Kurzweil Says a Machine Intelligence Will Pass the Turing Test by 2029. What Do You Think?
Kurzweil has a bet against Mitchell Kapor on Longbets.org in which he states: “By 2029 no computer – or “machine intelligence” – will have passed the Turing Test.” $20,000 is on the line for the loser. Who do you side with, Kurzweil or Kapor, and why?
Related Stories
Connect With Us
Subscribe by Email
Post a Comment
Get Our Newsletter
Debate Stats
Total Comments: (28)
Date Started: April 15, 2011 - 6:39 pm
Popular On The Hub
Martin Ford Asks: Will Automation Lead to Economic Collapse?
Written by: Aaron Saenz 716 days ago
Will the future be filled with cool technologies and endless opportunities or will our own creations lead to eventual doom? [...]
5 Axis Robot Carves Metal Like Butter (Video)
Written by: Aaron Saenz 605 days ago
Industrial robots are getting precise enough that they’re less like dumb machines and more like automated sculptors producing artwork. Case [...]
Designer Babies – Like It Or Not, Here They Come
Written by: Keith Kleiner 1009 days ago
Long before Watson and Crick famously uncovered the structure of DNA in 1953, people envisioned with both horror and hope [...]
Comments
part 1 hardware:
17 years? if Kurzweil\’s guesstimates are right and Moore\’s law follows a 18 month doubling time, we\’ll see machines 2400x as powerful as todays (and cloud computing will allow shipping containers full of these machines to be used for AI). If we combine that with better algorithms, two or three generations of memory better ((now, and at the \”flat\” part of the \”s\” curve)NRAM, (next)layered NRAM, (next+?)layered Memristor memory) that is solid state, fast and holds petabytes.
part 2 software:
part 2 software:
IBM\’s Watson ,Apple\’s Siri and google\’s self driving cars all leap to mind. There are companies around AI Numenta, VIA Science, Navia, vicarious and dozens of open source or university projects. Most of these will be dead ends, but some of them will create useful analogs for parts of the brain, with enough of these analogous pieces *working together* we will likely have a legitimate AI. I astrisked \”working together\” because I think that\’s the hard part.
I was under the impression that they had to modify the Turing test rules to be more stringent because computers were already fooling people.
2029 is conservative.
We will probably see the Turing test achieved many years before 2029 if Kurzweil is correct. The test does not represent the human-level strong AI cognition expected at all levels. The problem may be that once computer intelligences approach human levels they will soon afterwards exceed it. Compounding this issue will be the access most systems will have to data and information in the form of the cloud / internet. With Siri and other systems storing, rating and categorizing responses, humans might be fooled in that Turing test by virtue of the AI choosing the response to a question based more on statistical analysis than actual ‘thought.’
So prior to the AI system actually achieving human intelligence, the algorithms will be effective enough to fool the human counterpart, and thereby pass the test, however actual human-level intelligence may follow years behind.
Hmmm….
*looks at Watson*
*imagines 1000^2 times more powerful computer*
*imagines Watson becomes as good as other things as he is with natural language*
effectively this date gives computers after the reach the mark of human brain capacity for a $1000 another 20 years to put that technology to good use to create an intelligent being. Yeh I think it will happen far before that.
Seems quite possible. But will it be AI? Or will it be a well-programmed machine that understands native English? Still impressive, but there’s a big difference between the two.
I think 2029 is a quite conservative guesstimate, and I predict the Turing Test will come under close scrutiny (overnight dismissal and ridicule as if we should have known all along that it was a stupid test) the moment a machine passes it.
On a daily basis I hope that many of the humans , especially the newer generations are computers for the way they think and conduct is non human , self destructive , like a robot with a virus or mall functioning operating system .
First let’s get something straight about the Turing test, it is not a test of intelligence. There can be no such test, because there is no scientifically precise definition of intelligence. Saying “X is intelligent” will never be a scientifically falsifiable proposition and thus a “test” for it would be senseless.
My impression is that Turing was coming up with a pragmatic definition of person-hood (that is, broadly, sentience). It basically works on the premise that if you can’t tell the difference between a man and a computer (in terms of their behavior) empirically there IS NO DIFFERENCE and therefor the computer MUST BE TREATED AS A PERSON. This might sound absurd but pragmatically to deny this would be a form of solipsism (for how could I judge that ANYBODY is conscious like I am -I am I swear!- without referring to their behavior?)
I suspect that computer WILL be able to trick humans into thinking they’re also human by 2029, but only if the human is pretty dumb and, furthermore, doesn’t try very hard. If he does try he will soon pick the impostor, because (as a writer) I understand their is absolutely nothing superficial about writing (which the Turing test is all about). If a Man can bare his Soul in words then a computer shan’t be a Man until it too has a Soul to bare. Personally I suspect that level of sophistication isn’t a mere 18 years away. Watson, for instance, couldn’t even dream of dreaming about writing this comment. (And I don’t claim to be writing anything at all profound.)
Frankly, computers will pass the Turing test when and only when they become sentient. Why? Because computers will become sentient when and only when they pass the Turing test (or something like it). Unless someone can suggest a better definition?
I will side with Kapor , in the next 18 years , as previously compounded human knowledge has increased and sadly wisdom decreased , we are in the era of compounding idiocrisy , further more computers / AI , is having a greater impact on humans than humans could ever imagine having on or towards computing / AI . The logistical calculating compounded results from programming to self programming or teaching is the next factor needs to be taken into account. There are no dreams or imagination when it comes to this aspect , it is only visions and fears of the super ability machines that can occupy many humans at speeds that humans cant intercept or occupy them . Who runs this ship , who is your captain ???
First supercomputer with human-level intelligence will probably think much slower than real human. It will consume many megawatts of power and be able to perform equally complex reasoning that humans are, but maybe 10 or 50 times slower. It will not pass Turing test and will be useless as an AI, but it will shake the world with discussions about computer consciousness, ethics, religion and laws. And will probably trigger new arms-race to superintelligence.
This will probably happen before the end of this decade.
I’m going to say no (so I can never be disappointed, I admit).
However, I really hope so.
undoubtedly. Maybe even sooner. If a single supercomputer or something was used then yeah, but computers are becoming increasingly more network. And now there’s the cloud, moving network infrastructure on to other essentially borrowing their computer power. With cloud becoming increasingly more ubiquitous and connection speed increasing we will be using resources frequently above the capacity of our personal devices on a regular basis, borrowing the power of the whole. Now imagine this with AI’s, thousands of narrow AI’s may be loaned out to other companies, like Google’s Translator. Using this distributed approach it is possible that such AI may exist even earlier then predicted.
I,ll side with Kapor ,
in many perspectives machine intelligence / artificial intelligence has already surpassed the human levels of communication , we are just to daft to get it , think cautiously , who has who by the tennis elbow ,
if we had the communication skill , whats the hold up , why are we trying to get there ?
We just stumble upon a certain aspect or element , use and apply it best we can , and then we have the superior wisdom to site our ability as the limit of the discovery , it may be the other way round.
the turing test is useless, human are not intelligence
and by the way it will take less than 18 years
18 months ?
I agree, today’s chat-bot can confuse you for a second, no doubt. But try to ask what were you talking about a second ago, or something like “Why do you think cats get stuck on a tree” and its game over. It requires crazy amount of computation to comprehend abstract meaning, imagination, insight into other minds or understanding example scenarios. But, we are capable of that… our physical brain is doing that with calcium pumps and constantly reconnecting living logical circuits. No magic needed. I dont think we have to recreate complete human connectome to get working mind, I think derived logical rules and ability to change will be enough. 2029? Hell why not? Given some religious nut with nuclear arsenal wont start WW3…
I wonder what ratio of the human race can tell you why cats get stuck in a tree !!!
It could be much sooner. THere just needs to be a push from a well supported Watson type team. Step 1 load in chat info from a large network like face book or AIM. step 2 refine.
Even if the worst case (of classical brains) occured and we had to the equivalent of simulating entire brains at the detail of dendritic spines and short axon/dendrite sections at millisecond (sub action potential) refresh rates to do AI, then we would have to wait until about 2036 (based on what I learned in my Neuroscience and Neural Nets degrees respectively) – a whole 7 years after Kurzweils prediction – not a massive big deal IMO.
Frankly I haven’t seen dates much longer than that, that are actually based on calculations – in other words longer timescales seem to me to be plucked out of the air based on “unknown unknowns”.
The important point is when machines become better at designing machines than we are – because at that point we should see positive deviations from More’s law until we hit absolute physical limits.
IMO 2029 is looking better and better.
“Even if the worst case (of classical brains) occured and we had to the equivalent of simulating entire brains at the detail of dendritic spines and short axon/dendrite sections at millisecond (sub action potential) refresh rates to do AI, then we would have to wait until about 2036″
I don’t know when a computer will pass the turing test (and doubt the turing test is a good proxy for a machines ability to recapitulate humanities intelligence) but dumb simulation of the brain is likely harder than just simulating a connectome.
A simulation of any level connectome (micro, meso or macro) alone almost certainly wouldn’t recapitulate human intelligence. Connectomes don’t include synaptic weights, cotransmission of multiple neurotransmitters, volume transmission, gap junctions/electrical synapses , rules for how chemical connections change over time, rules for how electrical synapses change over time, or rules for the role of non-neuron cells in computation.
The strange thing about simulating nervous systems is we skipped the simple model systems and went straight for the hard systems. Normally in science you would expect people to start with the more tractable problems before moving on to the harder problems (we sequenced small viral genomes before we sequenced any mammalian genome). skipping over simulating the simple organisms is even more strange given that we have had a (mostly) complete connectome for C. elegans for 20 years. Yet to my knowledge no one has ever run a simulation using that data and shown biologically realistic behavior coming from the simulation.
OK I certainly was including synaptic weights in the estimate!
Co-transmission doesn’t actually add much to the computational load, and neither (counterintuitively perhaps) does volume transmission.
No I hadn’t included gap-junctions but was including rules for updating weights.
Glial cells however might be significant (in terms of computational load) if internal dendritic communication is significant, and I did think I should have mentioned them after posting.
They are the “Junk DNA” of the brain – in other words turning out to be really important and out numbering neurons 8 to 2.
which might mean that the brain is doing 4 times more comutationally than we thought.
My understanding was that the Blue Brain project is simulating brains in a more realistic way.
Personally I think if AI’s don’t have to go the brain simulation route (which would be awful) we could get them (significantly) sooner than the worst case (OK with glia doing lots then maybe 2042 -eek) – because we’d just be doing the abstract stuff and not having to actually simulate the detailed biology..
the thing about the Turing test is that it’s completely time dependent. a computer today could easily fool a person into thinking it’s real for a few seconds, but 20 minutes is another matter. I think that by 2029 a computer will likely be able to hold its own in a short conversation, but after that people will want it to fool us for longer.
Non experts do tend to ask rubbish questions.
Here’s an example of a question that takes a little imagination and insight.
On paper or in your imagination, draw a ball, underneath the ball (and touching it) draw a crucifix, underneath that and touching it draw an upside down V shape.
What did you draw?
But no – people ask: Who won the world series last year? – sigh.
And here’s a couple more for fun:
1.
oo 1111
000 11 11
0 02 11 41
00 11 11
00 10 16
30 11 11
00 11 11
01 1112
What number is one mostly made of?
ans: 0
2.
a) Stand up and hold out your left arm horizontally and straight.
b) Put your right arm on your hip.
c) Lift your left forarm up so it has an ange of about 45 degrees to the horizontal.
d) Point your left hand horizontally with your fingers flat but your thumb pointing down.
e) lean your torso to the left.
f) lean your torso back to vertical
g) repeat e and f a few times.
What little t are you?
ans: I’m a little teapot.
Er – you could do with a preview button!!!!
How does the first question mostly made of 0′s? There are clearly more 1′s (especially if one distinguishes between o and 0)
I think the reason that calls the fools bluff is the inherent increase in idiotic conceptual behavior that humans have , sad is it not .
I wasn’t so sure about this before IBM’s Watson came along. But if you look at what Watson did on Jeopardy, it would seem like it isn’t that crazy for an AI to take it to the next level of a passing a Turing test in the next 18 years.
Kurzweil of course! 2029 is almost slightly pessimistic it might be a few years earlier. But knowing Kurzweil’s predictions it will probably be the last months of 2028