§1. Following a thought that surfaced this morning between periods of a deep sleep, I read an article on KV-caches in LLMs for the memory research group, and talked to Claude about whether we could conceive of KV-caches and LLM-contexts as representations for action. We can, to an extent, he responded: KV-caches contain transformed representations, potentials for action (where action is next-token generation). Taking the analogy further, we can see prompt engineering as niche construction, and chain-of-thought reasoning (in which interim outputs are generated) as cognitive scaffolding—producing representations which themselves afford certain actions.
§2. I then asked Claude to characterise our present human-LLM interaction in the language of distributed cognition. Here Claude suggested that my prompt to consider an idea that resulted in (Claude-produced) speculative explanations which themselves made certain kinds of thought (and prompting) available, or possible. One of these thoughts was this: what modes of cognition does the prompt-response model (or in Philip Agre's terminology, the oracle-seeker model) preclude? Claude went through a few examples, unified by this theme: escaping premature coherence, which the prompt-response model cannot help but tend towards. Prompt-response privileges certainty/finality and linearity: there is a forward motion to the cognitive process, and the most recent response tends to assume the entire stage; there is a rigid convergence of conceptual space. "Sitting with uncertainty" is precluded.
§3. Uncertainty is not wholly impossible: I, separate from Claude, can produce and maintain the cognitive potentials that sustain it. I can carry the ideas we developed from our conversation to this notebook and develop and question them; I can test them against other ideas I have had and recorded, and I can compare them with ideas from other texts and conversations. These acts are co-ordinated internally; whereas I could offload the explanation of KV-caches to Claude, I do not have a means of offloading the entire exploration-orchestration process. I wonder what an instrument for this purpose, and not for the individual acts of research and questioning that it consists of, could look like.