Are we on the brink of passing the Turing test?

[zenhuman]

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

Individual intelligence-technologies (pace Feigenbaum's "expert systems") routinely outperform humans at individual complex tasks (e.g. calculators are better than humans at performing complex calculations.)

Conversation is immensely complex....The major "complex" tasks in which humans engage are learning and internal information sharing...dynamic movement ...socialization - communication, especially...and, arguably, invention. Honestly, the things you listed pale in complexity to any of the above.

Agreed. If we wanted to build a program that could hold a conversation (socialize/communicate), what test might we use? Passing a Turing test (i.e. having a program "trick" us into thinking it's human) represents a purely formal accomplishment. Twitter-bots probably pass Turing every day...in an established environment of brief exchange. On the other hand, something like SIRI gives us (the first inklings) of useful content...but then SIRI isn't trying to trick us out of knowing we're talking to a program.

[Griffin]

Individual intelligence-technologies (pace Feigenbaum's "expert systems") routinely outperform humans at individual complex tasks (e.g. calculators are better than humans at performing complex calculations.)

Which is why I asked for myriad rather than individual.

And I think the ones that don't get ahead of themselves by trying to pretend to be humans are going to be the best off.

A true AI shouldn't need to convince us its human - that just means its good at lying. It should need to convince us that it is /intelligent/.

[HungryHobo]

And I think the ones that don't get ahead of themselves by trying to pretend to be humans are going to be the best off.

A true AI shouldn't need to convince us its human - that just means its good at lying. It should need to convince us that it is /intelligent/.

[/quote]

Most don't. Few chess programs try to convince you you're playing against a human. yet they can beat the best human players.

AI is all around us already but the point of the turing test is that people almost never accept anything a machine does as genuinely intelligent.
By necesity a machine which could convince people that it's an intelligent human for an extremely extended period of time would have to be as intelligent as the sort of person it's pretending to be.

Things a machine is already able to do are rarely considered genuinely intelligent because once a machine can do it "obviously computers can do that, it's just automation".

[Griffin]

Or, perhaps, because the things computers have learned to do well rarely require intelligence?

I've /never/ considered playing Chess well to be particularly intelligent all by its lonesome. "Learning" to play Chess well, that's something different. Most of the "intelligent" bits of Deep Blue came from the people designing it - it's like calling a human intelligent for freaking out when surprised and punching an assailant in the face.

Sure, rapid threat assessment is a pretty difficult task, but an instinctual response is not intelligent.

The intelligence comes in applying those instincts towards mastering new tasks. (which obviously requires with a desire to master new tasks).

This isn't to say Deep Blue doesn't have any intelligence - his game watching and pattern matching skills are actually really impressive. But it is, at best, an "idiot-savant" - and to the extreme!

If they developed a computer, like Deep Blue but capable of playing /any/ game (even games not invented yet, and it doesn't need to be as good as Deep Blue) well, without any additional programming, then we'd have a machine that I think we could honestly call intelligent.

Not super intelligent - if intelligence is an average of aptitudes, it still rates pretty low. If intelligence is a measure of learning capabilities, limiting it to games still rates pretty low (though many things can be framed as games, so this depends on its knowledge acquisition methodology. Figuring out the 'rules' on its own, from observation, is much smarter than needing the rules to be explained). But definitely intelligent.

Like I said, I've met a few bits of software I'd consider to have some level of intelligence - the Creatures creatures come to mind. Deep Blue... does not. And it's certainly not because the Creature AIs have convinced me they are humans. (they still aren't super intelligent, but the guy who made them has a lot of other pretty smart bits of software around too)

[HungryHobo]

I've /never/ considered playing Chess well to be particularly intelligent all by its lonesome. "Learning" to play Chess well, that's something different. Most of the "intelligent" bits of Deep Blue came from the people designing it - it's like calling a human intelligent for freaking out when surprised and punching an assailant in the face.

most of the best game playing programs nowdays aren't programmed with strategies by humans, they teach themselves by playing millions of games against themselves starting with no knowledge other than the rules.

If they developed a computer, like Deep Blue but capable of playing /any/ game (even games not invented yet, and it doesn't need to be as good as Deep Blue) well, without any additional programming, then we'd have a machine that I think we could honestly call intelligent.

General Game Playing :

http://games.stanford.edu/

"General game players are systems able to accept declarative descriptions of arbitrary games at runtime and able to use such descriptions to play those games effectively. Unlike specialized game players, such as Deep Blue, general game players cannot rely on algorithms designed in advance for specific games. "

Figuring out the 'rules' on its own, from observation, is much smarter than needing the rules to be explained). But definitely intelligent.

You mean by, say reading the manual and using it to help figure out the rules?

http://web.mit.edu/newsoffice/2011/lang ... -0712.html

Creatures creatures

they weren't actually all that advanced by modern standards, they were an ANN tied in with a couple of other fairly standard machine learning techniques.

[Griffin]

Do note that I spent the first part of the thread arguing that we already some intelligent machines, and that I only saw them becoming more intelligent as time goes buy. I simply wouldn't call any of them human-level intelligence yet, because of a lack of a certain adaptability, versatility, and autonomy, but this doesn't mean they aren't intelligence. And I certainly don't require it to be humanlike to possess those qualities on par with a human.

And yes the Creature creatures are a very simple intelligence, and we've done better since then. Thanks for the links by the way.

The only thing we disagreed on was over the specific examples of machine intelligence, really.

[gmalivuk]

The only thing we disagreed on was over the specific examples of machine intelligence, really.

Your use of "if" also suggested disagreement over whether computers have *already* been developed capable of playing any* game (including games not invented yet) without any additional programming.

* Alright, yes, what exactly counts as a "game" for the purpose of general game playing machines is more limited than what we humans are likely to call by that name.

[Griffin]

I didn't know those specific programs had progresses that far, I was simply describing what I'd consider an intelligent (or far more intelligent) alternative to deep blue.

I had still said previously we already had computers that met my requirements for 'intelligence', this simply pushes one type up higher than it was before. I believe we have intelligent machines at this point, simply that they are, as of yet, still fairly simple, if often highly focused, intelligences.

That's the funny, thing, though - we've got some amazing advances in dozens of different fields - if you could somehow pull a whole bunch of these AI systems together, tie them together and let them interact, like the different parts of the brain... I think you'd actually have something pretty intelligent.

While we're expanding our knowledge, here, does anyone have any knowledge of attempts to tie multiple AI programs together? Since that's kind of how the brain works, I'd imagine at least a few people have had a go at it.

[freeTheInfo]

Humans are so generally incapable of understanding our own intelligence that to claim that we are anywhere near replicating our own intelligence is a supreme folly in my book.

I think some programs can already pass the Turing test readily, but only for laypersons who also might not be the brightest, and don't know what questions to ask.

[freeTheInfo]

While we're expanding our knowledge, here, does anyone have any knowledge of attempts to tie multiple AI programs together? Since that's kind of how the brain works, I'd imagine at least a few people have had a go at it.

Well, the most abstract AI's use neural networks, which is how the brain works exactly. There was an @Home project to simulate all the neurons firing in the brain in near-real time, I believe. Some reporters have claimed that the standard desktop will be capable of simulating a human brain in real time by 2025. It's important to note that simply simulating the neuron's firing means nothing about intelligence without a fully accurate memory map of some kind, something we have been totally unable to provide, and which is at least multiple decades off (Singularity evangelists aside, nobody believes we are closer than 30-60 years best case for brain dumps).
The downside to neural networks is their inability to provide a reason behind their answer, which is very much the same problem humans have. Fuzzy logic provides a better results, sometimes, but not a consistent result.

I imagine it's possible to run all the AI programs seperatly from each other, have each output a short list of possible items with probabilities, and have a central program (perhaps based on a neural network) select a singular output. If the highest probability output wasn't within a threshold (or if the CPU has some time before an answer needs to be reported), the central program could run all the other AI's outputs as seed inputs for each other,and see if anything better came up.

You'd have to train the central AI program to accept less likely answers from AI's with more relevant knowledge in the field you're dealing with though. So if the question has to do with Music, you give more weight to the Pandora AI even though it's answer is reported as not as probably as an answer from the Gutenburg AI based upon 16th Century English natural language processing.

[gmalivuk]

Some reporters have claimed that the standard desktop will be capable of simulating a human brain in real time by 2025. It's important to note that simply simulating the neuron's firing means nothing about intelligence without a fully accurate memory map of some kind

Why? You don't think a complete simulation of every neuron in a brain will result in the same answers a brain would give, and thus the same intelligence?

[morriswalters]

I believe that he is saying that the sum total of what we are is a product of of the programming of the brain by our experiences and as such each brain is slightly different. So to simulate the intelligence of a thirty year old man you would have to simulate the thirty years of inputs to gain his intelligence. This presupposes that you accept that, at least in part, what we classify as intelligence is the ability to manipulate data based on our experiences.

[gmalivuk]

No, because all of that experience shows up in how the neurons are connected and firing. So if we're simulating the whole brain, that's one of the things being simulated.

[morriswalters]

Then I would suggest that the possibility of simulating the human brain on the desktop is much further off than 2025. I took a quick look at some simple sources, and was quickly reminded of fusion research, very optimistic. At 100 trillion synaptic connections I wish them good luck. In terms of doing it on the desktop, unless they can bring the energy consumption down to something like that of human body than I would suggest that you will need a big fan.

[gmalivuk]

I agree that it's farther off than 13 years, but at the same time contend that simulating all of the neurons is the same as simulating everything relevant about the whole brain, thinking and memory included.

[Griffin]

Are we talking about just simulating the neurons, or simulating all the bits of the brain? Neurons firing isn't the only means of brain communication, after all. I mean, the neurons are probably the toughest bit, so maybe the rest isn't so hard. You might be able to skip simulating the others bits entirely, and just simulate their effects, buts its tough to know the effects that will have.

[morriswalters]

I agree that it's farther off than 13 years, but at the same time contend that simulating all of the neurons is the same as simulating everything relevant about the whole brain, thinking and memory included.

I don't disagree with that. Implementing it may be difficult if not impossible though. Danial Dennett wrote a book call Consciousness Explained which makes exactly that argument. Although I don't care a lot for Philosophy he makes a cogent, and even better, a readable argument.

[negatron]

At 100 trillion synaptic connections I wish them good luck.

With <100 integration steps per second per synapse, which is more than sufficient for low error with something like exponential euler, the computational requirements are below 10 Petaflops. Not that much at all.

[morriswalters]

At 100 trillion synaptic connections I wish them good luck.

With <100 integration steps per second per synapse, which is more than sufficient for low error with something like exponential euler, the computational requirements are below 10 Petaflops. Not that much at all.

I wish I had some idea of what you just said, I am envious beyond all doubt. Having said that and because I am not to bright I will posit a question to you. Perhaps it won't make sense. The synapses themselves seem to be the obvious data structures with the neurons themselves being the network wiring and power supply. In addition the brain must, assuming no more than normal behavior, do this while monitoring the activity of, and modulating the biochemical systems that operate your body, as well as maintain your location within the space you occupy. Manipulate the musculature of the body well enough to ambulate(640 muscles low count), interpret the visual field well enough to allow you to reproduce that field in a painting, hear well enough to understand speech and play classical music on the piano, as well as other things like smell and taste and speech. Notice that I did not say create scientifically since that is almost always a group effort requiring multiple people operating over time, as are almost any large job. Can 10 Petaflops do all that on an energy budget of less than 40 watts.

[negatron]

Can 10 Petaflops do all that on an energy budget of less than 40 watts.

Nope. There are inherent power inefficiencies in Von Neumann architectures for applications with high data locality (biological neural networks). Special neuromorphic architectures could achieve much higher power and cost efficiency, at perhaps some compromise to programmability. In under 20 years however, perhaps closer to 10, even conventional architectures should reach that level of power efficiency.

[morriswalters]

Can 10 Petaflops do all that on an energy budget of less than 40 watts.

Nope. There are inherent power inefficiencies in Von Neumann architectures for applications with high data locality (biological neural networks). Special neuromorphic architectures could achieve much higher power and cost efficiency, at perhaps some compromise to programmability. In under 20 years however, perhaps closer to 10, even conventional architectures should reach that level of power efficiency.

Cool.

[gmalivuk]

The synapses themselves seem to be the obvious data structures with the neurons themselves being the network wiring and power supply. In addition the brain must, assuming no more than normal behavior, do this while monitoring the activity of, and modulating the biochemical systems that operate your body, as well as maintain your location within the space you occupy. Manipulate the musculature of the body well enough to ambulate(640 muscles low count), interpret the visual field well enough to allow you to reproduce that field in a painting, hear well enough to understand speech and play classical music on the piano, as well as other things like smell and taste and speech. Notice that I did not say create scientifically since that is almost always a group effort requiring multiple people operating over time, as are almost any large job. Can 10 Petaflops do all that on an energy budget of less than 40 watts.

Way to completely move the goalposts, there. No one was saying the simulation of a brain would interact with the rest of the world exactly as a real brain. That's not what the word "simulation" means. It doesn't magically stop being a *simulation* of the brain just because it can't taste, or move a robot around perfectly, or run on as little power as the human brain.

Maybe next time you don't understand someone's post, you can just ask them what it means, without also adding a load of irrelevant bullshit.

[morriswalters]

I am not sure how to respond. What are you going to simulate then? The point of the question was to elicit the response I got. Any simulation of the brain will have to deal with this type of data. All of this in interwoven through the function of the brain. Since, unless I am mistaken, we have not localized any of the functional areas of the brain in more than a general fashion, that to me, is the whole point of the simulation. I'll give an example of what I am looking for. I can tell you that it is possible to document people that have grown to adulthood with no language, no sign language or text comprehension. How do these people process their experiences? And how do you define their intelligence without language. And how does their adult brain differ in function from mine?

The question about power is a purely practical one. Since the state you are looking at is a very organized state, if you can't get the efficiency up the power to produce the simulation will eat you alive. Entropy.

Now, let me make something clear from my personal point of view. In the absence of any meaningful input from the world I am not certain that intelligence makes any sense. Certainly a baby born with no sensory function would not in any meaningful way be human, and since the adult human brain is constructed on the fly as you age I am at a loss to understand how those areas of the brain which are programmed as you age would get done.

[gmalivuk]

Since the state you are looking at is a very organized state, if you can't get the efficiency up the power to produce the simulation will eat you alive.

Yes, this is a trivially true and utterly irrelevant point, which could literally be made about absolutely *anything*. "If you can't do it efficiently, your energy requirements will be greater than you want!"

Gosh, why hadn't I thought of that?

Certainly a baby born with no sensory function would not in any meaningful way be human, and since the adult human brain is constructed on the fly as you age I am at a loss to understand how those areas of the brain which are programmed as you age would get done.

If you were to simulate, say, every neuron in my brain, all that structure that required sensory perception would necessarily be included by assumption.

[morriswalters]

Since the state you are looking at is a very organized state, if you can't get the efficiency up the power to produce the simulation will eat you alive.

Yes, this is a trivially true and utterly irrelevant point, which could literally be made about absolutely *anything*. "If you can't do it efficiently, your energy requirements will be greater than you want!"

Gosh, why hadn't I thought of that?

Certainly a baby born with no sensory function would not in any meaningful way be human, and since the adult human brain is constructed on the fly as you age I am at a loss to understand how those areas of the brain which are programmed as you age would get done.

If you were to simulate, say, every neuron in my brain, all that structure that required sensory perception would necessarily be included by assumption.

I'm not really sure what your point is, or what your objection to my post is. We weren't talking about just anything. The question I asked negatron reflects my interests, if they don't interest you then don't respond. In terms of being able to program the initial state of a brain into a neural network that represents a functional brain, I am uncertain as to how you might do that. It would require understanding the brain well enough to be able to program the simulation accurately. Perhaps it would be useful for you to explain what this simulation that you are talking about is for. I have to be missing something.

[Technical Ben]

Is it the difference between the data on a hard disk, and the software running? Copying data off a hard disk could be a trivial task. Even to someone without any understanding of the software, it looks rather an easy job. However, without knowledge of the software, how can you turn the data into a running program?

So, we might be able to get the information or data of the brain and it's synapses. But can we come close to guessing at how to run the "software" is works with?

Lets just take a modern CPU for example. How many transistors does it have? If we were to reverse engineer one, how long would it take? I'd argue it would take at least as long as it has to develop the processor with that many transistors, if not longer! So how about 100 trillion synaptic connections? That's smaller than a petabyte, only 100 terabytes, but how long would it take to manage that amount of data in a reasonable time? I think that's the exponentially big problem Morriswalters was pointing to. Right now, it takes a lot of power, chips, rams etc to even conceive those numbers. We might get there, but how quickly? I'd not bet on it any time soon.

I think it would be conceivable to have the processing power to simulate a brain on the neural level by 2025. But that would only be with a perfectly efficient software. If our software has any overheads, were gonna need at lot more.

[negatron]

In terms of being able to program the initial state of a brain into a neural network that represents a functional brain, I am uncertain as to how you might do that.

It's an interesting problem. The human brain project uses a mix of approaches such as evolutionary and pattern inference. The neocortex is arguably a simpler problem because cortical architectures are highly independent of the functions which they acquire, and can likely be programmed simply through exposure to stimuli, however there's no apparent way that I'm aware of to determine the proper layer densities and connective pathways for multi-module cortical systems for a given task. Nevertheless generic cortical networks can acquire learned function, however suboptimal compared to evolved optimizations.

More task-specific regions of the brain, amygdala, etc, would be more problematic because connection patterns across the structure can be very irregular. There's no clear way to reliably infer the structure of the whole from small observations.

There are developments called "connectome projects" that have been started up fairly recently (human connectome project, open connectome project) which are in an incredibly primitive state of development however progressing well with the ultimate intent to produce high throughput automated scans of fixated tissues. Ultimately this would imply that we would be able to extract full functions intact, including learned behavior and memory.

[morriswalters]

It's an interesting problem. The human brain project uses a mix of approaches such as evolutionary and pattern inference. The neocortex is arguably a simpler problem because cortical architectures are highly independent of the functions which they acquire, and can likely be programmed simply through exposure to stimuli, however there's no apparent way that I'm aware of to determine the proper layer densities and connective pathways for multi-module cortical systems for a given task. Nevertheless generic cortical networks can acquire learned function, however suboptimal compared to evolved optimizations.

I take this to represent what happens when language is learned after the brain has lost its plasticity. That is that language is processed differently when learned after the structure is set. In the sense that learning a language as an adult is different then learning a language as a child. Or am I missing it?

[negatron]

Plasticity is never entirely lost and it's been observed recently that plasticity can be selectively restored via chemical intervention; Of course plasticity is not strictly a good thing, which is why it is lost, inducing plasticity in the brain would result in regression of function.

In the study of "second language acquisition" it has been shown through MRI scans that different brain regions process secondary languages depending on what age they were learned. It indicates that more "general purpose" brain regions are used if acquired later in life.

[morriswalters]

The response was more geared to clarifying what you actually said in your previous post. I restated what I understood you to say. Or to state it another way you zoomed over my head and I am not sure that I understand.

[zenhuman]

While we're expanding our knowledge, here, does anyone have any knowledge of attempts to tie multiple AI programs together? Since that's kind of how the brain works, I'd imagine at least a few people have had a go at it.

I've had a few conversations about this possibility, and while I (and I imagine you) may be looking for more robust examples of tying multiple AI's together, to avoid overlooking the obvious, an example that comes to mind is SIRI. SIRI blends voice recognition, location, search, other smartphone functions, and (probably above all) semantic analysis. Of course, the (still nascent) prestige for many is that they're talking to an "intelligence."

[zenhuman]

You'd have to train the central AI program to accept less likely answers from AI's with more relevant knowledge in the field you're dealing with though. So if the question has to do with Music, you give more weight to the Pandora AI even though it's answer is reported as not as probably as an answer from the Gutenburg AI based upon 16th Century English natural language processing.

Nice. Kind of like a university model for AI. Questions are redirected to experts--questions about music go to Pandora; movies, Netflix; books, Amazon; sociology, Facebook; pop culture, Twitter; classical languages, Perseus. I'm not sure we've got all the professors in the room yet, but it's certainly an exciting time. There's an API for that.