# Integration Queue > [!info] Legacy > New remembered items are created as individual tagged notes in `Notes/`. See [[Recent Captures.base]]. This note is preserved as historical archive. Unsorted passages and ideas. Route to projects during review. --- ### 2026-02-24 — Substack idea: meta-intelligence for managing AI agents > Quick idea for a Substack article Meta Intelligence about the intelligence required to manage The intelligences you are using via AI agensts like claude code. --- ### 2026-03-11 — LLM as substance: core framing against Hoel's "mere tools" > **The substance framing captures something important that "tool" misses.** > > Substances don't have proper functions — they have *properties* from which uses emerge. Iron isn't *for* anything; it's stuff with certain capacities (hardness, malleability, conductivity), and we discover uses by working with it. This is why the "What is ChatGPT for?" question doesn't have a good answer: it's not that kind of thing. > > **It reframes Hoel's oil analogy against him.** > Hoel drops "discovering oil" as if it's obviously transformative, then argues LLMs haven't been transformative, so they're "just tools." But the oil discovery was a *process*. For decades, oil was lamp fuel. The transformations came later — internal combustion, plastics, petrochemicals. > > If LLMs are a substance, we might be in the kerosene phase. We've discovered they can generate text. We haven't discovered the internal combustion equivalent. That's not a prediction of transformation; it's a point about the structure of Hoel's argument. He's evaluating too early. > > **It connects to "medium" in a useful way.** > Substance is what the stuff *is*. Medium is what you get when you take up the stuff into a *practice*. Oil is a substance; oil painting is a medium. The LLM is a new kind of stuff; text-exchange-with-an-LLM is a medium. > > This gives you a two-part move: > - Against "tool": LLMs are more like a substance (stuff with emergent capacities, not things with fixed functions) > - For the positive characterisation: when we take up this substance into a practice, it becomes a medium --- ### 2026-03-11 — LLM as substance: Aristotelian potentiality and Lockean substratum > **Aristotle (ousia):** Aristotelian substances have *potentiality* (dynamis). They're not purely passive. A stone has the potentiality to fall. Fire has the potentiality to burn. The acorn has the potentiality to become an oak. Capacities are *in* the substance, waiting to be actualised. > > Form and matter: matter is potentiality (the stuff that can become various things), form is actuality (what the thing is). A bronze statue is bronze (matter) shaped into a figure (form). The bronze has the potentiality for many forms; the sculptor actualises one. > > This fits the LLM case: the LLM is matter with potentialities. We don't yet know what forms can be actualised from it. Prompting is a kind of actualisation — drawing out what was latent. > > **Locke:** Skeptical about substance. We observe qualities (hardness, color, weight) but we posit "substance" as the unknown substratum that holds them together — the "something I know not what." > >> "The idea then we have, to which we give the general name substance, being nothing but the supposed, but unknown, support of those qualities we find existing..." > > This might fit the LLM case. We observe the capacities (text generation, pattern matching, response to prompts) but we don't know what the underlying thing *is* that produces them. The "substance" is the posited unity behind the observable capacities. --- ### 2026-03-11 — LLM as substance: epistemic openness as the strongest point > **The epistemic claim: we don't yet know what can be elicited from this stuff.** Hoel assumes we can evaluate LLMs by looking at writing. The substance framing says that evaluation is premature — we're still in discovery mode. Prompt engineering, in-context learning, chain-of-thought — these are *discoveries about what the substance can do*. The history isn't over. > > This doesn't commit you to hype. You're not saying "just wait, transformation is coming." You're saying: Hoel's argument assumes we already know what this stuff is for. We don't. That's why "bits in, bits out" is too quick. ---