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

Mon Apr 3 18:31:27 UTC 2023

It has been said and I agree that language models are like advanced
auto-completes. Based on common English usage on which they have been
trained extensively, they not only can predict which word will be most
sensible to us in a sentence, but also which paragraph will be most
sensible in an essay, which answer will be most sensible after a question,
and so on, constrained only by their token limit and the extent of the
datasets on which they have been trained and the imaginations of the humans
who write the prompts. It is an amazing feat of software engineering.

-gts

On Mon, Apr 3, 2023 at 12:17 PM Gordon Swobe <gordon.swobe at gmail.com> wrote:

>
> People keep attributing these claims to me, but as below, it is ChatGPT-4
> itself that says it has no true understanding of the world.
>
> It understands (so to speak) how language models work.
>
> -gts
>
>
>
>
>
> On Mon, Apr 3, 2023 at 12:09 PM Brent Allsop via extropy-chat <
> extropy-chat at lists.extropy.org> wrote:
>
>>
>> Yes, it is very exciting to see this much discussion about
>> colerness qualities, again.  In the old days, the term 'qualia' became a
>> dirty word on this list, and most people would just filter out any thread
>> containing that word.  Chat GPT is making it necessary that we explicitly
>> understand the difference between abstract systems and phenomenal systems
>> that represent knowledge with physical qualities (ok phenomenal qualities
>> for those of you that think qualities are not physical,  or not
>> approachable/sharable via science, apparently the majority of you?)
>>
>> But I disagree with your assertion that GPT4 has no idea about what it is
>> talking about, and I don't think GPT4 is saying that at all in these
>> answers.  Its large language model clearly has the ability to model, at
>> least abstractly, a large subset of what humanity talks about in very
>> powerful ways, and is on the verge of General Intelligence.  Just because
>> it doesn't know what redness is like, doesn't mean it can't know more about
>> redness than most humans.  Just like Black and White Marry, it can clearly
>> indicate in these kinds of answers that it has a very powerful grasp of
>> reality, and the difference between itself and us, and what Mary will learn
>> when she experiences redness  It is only a matter of time till an abstract
>> AI helps us discover which of all our descriptions of stuff in the brain is
>> a description of redness, then endows itself with that so it can finally
>> say: 'oh THAT is what redness is like." and thereby finally having a
>> grounded understanding, making it also, phenomenally conscious like us.
>>
>> I so look forward to the day when we finally have an interest to do the
>> same.  Finally discovering new phenomenal qualities no human has ever
>> experienced before, and endowing ourselves with those.  I'm so tired of
>> only knowing the qualities things seem to be.  Am I the only one that wants
>> to know what all that stuff out there is really like, and what phenomenal
>> uploading to phenomenal avatars will really be like in a few more years?
>>
>>
>>
>>
>> On Fri, Mar 31, 2023 at 5:41 AM Gordon Swobe via extropy-chat <
>> extropy-chat at lists.extropy.org> wrote:
>>
>>>
>>> Many people here seem interested in the question of whether LLMs can
>>> have conscious understanding of the words/symbols that it takes as inputs
>>> and generates as outputs. It seems that most people here want to think GPT4
>>> has something like a conscious mind, but I agree with GPT4  below that it
>>> has no idea what it is talking about.
>>>
>>>
>>> On Thu, Mar 30, 2023 at 1:55 PM Gordon Swobe <gordon.swobe at gmail.com>
>>> wrote:
>>>
>>> GPT-4 explained to me that it could not solve the symbol grounding
>>>> problem for itself (I reported this in another thread). I asked it to
>>>> explain the significance.
>>>>
>>>> My prompt: What is the significance of the fact that ChatGPT cannot
>>>> solve the symbol grounding problem for itself and can only simulate that it
>>>> has solved it? How does this put ChatGPT at a disadvantage compared to
>>>> humans?
>>>>
>>>> GPT-4: The symbol grounding problem refers to the difficulty in
>>>> connecting abstract symbols (such as words or concepts) to their real-world
>>>> referents. While ChatGPT can simulate understanding by generating
>>>> human-like text based on its training data, it doesn't possess the direct
>>>> experience or sensory input that humans have to ground these symbols in
>>>> reality.
>>>>
>>>> The significance of this limitation has several implications:
>>>>
>>>> Lack of true understanding: ChatGPT doesn't have a genuine
>>>> understanding of the world, as it relies on patterns and associations found
>>>> in the text it was trained on. It can't form new connections based on
>>>> direct experience, which may result in limitations in problem-solving or
>>>> abstract reasoning.
>>>>
>>>> Difficulty with context: Without grounding in real-world experiences,
>>>> ChatGPT may struggle to fully understand context or cultural nuances. It
>>>> may provide answers that are technically correct but inappropriate or
>>>> irrelevant in a specific situation.
>>>>
>>>> Inability to learn from sensory input: Humans learn through direct
>>>> experience and sensory input, which allows them to adapt and understand new
>>>> situations better. ChatGPT can't update its knowledge through real-world
>>>> experiences, making it less adaptable to novel contexts.
>>>>
>>>> Absence of common sense: Although ChatGPT can generate text that
>>>> appears to show common sense understanding, it might still make errors that
>>>> a human would easily avoid. This is because its responses are based on
>>>> patterns in the data, rather than a true comprehension of the concepts.
>>>>
>>>> Limited creativity: ChatGPT's creativity is confined to the scope of
>>>> its training data. While it can generate new text based on patterns it has
>>>> seen before, it can't create truly novel ideas or solutions in the same way
>>>> that humans, with their real-world grounding, can.
>>>>
>>>> In summary, the inability to solve the symbol grounding problem puts
>>>> ChatGPT at a disadvantage compared to humans in terms of true
>>>> understanding, adaptability, context-awareness, common sense, and
>>>> creativity. However, it is essential to remember that ChatGPT is an AI
>>>> language model designed to assist and augment human capabilities, not to
>>>> fully replicate human cognition.
>>>> ---
>>>>
>>>
>>> -gts
>>>
>>>>
>>>> _______________________________________________
>>> 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/20230403/be0360ae/attachment.htm>