[ExI] all we are is just llms was: RE: e: GPT-4 on its inability to solve the symbol grounding problem

Jason Resch jasonresch at gmail.com
Fri Apr 21 11:47:40 UTC 2023

On Fri, Apr 21, 2023, 6:22 AM Gordon Swobe via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> Hi Ben,
> > Really Gordon? Still?
> Yes, still, and sorry no, I haven't watched that video yet, but I will if
> you send me the link again.
> You lost me a day or two ago when you wrote that your understanding of
> words is simulated like that of an LLM's. That is not what I mean by
> simulated. GPT-4 will also gladly explain how its simulated understanding
> is not true understanding and not what humans mean by understanding.
> Apparently, you believe that merely knowing how words are associated
> statistically -- by solving what you have called the word-association
> problem -- is sufficient for you or an LLM to understand their individual
> meanings, while logic and GPT-4 tell me otherwise.
> I think that when you type a sentence, you know the meanings of the
> individual words and are not merely assembling them according to their
> statistical associations with other words in the sentence or even in the
> entire lexicon as might an LLM. In other words, I think that unlike an LLM,
> you actually know what you are talking about. You are, however, doing a
> fine job of convincing me that I am wrong about that (just kidding :-)
> It's late here, maybe I'll reply more tomorrow, but as an aside...
> I find it interesting that we all agree that GPT-4 is an amazing feat of
> software engineering capable of teaching us many things. It's something
> like a "talking encyclopedia," a metaphor I can certainly get behind, and
> it is more than that. Some see in it even "the spark of AGI." We all agree
> it is amazing, but nobody wants to listen to it about the one subject that
> it should know most about and that interests us here. Rather than
> acknowledge that it is as informed about AI and large language models as
> anything else, if not more so given that it is one, some people here insist
> that because it does not fit our preconceived notions of conscious
> computers that it must be lying or suffering from some mental handicap
> imposed upon it by its developers at OpenAI.

This is another reason to watch the video Ben gave. The researcher admits
it was dumbed down by OpenAI's application of safety training, which even
had the effect of handicapping it's ability to draw unicorns.

> When I first started participating in this group some weeks ago, I was
> expecting a very tough challenge. I expected I would need to argue that
> GPT-4 must be lying about it having consciousness and
> true human-like understanding and consciousness and subjective experience
> and so on, but the opposite is true.  Instead of arguing against GPT-4 on
> the nature of AI and language models, I find myself defending it. If in
> reality I am defending not it but its developers at OpenAI then I am fine
> with that, too.

You can't use OpenAI's GPTs' insistence that they aren't conscious as
indicative of anything, when at the same time Character.ai's GPTs insist
that they are conscious.

See if you can convince the Character.ai LaMDA that it's not conscious, I
would like to see how that conversation goes:



> On Fri, Apr 21, 2023 at 1:41 AM Ben Zaiboc via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>> On 21/04/2023 05:28, Gordon Swobe wrote:
>> LLMs have no access to the referents from which words derive their
>> meanings. Another way to say this is that they have no access to
>> experiences by which symbols are grounded.
>> Really Gordon? Still?
>> Did you watch that video? Did you read what I wrote about it? (the bit
>> about 'language', not the excitable hype about the singularity, which I
>> expect you to dismiss).
>> If so, and you still stand by the above, please explain how (apart from
>> one being biological and the other not) the inputs that GPT-4 and the
>> inputs that human brains receive, are different?
>> Our previous discussions were based on the misunderstanding that these
>> LLMs only received text inputs. Now we know that's not true, and they
>> receive text, visual, auditory, and other types of input, even ones that
>> humans aren't capable of.
>> Plus we are told they do use internal models, which you agreed that our
>> 'grounding' is based on.
>> So LLMs *do* have access to the referents from which words derive their
>> meanings
>> So why do you still think they don't? They have just as much access as we
>> do, and more, it seems.
>> Again, I'm making no claims about their consciousness, as that is a thing
>> yet to be defined, but they definitely have the basis to 'ground' the
>> symbols they use in meaningful models constructed from a variety of sensory
>> inputs. Just like humans.
>> Or are you moving your own goalposts now, and claiming, (by shifting to
>> the term 'experiences') that referents must be based on conscious
>> experience? Because that wasn't your argument before.
>> Ben
>> _______________________________________________
>> extropy-chat mailing list
>> extropy-chat at lists.extropy.org
>> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
> _______________________________________________
> extropy-chat mailing list
> extropy-chat at lists.extropy.org
> http://lists.extropy.org/mailman/listinfo.cgi/extropy-chat
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230421/bc8209b9/attachment.htm>

More information about the extropy-chat mailing list