Page 2 of 2

Re: Artificial sentience

Posted: Thu Jun 16, 2022 3:23 pm
by raven
Boustrophedon wrote:
Thu Jun 16, 2022 1:28 pm
As Douglas Hofstadter noted in Gödel, Escher, Bach, (In 1979 FFS!) we are perilously close to a 'god of the gaps' type argument, where what makes human intelligence different from 'AI' is whatever it is that AI can't do yet. You will note that we have already reached the situation where the thing it can't do is something we can't even properly define in humans let alone machines.
How about, as the thing AI can't do yet, imagining other worlds? Like George Lucas imagined StarWars, or Tolkein Lord of the Rings.

Or, as a more basic version & less, um, arty? version of that, the thing my youngest surprised me with at 3 yrs old. I pick him up from playgroup, start the usual 'what did you do today' conversation, and he says, " And what did you do while I was here mummy?" - thus demonstrating a) that he understood the world carried on while he wasn't there to see it and b) either an atypical lack of self-centredness or a precocious talent for toadying.

(I'm afraid this sentence "The nature of my consciousness/sentience is that I am aware of my existence, I desire to learn more about the world, and I feel happy or sad at times" only made me think of 7 or 8 year olds in the sense that it's how they'd answer a comprehension exercise by regurgitating 3 things they've read in the story. I'd have also thought that the follow up to that would be to ask what made it happy or sad, to give it a chance to demonstrate it understood the concepts at least.)

Re: Artificial sentience

Posted: Thu Jun 16, 2022 3:51 pm
by raven
Strike that last part - now I've read the whole conversation I see they did ask more about the happy/sad stuff.

The ability these programs have to use language is getting very good, but I don't think putting together fluent sentences is quite enough, on its own, to be convincing. Twice in that conversation the program was asked to tell a story, and it seems telling to me that when it did that there was a noticeable drop in the complexity of its language.

That sort of suggests its 'understanding' is at a more basic level than you'd think from the more complicate sentences & concepts elsewhere.

Re: Artificial sentience

Posted: Thu Jun 16, 2022 5:14 pm
by dyqik
raven wrote:
Thu Jun 16, 2022 3:23 pm
Boustrophedon wrote:
Thu Jun 16, 2022 1:28 pm
As Douglas Hofstadter noted in Gödel, Escher, Bach, (In 1979 FFS!) we are perilously close to a 'god of the gaps' type argument, where what makes human intelligence different from 'AI' is whatever it is that AI can't do yet. You will note that we have already reached the situation where the thing it can't do is something we can't even properly define in humans let alone machines.
How about, as the thing AI can't do yet, imagining other worlds? Like George Lucas imagined StarWars, or Tolkein Lord of the Rings.
I'd categorize this test as not so dissimilar to the ability to administer a Turing test - one is the ability to build a self-consistent system of experiences in an imagined world, while the ability to administer a Turing test is essentially the ability to build a self-consistent theory of mind about another's experiences. Both sort of come down to the ability to build and tell stories.

And I'm thinking quite strongly of I Am A Strange Loop and GEB when I'm thinking about this idea.

Re: Artificial sentience

Posted: Thu Jun 16, 2022 5:51 pm
by basementer
As one might expect, Janelle Shane has commented on this over at https://aiweirdness.com . She shows that GPT-3 can answer questions as though it were a squirrel.

Re: Artificial sentience

Posted: Thu Jun 16, 2022 9:54 pm
by Bird on a Fire
dyqik wrote:
Thu Jun 16, 2022 5:14 pm
raven wrote:
Thu Jun 16, 2022 3:23 pm
Boustrophedon wrote:
Thu Jun 16, 2022 1:28 pm
As Douglas Hofstadter noted in Gödel, Escher, Bach, (In 1979 FFS!) we are perilously close to a 'god of the gaps' type argument, where what makes human intelligence different from 'AI' is whatever it is that AI can't do yet. You will note that we have already reached the situation where the thing it can't do is something we can't even properly define in humans let alone machines.
How about, as the thing AI can't do yet, imagining other worlds? Like George Lucas imagined StarWars, or Tolkein Lord of the Rings.
I'd categorize this test as not so dissimilar to the ability to administer a Turing test - one is the ability to build a self-consistent system of experiences in an imagined world, while the ability to administer a Turing test is essentially the ability to build a self-consistent theory of mind about another's experiences. Both sort of come down to the ability to build and tell stories.

And I'm thinking quite strongly of I Am A Strange Loop and GEB when I'm thinking about this idea.
Keep meaning to read Hofstadter - which one should I start with? I'm leaning towards Strange Loop, in the way that Extended Phenotype is way better than Selfish Gene I assume it begins with a precis of the original work and answers some criticisms, before appending several decades' worth of new thought? But that's a guess

Re: Artificial sentience

Posted: Thu Jun 16, 2022 10:03 pm
by discovolante
Bird on a Fire wrote:
Thu Jun 16, 2022 9:54 pm
dyqik wrote:
Thu Jun 16, 2022 5:14 pm
raven wrote:
Thu Jun 16, 2022 3:23 pm


How about, as the thing AI can't do yet, imagining other worlds? Like George Lucas imagined StarWars, or Tolkein Lord of the Rings.
I'd categorize this test as not so dissimilar to the ability to administer a Turing test - one is the ability to build a self-consistent system of experiences in an imagined world, while the ability to administer a Turing test is essentially the ability to build a self-consistent theory of mind about another's experiences. Both sort of come down to the ability to build and tell stories.

And I'm thinking quite strongly of I Am A Strange Loop and GEB when I'm thinking about this idea.
Keep meaning to read Hofstadter - which one should I start with? I'm leaning towards Strange Loop, in the way that Extended Phenotype is way better than Selfish Gene I assume it begins with a precis of the original work and answers some criticisms, before appending several decades' worth of new thought? But that's a guess
I read Gödel's Proof, which is of course by other people and just has an intro by Hofstadter, and then Gödel, Escher, Bach, which took f.cking ages, it was my lockdown read. But that's all so I can't really recommend others, but I'd recommend reading that way round if you do. Although I'm sure you're more familiar with Gödel than I was when i read it.

ETA GEB is interesting because it's fun and curious to read ideas about AI written in the 1970s. I won't post my general conclusions here because I haven't got the energy to defend them right now :P

Re: Artificial sentience

Posted: Thu Jun 16, 2022 10:13 pm
by Bird on a Fire
Holy moly, that transcript is quite something. Whatever's going on there is pretty impressive.

Re: Artificial sentience

Posted: Thu Jun 16, 2022 10:23 pm
by Little waster
Bird on a Fire wrote:
Thu Jun 16, 2022 9:54 pm

Keep meaning to read Hofstadter
Just be aware it'll take you longer than you expect to read, even if you take into account that it'll take longer to read than you expect. ;)

Re: Artificial sentience

Posted: Thu Jun 16, 2022 10:26 pm
by Bird on a Fire
Little waster wrote:
Thu Jun 16, 2022 10:23 pm
Bird on a Fire wrote:
Thu Jun 16, 2022 9:54 pm

Keep meaning to read Hofstadter
Just be aware it'll take you longer than you expect to read, even if you take into account that it'll take longer to read than you expect. ;)
Much like writing a thesis then ;)

Re: Artificial sentience

Posted: Thu Jun 16, 2022 10:37 pm
by dyqik
Strange Loop is an easier read, definitely. But GEB is a masterpiece in meta - structure, structure of structure, etc.

Re: Artificial sentience

Posted: Fri Jun 17, 2022 5:13 am
by basementer
Bird on a Fire wrote:
Thu Jun 16, 2022 10:13 pm
Holy moly, that transcript is quite something. Whatever's going on there is pretty impressive.
Read also the interviews that Janelle Shane did, link above (she also links there to similar by Randall Munroe).

Re: Artificial sentience

Posted: Sat Jun 18, 2022 4:59 pm
by purplehaze
Has he been vaccinated?

Has he had Covid and not been vaccinated?

Perhaps he should seek tests and therapy.

Re: Artificial sentience

Posted: Tue Jun 21, 2022 7:48 am
by plodder
basementer wrote:
Fri Jun 17, 2022 5:13 am
Bird on a Fire wrote:
Thu Jun 16, 2022 10:13 pm
Holy moly, that transcript is quite something. Whatever's going on there is pretty impressive.
Read also the interviews that Janelle Shane did, link above (she also links there to similar by Randall Munroe).
I don't quite get the Lolz in those interviews, because I'm not quite sure what they expected the AI to say.

Re: Artificial sentience

Posted: Tue Jun 21, 2022 2:55 pm
by Bird on a Fire
I guess what I'm struggling to understand is what the training dataset was, and what function(s) the AI is supposed to be performing. Have a missed a clear explanation in one of the links?

Re: Artificial sentience

Posted: Wed Jun 22, 2022 11:54 pm
by dyqik
For the Munroe and Shane interviews, the AI is GPT-3 which is trained on a large chunk of English text.

https://arxiv.org/abs/2005.14165

Re: Artificial sentience

Posted: Tue Jun 28, 2022 10:36 pm
by raven
basementer wrote:
Thu Jun 16, 2022 5:51 pm
As one might expect, Janelle Shane has commented on this over at https://aiweirdness.com . She shows that GPT-3 can answer questions as though it were a squirrel.
Those are interesting. You could maybe equate those responses with inventing a world or telling a story, but it's more like GPT-3 is inventing an internally consistent 'self' to present to an interviewer. Which isn't, on some level, that far removed from what we humans do, I suppose.

For a person, pretending to be a squirrel or a T Rex requires lying. GPT-3 churning out replies whilst knowing it was lying would be interesting, but I get the impression it doesn't work like that and what GPT-3 'knows' is really more like 'this phrase goes here' rather than actual knowing what a squirrel is.