[ExI] Is the GPT-3 statistical language model conscious?

[Dave Sill]

On Tue, Oct 13, 2020 at 9:14 AM Stuart LaForge via extropy-chat <
extropy-chat at lists.extropy.org> wrote:

>
> Quoting Dave Sill <sparge at gmail.com>:
>
> > The test is a demonstration of what GPT-3 is, and isn't. It is good at
> > generating reasonable text. It isn't smart.
>
>  From what I have been able to see of its output, it actually is
> pretty smart when comes to writing stuff. It just seems to lack common
> sense which is understandable since GPT-3 has no sensory inputs except
> for text. This could cause it to underperform on tasks that would
> require it to associate text with sensory and motor experiences just
> as Bill Hibbard observed earlier.
>

GPT-3 is a statistical language model. It deep-learns from a massive amount
of written text. It has no mechanism whatsoever for understanding the text
or reasoning/thinking/planning/problem solving/self awareness. Any
intelligence you perceive in its output comes from the authors of the
training text.

> The original question of the thread was: is GPT-3 conscious. I think it's
> > clearly not.
>
> You have made that quite obvious. And while I do value your opinion, I
> am agnostic at this point barring further reliable data but very
> curious. Therefore, I have joined the waitlist to beta test GPT-3
> through an API for research purposes. If my request is approved, I
> think it would be interesting experiment to have GPT-3 set up to post
> to ExI's mailserver although I would need assistance with that from
> John Klos and perhaps you or one of the other tech gurus on the list.
> Then we could generate our own reliable data.
>
> Are you interested?
>

Sure, that would be interesting. It's a shame that OpenAI isn't really open
and that Microsoft "owns" GPT-3.

-Dave
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20201013/cfaefda8/attachment.htm>