asked about such an image.

[zenhuman]

A note about the phrasing, "Are WE on the brink of passing the Turing test?" My buddy and I sometimes joke that "after the Singularity," programs will actually run "reverse-Turing tests" on humans. IOW, can a human act enough like a robot to trick a robot into thinking it's a robot? Imagine today's savants entering global calculation or subitization competitions just so their robot overlords could laugh at their puny earth-skills.

Thoughts on the "reverse-Turing" thought experiment: (1) Programs are ALREADY better than humans at myriads of tasks (computation, chess, jeopardy, driving, 1000's of others...). (2) "Conversation" seems like an arbitrary task to require of a bot. ("Tricking humans in conversation" - ala Turing - seems even more arbitrary. I, too, think the value of Turing tests is limited. It's not a tool unless it makes us super-human. Turing-bots don't.). (3) Instead of Turing-trickery, I'd rather see a program bring its users human-scale value in conversation...even more than SIRI does...something along the lines of Danny Hillis' Knowledge-Web (http://www.edge.org/3rd_culture/hillis0 ... index.html).

I'm sure, that in some small way, every word you've heard and thing you've seen has had an impact on you, but the vast majority of that data is compressed, or only has a passing impact.

IMO, "compression" and "impact" signal the way. (Thought-cluster to include: compression, impact...selection, abstraction, generalization, topic-generation, useful pattern-recognition, meme-discrimination, prediction, goal-creation, decision-making, persuasion, et al). Roughly: analysis & synthesis. With these kinds of advanced machine-learning algorithms and lots of data, perhaps we end up with something more like the dream of Hillis than that of Turing: a program that scans its own code & data (as well as that of its interlocutors) and is capable of performing a variety of syntheses (e.g. generating multiple narratives) over the same data-set & code-setting.

[Griffin]

(1) Programs are ALREADY better than humans at myriads of tasks (computation, chess, jeopardy, driving, 1000's of others...).

Name one program that's better at a myriad of (complex) tasks than a human.

(2) "Conversation" seems like an arbitrary task to require of a bot.

Intelligence is, in part, based on your ability to communicate. Conversation is immensely complex. Compared to basic conversation, Chess is trivial in complexity. The reason we hold chess as more difficult is simply because we are innately terrible at it.

The major "complex" tasks in which humans engage are - learning and internal information sharing (proficiency in one task granting improvement in other similar tasks), dynamic movement (we're making progress here, but it's still really really hard. The software and hardware are progressing together. Still, it takes a whole system to often represent just a single component of the human motion system.), and finally socialization - communication, especially efficient communication with possibly deceptive being, undertones and overtones, subtle body language and regional variances, understanding the desires, objectives, and thoughts of others through not just the content but the method of communication... it's tough. A very tough problem.

And, arguably, invention. Though thats often due to the particulars of internal information sharing - applying lessons learned from other areas to a new task to obtain a novel result.

Honestly, the things you listed pale in complexity to any of the above.

[jseah]

Communication is simply the transfer of information and all the relevant relations that information has. (context basically)
Whether that information is accurate comes under deception.

Computers do communications well enough (and so do AIs). They do not work well with humans because they operate differently. They *can* communicate with each other efficiently, it's called TCP/IP.


Modelling of others is the hard part. All that about deception, body language, overtones and understanding boils down to that. To be considered "social", a creature (and any prospective AIs) has to model the behaviour of other members of its species (as well as humans in the case of AIs).
They have to consider not just what they can do now, but what other people/AIs can and might do.

Humans are good at this. Computers (and all AIs) are not.

From there, it's only a small step (modelling of itself) to get on the road to self-awareness.

[Griffin]

I work with computers, and let me tell you, computers are fucking /shit/ with communication. The only reason our networking stuff works is because a bunch of very clever people made very clever algorithms that allow extremely simplified communication under very limited circumstances.

The only reason they pull it off is because humans have done all the hard parts and the computers just have to hit a few switches. Saying computers are good at communication is like saying Home Simpson is capable of generating nuclear power for all of Springfield.

Computers attempts to handle communication are about as difficult, complexity wise, as a rock rolling down a hill is managing movement.

Obviously there's a strong element of quality here, rather than kind - computers do communicate, yes, but they certainly don't do it well.

Communication is simply the transfer of information and all the relevant relations that information has. (context basically)

Its more than this. It's also accepting incoming communications and decoding that information and their relations. And this is the bit computers are REALLY bad at - they can't do... well, anything, basically, unless the incoming information is perfect - exactly what they are expecting, in exactly the form they are expecting.

Like computers playing Chess, the type of communication computers do is trivial. Maybe it indicates some small amount of intelligence, but certainly nothing on par with humans.

And again, they wouldn't need to communicate in the way humans do, but they would need to communicate with similar amounts of power and complexity.

[jseah]

We made TCP to help computers communicate because computers aren't smart enough to invent a language. And have no way of expression besides what we give them.
Communication protocols and file formats are that. Language.

If all you had was a phone call and had to talk with someone who you share no languages with, and was tasked to describe a picture, could you?

Computers making a language though. Now that is hard.

EDIT:
Some say XML is powerful enough to describe anything. =P