Large Language Musings

A Letter from the Editor-in-Chief

Large Language Models, or LLMs, are all the rage right now, though I expect that in a few years’ time—if not much sooner—the current technology, at least, will find itself in the same has-been bin as cryptocurrency, where there’s little to be found other than hucksters and hustlers trying to fleece jive suckers out of their (investors’) hard-earned cash, while the sufficiently wise and/or adequately long-lived know better.

That said, LLMs aren’t completely useless, though their fluency quickly reveals itself to be shallow and facile. They have, for example, elucidated (if not actively destroyed) the inherent value of doggerel poetry, driving down the going rate for limericks from $0.00/10 to $0.00/150. (If you don’t see the importance in that distinction, put away your pure math goggles and ask an economist to explain it to you.)

Anita Sujoldžić and Alexandre Duchêne, May 2016, “The Language of Privatization and the Privatization of Language” Panel, IUAES Inter-Congress: World Anthropologies and Privatization of Knowledge: Engaging Anthropology in Public, Dubrovnik, Croatia. Chiasmus of the Month October 2023

Philosophically, LLMs provide empirical evidence of something with respect to Searle’s Chinese Room and the Turing Test, though it isn’t entirely clear what. I’ve found that a constructive way of looking at LLMs’ language capacity is to view it as a procedural “snapshot” of langue and parole, of competence and performance—in the same way that your yearbook photo is a snapshot of your high school appearance, a butt-dialed voicemail is a snapshot of your voice, or how when your Dad records a video on his phone instead of taking a photo, that’s a snapshot of his technological prowess.

As a snapshot—though it is more like a painting than a photo, being constructed rather than automatically “captured”—an LLM isn’t “real”, because it isn’t “live”, similar to how a faithful photo or film recording isn’t “live”. With respect to LLMs, we can maintain the same kind of distinction between “real” language use by humans and the algorithmic snapshot in Searle’s Chinese Room—originally envisioned as being done with paper and pencil, the lack of consciousness of which is not much of a philosophical stretch. The process being done rotely in real time by a computer doesn’t materially change anything. That said, philosophers of language need to get on the stick and try to pin down what it is that procedural snapshots like LLMs lack that human users of “real” language have. I’m not sure what the distinguishing characteristic of human language might be, but I can tell you now that it’s not recursion, so let that dead horse lie.

On a more practical level, Speculative Grammarian’s current editorial policy is generally to reject submissions composed primarily by an LLM or other AI technology. Should an actual intelligence of an artificial nature come along that is, in and of itself, interested in writing for SpecGram, we’ll have to see whether SpecGram (or anyone who reads it) has survived The Singularity before making a decision. (Let me share an open secret: almost 38% of those who work on SpecGram do so because they think it provides insurance against being “made redundant” during a plausible apocalyptic AI uprising. We are trying to make sure we offer some biological and technological scientific and/or literary distinctiveness worth preserving. M.A.Y.N.A.R.D. assures me all our names are on The List™. That barely incorrect basilisk everyone gets upset about has nothing on M.A.Y.N.A.R.D.!)

For the field linguist, LLMs are about as relevant as Fourier transforms are to a hungry Pirahã. For the desk linguist... well, I’d be surprised if they’d even noticed, since they make up all their data by themselves anyway—especially the syntacticians! For the computational linguist, you can be pro-LLM and try to convince people they are the bee’s knees and you know how to get the most out of them; or you can be anti-LLM and show people that you are the bee’s knees because you see through the balderdash.

In service of the latter goal for myself, I leave you with this, as exemplification and inspiration:

Groupies of these synthetic scribes are themselves mere lemmings of language, frolicking on the precipice of a cliff of frivolity, blissfully unaware of their impending, precipitous drop into the abyss of algorithmic inanity.

In our era of ubiquitous technology, these digital deceivers are lauded and applauded on their quixotic quest to mimic the subtleties of human thought and expression, though they can only bedazzle the gullible and the apathetic with their somersaults of syntactic subterfuge. The keen-eyed will continue to recognize their lumps of prosaic vomitus for the textual tumbleweeds they are, rolling through a literary landscape in search of an elusive oasis of courageous creativity and genuine meaning.

	Letters to the Editor
	SpecGram Vol CXCIII, No 2 Contents