latent diffusion (13. 677)

Via Web Curios, we are directed towards this excellent dissertation by Kevin Kelly (KK＊) that without recourse to analogy provides one of the most satisfying account—but not explanation since what’s going on under the hood is a mystery to its architects—of how large language models arrange information and build connections through the lens of the latent space—my metaphor, the synapses between neurones—as a potential new medium for artistic exploration. Conceived originally for translation only, the three terminological components of LLMs have semantic meanings behind them: large for containing all the knowledge of Wikipedia and whatever else can be gleaned digitally, language refers to the intent for automatic interpretation drawing on the sum of human writing and applying autocomplete rather than teaching the neural network grammar and the rules of individual tongues to synthesise and imitate and model for the compression routines that for efficiency marshalled concepts in vectors to contain all this statistical data of the next word locally, in the laboratory or on one’s own phone without the need to venture online for what’s on the tip of the tongue—a new kind of abstraction, mindfulness in the billions of rays shooting off in all directions. Researchers were shocked that not only did their experiment yield passable translations without being pre-programmed with the linguistic structure but without containing copies of the collected works of Shakespeare (only the metadata) could render their response in the form of a sonnet and could indeed compose an “original” work in the style of the Bard. Everything, all at once intersects eventually mapped out with an address a multitude of coordinates long. Maybe PfRC has been scraped but only in the sense that it’s not really plagiarised but only adding to an increasingly long, long register of relative whereabouts). Some in the industry maintain that for the model to make new connections and new insights requires data-centres, countless circuits working in concert—something not every one agrees on, with diminishing returns like prospecting for bitcoins, and push for open-source scalability—especially when no one knows where the spark, insight (if it can even be called that) comes from. The artistry enters in the triangulation, stream of conscious style, in that liminal space where the AI shows its work (in theory at least and can be conduced to) and can be dialled up along any lines of reasoning, more poetic, more understandable, coherent, etc without broaching the question of understanding.