Over the last couple of days I have been trying to make explicit the dynamics of how I have been interacting with AI/LLMs to develop Pond. The analogy that comes closest to capturing what it feels like was inspired by a Floating Points live set that I caught at a music festival in Southwark Park, London. He had a large modular synthesiser with him, and a drum machine and sequencer as well probably, from which he coaxed out lush FM landscapes, steely but spherical (this is the best word I can find) polyphonic melodies, warm yet wiry feedback, and fearsome, distorted-singed at the edges, barely-contained-within-their-sonic-cages bass lines.
It wasn't the individual sounds that I was spellbound by however, nor the way in which they felt like they couldn't ever have been separate from one another, that nature herself had necessitated their co-occurrence and co-existence and co-terminus. Rather it was that at all times the human in this arrangement appeared to be taming a set of formidable beasts, and that although he was undoubtedly in total command, and the performance appeared against every sane intuition effortless, the machines were wells of a latent, intimidating, seemingly wild force, which would occasionally slip out of containment, and Floating Points would reach towards the edges of his immense capabilities to shepherd it back. I think it is a testament to his virtuosity (as in being a virtuoso, not being virtuous, although the connection between these words is interesting) that this latent machine force was visible at all, let alone that it was guided into forming the most profound arrangements in electronic music.
In these words I've described Floating Points as shepherding, or taming, or stewarding a force that every once in a while reveals a wild (?). Without being dramatic, and with every sense of embarrassment at suggesting a comparison to what Floating Points is doing, this is what working with LLMs feels like.
One suggestion is that what is at play is the phenomenon of emergent complexity; that the performance is not reducible to, nor describable (or perhaps better: representable) in the terminology of, nor inferable or predictable by recourse to its individual parts. This is true, of course, but if this is emergence then it's a trivial phenomenon, and it is instantiated by nearly every system involving an agent co-ordinating with its environment. I cannot reduce a shower to its individual parts of a water-supply, shower-head, and human; to do so would be to rid showers of their conceptual significance. I suppose I can predict that a shower will occur based on the presence of these objects and the behaviour of a person interacting with them, but I can only do so because I have already grasped the notion of a shower, which is the supposedly emergent phenomenon.
To end abruptly with a summary: the modular synthesiser model of LLM interaction imagines LLMs as a source of loosely structured, immensely potent, capricious, and yet malleable (tameable, perhaps) flow of raw material. In this mode, we act not directly on the material, but on (the properties and configuration of) its source, and by architecting patterns of interaction between the material and various modules that respond in various ways to it.