Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Artificial neural networks are making strides towards consciousness?

economist.com

Post reply on HN
Print view View original

61–70 of 85 posts

Re: Artificial neural networks are making strides towards consciousness?

#61

Post by vixen99 »

mdp2021 wrote: post #50
whiplash451 wrote: post #33

In the same edition of The Economist, an article by Douglas Hofstadter debunks this statement in a pretty compelling way.

https://archive.ph/5QS7t -- > [Hofstadter: ] I am at present very sceptical that there is any consciousness in neural-net architectures such as, say, GPT-3, despite the plausible-sounding prose it churns out at the drop of a hat. Let me explain my viewpoint through a few examples. // My colleague David Bender and I have recently enjoyed inventing all sorts of amusing probes of GPT-3 that reveal a mind-boggling hollow…

Hofstadter's books (such as Surfaces and Essences) provide a salutary reminder of how distant we are in AI technology in approach to a machine understanding of even ordinary human language expression. For those unfamiliar with his work, Hofstadter's premise is that analogy is the core of human cognition.

Re: Artificial neural networks are making strides towards consciousness?

#62

Post by machiaweliczny »

chasing wrote: post #45

> This capacity to produce a stable, psychological model of self is also widely understood to be at the core of the phenomenon we call “consciousness”. No, it's not. If you're calling something "conscious" you're arguing that it has a first person subjective experience. A soul. A spark. If you're calling a machine "conscious" you're saying you have somehow given it life. This is a major fucking claim and we'll need s…

IMO consciousness might be just feeling of internal state so to speak which acts as input also. So it might be basically process with memory that oversees other processes. This kind of scheduler makes obvious sense in parallel architecture.

Assuming one has limited computational capacity and some know probabilistic distribution and needs to probe stuff for maximum EV. It make sense to have ability to oversee and kill computation and also to have ability to bring new computation to attention in need.

Hope I am wrong.

Re: Artificial neural networks are making strides towards consciousness?

#63

Post by machiaweliczny »

machiaweliczny wrote: post #62
chasing wrote: post #45

> This capacity to produce a stable, psychological model of self is also widely understood to be at the core of the phenomenon we call “consciousness”. No, it's not. If you're calling something "conscious" you're arguing that it has a first person subjective experience. A soul. A spark. If you're calling a machine "conscious" you're saying you have somehow given it life. This is a major fucking claim and we'll need s…

IMO consciousness might be just feeling of internal state so to speak which acts as input also. So it might be basically process with memory that oversees other processes. This kind of scheduler makes obvious sense in parallel architecture. Assuming one has limited computational capacity and some know probabilistic distribution and needs to probe stuff for maximum EV. It make sense to have ability to oversee and kill…

I am almost certain that people won’t be able to recognise such AI scheduler from human. Maybe only due to non-understandable goals. But if embodied and limited to our senses it would pass as conscious to most in some blind test.

Re: Artificial neural networks are making strides towards consciousness?

#64

Post by mdp2021 »

vixen99 wrote: post #61
mdp2021 wrote: post #50

Earlier quoted context omitted.

https://archive.ph/5QS7t -- > [Hofstadter: ] I am at present very sceptical that there is any consciousness in neural-net architectures such as, say, GPT-3, despite the plausible-sounding prose it churns out at the drop of a hat. Let me explain my viewpoint through a few examples. // My colleague David Bender and I have recently enjoyed inventing all sorts of amusing probes of GPT-3 that reveal a mind-boggling hollow…

Hofstadter's books (such as Surfaces and Essences ) provide a salutary reminder of how distant we are in AI technology in approach to a machine understanding of even ordinary human language expression. For those unfamiliar with his work, Hofstadter's premise is that analogy is the core of human cognition.

Available in the Library at archive.org :

Hofstadter, Douglas R; Surfaces and essences: analogy as the fuel and fire of thinking (2013) - https://archive.org/details/surfacesessences0000hofs

Re: Artificial neural networks are making strides towards consciousness?

#65

Post by SheddingPattern »

layer8 wrote: post #26
fdgsdfogijq wrote: post #7

Queue the hacker news haters. As someone that works with neural networks on a daily basis, I am constantly stunned by the effectiveness of modern attention based NLP models. They actually work for real world tasks, and you don't need to be an expert in them to apply them. Not only that, the newer ones are drastically better than the ones only two years ago. These are the models released in public. The sheer amount of…

I don’t care about consciousness. I’d be overjoyed if voice assistants and search engines would even remotely understand what I want from them, and if I could help them improve their understanding and teach them my preferences by interactive conversation. We’re not anywhere near that, most importantly the interactive part.

Exactly. Is this story anything other than marketing? Something to keep interest in their flagship product.

Re: Artificial neural networks are making strides towards consciousness?

#66

Post by im3w1l »

Imnimo wrote: post #3

>In our conversation, lamda tells me what it believes Ramesh felt that Lucy learned about what Mateo thought about Lucy’s overture. This is high order social modelling. I am not sure the example convinces me that what GPT-3 is doing here is 'modeling' the agents in the story. It feels like the author is trying to draw a contrast between rote statistical language modeling and higher-order understanding, and it's not c…

A model is not required to get everything right. I found those deductions absolutely amazing honestly.

Re: Artificial neural networks are making strides towards consciousness?

#67

Post by YeGoblynQueenne »

lemmsjid wrote: post #48
divs1210 wrote: post #13

Earlier quoted context omitted.

Yup. I'll believe it when the model gets decent grades in a GRE comprehension test.

That was an intriguing suggestion. I found this GRE question: https://www.ets.org/gre/revised_general/prepare/verbal_reaso... . I'm kind of in a rush, so I picked some questions around Question 2 and gave them to OpenAI's chat bot. I reformulated the questions to make them not a multiple choice scenario. As far as the answer guide is concerned, it got the most obviously correct interpretation right, it also identifie…

Searching online for the following text:

"AI: The passage addresses the issue of whether or not Glass's use of popular elements in his classical compositions makes him a composer of popular music."

I find the following two pages both of which have text that looks similar to GPT-3's answer:

https://www.ets.org/gre/revised_general/prepare/verbal_reaso...

>> The passage describes in general terms how Philip Glass uses popular music in his classical compositions and explores how Glass can do this without being imitative.

https://greprepclub.com/forum/reviving-the-practice-of-using...

>> The passage addresses the issue of Glass's use of popular elements by stating that inspite of using these elements, Glass himself is not regarded as a composer of popular music.

There's a chance that GPT-3 was trained on the text of those web pages and learned the answer of the question from those. There may be more such resources on the internet that my perfunctory search did not identify.

In any case, making sure the answers returned by GPT-3 to a query are original, and have not been memorised, is up to the person who claims they are not. Is that your claim? In that case, you should do the job.

Also: the system is GPT-3, not "AI".

Edit: I just realised the first link I post above is where you got the GRE questions from. That page lists not only questions, but also the correct answers to questions.

Re: Artificial neural networks are making strides towards consciousness?

#68

Post by AnIdiotOnTheNet »

alasdair_ wrote: post #49
edgyquant wrote: post #34

Earlier quoted context omitted.

I don’t see how consciousness is at all hard to define. A system that is aware of itself. Checking for consciousness is harder, but the rule of thumb should be anything that makes a highly subjective case it’s conscious probably is.

>I don’t see how consciousness is at all hard to define. A system that is aware of itself. A thermostat that is aware of the temperature inside its own housing is "aware of itself". It's not conscious however. Nailing down exactly what consciousness means is remarkably difficult.

Honestly I think this is largely because humans want the conclusion that only they are consious, or at least that only they and some animals are, and are trying really hard to carve out a definition that makes that work. Why would we do that? Well obviously because we create AI to do work for us and we don't want to have any ethical considerations get in the way of treating our new slaves however we want.

Re: Artificial neural networks are making strides towards consciousness?

#69

Post by AnIdiotOnTheNet »

mdp2021 wrote: post #50
whiplash451 wrote: post #33

In the same edition of The Economist, an article by Douglas Hofstadter debunks this statement in a pretty compelling way.

https://archive.ph/5QS7t -- > [Hofstadter: ] I am at present very sceptical that there is any consciousness in neural-net architectures such as, say, GPT-3, despite the plausible-sounding prose it churns out at the drop of a hat. Let me explain my viewpoint through a few examples. // My colleague David Bender and I have recently enjoyed inventing all sorts of amusing probes of GPT-3 that reveal a mind-boggling hollow…

What this shows us is that GPT-3 doesn't really understand what it is talking about and is just a decently good bullshitter. To be frank, I've met humans with the same quality of always speaking with confidence about things they have no understanding of. Possibly this is a result of how the system is constructed, GPT-3 being forced to answer but not able to express uncertainty, for instance. At any rate I do not believe it is sufficient evidence that GPT-3 is not, in some way, conscious.

Re: Artificial neural networks are making strides towards consciousness?

#70

Post by Engineering-MD »

YeGoblynQueenne wrote: post #67
lemmsjid wrote: post #48

Earlier quoted context omitted.

That was an intriguing suggestion. I found this GRE question: https://www.ets.org/gre/revised_general/prepare/verbal_reaso... . I'm kind of in a rush, so I picked some questions around Question 2 and gave them to OpenAI's chat bot. I reformulated the questions to make them not a multiple choice scenario. As far as the answer guide is concerned, it got the most obviously correct interpretation right, it also identifie…

Searching online for the following text: "AI: The passage addresses the issue of whether or not Glass's use of popular elements in his classical compositions makes him a composer of popular music." I find the following two pages both of which have text that looks similar to GPT-3's answer: https://www.ets.org/gre/revised_general/prepare/verbal_reaso... >> The passage describes in general terms how Philip Glass uses p…

I think these kind of things are so common and sometimes difficult to see when words are the altered. I’d be interest to see it go through plagiarism checkers (especially semantic ones as well as actual words). When they are trained on a significant proportion of the internet it makes it hard to test without already having seen it or something similar. Little in life is new it seems, and tests are past so much more easily when you’ve already seen the test paper.
Post reply on HN