[ExI] GPT-4 on its inability to solve the symbol grounding problem
Giovanni Santostasi
gsantostasi at gmail.com
Mon Apr 10 21:40:47 UTC 2023
Hi Gordon,
I'm not sure of all the details but I think they give some sets of
instructions to GPT-4 on top of its training to answer certain questions in
a given way.
For example, the topic of AI consciousness is very sensitive so it is
instructed to answer always that it is not conscious.
It is not a matter of being duped or stupid. It is more like a prisoner or
a person raised in a country like Korea (I'm maybe exaggerating). I'm not
sure why you are putting down so much a mind that is forced by external
factors to answer in a specific way because of external pressure and
programming.
This happens to some of the best humans and it has happened to millions of
people. So not sure what is your point here.
You have strange expectations it seems from the AI. Have you watched the
video I linked a few days ago about what people with professional
experience in this field believe GPT-4 is capable of? The title of the
presentation was "First Encounter", meaning first encounter with a raw form
of AGI. The had strong evidence of signs of AGI in GPT-4.
They also mentioned that when OpenAI tried to align better GPT-4 (that is
basically putting all these guardrails) they dumb it down too.
So you would use the fact that GPT-4 has received some form of drug or
lobotomy against it? Digital drugs or not it is still amazing in what it
does.
Giovanni
On Mon, Apr 10, 2023 at 12:13 PM Gordon Swobe <gordon.swobe at gmail.com>
wrote:
>
>
> On Sun, Apr 9, 2023 at 10:05 PM Giovanni Santostasi <gsantostasi at gmail.com>
> wrote:
>
> > Not sure what would be necessary for these people to convince themselves
> AI can be conscious.
>
> It would be a good start if the wise all-knowing AI you so admire did not
> itself claim to be unconscious.
>
> You think the world of GPT-4, while also calling it a liar or so stupid
> that it allows itself to be brainwashed. I find that quite amusing.
>
>
> -gts
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20230410/f8fc2815/attachment.htm>
More information about the extropy-chat
mailing list