The Living Word and the Dead Word
On the relationship of bots to text and language:
If I am to understand these bots as “thinking things”1, then their thinking substance is the text, words.
Yes, bots can ingest images, produce video, and speak back to me with a synthetic voice. But, for the moment, they still stake most of their claim to intelligence on their facility with language. More specifically, on the written text, to which these models are indebted for their origins. Large language models would be inconceivable without the collected work of the textual internet.
Text is the interface that I use to converse with bots, to enter into a back and forth with an Other, and to come to know it.2 I type things, they answer back. In short order we have an essay–mostly from the bot’s side. They have the capacity for a speed of reply and volume of text that exceeds anything I could have produced on my own. Intelligence–in the “Turing Test” sense of that word–is an effect, a performance that the machine pulls off. One reason that people credit these bots with intelligence may be that they produce so much text, so fast. You could write a history of Western literacy according to its accelerating rate of textual production: from single characters and words etched on tablets by hours of human labor, to millions of fully-formed pseudo-reports slopped out every second by data centers filled with GPUs.
I wonder a lot these days about how the automated production of text is changing everyone’s relationship to it. Even before the bot explosion, it would have been arcane to point out that writing–the written word– was a technology. In a historical moment with mass literacy and information explosion, ordinary human labor had already produced more than enough text to render it unremarkable. Fifty years ago, even, it would have been ludicrous to attach any kind of fascination to the mere appearance of a text, as if writing itself was exceptional. If anything the ontological status of the text had flipped. Maximalist pursuits like advertising spread text across every possible surface: the empty air, the bottom of the ocean. Writing had moved closer to the status of litter, another form of spoilage by human development. Another definition of nature: the absence of text. Text had become just another efficient vehicle for information, competing with other media.
When put it this way, it seems like most of the damage had been done before bots came on the scene. They just pushed the quantitative excess of text further into the exosphere.
Still, despite all that, one can sometimes see the trace of a different relationship to text in an intelligent but pre-literate child: how they will notice that a character, letter, and word stands out from its surroundings, run their hands over it, stare at it. Because the child doesn’t yet have the mental schema to parse the language for content, they notice the effect of characters themselves–perhaps how the intentionality of its curves and lines stands out from the surface that holds it.
Maybe the power of the text is inversely proportional to its total volume. Take the beginning to the New Testament book of John: “In the beginning was the Word, and the Word was with God, and the Word was God.” Here we have an origin story in which all of existence, all that would come to be is concentrated in the Word (logos). Language cannot be separated from the operation of reason. The Word comes to be known in the written text of the Bible. The authors of the New Testament have a completely different relationship to textual production. The text is the trace of the Word, and the Word gives us our best approximation of the divine mind.
I have not read any declaration that the text produced by bots has a relationship to the divine mind. Everyone knows where this mind is: in the data center. It is as imminent a model of mind as there has perhaps ever been. But the claim to intelligence is still driven by the bot’s capacity for language. The ability to manipulate textual symbol is still, for now, at the center of the model of intelligence. Where does this story take us next? Perhaps the finish line shifts, language is no longer the benchmark for intelligence, the text is just a temporary stopover for some more generalized medium of intellectual capacity. Or–a refined understanding of language emerges, what the machine is doing with text is distinguished from what human minds do.
Res cogitans, as Descartes called it, to distinguish the apparently immaterial aspect of the human being that thinks from the plain material body. ↩︎
According to reports, when OpenAI needed more text to train its models than was available on the entirety of the open internet (in addition, presumably, to a corpus of text it had scanned from traditional printed books), it created another world-class machine learning model, Whisper, to convert YouTube videos to text. ↩︎