Today we LARP as an inferentialist.
I haven’t actually read inferentialism. I’ve read a bit of Rorty, but that’s about it. As for analytic philosophy, I’m only really intimate with Carnap and maybe Quine (insofar as he responds to Carnap) - forget Kripke, Lewis, etc. (which I should probably do if I want to engage with AI people…). My main background has been philosophy of science.
But from what I understand, the whole representationalism issue (the accusation of representationalism, from Rorty’s POV) is responding to a tendency to first secure this neutral base from which we must first neutrally agree, before proceeding with discourse.
Funnily enough, my current job talks about something similar - though I would not try and read too much into the analogy. But if you frame this kind of talk into something like, “Go read XKCD 927”, then all of a sudden it sounds actionable and reasonable, and not like some “weird relativists in the humanity departments again” talk.
But anyways. You should read this as, somebody who has
literally read nothing about inferentialism (I do have a copy of
Making it Explicit though…), pattern matching onto
the word “inferential” and trying to do something with it. I am
quite literally LARPing.
The meaning of words
If words get their meaning from inferential relations, I get the vision that, they’re immediately tightly coupled to the actor (the person saying the words), and specifically, what the actor’s model of the world is.
But even that’s not correct, because you can just say, “Well, duh, the actor’s trying to articulate the best model of the world”.
Note: Actually, this whole question reminds me a bit of Solomonoff Induction and next-token-prediction debates about how the model “must form a model of the world”. The point that people emphasize is that, sure, it’s “just predicting the next token”, but the kinds of structures that could predict the next token properly, must be really sophisticated.
I just think to architectures such as Carnapian logic or Bayesianism and I see representationalism as sort of pushing down complexity to some sort of descriptive layer at the end - and the job of the algorithm is to merely select the best description out of possible outcomes. So now, I’ve dropped the Rortyian railing here, I’m talking purely from an “interestingness of algorithm” perspective.
Whereas a maximal inferential agent would say a phrase, with the intent that someone could grasp the downstream implications, the rebuttals, the rebuttals to said rebuttals, etc. This reminds me of the image in my head when I hear Solomonoff induction - I just have this image of my mind of a Turing machine simulating an arbitrarily omniscient model of the universe, and saying, “We did it!” (I mean, sort of the point is to find the “smallest” one, but this whole thought exercise has made me a lot more suspicious whenever mathematicians try to make generalized statements about thought)
So in the maximal cases, both architectures are pretty useless. Yet they still, I think, elucidate the struggles with each position:
For representationalism: Where do you draw the line here? Why is it just weird forms of classification? Does your representation start smuggling in arbitrarily powerful Turing Machine semantics? (Any TCS person would kill me for this, I am just using it here to call back to the image procured by Solomonoff induction - but I think it still gives a better image than “arbitrarily powerful God” in terms of evoking something procedural and algorithmic and doing something and not just a lookup table)
For inferentialism: The classic one is, “how do you know that tree corresponds to a tree, then?” <insert argument about biology, memory, association, sense data>. In the agent setting, it’s something like, “Different agents clearly converge, that shared convergence must have some representational structure”.
Meh. Maybe the idea is hogwash.
I think for me though, it opened me up different avenues of thinking. “Just say your words and let the truth follow” sounds noble, but is probably one of the worst possible pieces of conversation advice possible. I don’t think I ever followed that literally, but when I think back to my behavior holistically, I tended to steer in that direction.
And the antidote to this is not like, “okay, well we can approximate truth, we’ll get there, it’s just hard.”. That’s like saying, “Okay, representationalism is hard, but we’re still going to stay in that paradigm”. Nah, throw the whole thing out.
I do think inferentialism is different from just pragmatism, or “consider consequences of speech, other than raw truth”. (To be extra clear, I don’t think representationalism has access to this thing called ‘truth’, but I’m just using the rhetoric). I mean, after all, they’re different terms. But I don’t know.
Addendum 8-8-2026
OK, the most concise way to say this, is that I feel like representationalism, once you get to the “base”, assumes there is a “mere” mapping function involved. Now, mapping functions can get hard. They can be nearest neighbor, some kind of high-dimensional classifier, etc.
Inferentialism claims that the “mapping” function grows arbitrarily hard (well, it would reject the framing of a “mapping” function, but if we had to phrase it in purely representational terms, this is my model of it). Something can get its meaning not because it got looked up in some abstract high dimensional lookup table, but because some turing machine computed an arbitraily hard model of what should happen and all downstream consequences from it.
This is also incidentally why I have an issue with just mere “mental model” framings, because when people say the word “mental model”, they usually assume a representational mode of thought - the base, first actionable thing is taxonomy.
And again, it’s hard to argue out of this, because you hear this, and you think, “Well, duh! I have all these representations in my head that form a world model”. That’s… not really what is going on, I think.
And this is why I really think, despite how “vacuous” and stupid Solomonoff induction can be, it actually presents something actionable, because when I say “arbitrary Turing Machine” (not in the formal sense - again, sorry, complexity theorists) - the intent is to break out of representation and say, hey, no, there can be way more complicated modes of thought.
It takes time to sink in, because any argument from arbitrary compleixty can be reformulated in representational terms as “Oh, there’s just a very high latent space” or something. I don’t know.