[ home ] [ pony / townhall / rp / canterlot / rules ] [ arch ]

/townhall/ - Townhall

A place for civilized animals
Name
Email
Subject
Comment
File
Flags  
Embed
Password (For file deletion.)

[Return][Go to bottom]

 No.4718

File: 1580446654604.jpg (20.42 KB, 425x520, 85:104, thinker.jpg) ImgOps Exif Google

How would you define "general intelligence"?

I think most people would agree that, for example, AlphaGo Zero is intelligent but not generally intelligent (i.e., it is intelligent only at the specific task of playing Go).  GPT-2 is a little more general, but still nowhere close to human general intelligence.  Do you think existing deep machine learning will ever lead to general intelligence, or is a completely new paradigm needed?

 No.4719

The ability to master new tasks. I.E general rather than specific.

I'm not really sure a new paradigm is needed. Improvements for sure though.

 No.4720

Good topic, moved it to /townhall/.

 No.4721


>>4718
>new paradigm

Thats the problem with "measuring" intelligence.  You can only judge it from the stale current perspective, not from the not-yet-current new paradigm, whatever it is.

 No.4724

"General intelligence" is the ability to react to new situations.  Perhaps every situation.  It would be supremely difficult to program something like that, but I can't say that it's necessarily impossible.

 No.4725

>AlphaGo Zero is intelligent but not generally intelligent

AlphaGo Zero is capable of solving a problem (optimizing Go moves). A long stick is capable of solving a problem (moving a rock) that AlphaGo is not capable of solving in any capacity. I wouldn't call either one intelligent for it. They are tools.

 No.4727

File: 1580625508908.jpg (110.48 KB, 700x700, 1:1, 1529720371746.jpg) ImgOps Exif Google

Very related to this thread's topic, I happened to read https://arxiv.org/pdf/1703.10987.pdf today.  (Note the month and day of the publication date, as noted directly above the abstract.)

>>4725
>I wouldn't call either one intelligent for it.
Your argument seems to prove too much.  By your argument, it seems that no machine processing unit could ever be said to be intelligent.

 No.4728

>>4727
>By your argument, it seems that no machine processing unit could ever be said to be intelligent.

I don't think that's a huge stretch.  We have an definative basis for "intelligence" in humans, with maybe up to a dozen other animals as lesser examples.  If a definition doesn't include those then it would be more questionable, but I see no reason the definition must include any kind of machine.

 No.4730

File: 1580660526613.jpg (94.87 KB, 1280x720, 16:9, 1516510176463.jpg) ImgOps Exif Google

>>4728
I think you're using the word "intelligence" differently than most people.  I'd say that a vast majority of people would say that the term "artificial intelligence" applies to, e.g., self-driving cars and AlphaGo Zero.

 No.4731

>>4730

Ah, sorry, I think I've misread things.  When I said intelligence, I was still talking about "general intelligence" and presumed others were as well.

 No.4732

>>4728
Humans are machines using digital memory storage.

 No.4734

File: 1580756419347.jpg (22.34 KB, 220x293, 220:293, 220px-Platovterm1981.jpg) ImgOps Exif Google

>>4718
Hmm...measuring general intelligence in humans is controversial since it attempts to rank humans linearly, especially on something with no known physical cause, so I expect as AI gets better, things will get muddier.  But for now, it's good enough to say a generally intelligent AI would be able to do a variety  of tasks equal or above what a typical human can do.  Right now, machine learning, as far as I understand, requires massively more training data than a human, so you might further stipulate that a generally intelligent AI require about the same amount of training, although general intelligence in machines is far enough off that with any amount of data, it would be an accomplishment.

 No.4735

File: 1580757916052.png (6.89 KB, 678x699, 226:233, equation.png) ImgOps Google

>>4734
It occurs to me, we may not be that far from AI similar to that which won Jeopardy! scoring well or even exceptional on a standard IQ test.  It it would be an odd savant that could score an  IQ of 180 but could not wash dishes or write a sentence, and more than anything such a feat would call into question the validity of the test.

 No.4737

>>4735
>It it would be an odd savant that could score an  IQ of 180 but could not ... write a sentence,
Oh, you must have missed the GPT-2 thread!  It can definitely write sentences that look very human-like!  See https://TalkToTransformer.com/

>and more than anything such a feat would call into question the validity of the test.
OK, I think you have a major misconception of IQ.  It isn't some sort of Platonic form that is supposed to be universal among all intelligent life.  It is an imperfect measurement that empirically happens to work reasonably well for most humans.  You wouldn't say that existence of eyeless sapient aliens calls into question the validity of the Glasgow Coma Scale.  Similarly, it would be a misunderstanding to say that an 'idiot savant' machine calls into question the validity of IQ tests.

>pic
That picture violates at least one of the following conventions:
(1) Strings of digits denote a decimal integer.
(2) "+" indicates addition.
(3) "=" indicates equality.
Let's investigate, maybe it breaks convention #2. Maybe "+" denotes something wacky like
LEA
in x86?
I'm going to write "#" to denote the operation denoted by "+" in the picture, and I will reserve "+" for normal addition.
So, (1 # 4) = (1 + 4).  I'm gonna guess that this generalizes to
(1 # x) = (1 + x).  What's such a possible operation?

Let me guess: (y # x) = (y + y*x).  Does that work?  
2 # 5 = 2 + 2*5 = 2 + 10 = 12
3 # 6 = 3 + 3*6 = 3 + 18 = 21
Oh, a lucky guess!
So, for the final row:
8 # 11 = 8 + 8*11 = 8 + 88 = 96

 No.4739

File: 1580789769262.png (911.39 KB, 790x807, 790:807, 4aecbceb7a7f8f87a7852d79e8….png) ImgOps Google

>>4737
The pattern dictates that the answer would be "40"

 No.4740

>>4739
Tracer.

 No.4741

>>4737
You seem to object to my saying 'the validity', to mean a validity currently exists.  Ok, I can see that.

I think you'be got what the pic-maker was going for.  It was pulled from the Web without much attention because the exercise is not important to the discussion, and you seem to like precision, so I guess I won't know the right answer for sure.

 No.4742

>>4739
Ah, adding the new numbers to the last sum.  Well, that works as well.

 No.4745

File: 1580869350875.jpg (173.51 KB, 600x600, 1:1, Chen.full.1072142.jpg) ImgOps Exif Google

>>4741
>You seem to object to my saying 'the validity', to mean a validity currently exists.
Huh?

 No.4746

>>4745
Oh, not that then.  Well, then we'll have to agree to disagree because I won't consider a computer able to only score well on a IQ test to have general intelligence.

 No.4747

File: 1580898988980.jpg (483.58 KB, 1200x729, 400:243, touhou-a15.jpg) ImgOps Exif Google

>>4746
> I won't consider a computer able to only score well on a IQ test to have general intelligence.
I wouldn't either.  I think we agree on that.

 No.4748

>>4747
Well, then I somewhat don't understand what you were saying.  But it's OK, we seem to have come to the same result.

 No.4750

>>4747
> I won't consider a computer able to only score well on a IQ test to have general intelligence.

I wouldn't consider a human able to score well on IQ tests to have general intelligence.

Just a little education in test prep.  Tests nothing more.  Anyone who thinks IQ tests are valid lacks the intelligence to understand the difference.

 No.4751

>>4750
Do you have any research to support your claims?  Most psychological research that I'm aware of shows a significant correlation between IQ and various life successes such as being able to do well in college and graduate.

Of course, Goodhart's law (https://en.wikipedia.org/wiki/Goodhart%27s_law) applies, and IQ tests are not resistant to people downloading the answers into their brain in advance of the test.

 No.4752

1) Correlation does not imply causation.
2) Proper scientific reasoning implies skepticism to positive claims is generally a default assumption. If one is not convinced of arguments in favor of general intellect, it's not on the skeptics to find research to "prove" their doubtful position. All they need is to explain their skepticism, from there it's up to those arguing the position being doubted to answer those reasons.

 No.4753

File: 1581222741932.jpg (18.03 KB, 290x500, 29:50, 416KUpwy35L._AC_.jpg) ImgOps Exif Google

>>4752
>1) Correlation does not imply causation.
Are you replying to >>4751?  If so, let me reply as follows:
If you're even considering IQ causing something, then let me repeat what I said in >>4737: you're thinking of IQ all wrong, just as if you were thinking that one's position on the Glasgow Coma Scale might cause something.  IQ is widely used and studied precisely because of its well-established correlations.  Yes, they are only correlations, but that doesn't mean that IQ is useless.

>>4752
> If one is not convinced of arguments in favor of general intellect
I'm not sure what you mean by this.  'General intelligence' exists in humans pretty much by definition.  When AI researchers talk about machines acquiring 'general intelligence', they mean the sort of intellectual faculties that humans have.

 No.4754

>>4753
Hmm...so your point of view is IQ correlates to success, that is the extent of the purpose of these tests, not so much to measure general intelligence, at least to the extent general intelligence is different than success.

 No.4755

>>4754
A high level of intelligence (e.g., roughly IQ ≥ 115) is a necessary condition but not a sufficient condition for achieving many kinds of success.  Other kinds of success (e.g., doing scientific research worthy of winning a Nobel prize) requires even higher intelligence.  But on the other hand, there are also people who are highly intelligent but who don't accomplish much.

 No.4756

>>4755
So...IQ measures a ceiling of success.  If a successful group scores low, the test is to be revised.  If high scorers are not successful, the test is acceptable.

 No.4757

File: 1581268676659.png (17.32 KB, 400x400, 1:1, 1414520065421.png) ImgOps Google

>>4756
>So...IQ measures a ceiling of success.
Only certain types of success, in particular, success that requires intelligence.  And IQ isn't a perfect measure.  For most humans it is a decent measure of intelligence.  But it doesn't work well for everyone.  Just like the Glasgow Coma Scale doesn't work as well for people who have an eyelid infection that makes it hard for them to open their eyes.

 No.4758

>>4757
OK, IQ measures a ceiling of success, in tasks involving intelligence, for many people, but not everyone.  (Perhaps mostly for people who are neurotypical, if that's the psychological analog to having eyebrows?)


[]
[Return] [Go to top]
[ home ] [ pony / townhall / rp / canterlot / rules ] [ arch ]