[ExI] GPT-4 on its inability to solve the symbol grounding problem
jasonresch at gmail.com
Thu Apr 13 10:21:16 UTC 2023
On Thu, Apr 13, 2023, 3:13 AM Gordon Swobe <gordon.swobe at gmail.com> wrote:
> On Wed, Apr 12, 2023 at 3:54 PM Jason Resch via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>> Let's see if when we agree on a premise that we can reach the same
> If we assume there's not something critical which we have yet to discover
>> about the brain and neurons, would you agree that the inputs to the brain
>> from the external world are ultimately just nerve firings from the senses,
>> and from the brain's point of view, the only information it has access to
>> is the timings of which nerves fire when? If you agree so far, then would
>> you agree the only thing the brain could use as a basis of learning about
>> the external world are the correlations and patterns among the firing…
> I’m not sure about “correlations and patterns,” but yes only if I
> reluctantly make that assumption that there is nothing more to the brain
> and mind then I can agree. Recall that you already asked me this question
> and I replied that I am not a strict empiricist.
Okay this is some progress, we have identified the point where our
assumptions differed which explains our disagreement.
> Also, even if that is all there is to the brain and mind, in my view and
> in agreement with Nagel, no *objective* description of these neural
> processes in the language of science or computation can capture the facts
> of conscious experience which exist not objectively, but only from a
> particular point of view.
We agree on this. The reason we can agree on and describe physical facts is
because we have shared referents in the physical world and a shared
understanding of math. We can both point to a meter stick and hold it and
see how long it is. This is why colors and sounds as they feel to us are
not describable, I cannot see into your head anymore than you can see into
mine. We have no common reference points on which to establish
> You might want to argue that my position here leads to dualism, but that
> is not necessarily the case. The dualist asserts a kind of immaterial
> mind-substance that exists separate from the material, but that supposed
> mind-substance is thought to exist objectively. The dualist makes the same
> mistake as the physicalist.
Did you see my thread on how computationalism leads to and recovers many
aspects of consciousness that have traditionally been ascribed to souls? I
wrote that in the hopes it might help serve as a bridge to connect our two
world views. Below, I think I can offer another:
That first person (non objective) properties emerge, counterintuitively
does not imply they cannot emerge from a system that is ultimately
objectively describable. As I understand it, this is your main motivation
for supposing there must be more going on than our objective accounts can
explain. In a sense you are right. There are first person properties that
we cannot access from our vantage point looking at the system from the
But recently it's been shown, somewhat technically, how for certain complex
recursive systems, these first person properties naturally emerge. This
happens without having to add new neuroscience, physics, or math, just
applying our existing understanding of the mathematical notion of
“In this paper I have argued that human brains can have logical properties
which are not directly accessible to third-person investigation but
nevertheless are accessible (at least in a weak sense) to the brain itself.
It is important to remember that these properties are not metaphysically
mysterious in any way; they are simply logical properties of neural
systems. They are natural properties, arising entirely from the processing
of information by various subsystems of the brain. The existence of such
properties can pose no threat to the scientific understanding of the mind.”
“The existence of these logical properties contradicts the widespread
feeling that information processing in a machine cannot have features
inaccessible to objective observers. But despite this offense against
intuition, these findings support a view of first-person access which may
be far more congenial to a scientific understanding of the mind than the
alternative views that first-person character is either irreducible or
unreal. Our conclusion suggests a way to bypass an important obstacle to a
reductionistic account of consciousness. Indeed, it suggests that
consciousness may be reducible to information processing even if experience
does have genuine first-person features.”
I hope this paper might show that we can keep our inaccessible,
irreducible, real first person properties *and* have a rational description
of the brain and it's objectively visible behavior. We don't have to give
up one to have the other.
> I agree the human brain is not akin to a LLM.
>> But this is separate from the propositions you also have disagreed with:
>> 1. That a digital computer (or LLM) can have understanding.
>> 2. That a digital computer (or LLM) can be conscious.
> Yes, and it is best to combine the two. I disagree that LLMs have
> conscious understanding.
> I give the LLM some instructions. It follows them. I concluded from this
>> the LLM understood my instructions. You conclude it did not.
>> I must wonder: what definition of "understand" could you possibly be
>> using that is consistent with the above paragraph?
> As above, LLMs have no *conscious* understanding, and this LLM called
> GPT-4 agrees.
> As I’ve written, the sort of unconscious understanding to which you refer
> is trivial and uninteresting. My smart doorbell “understands” when there is
> motion outside my door. I am not impressed.
Let's say you update your understanding after reading the above page I
linked, and you decide that the LLM has the necessary recursive logical
structures to have internally inaccessible first person properties that are
inaccessible from the outside. Would this change your opinion on whether
the LLM could be consciousness?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the extropy-chat