glen wrote:
It prolly won't surprise you that I disagree (I think). Those intuitions that we develop may be a) interesting to like-minded people, b) valid to those who hold the same value/logic systems [⛧], and c) useful for sussing out us-vs-them [in|out]groups.
I appreciate (and depend on?) this disagreement
There is /Reality/ and there is /Intersubjective Reality/ (similar but different to KellyAnne's "alternate facts")... I don't think they track reality *any better* than the collective corpus of human expression they were trained on?But they don't necessarily track reality.
You might even say (ala the Interface Theory of Perception) those intuitions are inversely proportional to one's ability to track reality, the stronger they are, the less they track. This is adjacent to Eric's full tea cup.My only (intended) claim is that machine and human intuitions are similar in that they nominally interpolate and extapolate the high-dimensional space of all the linguistic/symbolic artifacts they have been exposed to. In deference to my partial agreement with DaveW, I recognize that written/linguistic "symbolic artifacts" are a tiny subset of the world even if we acknowledge the linguistic artifacts (writing) about the world.
E.g. someone like Denis Noble, whose had a fantastic career in science. But now that he's old and out of his lane, his confidence puts him out in front of his skis: https://en.wikipedia.org/wiki/Denis_Noble#The_Third_Way_of_EvolutionI'm definitely old and out of any lane I might have ever held or been in for more than a few wild slaps of my arms.
If we allow something like an intuition in LLMs, it should be clear that in order for them to track reality, they need "online" learning (as Marcus has proposed) and/or robotic embodiment to be able to interact with the reality we expect/want those intuitions to be about.Agreed, their "intuition" is currently limited to intuition about the things they have "heard" us talking about... they are the ultimate gossip-mongers, listening and passing on and making up new juicy bits from the old bits. I find most pre-enlightenment "science" to be the same: rumours and gossip about "how rigid bodies in motion act" without the empirical grounding of say "dropping a couple of items off the tower of pisa and seeing which one hits the jerk staring up at them first" or <pick your favorite story from alchemy, etc>.
But where you could argue with me might be on something like "muscle memory". Turns of phrases in a language should probabilistically constrain the response from the LLM. This might be similar to the way some words and phrases roll off the tongue. But in that sort of case, it's not *intuition* as we might normally think of it ... it's more like habit or practice. Again the emphasis is more on the doing than the thinking.Just a facility for the glib? A propensity for riffing without comprehension? Like me. here. now? (sorta)
[⛧] Indeed, the only way "valid" has any meaning at all is in the context of a language system ... if you fail to say what logic you're working with, the use of "valid" is invalid. 8^D ... sorry for the poetic license.
I'm not sure this impugnes language models (or other ML/AI systems) as much as it does the human capabilities they purport to emulate? My point isn't that LLMs do "really cool and meaningful things" but rather that what they do *might* not be that dissimilar from what *we* do in many contexts soe of which we want to claim is cool and/or meaningful).
<anecdote about hand-grinding lenses to make a point about embodied learning>
I was trained in my physics classes to believe that the actual hands-on-practice in the lab was important, that there were things I could learn (ways my intuition could be informed) by that embodied experience that no amount of reading and/or simple observation could achieve. I'm still not clear on whether grinding my own objective reflector and silver sputtering it in a vacuum chamber improved my understanding of the things in the sky I then observed with it, but it did do something for me in terms of an "embodied appreciation" for the statistics and geometry of spherical and parabolic cross-sections and the implications of different levels of vacuum, boiling of elemental silver, thin-films, etc. My lab-professor wasn't interested in my questions about the geometric implications (as I saw them) of the "figuring" techniques which best I could tell took the basic shape of a very large radius spherical section and added (central deepening and edge flattening) two more spherical sectionings such that the final mirror was a piecewise combination of 3 spherical sections with "maybe" a little feathering or blending between them. I also worried the question of whether the distance and *angle* from the boiling silver (strip?) and the lens) mattered for the thickness/shape of the surface layer. He (understandably) held that all that maundering and mentating was masturbatory and "all that mattered" was "that it works". I'm glad I did those things and even though "george" can talk me through and around and up one side and down the other about DIY reflector construction and be more patient with me about whether the off-axis silver-sputter layer is "good enough and nuances of various methods of "figuring" the psuedo-sphere the primary grinding yields? </anecdote>
OpenPGP_0xD5BAF94F88AFFA63.asc
Description: OpenPGP public key
OpenPGP_signature.asc
Description: OpenPGP digital signature
.- .-.. .-.. / ..-. --- --- - . .-. ... / .- .-. . / .-- .-. --- -. --. / ... --- -- . / .- .-. . / ..- ... . ..-. ..- .-.. 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/
