Chomsky and AI

General discussion about learning languages
Cainntear
Black Belt - 3rd Dan
Posts: 3449
Joined: Thu Jul 30, 2015 11:04 am
Location: Scotland
Languages: English(N)
Advanced: French,Spanish, Scottish Gaelic
Intermediate: Italian, Catalan, Corsican
Basic: Welsh
Dabbling: Polish, Russian etc
x 8601
Contact:

Re: Chomsky and AI

Postby Cainntear » Thu Mar 23, 2023 7:38 am

Oops. Had forgotten about "hard AI" vs "soft AI", by analogy to hard vs soft behaviorism.

Soft AI says if it acts intelligent on the surface, it's intelligent; hard AI says that unless you can analyse how it "thinks", it's not intelligent.

As I say, this originated in behaviorism, which tended to hold the view that only behavior is observable and therefore thinking about thinking is not scientifically valid.

I'm not big on behaviorism, and that should be no surprise given what I've talked about in the correction thread: judging learning based on learner behaviour leads to incorrect conclusions.

Based on what I wrote, I think the logical conclusion is that the AI department in my uni was largely of a "hard AI" philosophy. Probably because they were totally creeped out by Skinner and the apparent inhumanity of his experiments.
2 x

User avatar
Iversen
Black Belt - 4th Dan
Posts: 4759
Joined: Sun Jul 19, 2015 7:36 pm
Location: Denmark
Languages: Monolingual travels in Danish, English, German, Dutch, Swedish, French, Portuguese, Spanish, Catalan, Italian, Romanian and (part time) Esperanto
Ahem, not yet: Norwegian, Afrikaans, Platt, Scots, Russian, Serbian, Bulgarian, Albanian, Greek, Latin, Irish, Indonesian and a few more...
Language Log: viewtopic.php?f=15&t=1027
x 14924

Re: Chomsky and AI

Postby Iversen » Thu Mar 23, 2023 9:02 am

Hard AI by definition excludes any way of reaching conclusions that isn't done just as it would have been done in a human brain - and that is in fact still not known in detail. The status is that you can study one human brain in such detail that you from scans can guess which out of a number of items or notions the brain is dealing with - but at that level you can't expect that another person's brain functions exactly as the brain of no. one. And guessing a picture is far easier than found out how a brain draws a conclusion or gets an idea - we just know that the preliminaries occur before the result pops up as a conscious thought. So asking for an exact replica of human brain activity is asking for something that can't be delivered.

And asking for an approximate replica is also fraught with problems: that would as a minimum demand that you structured some kind of electronic micro-component as a human neuron (and no, even quantum computers don't have such structures), and the big question is whether that is possible. But even if it is, trying to build the chaotic nature of neurons into a supercomputer might jeopardize the advantages of electronics, namely the precision that comes with complete knowledge of the structure of each chip. This leads to the question: WHY?? To achieve the good things about human throught like fantasy and selfawareness you might have to sacrifice the ability to be consistent in the treatment of large amounts of data in favour of a system that basically works on probabilities and lofty aspirations. I trust that my external harddisks can store 4-5 terabytes of data without losing a couple of megabytes here and there - I wouldn't ask the same of my biological brain. And the machine may work on rubbish and then it delivers rubbish, but we can trust that it does so in a consistent way - even though it might be too complicated for us to explain the details.

So the most logical strategy would be to drop hard AI and forget about the instinctive disgust with Skinner some people had fifty years ago ... and go for soft AI: “If it looks like a duck, quacks like a duck, then we don't need to have a real duck” The next problem is then solvable, namely how to give our electronic superduck some of the features that it lacks, namely better model making skills so that it can learn about the real world instead of just reading fake new on the internet. And maybe also a self awareness component, but maybe it's better not to give it that (and cross fingers that selfawareness doesn't appear as a consequence of the immense complexity in itself - after all that's what happened in humans, and now you can see the result .... :shock: :mrgreen:
3 x

Cainntear
Black Belt - 3rd Dan
Posts: 3449
Joined: Thu Jul 30, 2015 11:04 am
Location: Scotland
Languages: English(N)
Advanced: French,Spanish, Scottish Gaelic
Intermediate: Italian, Catalan, Corsican
Basic: Welsh
Dabbling: Polish, Russian etc
x 8601
Contact:

Re: Chomsky and AI

Postby Cainntear » Thu Mar 23, 2023 11:26 am

Iversen wrote:So asking for an exact replica of human brain activity is asking for something that can't be delivered.

Indeed, but my counterpoint is that asking for something informed by human brain isn't itself asking for the impossible.

I believe (on the grounds of watching plenty of explainer videos more recently) that modern neural nets are typically very shallow. When I was training neural nets on stuff, we typically only had three levels/layers of neurons/perceptrons because anything more would be ridiculously slow in the age when Pentium II was the top of the line, and we were still using Sun workstations with Solaris as an operating system.

The thing is, the modern "A.I." seems to have been built on the general finding that shallow networks with more perceptrons per layer seem to be more efficient than deep, multilayered networks with the same total number of perceptrons (i.e. return a result quicker).

I think that this is probably something that I've heard discussed in people trying to decide whether it classes as "AI" or "ML". The origin of the neural net is definitely in AI, but it's now being used in a way that is more ML in nature -- analyse the data, but get a reasonable result just by probabilities and not through structured modelling.

It's potentially the worst of several worlds, because NNs by their nature are difficult to analyse, unlike other machine learning systems that generate moderately explicit rules; and NNs open up companies to serious liability, because the networks are unauditable, and someone's going to get a massive lawsuit based on "computer says no" and the decision not being justifiable in court.

So the most logical strategy would be to drop hard AI and forget about the instinctive disgust with Skinner some people had fifty years ago ... and go for soft AI:

Ah, but dropping hard AI and defining AI as being soft AI is really something that favours prescriptive linguistics over descriptive.

When most people receive "AI" they don't perceive "soft AI"; at best they perceive fuzzy magic, at worst they perceive "hard AI" because of exposure to sci-fi etc.

Particularly:
“If it looks like a duck, quacks like a duck, then we don't need to have a real duck”

This will likely trigger the common cliche: "...then it is a duck" and it therefore subtly plays down the importance of the difference between not having a duck and not having a real duck.

When given the "if it looks like a duck" line during my undergrad years, my typical response would be "if it doesn't taste good with orange sauce, it's not a duck".

I don't know if I fully got my own point consciously back then, but was instead working on intuition.

I think my real point is that if you let yourself select the criteria, you can always prove whatever you want, and if you arbitrarily throw out criteria because you don't think they're important in your case, you undermine your own argument.

cf recent discussions about teaching and student blame -- to me, this is where hard behaviorism inevitably leads: "I can't see inside their brains, so the measurement of their success or failure is a valid endgoal." The way I see it, the measurement of behaviour is a cop-out if it's not used to support or disprove a hypothesis.
3 x

User avatar
Iversen
Black Belt - 4th Dan
Posts: 4759
Joined: Sun Jul 19, 2015 7:36 pm
Location: Denmark
Languages: Monolingual travels in Danish, English, German, Dutch, Swedish, French, Portuguese, Spanish, Catalan, Italian, Romanian and (part time) Esperanto
Ahem, not yet: Norwegian, Afrikaans, Platt, Scots, Russian, Serbian, Bulgarian, Albanian, Greek, Latin, Irish, Indonesian and a few more...
Language Log: viewtopic.php?f=15&t=1027
x 14924

Re: Chomsky and AI

Postby Iversen » Fri Mar 24, 2023 9:47 am

Cainntear wrote:(...) dropping hard AI and defining AI as being soft AI is really something that favours prescriptive linguistics over descriptive. (...)
cf recent discussions about teaching and student blame -- to me, this is where hard behaviorism inevitably leads: "I can't see inside their brains, so the measurement of their success or failure is a valid endgoal." The way I see it, the measurement of behaviour is a cop-out if it's not used to support or disprove a hypothesis.


I see the opposite consequence: going by Cainntear's own definition of soft AI it permits you to judge (and react on) behaviours without caring whether it's a real duck or a fake duck that produces them - except that they to conform to the behaviours of real ducks. And then you can analyse the quacks of the potentially fake duck and build models, in which there will be ample opportunities for hypothesis testing. The problem is of course that a fake duck might follow other rules than a real duck, but precisely that is something that can - and should - be tested by a qualified descriptivist. After all, you need to do that test to say in the first place that 'it' quacks like a duck. But doing that testing doesn't imply that the fake duck will continue to quack like an oldfashioned real duck - it may invent its own way of quacking (as Donald Duck did) - and then the job of a descriptive linguist is to describe that new dialect. In contrast, a prescriptivist linguist would tell the fake duck to revert to the good old ways of real ducks. Or tell the real duck to learn the modern new way of quacking - or end up as canard à l'orange...

The real problem here is of course that fake ducks with time will become the dominant producers of quacks on this planet, and we are already sliding in that direction.

PS and if I had been a behaviourist OR a purebreed descriptivist I wouldn't care, but I see myself in the role of potential canard à l'orange. And then it does matter who quacks...
0 x

Cainntear
Black Belt - 3rd Dan
Posts: 3449
Joined: Thu Jul 30, 2015 11:04 am
Location: Scotland
Languages: English(N)
Advanced: French,Spanish, Scottish Gaelic
Intermediate: Italian, Catalan, Corsican
Basic: Welsh
Dabbling: Polish, Russian etc
x 8601
Contact:

Re: Chomsky and AI

Postby Cainntear » Fri Mar 24, 2023 11:20 am

Iversen wrote:
Cainntear wrote:(...) dropping hard AI and defining AI as being soft AI is really something that favours prescriptive linguistics over descriptive. (...)
cf recent discussions about teaching and student blame -- to me, this is where hard behaviorism inevitably leads: "I can't see inside their brains, so the measurement of their success or failure is a valid endgoal." The way I see it, the measurement of behaviour is a cop-out if it's not used to support or disprove a hypothesis.


I see the opposite consequence: going by Cainntear's own definition of soft AI it permits you to judge (and react on) behaviours without caring whether it's a real duck or a fake duck that produces them - except that they to conform to the behaviours of real ducks. And then you can analyse the quacks of the potentially fake duck and build models, in which there will be ample opportunities for hypothesis testing. The problem is of course that a fake duck might follow other rules than a real duck, but precisely that is something that can - and should - be tested by a qualified descriptivist. After all, you need to do that test to say in the first place that 'it' quacks like a duck. But doing that testing doesn't imply that the fake duck will continue to quack like an oldfashioned real duck - it may invent its own way of quacking (as Donald Duck did) - and then the job of a descriptive linguist is to describe that new dialect. In contrast, a prescriptivist linguist would tell the fake duck to revert to the good old ways of real ducks.

I think you've missed the point of me bringing up prescriptivism vs descriptivism (which is fair enough as I really didn't explain myself well enough). To requote myself (before expanding on it):
Cainntear wrote:Ah, but dropping hard AI and defining AI as being soft AI is really something that favours prescriptive linguistics over descriptive.

When most people receive "AI" they don't perceive "soft AI"; at best they perceive fuzzy magic, at worst they perceive "hard AI" because of exposure to sci-fi etc.

I think my point is people are generally too quick to say the dictionary agrees with them "so I'm right" (maybe "pedantry" would be more useful a word here than "prescriptivism"...?), and are unlikely to take on the descriptivist approach of recognising that most people don't use the term as defined in the dictionary.

I am personally of the opinion that the popular understanding of AI is too far removed from the intended meaning for the term AI itself to be useful. Even if unintentional, its usage misleads people.

I think that's the only respect in which me and Iversen's views differ, because he's using the "if it looks like a duck..." line here to describe something that I agree with: investigating and expanding human knowledge and understanding.

Assumptions are what real science uses all the time, and always looks to challenge.

Assumptions are what pseudoscience uses all the time, and never tries to challenge.

As I said, the danger in "if it looks like a duck..." is that it risks people thinking there's a justification for certainty.
0 x

User avatar
Iversen
Black Belt - 4th Dan
Posts: 4759
Joined: Sun Jul 19, 2015 7:36 pm
Location: Denmark
Languages: Monolingual travels in Danish, English, German, Dutch, Swedish, French, Portuguese, Spanish, Catalan, Italian, Romanian and (part time) Esperanto
Ahem, not yet: Norwegian, Afrikaans, Platt, Scots, Russian, Serbian, Bulgarian, Albanian, Greek, Latin, Irish, Indonesian and a few more...
Language Log: viewtopic.php?f=15&t=1027
x 14924

Re: Chomsky and AI

Postby Iversen » Fri Mar 24, 2023 11:59 am

.. and I second that
0 x

Nogon
Green Belt
Posts: 305
Joined: Sat May 13, 2017 6:21 pm
Languages: German (N), Swedish (C), English (?), French (A2), Esperanto (A2). Reading Danish, Norwegian, Dutch, Afrikaans. Wanting to learn Polish, Yiddish
Language Log: https://forum.language-learners.org/vie ... 15&t=16039
x 1068

Re: Chomsky and AI

Postby Nogon » Fri Mar 24, 2023 1:16 pm

Iversen wrote:PS and if I had been a behaviourist OR a purebreed descriptivist I wouldn't care, but I see myself in the role of potential canard à l'orange. And then it does matter who quacks...

Quack
0 x
Assimil French : 65 / 113
Active wave : 15 / 113


Return to “General Language Discussion”

Who is online

Users browsing this forum: No registered users and 2 guests