By 2029 no computer will have passed the Turing Test
longbets.org
longbets.org
They maybe passed the turing test But that just convinced me that Turing test really means nothing. You can't really determine anything for sure with turing test but only probability. (many humans acts or are stupid, if machine is stupid some may decide that machine is like human)
What i would like to see more instead of machine passing turing test is building machine species that could survive in nature, if it can survive autonomously like some kind of animal than this machine species could be said to be intelligent.
If we know what we are made of (the molecules and how they are arranged) and how these molecules can be modeled (i.e. quantum mechanics) then it is only a matter of time until an entire human can be modeled on a computer. Once you can model the entire body, then you can have a computer that can pass the Turing test, because it more or less is human.
If humans can be modeled as deterministic systems that follow physical laws, then computers will simulate them at some point in the future.
First, the complexity in the physics/chemistry of our molecular machines is such that "only a matter of time" may extend longer than the time span of our species' existence.
Second, it is completely possibly that we are more than molecular machines. There may be an aspect to our functioning that is beyond the physical. That aspect of our existence may not be possible to replicate.
What does "beyond the physical" even mean? I assume you mean something "supernatural," meaning something that is "true/real," yet impossible to study or even verify empirically.
I would assume they would be possible to verify in some way once we've reached the stage where hard AI is nearly achieved.
After that, it's just a question of engineering.
In particular I would be interested to hear what some of these phenomena are. See, the lack of a "concrete explanation" isn't really a useful thing to say about something. We don't yet have anything that could be called a "concrete explanation" of even things like gravity. What we do have are theories that allow us to make useful predictions about how gravity will effect things. Imperfect as those, like all, theories are, we've still put stuff on other planets with them.
There certainly are many "open problems" in science, but to my knowledge we have yet to ever encounter anything that is fundamentally beyond the grasp of the physical sciences. Take for example the Coronal Heating Problem: On the surface it seems to violate the 2nd law of thermodynamics (which is some rather heavy shit...) but even in extreme cases like that science solders on. There are numerous hypothesis that have been proposed to explain it, and over time our understanding of it will only become more robust.
Anyway: To address this particular concern I'd like to say that until we have any strong evidence to suggest that humans possess a more powerful model of computation than we already regularly reason about, we have to assume that they do not.
It's possible that there are hard limitations on our ability to recreate some of the very fundamental phenomena we experience in everyday life. Intelligence may be among them.
Every other natural phenomenon that doesn't obscure infromation that way can be figured out. There may of course be soft limits like it taking a billion years to figure out.
Anyway, comparing creating intelligence to creating gravity is nonsensical. I literally cannot make sense of that.
Serendipitously to you mentioning this now, only a couple of weeks ago they announced they have a good candidate for the explanation, which is Type II spicules.
Biology is really not much more than a self-replicating subset of chemistry. Without evidence to suggest something sort of psuedo-supernatural difference between the simplest of single cell organisms and homo sapiens, I think the only reasonable position is that whatever it is that "preforms" intelligence falls under the realm of the physical sciences.
Until we acquire evidence to suggest otherwise, it is silly to assume anything else.
Longbets is fine, but such unfounded speculation doesn't have much use in serious conversation.
"proving the position proving the position"
It seems like you might be misunderstanding the role of science.
I seriously doubt it. Here's a conversation I just had with it:
Me: What's your name?
CB: My name?
Me: Yes.
CB: What's your name?
Me: Suzan.
CB: Nope.
Me: Nope what?
CB: Nope, allessander is not my name.
This is sub-Eliza quality...
EDIT: http://cleverbot.com/human
"Cleverbot was given more processing power for this test than it can be online. It had two dedicated, fast computers with solid state drives while talking to just 1 or 2 people at once. Online there are often 1000 people talking to each machine. We know you'd all love to talk to it the powerful version, but we need a lot more servers first! "
That's all the Turing Test is, it's fooling humans, and humans are incredibly easy to fool--you don't need human level intelligence to fool one. Just several months ago I only realized that a response I got on a craigslist posting was made by a bot after I replied and it replied back with an almost word-for-word copy of its original message (and urged me to fill out a form of personal info). Simple hacks like avoiding repeating the same things, remembering information shared to whom and to you from whom, and the occasional intentional grammatical error go a long way to trick the human. I think we could have had multiple Turing Test passes over the past couple years if that was actually an important goal. But I think most people in AI realize there are more interesting problems to work on than AI PR so projects like CleverBot aren't given top priority.
We haven't?
We got the AI winter because DARPA was for a while willing to fund projects that made outrageous promises, so they got outrageous promises that nobody could possibly deliver on and eventually gave up.
I mean, they were funding people who wanted to do things like build semi-autonomous robots that could drive a car to deliver supplies or fly a plane to do air reconnoissance or wheel around to sweep a field for mines. They funded claims that computers would be able to do speaker-independent voice recognition. Computer programs could beat the best human players at a game of chess. Or imagine a computer program that had enough common knowledge indexed and accessible that it could win contests that involve wordplay and trivia questions.
Wait, all that stuff has happened already, much of it just in the last few years.
The real problem with AI is it's defined to be "stuff we can't do yet". As soon as we manage to do one of those things, it stops being called "AI".
In short, I like Kurzweil's odds on this one. 2029 is a long time away in computer years and we've made a heck of a lot of progress. And he's right that humans think linearly. Exponential growth curves just aren't intuitive to us, so we underestimate what a reasonable amount of progress on a long-term goal looks like. (IBM's win at Jeopardy, Apple's Siri launch, and Google's self-driving cars are all things that happened since Ray made his prediction. Are these not AI progress?)
I don't believe that the von neumann architecture, of which all the applications you mentioned are a part, is ever going to yield anything more then applications with sophisticated symbol shuffling systems which are actually totally stupid behind the scenes. At least before the AI winter, we had impressive AI hardware, including the Lisp machines.
Now just because we have done a pretty poor job doesn't mean things have to continue to be this way in the future. DARPA is currently working to develop memristors, which may become the future of AI. Sadly, DARPA isn't getting nearly enough funding.
http://www.engadget.com/2009/07/14/are-memristors-the-future...
To be realistic, Ray Kurzweil's conviction that there will be strong AI in his lifetime is just wishful thinking. Feel free to prove me wrong.
Well that's what the bet is about.
The fact that people can think proves thinking is possible for mechanical systems; in the worst case we'll re-implement a human brain without understanding it.
> At least before the AI winter, we had impressive AI hardware, including the Lisp machines.
Oof. You're talking about machines with less processing power than an iPhone. Yeah, my dad programmed Shakey the Robot in Lisp on pretty good hardware for the time. If you told Shakey to "push the block off the platform" it could do the task, but would take 20 minutes of thinking about it beforehand. Whereas the Google cars can drive in real time at freeway speeds and robots made by high school kids today can play soccer. How is that not more impressive?
Given how pathetic the Lisp machines were compared to what we've got now, what makes them so impressive to you?
Try to beat even amateur players in 19x19 go. Von Neumann machines are inherently stupid, the domains you mention have very simple rules which allow symbol shifting systems to succeed. You may be able to convince a gullible person of the intelligence of a machine, but that is just cheating the test.
> Oof. You're talking about machines with less processing power than an iPhone.
I tend to agree with John McCarthy that "the computers of 30 years ago were fast enough if only we knew how to program them. Of course, quite apart from the ambitions of AI researchers, computers will keep getting faster." [1] The computers of the late eighties were a sufficiently fast platform for AI if just knew how to program intelligent behaviours in them properly. As such, the Lisp machine hardware was adequately fast, and since the software on them had considerable advantages over what we have today, they were able to do many things at about the same level as modern computers.
> The fact that people can think proves thinking is possible for mechanical systems; in the worst case we'll re-implement a human brain without understanding it.
The human brain has over a septillion atoms [2]. Someday we will duplicate the behaviour of these atoms in a computer, but I don't see that happening anytime soon. In general, I highly doubt there will be significant technological progress well our system of production for profit continues to produce recessions, depressions, and winters.
> Given how pathetic the Lisp machines were compared to what we've got now, what makes them so impressive to you?
The Lisp machines had a consistency and clarity of behaviour that resulted from the use of Lisp all the way down that is absolutely unmatched. You could modify the behaviour of any object in memory down to the machine level using just Lisp. Every object was stored in a single address space. What modern computer system compares to the Lisp machines in these respects? I would love to know, I will be the first to adopt a sanely designed computer platform for my own uses.
However, what I have seen so far is that companies like Microsoft, Apple, and Google are actively trying to replace the programmable computer with displays for external cloud services. The prevalence of these private corporations which are devoted to the pursuit of short-term profit, as opposed to government research agencies such as DARPA, has significantly held back technological progress.
[1] http://www-formal.stanford.edu/jmc/whatisai/node1.html
[2] http://www.quora.com/Human-Brain/How-many-atoms-are-in-the-h...
Computers already play go at an amateur level now - 2-5 dan. And are gradually improving with better algorithms and better CPUs. But go is a game with extremely simple rules - simpler than chess - and an inherently simple possibility space - simpler than Jeopardy - so if somebody were to build a program tomorrow that could beat all the world's best go players, wouldn't you then just dismiss this as not demonstrating true intelligence because we'd then know that go has "very simple rules which allow symbol shifting systems to succeed"?
Winning at Jeopardy is far more impressive with regard to demonstrating "intelligence" than winning at go would be. Jeopardy questions are limited by the english language and human creativity; the relevant info that could be brought to bear in answering these questions is the entirety of knowledge the humans playing the game might have, including knowledge of puns and wordplay. Whereas the go search space, while quite large, is something any programmer could model - just mappings of one 19x19 three-state grid to another. Go is theoretically solvable in a way that Jeopardy is not.
>The human brain has over a septillion atoms
If the individual atoms were all individually crucial to producing thought that would be relevant. But it seems rather unlikely that they do. We need to replicate the relevant properties of neurons, not their exact makeup. Suppose you need to replace a broken hipbone or jawbone. There's a whoooole lot of atoms in a piece of bone, but you can replace it with any compound that has suitable physical properties. What matters when replacing bone is characteristics such as strength, flexibility, and wear resistance. Whatever you replace the bone with will also have a lot of atoms, but knowing the exact number of atoms and where they were wasn't necessary to replacing the functionality.
It sounds like you're using the unfortunate definition that essentially defines any task as "not requiring intelligence" the instant a machine is able to perform it well. This has been done with voice recognition, facial recognition, music composition, etc., and is actually one of the main reasons we even had the AI winter.
I've never heard anyone say that music requires intelligence. Creativity maybe. But I can make a very simple program that has a bunch of hard-coded music patterns that it picks from and nests randomly. It can 'creatively' output trillions of different songs and clearly has nothing even approaching intelligence.
Intelligence is not about what an algorithm can mechanically do. It's about comprehensive world-modeling that can predict and communicate with other agents. If something can be coded in a month by a grad student to run on an 8086 then I feel comfortable saying it's not AI.
Edit: added the word mechanically to be clearer
It's not knowing the algorithm at all. I'm just pretty sure that the algorithm for intelligence is more complex/further from our knowledge than a contemporary grad student could do. I'd love to be proven wrong by a PhD thesis with attached AI.
For example, if someone gave me a ton of rules on how to convert sentences from English to another language, and the output was amazing, would that mean I'm intelligent? Not likely, just that I can follow the language conversion rules.
I think the issue is that too much emphasis is being put on Turing tests as a measure of intelligence when in fact it's really more a measure of how well a computer can follow conversations and social norms. Just because you can fool a real person, doesn't mean you're intelligently interacting with that person. Just like if I can fool another person I can speak another language by following translation rules, it doesn't actually mean I can speak the other language at all!
I actually agree, but for different reasons. I don't think human judges (That is, NORMAL human judges, non-geeks.) are a good measure of an entities intelligence. I personally have managed to convince at least one person that I'm a machine.
Sufficiently good pattern matching to produce reasonable-enough sounding sentences would probably fool most casual observers. For me, the validity to a turing test relies heavily on how long and under how much pressure the A.I has to keep up it's illusion of humanity.
A more objective test of intelligence would be nice though.
One test I heard of that sounded pretty promising is the ability of the computer to discover new patterns in existing data. Not just pattern matching, because that's possible without intelligence, but to discover new patterns from the existing data and use those new patterns to correctly predict the future. Now that's incredibly hard to do!!
If you accept that humans are intelligent, and that they can judge that another human is intelligent by conversing with them across a text-only channel, then you run into a big problem by stating that a Turing-test-passing algorithm is unintelligent. To do so would expose the fact that your definition of "intelligence" secretly includes the class "...and is a human," which makes "intelligent machine" a contradiction of terms. It is essentially an example of the No True Scotsman fallacy, because you're revealing a new facet of your claim when faced with an apparent counterexample.
If you're defining "intelligence" to be a purely human trait, then come right out and say so, and everyone will agree that on your terms a machine cannot be intelligent. Of course, I would argue that such a definition isn't very useful, since it basically means that the adjectives "intelligent" and "humans" are synonyms.
I would say that passing the Turing test is a necessary but insufficient measure of complete human intelligence. It would be astounding, and a major feat in the field, but there must be more to the definition of human intellect than simply carrying on a text conversation.
The concept itself that is presented is quite sound: If you cannot tell that it's not intelligent, how can you say that it is not?
I cannot think of a better test. The only weak point as I see it does not say anything one way or the other about intelligence that is fundamentally different from our own (for example: doesn't happen to use natural language).