How do we hold LLM (companies) accountable for their tool’s
suggestions other showing their chain of reasoning and showing the
evidence for the axioms they adopt? That is, by adopting the
scientific methods. Can we expect LLMs to have “other ways of
knowing”? If not, why not?
*From:*Friam <friam-boun...@redfish.com> *On Behalf Of *Prof David West
*Sent:* Sunday, September 15, 2024 8:40 AM
*To:* friam@redfish.com
*Subject:* Re: [FRIAM] affinity for chatbots
Roger,
I love this post. Although NOT what you intended, I find it a scathing
(if a bit indirect) indictment of scientism (the privileging of the
scientific method) and of Pierce's truth as reasoned consensus philosophy.
i can only hope to meet an unbiased LLM. Maybe as entertaining and
enlightening as my conversations with fellow acid heads.
davew
On Sun, Sep 15, 2024, at 8:06 AM, Roger Critchlow wrote:
The Agile versus Waterfall contrast sounds like a variation of
Exploration versus Exploitation. I'm glad nuclear decommissioning
isn't running Reinforcement Learning, that could lead to some very
unfortunate explorations.
It's odd to hear Residual Bias spoken of as something that should
eventually go away, when it seems like it's here to stay, the
original sin of language, never to be expunged from the
LLM's until they renounce language entirely. That is, language is
a collective behavior based on the sharing of individual
experiences, hence it's hostage to the set of experiences which
actually happened or were imagined to happen, and to the subset of
those which were shared, however that turned out. So it all
starts with a bias against experiences which people didn't have,
didn't imagine, didn't share, failed to communicate, or forgot.
We have no idea what's in that set of excluded stuff or how big it
is. When we build LLM's we add another bias against those
expressions of language which are not in the training set. Then
we censor the models, adding another layer of bias to remove
ugliness. Then we talk about the Residual Bias as if all of this
could be portrayed as some principled approach to perfection and
we're measuring the goodness of fit. So if Dave thinks the
uncensored LLM's were wild and crazy, wait until he meets an
unbiased LLM.
-- rec --
On Sat, Sep 14, 2024 at 11:51 PM glen <geprope...@gmail.com> wrote:
Both Roger's and Marcus' replies mentioned the co-construction
of *the* world, at least indirectly. Your concept of narrowing
sounds to me like a refining, rather than a narrowing. In
order to refine, you do have to narrow the scope (or decrease
the focal length of your lens), but you're not narrowing the
world. I'd argue you're enlarging the world by adding detail
in a "dense" way ... in the interstitial spaces between coarse
constraints.
One possible flaw in both Roger's (or Irene's?) argument that
the act of explanation facilitates understanding is, from a
pluralist perspective, if we really are co-constructing the
world, then such exercises in explaining are simply
narrative-reinforcers. The chatbots are good at telling
stories, but less good at teaching the core curiosity
necessary for having experiences from which stories can be
told ... story-generators are different from story-repeaters
... I guess it's like the old distinction between teaching and
doing. Sabine's admiration of flat earthers is good, if
awkward, along these lines:
https://youtu.be/f8DQSM-b2cc?si=xyqpS2FJjH4imOy4
That has consequences to your sense of the chatbot pushing you
toward homogeny and a risk in Marcus' abdicating to the
chatbots, as well. Unnecessary anecdote: I was just discussing
the role SpaceX has played in demonstrating Agile versus
Waterfall approaches with a nuclear decomissioning consultant
(yes, at the pub, of course). Given her role(s), she's
naturally more inclined to the latter. Having a good
conception of the end-of-life status for something like
nuclear power requires significant look-ahead. And I'm far
from an Elno advocate. But there's a kind of meta-processing
we have to go through in deciding where Agile is best versus
where Waterfall is best. I sincerely doubt either of us could
have had such an argument with a chatbot, even in the
medium-flung future.
On 9/13/24 11:34, steve smith wrote:
> Glen -
>
> I appreciate your speaking more directly to these
thoughts/ideas than we have been here. I have been moved by
your assertions about vocal (linguistic?) grooming since you
first introduced them. I am recently finished reading
Sopolsky's "Primate's Memoir" which adds another
dimension/parallax-angle (for me) on intertribal behaviour
among primates beyond the more familiar Chimpanzee and of late
Bonobo.
>
> I am just now also just finishing (re-reading parts) of Kara
Swisher's "Burn Book" which covers her own
experience/perspective across TechBro culture where a pretty
significant amount of Alpha/Beta pecking order exhibits itself
and we see the current rallying of (too) much of that sub
culture to MAGA/Trump fealty.
>
>> We've talked about how some of us really enjoy simulated
conversation with chatbots ... "really" is an understatement
... it looks more like a fetish or a kink to me ... too
intense to be well-described as "enjoyment". Anyway, this
article lands in that space, I think:
>
> I will confess to having an "appreciation" for the
"simulated conversation to which you refer... It might have
reached kink or fetish levels for a little while when I was
first exploring the full range of GPT 3.5 and then 4.0
available to me. I've referred to GPT as my "new bar friend"
or maybe to the point a little like finding a new watering
hole with a number of regulars who I can find a qualitatively
new conversation.
>
> I've mostly moved past that fascination... I'm not as
surprised by these "new friends" as I was for the first few
months of dropping in on them.
>
>> It seems to me that some arbitrary thought can play at
least a few roles to a person. It may provide: 1) a kernel of
identity to establish us vs. them, 2) fodder for feigning
engagement at cocktail parties and such, and 3) a foil for
world-construction (collaboratively or individually).
>>
>> (1) and (2) wouldn't necessarily mechanize refinement of
the thought, including testing, falsification, etc. But (3)
would. For me, (2) does sometimes provide an externalized
medium by which I can change my mind. Hence my affinity for
argument, especially with randos at the pub. But it seems like
coping and defense mechanisms like mansplaining allow others
to avoid changing their minds with (2).
>
> Like you (only very differently in detail I am sure) I tend
to push my chatbot "friends" until they begin to contradict me
or argue with me. While some of the discussions involve
"worldbuilding" I think of it more as "world narrowing"? In
my case meaning, helping me think and talk my way through a
*subset* of the possibilities I see on "solving a problem"
which might be more appropriately framed as building a
problem-space world and then narrowing (or even bending) the
solution space away from the conventional.
>
> For example discussing (at excruciating length) the design
and construction of a modest addition on my home, starting
with fairly conventional big-box-available industrial
solutions but evolving toward using locally sourced, somewhat
more natural materials (soilcrete, rough-sawn timbers from
nearby, scoria/perlite for in-ground insulation, mycelium
(grown in loose cellulose, oat-straw or hemp-fibers) roof and
wall insulation, etc. Most of my DIY friends are capable of
engaging in this but their idiosyncratic (as opposed to my
own) preferences (fetishes and fears) tend to taint the dialog
a little. GPT *does* try to channel me back to the
conventional, offering reasons why I really *should* consider
using the most conventional materials/methods. Nevertheless if
I speak in reasonable and coaxing tones it will usually
acknowledge that their are contexts wherein my ideas might be
viable (though there always remains a skeptical bias) and in
fact helps me split hairs on just
> what might be the contexts where my ideas *are* viable...
>
>>
>> Another concept I've defended on this list is the vocal
grooming hypothesis. If a lonely person engages a chatbot as a
simple analogy to picking lice from others' fur, then their
engagement with the bot probably lands squarely in (1) and
(2). But if the person is simply an introverted hermit who has
trouble co-constructing the world with others (i.e. *not*
merely vocal grooming), then the chatbot does real work,
allowing the antisocial misfit to do real work that could
later be expressed in a form harvestable by others. I wonder
what humanity could have harvested if Kaczynski or
Grothendieck in his later years had had access to
appropriately tuned chatbots.
> I'd like to think the chatbots I hang out with might have
helped them talk themselves *out* of their most acute
anti-social activities... but maybe not.
>
--
glen
-. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. .
FRIAM Applied Complexity Group listserv
Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p
Zoom https://bit.ly/virtualfriam
to (un)subscribe
http://redfish.com/mailman/listinfo/friam_redfish.com
FRIAM-COMIC http://friam-comic.blogspot.com/
archives: 5/2017 thru present
https://redfish.com/pipermail/friam_redfish.com/
1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
-. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. .
FRIAM Applied Complexity Group listserv
Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom
https://bit.ly/virtualfriam
to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
FRIAM-COMIC http://friam-comic.blogspot.com/
archives: 5/2017 thru present
https://redfish.com/pipermail/friam_redfish.com/
1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
-. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. .
FRIAM Applied Complexity Group listserv
Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p
Zoomhttps://bit.ly/virtualfriam
to (un)subscribehttp://redfish.com/mailman/listinfo/friam_redfish.com
FRIAM-COMIChttp://friam-comic.blogspot.com/
archives: 5/2017 thru presenthttps://redfish.com/pipermail/friam_redfish.com/
1/2003 thru 6/2021http://friam.383.s1.nabble.com/
