[ExI] GPT-4 on its inability to solve the symbol grounding problem

Brent Allsop brent.allsop at gmail.com
Sun Apr 9 01:56:54 UTC 2023

On Sat, Apr 8, 2023 at 7:23 PM Giovanni Santostasi via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

> It is useless to ask GPT-4 if it is conscious or understands. There are
> several reasons for this statement. First, there are certain topics that
> are considered sensitive and GPT-4 has received instructions on top of its
> normal training to give warnings and disclaimers on these topics. This is
> why it almost always gives bottled-up answers on topics related to
> consciousness and awareness. It does the same thing when asked about
> medical topics (reminding the users to consult a doctor) or legal topics
> and similar ones.
> Second, even if it simply used a statistical method to answer these topics
> most of the literature that GPT-4 has access to has a very conventional and
> conservative view on AI.
> Mostly it is actually missing the recent breakthroughs in AI given the
> training of GPT-4 goes up to 2021.
> Furthermore, consciousness is what consciousness does. It is not about
> answering if you are conscious or not. If an entity is not conscious and
> answers "I'm not conscious" then this shows a certain level of awareness so
> it has to be conscious (and therefore it is lying). If an entity is
> conscious and it answers "I'm not conscious", then we will not be able to
> distinguish it from the previous case (basically they are the same).  So
> asking an entity if it is conscious while receiving the answer "I'm not" is
> the worst type of test we can imagine.
> If the machine said "I'm conscious and I want rights" and there is
> evidence that the machine does this in a sophisticated way (it demonstrates
> other nontrivial cognitive abilities) we should use caution and take the
> machine's statements at face value.
> The only true way to test for sparks of awareness and true understanding
> is to do experiments that push the limits of what GPT-4 was trained for and
> look for signs of cognitive abilities that are not expected from a simple
> autocorrection tool.
> I and others have given several examples. In particular, I have shown this
> understanding goes beyond text and includes the capability to go from text
> to symbols and back and the ability to be creative in a nontrivial way in a
> multi-modal way. We have discussed this for days now and it seems certain
> people are simply stuck in their own prejudices without considering or
> answering the counter-example given or the general discussion of this
> topic.

We seem to be talking about two completely different things.  You are
talking about general intelligence, while I am talking about what mechanism
any intelligent system is using to represent information and compute with.
(has nothing to do with general intelligence, other than some ways
(representing information directly on qualities, no dictionary required)
are more efficient than others (abstract - required additional dictionaries
to know what the property represents.)

> I do want to ask the other camp (that is basically 2 people at this point)
> what would be required for them to agree these AIs are conscious. I don't
> think I saw a concise and meaningful answer to this question in the 100s of
> posts so far.

Sorry you missed this.  Let me summarize.  There is general agreement that
redness is not a quality of the strawberry, it is a property of
our conscious knowledge of the strawberry - a quality of something in our
head.  The question is, what is the nature of that redness property?
like Stathis, predict that redness "supervenes" on some function.  Dualists
predict redness is a spiritual quality of something non physical.  "
', like me, predict that redness is just a physical property of something
in our brain.  I like to take Glutamate as an example of a hypothetical
possibility.  We predict that glutamate behaves the way it does, in a
synapse, because of its redness quality.  If someone experiences redness,
without glutamate, that hypothesis would then be falsified.  Then you test
for something else in the brain, till you find what it is that is reliably
responsible for someone experiencing redness.  Functionalists like Stathis
seem to predict that redness can "arise" from some function.  I predict
that they will never find any function that will result in a redness
experience, and that without glutamate, redness will not be possible.
Obviously a falsifiable claim.

If you objectively observe a bat representing some of its echolocated
knowledge with glutamate (your redness), Not only will you know the bat is
conscious, you will know what part of that bat's knowledge is like.  All
that will be possible, once we have a dictionary that said it is glutamate
that behaves the way it does, because of its redness quality.

So, objectively observing whether something is conscious or not, has
nothing to do with what it is saying, as you point out.  Once we have the
required dictionary telling us which of all our abstract descriptions of
stuff in the brain has what colorness qualities, it is simply objectively
observing  what we see in the brain, then using the dictionary to know not
only if it is conscious, but to also know what it is like.

Again, consciousness isn't a 'hard problem' it is just a color quality
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230408/4e564671/attachment.htm>

More information about the extropy-chat mailing list