I've recently developed a taste for judging people by the content of their character, something I used to and
kindasorta still do denigrate as hubris (because our models of others' character are models, always wrong,
rarely useful). And one of the best measures of character is how someone responds when presented with a
"learning opportunity". ChatGPT is an extraordinary mansplainer. And even though, when you show it
facts that contradict it's prior opinion, it gives lip-service with words like "sorry", it will
continue to *confidently* spout half-truths and rhetorical bullsh¡t (even if you ask it to, say, write a LCG
pRNG in C). Just like the tendency of apps like Stable Diffusion to "pornify" women, ChatGPT
encodes the culture of its input. So if ChatGPT is a mansplainer or evil, in any sense, it's *because* the
culture from which it draws its input is that. I.e. It's a mirror. We are evil. We are dreadworthy.
BTW, I did a back of the envelope calculation and the cost of operating one
(small) 777 per day seems to be about the same as operating the one ChatGPT per
day (~$100k). Presumably, if/when OAI begins distributing the model, such that
there are several of them out there (like the fleets of 777s), the costs will
be lower. At that point, the semantic content of one 777 might exceed that of
one ChatGPT instance.
On 2/8/23 09:08, Santafe wrote:
It’s funny. I was reading some commentary on this last week (can’t even
remember where now; that was _last week_!), and I remember thinking that the
description reminded me of Williams Syndrome in people. They have a
grammatical sense that is at the stronger end of the human range, but their
train of meaning has come to be characterized (again, a now-tropish short-hand)
as “word salad”.
That there should be several somewhat-autonomous processes running in parallel
in people, and coupled by some kind of message-passing, as Ray Jackendoff
proposes, seems quite reasonable and in keeping with brain biology, and if
there is, it would be a compact way to account for the seeming independence in
refinement of grammatical sense and whatever other part of sentence-coherence
we have come to term “semantics”.
Last year, too, someone (I think my boss at the time, which would make it two
years ago) told me about some nature paper saying that a comparative genome
analysis of domestic dogs and wolves had shown a mutation in the dogs at the
cognate locus to the one that results in Williams Syndrome in people. That
would be an easy indulgent interpretation: the greater affectionateness
preserved into adulthood, and the increased verbal-or-other communicativeness.
Though Barry Lopez, I think it was, argues that wolves have higher social
intelligence, which I guess would be making some claim about a “semantics”.
The chatbot has, however, a knd of pure authentic evil that Philip K. Dick
tried to mimic (the argument with the door), and came close enough to be
laughing-through-tears, but could not truly simulate as it shines through in
the Ginsparg exchange. Or dealing with the maddening, horrifying computer
interfaces that every company puts up to its customers, after they have fired
all the human problem-solvers. Few things put me in a real dread, because I am
now fairly old, and getting older as fast as I can. But the prospect of still
being alive in a world where that interface is all that is left to any of us,
is dreadworthy.
Eric
On Feb 8, 2023, at 11:51 AM, glen <geprope...@gmail.com> wrote:
I wrote and deleted a much longer response. But all I really want to say is that these
*models* are heavily engineered. TANSTAAFL. They are as engineered, to intentional
purpose, as a Boeing 777. We have this tendency to think that because these boxes are
opaque (more so to some than others), they're magical or "semantic-less". They
simulate a human language user pretty well. So even if there's little structural analogy,
there's good behavioral analogy. Rather than posit that these models don't have
semantics, I'd posit *we* don't have semantics.
The problem with communication is the illusion that it exists.
On 2/7/23 14:16, Steve Smith wrote:
DaveW -
I really don't know much of/if anything really about these modern AIs, beyond what pops
up on the myriad popular science/tech feeds that are part of *my* training set/source.
I studied some AI in the 70s/80s and then "Learning Classifier Systems" and
(other) Machine Learning techniques in the late 90s, and then worked with folks who did
Neural Nets during the early 00s, including trying to help them find patterns *in* the NN
structures to correlate with the function of their NNs and training sets, etc.
The one thing I would say about what I hear you saying here is that I don't think these modern
learning models, by definition, have neither syntax *nor* semantics built into them.. they are
what I colloquially (because I'm sure there is a very precise term of art by the same name) think
of or call "model-less" models. At most I think the only models of language they have
explicit in them might be the Alphabet and conventions about white-space and perhaps punctuation?
And very likely they span *many* languages, not just English or maybe even "Indo
European".
I wonder what others know about these things or if there are known good
references?
Perhaps we should just feed thesemaunderings into ChatGPT and it will sort us
out forthwith?!
- SteveS
On 2/7/23 2:57 PM, Prof David West wrote:
I am curious, but not enough to do some hard research to confirm or deny, but
...
Surface appearances suggest, to me, that the large language model AIs seem to
focus on syntax and statistical word usage derived from those large datasets.
I do not see any evidence in same of semantics (probably because I am but a "bear of
little brain.")
In contrast, the Cyc project (Douglas Lenat, 1984 - and still out there as an
expensive AI) was all about semantics. The last time I was, briefly, at MCC,
they were just switching from teaching Cyc how to read newspapers and engage in
meaningful conversation about the news of the day, to teaching it how to read
the National Enquirer, etc. and differentiate between syntactically and
literally 'true' news and the false semantics behind same.
davew
On Tue, Feb 7, 2023, at 11:35 AM, Jochen Fromm wrote:
I was just wondering if our prefrontal cortex areas in the brain contain a
large language model too - but each of them trained on slightly different
datasets. Similar enough to understand each other, but different enough so that
everyone has a unique experience and point of view o_O
-J.
-------- Original message --------
From: Marcus Daniels <mar...@snoutfarm.com>
Date: 2/6/23 9:39 PM (GMT+01:00)
To: The Friday Morning Applied Complexity Coffee Group <friam@redfish.com>
Subject: Re: [FRIAM] Datasets as Experience
It depends if it is given boundaries between the datasets. Is it learning one
distribution or two?
*From:* Friam <friam-boun...@redfish.com> *On Behalf Of *Jochen Fromm
*Sent:* Sunday, February 5, 2023 4:38 AM
*To:* The Friday Morning Applied Complexity Coffee Group <friam@redfish.com>
*Subject:* [FRIAM] Datasets as Experience
Would a CV of a large language model contain all the datasets it has seen? As
adaptive agents of our selfish genes we are all trained on slightly different
datasets. A Spanish speaker is a person trained on a Spanish dataset. An
Italian speaker is a trained on an Italian dataset, etc. Speakers of different
languages are trained on different datasets, therefore the same sentence is
easy for a native speaker but impossible to understand for those who do not
know the language.
Do all large language models need to be trained on the same datasets? Or could many large
language models be combined to a society of mind as Marvin Minsky describes it in his
book "The society of mind"? Now that they are able to understand language it
seems to be possible that one large language model replies to the questions from another.
And we would even be able to understand the conversations.
--
ꙮ Mɥǝu ǝlǝdɥɐuʇs ɟᴉƃɥʇ' ʇɥǝ ƃɹɐss snɟɟǝɹs˙ ꙮ
-. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. .
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/
