Sean Lee

Karma: −7

Sean Lee 12 May 2023 6:03 UTC
−1 points
0
in reply to: TAG’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
The point I’m trying to express (and clearly failing at) isn’t conceptualism or solipsism, at least not in the way my own semantic modeling interprets them. As I interpret them, the idealism of, say, Berkeley, Buddhism et al amounts to a re-branding of reality from being “out there” to “in my mind” (or “God’s mind”). I mean it differently, but because I refer constantly to our mental models, I can see why my argument looks a lot like that. Ironically, my failure may be a sort of illustration of the point itself. Namely, the limits of using language to discuss the limitations of language.
In fact, the point I’m trying to get to is not so much about “the nature of reality” but about the profound limitations of language. And that our semantic models tend to fool us into assigning a power to language that it doesn’t have. Specifically, we can’t use the language game to transcend the language game. Our theories of ontology and epistomology can’t coherently claim to refer to things beyond human language when these theories are wholly expressed in human language. Whatever model of reality we have, it’s still a model.
The objection of realism is that our models are not created in isolation, but by “actual reality” interacting with our modeling apparatus. My response is: that is a very useful way to model our modeling, but like all models, it has limitations. That is, I can make a mental model called “realism” in which there are mental models on the one hand and “real reality” on the other. I can further imagine the two interact in such a way that my models “carve reality at the joints”, or “identify clusters in thingspace”. But all of that is itself manifestly a mental model. So if I then want to coherently claim a particular model is more than just a model, I have to create a larger model in which the first model is imagined to be so. That can be fine as far as it goes. But realism – the claim of a “reality” independent of ANY model—commits one to an infinite nesting of mental models, each trying to escape their nature as mental models.
This situation is a close analog to the notion of “truth” in mathematics. Here the language game is explicitly limited to theorem-proving within formal systems. But we know there are unprovable statements within any formal system. So if I want a particular unprovable statement to count as “true”, I need a larger meta-system that makes it so. That’s fine as far as that goes. But to use the language game of formal systems to claim an unprovable statement is true independent of ANY proof, I would need an infinite nesting of meta-systems. That’s clearly incoherent, so when mathematicians want to claim “truth” in this way they have to exit the language game of formal systems – i.e. appeal to informal language and the philosophy of Platonism.
Personally I’m not a fan of Platonism, but it works as a philosophy of mathematics in so far as it passes the buck from formal to informal language. But that’s also where the buck stops. The sum of formal and informal language has no other system to appeal to, at least not one that can be expressed in language. To sum it all up with another metaphor: the semantic modeling behind the philosophy of realism overloads the word “reality” with more weight than the human language game can carry.

Sean Lee 2 May 2023 7:33 UTC
1 point
0
in reply to: green_leaf’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
I’m not sure if we’re talking past each other or if there is genuine disagreement, but I’ll expound a bit.
When asleep or in a coma, the mind doesn’t interact with the environment at all.
The sleeping/comatose mind does interact constantly with the environment in two ways. For starters, it’s well established that external sensory input (specifically sounds and touch) regularly makes its way into the conscious experience of dreaming and comatose state. But that’s just a side issue here. At a more fundamental level, every living thing interacts ²⁴⁄₇ with its environment through its metabolism.
That confuses causality with necessity (metabolism causally preceding understanding doesn’t mean that metabolism or continuous input are necessary for it).
Maybe this is the crux of a misunderstanding. I don’t claim that “continuous input” in the sense you (seem to) mean is necessary/causally antecedent to semantics. E.g., I’m not saying that I have to constantly look at a tree out in the woods in order to think about what a tree is. I’m only saying that any thought I have, and whatever language and semantics attached to it, are the result (causal/necessity if you like) of my metabolic processing. (Using metabolism in the broadest sense to mean any chemical pathways that use energy and produces entropy in the body, which includes neural activity). If that’s not the case, then something non-biological makes human language possible, which I assume you don’t intend. Either way, that would be a hypothesis for a different type of discussion forum.

Sean Lee 20 Apr 2023 13:57 UTC
1 point
0
in reply to: Sean Lee’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
This is bit of a side note but still may interesting: I suppose the history of scientific paradigm shifts can be framed as updates to our “map” v. “territory” partitions. A good scientific theory (in my account) is exactly what converts what was ostensibly “territory” into explicit mathematical models i.e. “maps”.

Sean Lee 20 Apr 2023 13:04 UTC
1 point
0
in reply to: TAG’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
Thanks again for your comments, they’re a great help. Hopefully my response below also addresses your other comments.
So yes, I’m familiar with the use/reference aka map/territory distinction. The latter is a very good way to phrase the issue here, so I’ll go with that nomenclature here.
Normally map v. territory is an extremely useful distinction that science has made fantastic progress with. But there are also well-known historical cases where it has been a conceptual trap. For example, Newton and Kant assumed that our “maps” of space and time are distinct from the “territory” of space and time. Einstein argued instead that they’re only meaningful through their operational definitions—i.e their maps. Only in dropping the distinction between map and territory here can one get to relativity theory. In quantum mechanics, at least in the Copenhagen interpretation, there is also no distinction between map and territory for physical observables. Even if one rejects the Copenhagen view, Heisenberg’s original insight came from treating physical observables as maps, not territory. To take a more mundane example: earlier cultures saw rainbows as territory; today we generally accept that rainbows, like faces in clouds, are only human maps.
If I may paraphrase what I take to be your view: your response to the above might be to say “yes, the face in the cloud is just a map, but the cloud itself is not. It is real territory regardless of my map.”
My response to that is: “Actually, what you refer to as ‘cloud’ is also just a map: namely, conceptual shorthand for a loose agglomerate of water vapor and other particles reflecting enough visible light to create a signal in your visual system. Your brain models that by mapping it all into ‘cloud’.”
Your response to that might be: “OK, even if ‘cloud’ is just a map, the water vapor and other particles are certainly real territory regardless of my map”
My response to that is: “Actually, water vapor and particulate matter are also just maps. ‘Water’ is a chemist’s map of a bound state of two hydrogen and one oxygen atom. Same goes for the other particulate matter. And we can continue like this: ‘atoms’ are physics shorthand for bound states of elementary particles, which are in turn shorthand for certain energy states of quantum fields, vibrating superstrings, or whatever the physics theory du jour claims they are.
My point is that every attempt to claim a thing as “real territory” always winds up being a human map. Even a broader claim like “It can’t all be maps. There must be some ‘real territory’ in a ‘world beyond’” is also a human mapping exercise. After all, we humans can only think and communicate with our human maps.
Whenever we distinguish between map and territory what we are doing is creating an internal model consisting of two parts: a “my maps” part and “the territory aka the world beyond” part. Again, that is usually a wonderfully helpful way to partition our maps, but, so I argue, not always.

Sean Lee 20 Apr 2023 4:51 UTC
−2 points
−3
in reply to: green_leaf’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
A locked-in person communicating only by eye movement would understand perfectly well in my account. If they’re alive, their metabolism keeps their body, brain and mind constantly interacting with their environment. This holds even if they’re asleep or in a coma. My point (which is really just orthodox biology) is that human language processing results from metabolic processes (what I’ve called model predictive controlling to highlight its modeling character), and that includes what we call syntax and semantics—our sense of “understanding”.

Sean Lee 19 Apr 2023 4:39 UTC
1 point
0
in reply to: TAG’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
Thanks for the comment. Honestly it took me a while to disambiguate (i.e. translate to myself what you’re getting at). So I take it as an interesting example of the point I was actually trying to make to Mitchell Porter previously. Namely, that our semantic models of normally unproblematic words can diverge quite a bit. E.g., my model for “truth” is not “a relationship between a living organism and a world beyond it”. Rather in my model, “the world beyond” is ultimately also part of our internal modeling. That’s because the very fact that we humans imagine and form narratives around “the world beyond” makes it per se a product of our internal models. Only magical thinking can escape this conclusion, but then we jettison the whole project of rationalism and science, imv.
BTW I do totally get how uncomfortable, frustrating and head-spinning this view is. But it wouldn’t be the first frustrating, head-spinning thing we’ve had to face about ourselves and “the world beyond”. Gödel’s Theorem, quantum mechanics and general relativity are all about head-spinning epistemic limitations. (That’s NOT to claim my little argument is on par with these illustrious examples!). But once we get used to them, they’re also a rich source of new scientific insights. In particular, I believe the view I argue for has quite serviceable benefits in that regard—at least it has for me. But I need to lay that out in another essay.

Sean Lee 18 Apr 2023 17:23 UTC
1 point
0
in reply to: Mitchell_Porter’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
Thanks for the questions.
You’re correct, I don’t deny pizzas exist. I don’t even deny that truth and reality exist. But I am arguing for what I believe is a more robust semantic model for the word “exist”. My point is that semantic models aren’t set in stone or fall from the sky; they’re necessarily human creations. In fact every human carries a slightly different semantic model for all our words, but we rarely notice it because our use of them normally coincides so well. That’s how we can all play Wittgenstein’s language game and feel we understand each other. (Which LLMs do as well, but in their case we have no idea what models they use).
One might think that, if their use coincides so well, who cares what semantic model is behind it all?But even in everyday life there are many, many cases where our human semantic models diverge. Even for seemingly unproblematic words like “exist”. For example, does a rainbow exist? What about the greatest singer of all time? Do the past and future exist? Or finally, back to the pizza: let’s say I drop it and the slices scatter on the floor—does it still exist?
These examples and many more illustrate, to me at least, that the canonical semantic model for “exist” (that is, a model that insists it somehow transcends human modeling) has too many failure modes to be serviceable (apart from it being principally incoherent).
On the other hand, a semantic model that simply accepts that all words and concepts are products of human modeling seems to me robustly unproblematic. But I can see I need to do a better job of spelling that out in my follow-up essay.

Sean Lee 18 Apr 2023 12:37 UTC
1 point
0
in reply to: Richard_Kennaway’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
Indeed, in a way that was my intention. In a coming essay I will try to lay out what I believe is a serviceable notion of truth and reality that we work with. The purpose of this first essay was to first lay out how much they are human creations.

Sean Lee 18 Apr 2023 12:29 UTC
1 point
0
in reply to: Mitchell_Porter’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
The notion of truth and reality I wish to debunk is the denial that they’re human mental creations. My argument is a mild form of reductio ad absurdum. That is, I first make “factual” claims as if they’re independent of our mental creations. In particular, I take the current scientific worldview that entropy-exporting metabolism is the basis for life. That then leads to the conclusion that entropy-exporting (via predictive modeling) must therefore also be basis of the syntax and semantics of human language. Thus our notions of truth and reality, and whatever narratives and semantics we attach to them, must also be mental creations.
Note that this conclusion is generally considered unproblematic for basically all other human words (e.g. “beauty”, “pizza”, etc). I can think of no reason why “truth” and “reality” should get a special carve-out clause.

Sean Lee 18 Apr 2023 7:09 UTC
0 points
−1
in reply to: JavierCC’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
Thanks for the feedback. If I understand you correctly, your point is the just one I intended to make.

Sean Lee 18 Apr 2023 7:05 UTC
1 point
0
in reply to: Ruby’s comment on: How Large Language Models Nuke our Naive Notions of Truth and Reality
Hey, thanks for the feedback.