I tripped over (in my Gnewsfeed) an article that seemed to speak more
clearly
<https://www.marktechpost.com/2023/04/06/8-potentially-surprising-things-to-know-about-large-language-models-llms/>
to some of my maunderings:
8 Potentially Surprising Things To Know About Large Language
Models LLMs
<https://www.marktechpost.com/2023/04/06/8-potentially-surprising-things-to-know-about-large-language-models-llms/>
And the paper it summarizes (with a similar title, more detail and
references):
8 Things to know about Large Language Models - Samuel R Bowman
<https://arxiv.org/pdf/2304.00612.pdf>
And in particular this point made:
3. *LLMs frequently acquire and employ external-world representations.*
More and more evidence suggests that LLMs build internal
representations of the world, allowing them to reason at an abstract
level insensitive to the specific language form of the text. The
evidence for this phenomenon is strongest in the largest and most
recent models, so it should be anticipated that it will grow more
robust when systems are scaled up more. Nevertheless, current LLMs
need to do this more effectively and effectively.
The following findings, based on a wide variety of experimental
techniques and theoretical models, support this assertion.
* The internal color representations of models are highly
consistent with empirical findings on how humans perceive color.
* Models can conclude the author’s knowledge and beliefs to
foretell the document’s future course.
* Stories are used to inform models, which then change their
internal representations of the features and locations of the
objects represented in the stories.
* Sometimes, models can provide information on how to depict
strange things on paper.
* Many commonsense reasoning tests are passed by models, even ones
like the Winograd Schema Challenge, that are made to have no
textual hints to the answer.
These findings counter the conventional wisdom that LLMs are merely
statistical next-word predictors and can’t generalize their learning
or reasoning beyond text.
On 4/6/23 8:27 AM, Steve Smith wrote:
I have been reading Jeff Hawkins' _1000 Brains_ which is roughly *his*
take on AI from the perspective of the Neuroscience *he* has been
doing for a few decades, including building models of the neocortex.
What struck me strongly was how much *I* expect anything I'd want to
call artificial *consciousness* to engage in "co-munnication" in the
strongest sense. Glen regularly admonishes us that "communication"
may be an illusion and something we don't actually *do* or maybe more
to the the point "it doesn't mean what we think it means"?
So for all the parlor tricks I've enjoyed playing with chatGPT and
DALL-E and maybe even more spectacularly the myriad examples *others*
have teased out of those systems, I am always looking for what sort of
"internal state" these systems are exposing to me in their
"utterances". And by extension, I am looking to see if it is in any
way apprehending *me* through my questions and prompts.
Dialog with chatGPT feels pretty familiar to me, as if I'm conversing
with an unusually polite and cooperative polymath. It is freeing to
feel I can ask "it" any question which I can formulate and can expect
back a pretty *straight* answer if not always one I was hoping for.
"It" seems pretty insightful and usually picks up on the nuances of my
questions. As often as not, I need to follow up with refined
questions which channel the answers away from the "mundane or obvious"
but when I do, it rarely misses a trick or is evasive or harps on
something from it's own (apparent) agenda. It only does that when I
ask it questions about it's own nature, formulation, domain and then
it just seems blunted as if it has a lawyer or politician intercepting
some of those questions and answering them for it.
I have learned to "frame" my questions by first asking it to defer
it's response until I've given it some ... "framing" for the actual
question. Otherwise I go through the other series of steps where I
have to re-ask the same question with more and more context or ask a
very long and convoluted question. At first it was a pleasure to be
able to unlimber my convoluted-question-generator and have it (not
mis) understand me and even not seem to "miss a trick". As I learned
to generate several framing statements before asking my question, I
have found that I *can* give it too many constraints (apparently) such
that it respects some/most of my framing but then avoids or ignores
other parts. At that point I have to ask follow-up, elaborating,
contextualizing questions.
I do not yet feel like I am actually seeing into chatGPT's soul or in
any way being seen by it. That will be for a future generation I
suspect. Otherwise it is one hella "research assistant" and
"spitball partner" on most any topic I've considered that isn't too
contemporary (training set ended 2021?).
- Steve
On 4/4/23 5:54 PM, Prof David West wrote:
Based on the flood of stories about ChatAI, it appears:
- they can 'do' math and 'reason' scientificdally
- they can generate essays, term papers, etc.
- they can engage in convincing dialog/conversations
- as "therapists"
- as "girlfriends" (I haven't seen any stories about women
falling in love with their AI)
- as kinksters
- they can write code
The writing code ability immediately made me wonder if, given a
database of music instead of text, they could write music?
The dialog /conversation ability makes me wonder about more real-time
collaborative interaction, improv acting / comedy? Or, pair
programming? The real-time aspect is critical to my question, as I
believe there is something qualitatively different between two people
doing improv or pair programming than simply engaging in dialog. I
think I could make a much stronger argument in the case of improv
music, especially jazz, but AIs aren't doing that yet.
davew
-. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. .
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 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/
