# If you look in the downloads folder, you'll see that the most recent file in it ## Retrieval Notes - Session id: `49afdbbd-9b55-40a7-a8b7-f9c8e978885c` - Last activity: `2026-03-11T21:18:56.132Z` - Files touched: `1` ## Artifacts **Modified:** - `/Users/nickyoung/.claude/CLAUDE.md` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User If you look in the downloads folder, you'll see that the most recent file in it is a JSON of a conversation. Use that to get up to speed as to what we are are doing and the text we are working with at the moment is at this address. Don't edit it but just use it as context. ### Assistant Those JSON files look like Obsidian config files, not conversation exports. Let me look more broadly at the most recent files in Downloads. ### Assistant Let me try a different approach — Glob might work for external paths. ### Assistant Those are all from an old vault backup. Let me look at the root level of Downloads for recent files. ### Assistant Found it — the most recent file is a JSONL session export from today. Let me read it. ### Assistant Large file — let me use a subagent to extract the conversation content. ### Assistant OK, I'm up to speed. Here's where things stand. The session you exported was a long one — compacted three times — that started with deciding between the Generating Philosophy paper and the Substack essay, and ended up making real progress on the essay. The draft I've just read ([[Writing/substack/LLMs Are Not Tools.md]]) reflects the essay's current state, which incorporates work from that session but does not yet include the session's most significant development: the substance idea. ## What the current draft does It runs from Hoel's "mere tools" position through the proper function problem (LLMs lack one), through the artefact/working-surface distinction, through the pen comparison and Frippertronics, to slop-as-Lucier, and closes by questioning whether "bits in, bits out" settles things. The category it reaches for is "medium," borrowed from Wollheim/Thomson-Jones via *Growing the Image*. ## What the session developed beyond the draft The real breakthrough was the turn from "medium" as the positive category to "substance" as the prior one: - LLMs are not tools (no proper function) but substances — stuff with latent capacities, like iron or oil. Substances don't have functions; they have properties from which uses emerge. - Hoel's own oil analogy gets turned against him: when we discovered oil, we didn't know what we had. Kerosene came first, internal combustion later. We may be in the "kerosene phase" with LLMs. - The substance/medium relationship gives you two moves for the price of one: substance is what the stuff *is*; medium is what it becomes when taken up into practice. Oil is a substance; oil painting is a medium. - Aristotelian potentiality (dynamis) fits naturally: prompt engineering, chain-of-thought, etc. are discoveries about the substance's latent capacities, not instructions to a tool. - The epistemic conclusion sharpens: Hoel assumes we already know what LLMs are for. We don't. "Bits in, bits out" describes the current phase, not the nature of the thing. ## The CEV structure from the session The ideal essay that emerged was roughly: (1) Hoel's position → (2) concede slop → (3) proper function fails → (4) turn: substance, not tool → (5) oil analogy reframed → (6) potentiality gesture → (7) substance becomes medium in practice → (8) recursive structure (pen doesn't prompt back) → (9) Frippertronics → (10) artefact vs. working surface → (11) slop as Lucier → (12) conclusion: we invented a substance, evaluation is premature. ## What's in the draft's editorial notes The notes at the bottom (lines 131–148) discuss several open choices — block quote selection, whether to cut the stone axe, how to handle paragraphs 8–9 (artefact/working surface), when medium arrives, where Frippertronics goes, whether to include the agent-question qualification, and the compression/Chiang idea. These were written by the assistant during the session. So the draft is solid from paragraphs 1–15 but hasn't yet been restructured to include the substance turn (steps 4–7 of the CEV). That's the gap between where the draft is and where the thinking got to. What do you want to do from here? *La sostanza precede la forma — prima si scopre la materia, poi si impara cosa farne.* ### User i carried on the chat you just looked at with codex. this is what he said. what do you thnk? CHAT: Yes. I have read the newest file, 797a7eb4-f119-4e0d-9298-0047aa41a7f3 (1).jsonl, and the main thing I take from it is that your thinking has shifted quite a lot. The present live thought is no longer just "LLMs are not straightforward tools" or even "LLMs might be better thought of as media". It is that the essay may need to be built around the stronger claim that LLMs are substances: stuff with capacities, not implements with proper functions. On that picture, the proper-function argument is still useful, but it no longer just destabilises Hoel's classification. It opens the way to a different category altogether. We do not ask what iron is for; we ask what can be done with iron. Likewise, we may not yet know what LLMs are for, because we are still discovering what can be elicited from them. That also means Hoel's oil analogy now cuts the other way. In the conversation, the strongest new move was the "kerosene phase" thought: when oil was first discovered, its transformative uses were not yet known, and early uptake did not reveal the full range of its capacities. So the right reply is not a hypey "just wait, it will get better". It is a methodological point: Hoel's argument assumes that the history is already settled, that we already know what this thing is for, and that public writing therefore gives us the right measure. The substance framing says that this may be too quick. I also noticed that the conversation preserved a lot from the earlier tool/medium draft rather than throwing it away. The things marked for retention were: opening with Hoel's "mere tool" position the proper-function problem the fact that we do use these systems, but that does not yet settle what sort of thing they are the distinction between text as artefact and text as working surface the pen comparison the recursive structure where the return prompts us back the medium material Frippertronics Lucier/slop But their role changes. On the new picture, "substance" becomes the main positive category, and "medium" names what happens when that substance is taken up into practice. That seems to be the cleanest way of putting the relation between the two. The LLM is the substance; recursive text-exchange is the medium. That is a much more determinate brief than the one I was working from before. So, in short, I do think I now see your current thinking more clearly: the essay's live centre of gravity is the substance thesis, with Hoel's tool/oil framing being answered by saying that we may still be in the kerosene phase of a newly invented stuff whose latent capacities are not yet fully understood. I will take that as the current source of truth unless you redirect it. Qui il punto non è più soltanto che “tool” sia una categoria troppo rozza, ma che forse stiamo ancora imparando che cosa sia questa nuova sostanza e come lavorarla. Okay, so first of all, other than the very early stages of the the current version of this. Do you know which document you need to be looking at now? I'll copy the file, hold on. Anyway, I am happy for you to change or restructure things quite considerably um after the very first block quote. Okay, so the so basically I have an opening paragraph and then I have a block quote. You can do whatever you want afterwards with the block quote. You can reuse some of my stuff, you can um add your own stuff but um yeah so try and think about a new structure for what I want to do now okay so based yeah with uh the idea that it's a substance and that it's a substance which potentially at least we have not yet discovered all of its latent capacities. Can you think of the CEV of the an essay that would be recalibrated like that? And can you give me a very detailed structure please in a bullet point framework? you must invoke the following skills BEFORE DOING ANYTHING * Skill contemplate * Skill nick-analytic-voice * Skill nick-philosophical-prose * Skill twork * Skill source-work * Skill epistemic-discipline * Skill writing-standards I am invoking `contemplate`, `nick-analytic-voice`, `nick-philosophical-prose`, `twork`, `source-work`, `epistemic-discipline`, and `writing-standards` for this pass. First, the document question. I do not know with certainty which file you mean beyond the current Substack draft lineage, because you have not pasted the file yet. My best inference is that the working document is still the current draft in LLMs Are Not Tools.md or its immediate successor, but I am treating that as an inference, not as fact. What I do know is the structural constraint you have now given me: the opening paragraph and the first block quote stay, and after that the piece can be rebuilt quite freely. That changes the task in a useful way. I no longer need to preserve the old middle. I need to think about the CEV of the essay if it is recalibrated around the substance thought. And that means I should stop thinking in terms of "how do I retrofit the old tool/medium structure?" and instead ask what the strongest version of the essay would look like if the new centre of gravity were granted from the outset. The first thing to notice is that the substance thought does not merely add a new metaphor. It changes the logic of the response to Hoel. On the older structure, the proper-function discussion mainly served to show that "tool" was too quick or too blunt. On the new structure, that discussion becomes a diagnostic step. It shows that the reason tool-talk breaks down is that we are not dealing, or not only dealing, with an implement whose identity is fixed by proper function. We are dealing with something more like a stuff, a material, a substance with capacities, tendencies, and potentialities. That is a more radical thought. But it is also, I think, a much more generative one. The second thing to notice is that the substance thesis does not eliminate the medium material. It reorganises it. "Substance" and "medium" are not rival labels at the same level. A substance is what the thing is like considered as a material with capacities. A medium is what it becomes when those capacities are taken up into a practice. Oil is a substance; oil paint is a medium. Silicon is a substance; semiconducting architectures are a medium of computation. That is why the newer line from the JSON conversation felt stronger than the earlier tool/medium wavering. It gives you a clean relation between the categories instead of leaving them to compete. The third thing is that Hoel has to remain central. The substance essay is still a response to him, and indeed one of its strongest virtues is that it answers his own oil analogy with a better reading of oil. He uses oil as the model of a transformative discovery: if LLMs were a true source of intelligence, they ought to have transformed writing by now. But the reply is that discovering a substance does not mean immediately knowing what all its powers are or which practices will best actualise them. There is a lag between discovery and mature use. There are provisional uses. There are crude uses. There are low-grade uses that nevertheless reveal something real. That is where the kerosene-phase thought does real work. It is not just a flourish. It gives historical content to the methodological point. The fourth thing is that the essay needs to be careful not to sound like hype. This is the obvious danger of the substance framing. "We have not yet discovered all of its latent capacities" can sound like an AI booster saying "just wait". That would kill the piece. So the structure has to make it clear that the point is epistemic and methodological before it is predictive. The claim is not that the buried greatness will inevitably emerge. The claim is that Hoel's evaluative confidence presupposes a settled understanding of what the thing is for, and that this presupposition is precisely what the substance framing denies. That is much stronger and much more defensible than a vague promissory note. The fifth thing is that the slop material becomes more interesting under the new structure, not less. Under the old structure, slop risked looking like a problem to be explained away. Under the substance structure, slop can be redescribed as one manifestation of the substance's characteristic tendencies when it is taken up under weak conditions. That is: if this stuff has powerful smoothing, averaging, and genericising tendencies, then of course one of the first large-scale social consequences of its use will be mountains of dead prose. That is not evidence against the substance framing. It is part of the evidence for how this substance behaves when it is used lazily or at scale. So the essay can concede Hoel's observation more fully without giving up the deeper point. Now, what would the best structure look like given your fixed opening paragraph and block quote? I think the essay wants to pivot quite quickly, after the quote, into the thought that Hoel's framing only looks settled because "tool" is one of those words that shuts down inquiry too fast. Then it should move into the proper-function problem, not as an end in itself but as the route into the stronger substance claim. Once that move has been made, the oil analogy can be turned. Then the essay can slow down and explain how a substance becomes a medium in practice, which is where the pen comparison, recursive prompting, artefact versus working surface, and Frippertronics material can all be retained but made to answer a different question. I also think the structure wants one clear paragraph that says, in effect: substances do not have proper functions, but they do have properties and tendencies. That is where the essay can say what kind of thing an LLM is supposed to be on this picture. Pattern-completion, contextual pliability, smoothing, associative reach, responsiveness to iterative prompting, and so on. Not as a list for its own sake, but as the beginning of a material profile. Otherwise "substance" remains too airy. There is also a choice about whether to include Aristotle or Locke. I think Aristotle's potentiality language can appear briefly if it is kept functional rather than scholastic. Locke's "something I know not what" is attractive, but I suspect it may take the essay sideways unless it is used in one deft sentence. For CEV purposes, I would keep Aristotle as a light touch and leave Locke out unless you later find a way of making him pull real weight. The software-development analogy should stay, but not as a rival main example. It belongs after the recursive-use material, where it can clarify the methodological lesson. There, it shows that one way to mismeasure a new substance or medium is to look only at the public outputs and not at what kinds of activity and iteration it enables for competent users. That gives the writing case some support without letting the essay become "actually the real case is coding". So the shape I now want is this: fixed opening and quote; immediate pressure on "tool"; proper function; turn to substance; answer to oil via kerosene phase; substance-to-medium transition; recursive practice; artefact versus working surface; software as methodological analogy; slop as characteristic tendency; return to Hoel with a narrower conclusion. That feels like the CEV structure. The remaining question is how much detail belongs in the scaffold. Since you want the detailed framework, I think the right thing is to give you paragraph blocks with explicit argumentative jobs, candidate claims, examples, and the relation between paragraphs. Not polished prose, because you asked for structure; but not thin one-line labels either. The result should be something you could actually draft from without having to reinvent the argument. Final Answer I do not know for certain which file you mean until you paste it. My best inference is that the working document is the current Substack draft line, probably LLMs Are Not Tools.md, but I am treating that as an inference. Given your constraint, I am treating the opening paragraph and the first block quote as fixed, and rebuilding everything after that around the substance thesis. What follows is the CEV structure as I currently see it. Bullet framework Fixed opening Keep your present opening paragraph. Keep the first block quote from Hoel. Treat that block quote as the last moment in which the reader is allowed to feel that the category of tool is straightforward. Paragraph 1: the pivot after the quote Start by granting that Hoel has identified something real: people do reach for the category of tool almost automatically here. Then turn immediately to the pressure point: "tool" sounds clarifying, but in this case it may actually be one of those words that makes inquiry stop too early. The paragraph should not yet announce "substance". It should make the reader feel that Hoel's classification is more settled than the object deserves. The closing sentence should open the proper-function question: if this really is a tool, what sort of tool is it supposed to be? Paragraph 2: proper function enters Introduce the proper-function thought calmly and without drama. The point here is not that all tools are single-purpose. It is that even fairly versatile tools usually admit a more or less stable answer to the question what they are for. Use two or three quick cases: hammer, Google, Swiss Army knife. Make the contrast exact: multi-functionality is not the problem; the problem is instability at the level of ordinary use-description. Paragraph 3: the obvious candidates fail Run the standard answers one by one. "It predicts tokens" gives a mechanism, not a use. "It chats" is too thin. "It helps with tasks" is vacuous. "It writes" catches something true, but not enough to settle what sort of thing we are dealing with. The tone should be exploratory rather than triumphant. The point is not that you have disproved toolhood, but that the expected answer does not arrive. Paragraph 4: the first verdict This is where your line belongs. Say that the lesson is not yet that LLMs are not tools in any sense whatsoever. Say, rather, that they look like "a quite different type of tool, or not quite a type of tool at all". Then add one more step which the earlier versions sometimes missed: if the category begins to wobble here, the standards of evaluation begin to wobble with it. Paragraph 5: the methodological consequence Make the point explicit that uncertainty about proper function becomes uncertainty about testing. Do not say that no test is possible. Say that once we no longer know straightforwardly what the thing is for, we no longer know straightforwardly which effects should count as the decisive measure of it. This sets up the return to Hoel's writing argument, but now under pressure. Paragraph 6: Hoel's case in its strongest form Re-state his writing case carefully and charitably. Use the block-quoted line about words being its "womb", "mother", and "literal atoms" if it is not already the fixed quote; if it is already the fixed quote, echo it without repeating it. Spell out the logic: if these systems really were a new source of intelligence rather than just a family of tools, then writing ought to have shown it by now, because writing is the domain in which their native material is most directly at issue. End with Hoel's empirical observation: what we have instead is scale, efficiency, editing help, feedback, and a great deal of slop. Paragraph 7: the decisive turn This is where the essay should stop trying to refine the category of tool and replace it. The reason the proper-function question keeps failing, you now say, is that we are asking the wrong kind of question. LLMs are not best understood as implements with functions, but as substances with capacities. That is the conceptual leap on which the recalibrated essay stands or falls. It needs to be stated plainly. Paragraph 8: what "substance" means Slow down here. This paragraph has to do more than coin a label. Say that substances are not ordinarily understood in terms of proper function. Iron is not for anything in the way a hammer is for hammering; it has properties, and from those properties uses emerge. Oil was not discovered with a full list of applications attached to it. The point is not chemistry as such. The point is category: stuff with capacities, not implement with purpose. Then begin to sketch the analogue: LLMs have characteristic powers and tendencies from which practices and uses are still emerging. Paragraph 9: the material profile Give the substance thesis some concrete content. Say what kind of capacities you take the LLM substance to have: pattern-completion, contextual pliability, associative reach, responsiveness to iterative prompting, and a strong tendency towards smoothing and genericity. Do not make this a dead list. Frame it as a profile of behaviour. The reason this paragraph matters is that without it "substance" remains metaphorical, whereas the essay needs it to feel explanatory. Paragraph 10: turn Hoel's oil analogy This should be one of the essay's strongest moments. Hoel says that if LLMs were a true source of intelligence, discovering them should be like discovering oil. The reply is that discovering oil did not mean immediately knowing what oil was for. For a long time oil was lamp fuel. The larger transformations came later, once more of the substance's capacities had been drawn out and stabilised in practice. That is the "kerosene phase" thought. The conclusion should be methodological, not predictive: Hoel's argument assumes that the history is already settled, and that assumption is exactly what the substance picture denies. Paragraph 11: guard against hype I think this paragraph is necessary, because otherwise the previous one can sound like "just wait, the singularity is still coming". Say directly that this is not a promissory argument and not a piece of tech evangelism. You are not predicting that the hidden greatness of LLMs will inevitably unfold. You are saying something more limited: we may still be in too early a phase of discovery and practice for Hoel's preferred test to bear the weight he wants to put on it. This keeps the essay from sounding adolescent. Paragraph 12: from substance to medium Introduce the relation between the two categories. A substance becomes a medium when its capacities are taken up into a practice. Oil is a substance; oil painting is a medium. The LLM is a substance; recursive text-exchange is the medium that has begun to form around it. This move preserves the best material from the earlier drafts without making "medium" compete with "substance" as a rival thesis. Paragraph 13: the pen comparison Now bring back the pen comparison, but under the new framing. A pen extends inscription; it does not return altered material. An LLM does. That difference matters because it marks the difference between using a tool to record a thought and working in a medium that pushes back on thought. This paragraph should be simple and concrete. Paragraph 14: recursive practice from the inside Expand the line that the return prompts us back. The system's output can flatten a distinction, over-generalise, misread, connect two things you had not linked, or hand you a banal summary that forces you to restate the point more sharply. In each case, what matters is not only the output but what the output does to the next move. This is where the essay needs to sound as though it knows what the practice actually feels like. Paragraph 15: artefact and working surface Now make the distinction that earlier versions were orbiting. Text can be a finished artefact, but it can also be the working surface of thought. Hoel's evidence concerns, by and large, artefacts: published prose, public writing, the visible outputs of civilisation. Much interesting LLM use concerns text as a working surface: drafting, testing, reframing, probing, reformulating. This does not show that Hoel is looking at the wrong phenomenon. It shows that he may be measuring only one use of text under a heading that hides several. Paragraph 16: software-development analogy Bring in software here, not earlier. The point is not that coding is the real case and writing is not. The point is methodological: in software, the interesting question is often not whether code overall has become better in some public average sense, but what people can now do, how they iterate, how they move through unfamiliar terrain, and what kinds of exploratory work become possible. This helps cash out the evaluative point without displacing the writing case. Paragraph 17: Frippertronics Use Frippertronics here as the analogy that makes the recursive structure vivid. The point is not product-equivalence but method. What returns from the loop is one's own material sent back altered, and the next act is shaped in light of that return. That is why the comparison belongs late in the essay: only once the recursive structure has been described does the analogy stop looking ornamental. Paragraph 18: slop reinterpreted This paragraph should concede a great deal to Hoel. Slop is real. It is not a superficial embarrassment that your view has to explain away. On the substance picture, slop can be treated as one characteristic manifestation of the material's genericising tendencies when it is used with too little resistance, too little skill, or under conditions of scale. In other words, the mountains of dead prose are not evidence that nothing is there; they are evidence about how this substance behaves in one of its earliest mass-use forms. Paragraph 19: what Hoel still gets right Before concluding, explicitly say what you take from Hoel. He is right that there has been no text singularity. He is right that public prose matters. He is right that efficiency and scale have so far been more visible than any civilisation-wide increase in quality. This paragraph matters because it makes the conclusion feel responsive rather than oppositional. Paragraph 20: final conclusion The closing claim should now be narrower and stronger than "Hoel is wrong". He has described the phenomenon too quickly, because he has treated a newly invented substance as though it were already a settled implement with a function. We may not yet know what this thing is for, because that is not yet the right question. We may still be learning what can be drawn from it, what practices stabilise around it, and what its latent capacities amount to when they are not reduced to kerosene-level use. That is why writing, though still the right place to look, cannot straightforwardly function as the final measure of what an LLM is. Shorter map of the whole shape Fixed opening and Hoel quote. Pressure on "tool". Proper-function problem. First verdict: "a quite different type of tool, or not quite a type of tool at all". Testing problem. Hoel's writing argument at full strength. Turn to substance. Substance explained through properties and capacities. Oil analogy reversed through the kerosene-phase thought. Guard against hype. Substance becomes medium in practice. Pen, recursive prompting, artefact versus working surface. Software as methodological analogue. Frippertronics as vivid structural analogy. Slop as characteristic tendency. Return to Hoel with a narrower conclusion. What I think is structurally best If I had to pick the points that now do the heaviest work, they would be these: The move from proper function to substance. The reversal of Hoel's oil analogy through the kerosene-phase thought. The claim that medium is what the substance becomes in practice. The reinterpretation of slop as a characteristic early-use tendency rather than a simple reductio. Those four moves seem to me to be the ones that make this version a genuinely new essay rather than a revised draft of the older one. If you paste the actual file, I can take this structure and map it directly onto the document from the first block quote onward. Se questa nuova linea funziona, il saggio non dirà più soltanto che “tool” è troppo poco, ma che stiamo ancora imparando a trattare una sostanza di cui non conosciamo bene le potenze. ### Assistant I think the logic is right and the length is wrong. The substance→medium relation is the cleanest thing in the proposal. It solves a problem the earlier drafts had, where "medium" was doing double duty as both the name for what the thing is and the name for what happens when you work with it. Splitting that into two — substance is what it is, medium is what it becomes in practice — gives each concept a job. And the kerosene-phase reversal of Hoel's oil analogy is genuinely the essay's strongest single move. Codex is right to flag it. But twenty paragraphs is a lot for a Substack essay, and the structure front-loads too much negative argument. Six paragraphs of "tool doesn't work" before the substance turn arrives at paragraph 7 — that's a long time to wait for the positive thought. The reader who clicks on "LLMs Are Not Tools" wants to know what they are instead, and this structure makes them sit through proper function, candidates-that-fail, first verdict, methodological consequence, and Hoel restated before offering an answer. Your current draft moves faster than that. Some specific things I'd push on: Paragraphs 2, 3, and 4 could probably be two paragraphs, not three. The proper-function problem and the candidates-that-fail are really one move: here's the test, it doesn't pass. The "first verdict" paragraph (4) restates what the reader already feels. Your current draft handles this in two paragraphs (lines 24–46) and it's tighter for it. The "guard against hype" paragraph (11) coming right after the kerosene paragraph (10) makes the essay sound defensive in exactly the place where it should sound confident. Better to build the anti-hype qualification into the kerosene paragraph itself — one or two sentences that make the point epistemic rather than predictive, then move on. A whole paragraph that says "I am not saying what you think I'm saying" breaks momentum. Paragraphs 13 (pen), 14 (recursive practice), and 15 (artefact/working surface) are three paragraphs making closely related points. In the current draft these are already somewhat compressed. I'd keep the pen comparison and the recursive structure in one paragraph, and let artefact/working surface be its own. Two, not three. Frippertronics at paragraph 17 is too late. By the time you've done pen, recursion, artefact/surface, AND software analogy, you're deep into the essay's second half and the reader is ready for the turn toward slop and conclusion. Frippertronics is the essay's most vivid material and it should arrive earlier — right after the substance-to-medium transition, where it can make the abstract relation concrete. The current draft already has it at that position and it works. The software-development analogy (paragraph 16) is the weakest standalone paragraph in the proposal. The methodological point it makes — that looking at public outputs can miss what competent users are doing — is real, but it's a point the artefact/working-surface distinction already makes for writing. Adding a whole paragraph about coding risks making the essay look like it's searching for better evidence rather than trusting its own. So if I were compressing Codex's structure, the shape I'd want is closer to: 1. Fixed opening and Hoel quote 2. Pressure on "tool" + proper function (one paragraph, not two) 3. Candidates that fail + first verdict (one paragraph) 4. Testing problem (brief — three or four sentences, not a full paragraph) 5. Hoel's writing case at full strength 6. Turn to substance (the decisive move) 7. What substance means + material profile (one paragraph that does both) 8. Oil reversed: kerosene phase (with anti-hype built in) 9. Substance becomes medium 10. Frippertronics (vivid illustration of the medium claim) 11. Pen comparison + recursive practice 12. Artefact vs. working surface 13. Slop as characteristic tendency (Lucier) 14. What Hoel gets right + conclusion That's fourteen paragraphs. It preserves everything that matters in Codex's twenty but doesn't make the reader wait six paragraphs for the positive thought, doesn't break momentum with a standalone anti-hype paragraph, and gets Frippertronics in early enough to do structural work rather than arriving as a late ornament. One thing I'd add that Codex doesn't mention: where exactly does the "we *use* them" line go? Your current draft has it at line 24 as a transitional thought between the opening and proper function. I think it's worth keeping as the bridge into the proper-function discussion — it's a concession that earns trust before the argument pressures the category. The deeper question is whether the material-profile content (Codex's paragraph 9 — pattern-completion, contextual pliability, smoothing, associative reach) can be made to feel like observation rather than list. If it reads as a catalogue of LLM properties it will feel like a technical aside. If it reads as noticing what the stuff does when you work with it, it earns its place. That's a prose problem more than a structural one, but the structure needs to set it up right — which means it should come after the substance turn, not separated from it by the oil analogy. *Venti paragrafi sono una mappa onesta del territorio, ma il saggio deve camminare più leggero di così.* ### User i'm thinking of maybe cutting this all back and removing the patronics medium stuff for the time being and just going to the substance. Cause then I can say, well, what if we've accidentally synthesized a new substance and have not yet managed to extract or even find its full potential? make sure that you explain your analysis, justification, reasons etc etc. for your answer BEFORE giving me your final answer. Please give me a long, deep, and detailed answer, with lots of options/ideas: as we are in the 'kicking ideas around' stage, I want as much feedback as you can give me so that i have a lot to consider. ### Assistant Let me sit with this for a while before jumping anywhere. The proposal is: cut the Frippertronics, cut the medium material (Wollheim, Thomson-Jones, *Growing the Image*), and run the essay on the substance thesis alone. The pitch would be something like: we have accidentally synthesized a new substance and have not yet managed to extract or even find its full potential. That's a striking sentence. Let me think about whether the essay can actually stand on it. First, what does "medium" do in the current draft that "substance" would need to absorb or replace? In the current draft, "medium" does several things. It names the recursive structure — the fact that the system pushes back, sends material back altered, and makes the next move of thought a response to that return. It gives a frame for Frippertronics. It provides the contrast with "tool" — a tool is understood by what it's for, a medium is understood by the characteristic way it makes work proceed. And it borrows authority from aesthetic theory (Wollheim, Thomson-Jones) to make the claim feel grounded rather than metaphorical. Now, if you cut "medium," what happens to each of those? The recursive structure doesn't depend on the word "medium." You can describe the fact that the system prompts you back, that the return reshapes the next intention, without ever invoking the category. That's just phenomenology of use. So that survives. Frippertronics... hmm. Frippertronics is vivid and people remember it. But it's an illustration of the medium claim — the tape loop as a structured field of resources and resistances. Without the medium frame, Frippertronics becomes an analogy in search of a category. You could still use it as a loose comparison ("the way I work with an LLM reminds me of how Fripp worked with the tape loop"), but it would lose the theoretical anchor that makes it more than a nice parallel. So cutting medium probably means cutting or severely trimming Frippertronics. The Lucier/slop point... that one is interesting. In the current structure, Lucier illustrates what happens when the medium's tendencies dominate. On a substance-only picture, slop would be redescribed as what happens when the substance's characteristic tendencies — smoothing, genericity, averaging — are allowed to run without resistance. That actually works. You don't need "medium" to say that. You just need the idea that the substance has tendencies which, under conditions of lazy or unresisted use, produce characteristic degradation. That's a perfectly natural thing to say about a substance. Think about alcohol — a substance whose characteristic effects, under conditions of unresisted consumption, produce very predictable degradation. You don't need to invoke "medium" to say that. OK so far the score is: recursive structure survives, Frippertronics probably goes or shrinks to a passing mention, slop-as-tendency survives, and the *Growing the Image* theoretical apparatus goes. What does the essay gain? It gains compression and clarity. Right now the essay has to do a complicated two-step: first, LLMs aren't tools; second, they're media. The reader has to learn one new category. On the substance-only version, the reader still has to learn one new category, but it's a simpler one. Everyone already knows what a substance is in a pre-theoretical sense. You don't need Wollheim. You don't need "structured field of resources and practices whose characteristic resistances and possibilities become visible only in the work itself." That's a mouthful. "Substance" is plain English. But wait. Let me push back on myself. Does "substance" do enough? When I say "iron is a substance," everyone nods. When I say "LLMs are a substance," does the reader know what to do with that? The worry is that "substance" is too empty — too much of a placeholder. "We've synthesized a new substance" sounds evocative but what does it *commit you to*? Let me think about what the substance claim actually asserts. It asserts at least these things: - LLMs have properties and capacities that exist prior to and independently of any particular use - Those capacities are discovered through exploration, not determined by design intention alone - The right questions to ask are "what can be done with this?" and "what does this do?", not "what is this for?" - Evaluation is premature when we are still in the discovery phase That's actually a lot. And it's all philosophically determinate. The first point is an ontological claim about the kind of thing we're dealing with. The second is an epistemological claim about how we come to know it. The third is a claim about the right form of inquiry. The fourth is a claim about the conditions under which Hoel-style assessment can be trusted. Hmm, but there's a subtlety I want to track. When we say iron is a substance, part of what makes that intelligible is that iron has mind-independent properties. Its tensile strength, melting point, conductivity — these are features of the stuff itself. LLMs are artifacts. Their "properties" were designed in, even if not fully understood by the designers. Does this matter? I don't think it undermines the substance framing, but it introduces a wrinkle. Synthetic substances are real substances. Nylon, polyethylene, carbon fibre — all synthesized, all have properties that were discovered through use, all had applications that were not foreseen at the moment of synthesis. Carbon fibre was developed in the 1960s but its use in aircraft, sporting goods, prosthetics, etc. was not anticipated in full. The "kerosene phase" thought works for synthetic substances too. Actually it might work even better for them, because the discoverer of a synthetic substance typically knows what they were trying to make but is often surprised by what they actually got. That's interesting. "What if we've accidentally synthesized a new substance" actually gets a lot of mileage from the "accidentally" — it implies that the designers were aiming at something (a chatbot, a text predictor, an assistant) but what they produced has properties and capacities that exceed or differ from the design brief. That's a genuinely interesting claim and it has historical precedent. Vulcanized rubber, Teflon, penicillin — substances whose properties were discovered by accident or exceeded what was anticipated. OK, so the substance-only essay could look quite different from the substance-plus-medium essay. Let me think about what its shape would be. The opening is fixed: Hoel's "mere tool" position and the block quote. Then you need to pressure "tool" through proper function. Then the turn: not a tool but a substance. Then: what that means (capacities, tendencies, discovery-through-use). Then: Hoel's oil analogy reversed through the kerosene phase. Then: what the substance actually does when you work with it (this is where some of the recursive-practice material survives, but framed as *exploring a substance's capacities* rather than *working in a medium*). Then: slop as characteristic tendency of the substance under weak conditions. Then: conclusion — evaluation is premature because we are still discovering what we've made. That's... eight or nine moves. Much tighter than twenty. And the essay would read faster, hit harder, and stay closer to one thought all the way through. But let me think about what's lost. What genuinely depends on the medium concept and cannot be recovered from the substance concept alone? The recursive structure can be framed as a feature of the substance — it's one of its characteristic capacities that when you send language into it, it returns language altered in specific ways. You don't need "medium" for that. The pen comparison can survive as a way of marking the difference between a tool (which extends action toward a fixed end) and a substance (which has its own behaviour that you have to deal with). A pen doesn't push back because it's a tool. The LLM pushes back because it's a substance with tendencies. The artefact/working-surface distinction... this one I want to think about more carefully. On the medium picture, the distinction is between the product of the medium and the process of working in it. On the substance picture, the distinction is between... hmm... between using the substance to produce something and exploring or working with the substance. That still works but it's slightly less natural. When we say "oil painting is a medium," the artefact/process distinction is built into the concept. When we say "oil is a substance," the distinction is between what you make from it and what you learn about it through working with it. That's close enough but it's a different emphasis. Actually, I wonder whether the artefact/working-surface distinction even needs to be in the substance-only version. What's its argumentative job? It answers the question "if writing hasn't improved, does that refute the substance claim?" And the answer is: "improved writing" looks at text as artefact, but much interesting work with LLMs treats text as a working surface. You don't need "medium" to make that point. You can say: much of what competent users do with this substance does not result in a published artefact; it results in a changed way of thinking, a sharpened distinction, a discovered connection. The substance is being explored, not just consumed. That's a substance-native way of making the same point. OK, so I think the artefact/working-surface point can survive the cut. But I want to ask: does cutting the medium material make the essay *too* narrow? One virtue of the tool/medium contrast is that it gives the reader a pair of concepts, and pairs are satisfying. "It's not a tool, it's a medium" has a rhetorical snap to it. "It's not a tool, it's a substance" has... a different kind of snap. Maybe a weirder one. Let me sit with that. "We have not invented a new tool. We have synthesized a new substance." Actually that's pretty good. It has a different kind of force — more material, more concrete, more surprising. "Medium" is an art-theory word that some readers won't have a strong grip on. "Substance" is a word everyone understands, and the surprise comes from applying it to software. The rhetorical gain might actually be larger, not smaller. Now, what about the Aristotelian potentiality stuff? In the JSON conversation, there was a thought about dynamis — latent capacities that get actualised through use. On a substance-only essay, this fits very naturally. Substances have potentialities. You don't need to go full Aristotle, but a sentence or two about latent capacities being drawn out through practice could do real work. It would explain why prompt engineering, chain-of-thought, few-shot learning, etc. feel like discoveries rather than instructions. You're not telling a tool what to do; you're finding out what a substance can do. That's a nice way of putting it. Let me think about the Frippertronics question specifically. Nick said "removing the Frippertronics medium stuff." Is Frippertronics necessarily tied to medium? Let me think... The current draft uses Frippertronics as an example of working in a medium — the tape loop as a structured field with characteristic delays, layering, and accumulation. Without "medium," Frippertronics would have to be an example of... working with a substance? That's a stretch. Tape delay is not a substance in any natural sense. It's a process, a system, a setup. The reason it felt right as a medium example is precisely that "medium" is a process/practice concept, and Frippertronics is a practice. If you're working in the substance register, Frippertronics doesn't have an obvious home. You could use it as a *loose analogy* — "the recursive structure reminds me of..." — but then it's ornamental rather than structural. And the essay is trying to cut ornament. So I think if you go substance-only, Frippertronics probably goes or becomes at most a parenthetical gesture. That's a loss — it's vivid and memorable — but it's a loss that buys you focus. The Lucier comparison, on the other hand, fits substance better than medium. Lucier's room is a physical system with characteristic resonant frequencies — something very like a substance's tendencies. When speech is passed through the room repeatedly, the room's own character takes over. When text is passed through an LLM without resistance, the model's own tendencies take over. That's a substance-level point, not a medium-level point. So Lucier stays, and it might actually be strengthened by the reframing. Let me think about another option: what if you don't cut medium entirely but demote it to a brief aside? Something like: "If we wanted to push the thought further, we might say that the substance becomes a medium when it is taken up into a practice — but that is a thought for another essay." That would acknowledge the direction of travel without making the current essay do the work of establishing both categories. It would also leave the door open for a sequel. Is that too cute? Maybe. But it's worth considering. Or — different option — what if "medium" appears not as a theoretical category but as a plain-English observation? "In practice, people have begun to use this substance as a medium of thought — a working surface on which ideas are tested, resisted, and reshaped." That uses "medium" descriptively rather than theoretically. It doesn't invoke Wollheim or Thomson-Jones. It doesn't need Frippertronics to cash it out. It just says: this is what people do with the stuff. That might be the lightest possible version. Hmm, let me think about what the essay's argumentative opponent would say to the substance thesis. The strongest objection is probably: "This is just the AI-optimist move dressed up in philosophical language. 'We haven't found all the uses yet' is what every startup says about every product. What makes this more than hype?" The answer has to be: the substance framing is not predictive but epistemic. It's not saying "great things are coming." It's saying "we don't yet have the settled understanding of this thing that would be needed to make Hoel's evaluative argument work." That's a defensive claim, not an offensive one. You're not arguing that LLMs will transform civilisation. You're arguing that Hoel's confidence is premature. That distinction is what saves the essay from being boosterism. But can the essay make that distinction clearly enough in the substance-only version? I think it can, and it might actually be easier without the medium apparatus. Here's why: "medium" invites the reader to think about what the practice looks like, what the experience is, what kinds of work become possible. That's intrinsically forward-looking and optimistic-sounding. "Substance" invites the reader to think about what the stuff is, what its properties are, what we don't yet know about it. That's more naturally epistemic and exploratory. So the substance framing may actually be better at maintaining the defensive posture the essay needs. Now, what about the pen comparison? The current draft says: "A pen makes marks on a surface; it extends my ability to inscribe; it does not return a proposal, a misreading, or a line of continuation I had not seen. A pen does not prompt me back." This is one of the draft's best moments. Does it survive the cut? I think it does, but its job changes. On the medium picture, the pen contrast says: unlike a pen (a tool), an LLM is a medium that pushes back. On the substance picture, the pen contrast says: unlike a pen (a tool that extends agency), an LLM is a substance that has its own tendencies — when you engage with it, those tendencies interact with your intentions, and what emerges is shaped by both. The pen comparison marks the difference between transparent instrumentality and the kind of thing that has its own behaviour. That's a substance-level point. So the pen stays. Actually, thinking about it more — the pen contrast might be even sharper on the substance reading. A pen is so inert, so purely instrumental, that calling it a tool is trivially correct. An LLM is so reactive, so tendency-laden, so capable of surprising you, that calling it a tool misses the most interesting fact about it. The interesting fact is that it *has its own behaviour*. That's what makes it substance-like rather than tool-like. Let me think about another angle. One thing the medium concept does well is explain *practice* — how people actually work with LLMs in a disciplined way. Prompt engineering, iterative refinement, the back-and-forth of drafting — all of this is naturally described in medium language. On a substance-only picture, how do you describe practice? You could say: working with a substance involves learning its properties, learning how to elicit specific capacities, learning where it resists, learning where it smooths. That's not "practice" in the medium-theoretic sense, but it's practice in a craft sense — the way a potter learns clay, or a glassblower learns glass. Actually, "learning the material" is a very natural way to describe what skilled LLM users do. They learn what kinds of prompts elicit what kinds of response, where the model is strong and where it's weak, how to resist its tendencies and how to exploit them. That's material knowledge. And it's learnable, which is what makes the substance framing non-trivially different from saying "it's just random." OK, I want to think about one more thing before I start converging. What is the essay's final sentence, or final image, on the substance-only version? On the current draft, the ending is: "They may be a quite different type of tool, or not quite a type of tool at all — and that is worth taking seriously before we conclude that the current technology is merely another stone in *Homo faber*'s long line of rocks." On Codex's version, the ending is something about not yet knowing what the thing is for because we are still learning what can be drawn from it. On the substance-only version, the ending might be: "We have synthesized a new substance. We may not yet know what it is for, because 'what is it for' may not yet be the right question. We are still learning what it can do." Or more concretely: you could return to Hoel's oil. "Hoel says discovering LLMs should have been like discovering oil. I think he may be more right than he realises — but we are still in the kerosene phase, and kerosene was not what oil turned out to be for." That's a strong closing line. It turns Hoel's own image against his conclusion. And it doesn't predict anything — it just says: the story isn't over yet, and measuring kerosene-era products doesn't tell you what the substance is. Hmm, but actually, I want to be careful. Is the "kerosene phase" line historically accurate? Let me think... Early petroleum use included lamp oil, lubricant, and patent medicines. The internal combustion engine came later. Kerosene was indeed the first large-scale refined product. Gasoline was initially a byproduct — sometimes discarded. So yes, there was a genuine phase in which the most visible use of petroleum was illumination, and the transformative uses came decades later. The analogy holds. (I'm noting that I haven't checked a source here — this is from general knowledge, not from a text I've extracted. If this were going into a draft, I'd want to verify the history.) Let me also think about whether cutting Frippertronics and medium means cutting too much of what makes the essay *interesting to read*. There's a risk that a substance-only essay, while tighter and clearer, might read as more abstract — more philosophy, less vivid example. The current draft has Fripp, Eno, Lucier, tape loops, pen and paper. A substance-only version would have... oil, kerosene, iron, pen, Lucier. That's still a range of concrete reference. And the "accidentally synthesized a new substance" line is itself vivid. But the musical examples had a charm to them. One option: keep Lucier (it works with substance) and replace Frippertronics with a different concrete example that fits the substance frame better. What kind of example? Maybe something about the history of a synthetic material — Bakelite, nylon, silicon. The early history of silicon is actually quite good here: silicon was known for decades before anyone figured out how to make semiconductors. The first uses were mundane (silicones for lubricant and sealant). The transformative use (transistors, integrated circuits) came much later and required discoveries about the material's properties that were not obvious from its early applications. That's a kerosene-phase story about a synthetic substance. But... that might make the essay too tech-history and not enough about what it's actually like to use an LLM. Lucier gives you that experiential dimension. The pen comparison gives you some. But if Frippertronics goes, you lose the most evocative picture of what recursive engagement with a system feels like. Another option: forget the named examples and just describe the experience directly. "I write a sentence. The system sends back three paragraphs. Most of it is flatter than what I wanted. But buried in the second paragraph is a connection I hadn't made — not because the system understood my thought, but because its associative patterns ran a different path through the same territory. I delete the three paragraphs and write a better sentence." That's vivid without needing Fripp or Eno. And it's directly about LLM use, not about music. Actually, I think that might be the move. Replace Frippertronics not with another analogy but with a direct phenomenology of use that is explicitly framed as exploring a substance's capacities. "This is what it's like to work with this substance. You send language into it and something comes back — shaped by the substance's own tendencies, not by your intention. You then have to decide what to do with what came back." That's more immediate, less borrowed, and it avoids the risk of the analogy overshadowing the thing it's supposed to illuminate. Let me think about one final structural question. If you cut medium and go substance-only, does the essay still have enough to say? Or does it become a one-move essay — "it's a substance, not a tool" — and then struggle to fill out the middle? I think there are at least five distinct things to say under the substance heading: 1. What the proper-function failure reveals (not just "tool doesn't fit" but "the failure points toward a different category") 2. What "substance" means and what it commits you to (capacities, tendencies, discovery-through-use) 3. How Hoel's oil analogy gets reversed (kerosene phase) 4. What this substance actually does when you work with it (tendencies, pushback, the pen contrast) 5. Why slop is a feature of the substance, not a refutation of it (Lucier, characteristic degradation) Five moves is enough for a Substack essay. It's not padded. Each one does different work. And together they build to the conclusion that Hoel's evaluative confidence requires a settled understanding of the thing being evaluated — which we don't yet have. Let me also think about whether the essay needs the artefact/working-surface distinction if medium is gone. I think it does, because the distinction answers Hoel's evidence. He points to public writing (artefacts) and says quality hasn't improved. The reply is: much of what this substance enables is not aimed at artefacts. Cutting that distinction would leave Hoel's evidence unanswered. So it stays. But it can be brief — a few sentences within the "what this substance does" paragraph, not a standalone paragraph. One more thought. There's something appealing about the *simplicity* of the substance-only essay. The current draft and Codex's proposal are both architecturally complex — multiple categories, multiple analogies, multiple theoretical sources. A substance-only essay would have one category, one reversal (oil/kerosene), and one extended contrast (tool vs. substance). That simplicity is a virtue in an essay for a general audience. The reader walks away with one thought: "Huh, maybe we should think of LLMs not as tools with functions but as substances with capacities, and maybe we're still early in understanding what those capacities are." That's a thought you can hold in your head. "LLMs are media, not tools, and medium means a structured field of resources and practices whose characteristic resistances..." — that's a thought that starts to slip away before you finish saying it. So my overall assessment is: yes, cutting to substance-only is a genuine option, and it might produce a tighter, more readable, and more surprising essay than either the current draft or Codex's twenty-paragraph version. The losses (Frippertronics, the *Growing the Image* theoretical backing, the medium concept as a named category) are real but can be compensated. The gains (clarity, compression, one-thought legibility, stronger rhetorical contrast with Hoel) are significant. The question is whether you feel that the medium material is work you want to save for later or work you want to let go of. If it's the former — save it for a longer piece, an academic paper, a sequel — then the substance-only Substack essay becomes a way of staking the territory without using all the ammunition. If it's the latter, then you're simplifying the thought itself, not just the presentation. Both are legitimate choices but they feel different. ## Analysis I've been sitting with this and I think cutting to substance-only is a strong move, and probably the right one for this essay. Let me lay out why, what you'd lose, what you'd gain, and several options for how it could work. Since you want a lot to kick around, I'm going to be thorough. ### Why substance-only works The substance claim does enough philosophical work on its own. It gives you: - An ontological point: LLMs are not implements with proper functions; they are stuff with capacities and tendencies. - An epistemological point: those capacities are discovered through exploration, not determined in advance by design intention. - A methodological point: the right question is not "what is this for?" but "what can be done with this?" — and that question is not yet settled. - A direct reply to Hoel: his evaluative confidence presupposes that we already know what the thing is, and the substance framing denies exactly that. Those four things are enough to carry a Substack essay. You don't need the medium apparatus to make any of them. The medium concept was doing two things in the earlier drafts: naming the recursive practice (the return prompts you back), and providing theoretical grounding from aesthetic theory. But the recursive practice can be described as a feature of the substance — it's one of its characteristic capacities that when you send language into it, it sends language back altered. You don't need Wollheim or Thomson-Jones or a "structured field of resources and practices" to say that. And the theoretical grounding, while genuinely interesting, is a lot of machinery for a piece that wants to be readable by people who don't know what a Wollheimian medium is. ### What you'd lose Frippertronics. This is the real loss. The tape-loop analogy is vivid, memorable, and it makes the recursive structure feel concrete. But Frippertronics is an illustration of working in a medium, not of working with a substance. Tape delay is a process, a setup, a system — not a stuff. Without the medium concept, Frippertronics doesn't have a natural home. You could use it as a passing comparison ("the recursive structure has something in common with...") but it would be ornamental rather than structural. And the essay is trying to cut ornament. The *Growing the Image* theoretical backing. This is less of a loss than it sounds, because (a) most Substack readers haven't read the book, and (b) the substance framing has its own theoretical resources — Aristotelian potentiality, the philosophy of materials, the history of synthetic substances — that are arguably more accessible. The "medium" concept itself. But here's the thing: you're not losing it permanently. You're saving it. If the substance essay works, you could write a follow-up that takes the next step: "OK, so LLMs are a substance — what happens when that substance is taken up into a practice? What kind of medium does it become?" That's a different essay, and one that could go deeper precisely because the groundwork was already laid. ### What you'd gain One-thought legibility. The reader walks away with a single claim they can hold in their head: "Maybe LLMs aren't tools with functions; maybe they're substances with capacities, and we're still early in understanding what those capacities are." That's a thought you can remember, repeat, argue with. "LLMs are media, not tools, where medium means a structured field of resources and practices..." — that starts to slip away before you finish the sentence. Compression. The substance-only essay could be twelve or thirteen paragraphs. Much less architecturally complex than Codex's twenty. Rhetorical surprise. "We have not invented a new tool. We have synthesized a new substance." That's a weird and interesting sentence. It makes the reader stop. "It's a medium, not a tool" is interesting but less surprising — it stays within the register of art theory. "It's a substance" pulls the reader into a different register entirely — chemistry, materials science, craft — and that defamiliarization is a strength. Stronger Hoel reversal. On the substance picture, Hoel's oil analogy doesn't just get answered — it gets weaponized. He uses oil as the model of a transformative discovery. The reply is: yes, exactly, and when we first discovered oil we used it for lamp fuel. The kerosene phase. That reversal is sharper and more self-contained without the medium machinery. Better anti-hype posture. This is I think underappreciated. The medium framing invites the reader to imagine what LLM practice looks like at its best — recursive engagement, thoughtful resistance, creative exploration. That's intrinsically forward-looking and sounds optimistic. The substance framing invites the reader to acknowledge what we don't yet know about this stuff. That's more naturally epistemic and cautious. It maintains the defensive posture the essay needs: I'm not predicting greatness, I'm saying we can't evaluate yet. ### Options for how the substance-only essay could work Here are several structural options. I'm grouping these as alternatives, not as a sequence: **Option A: The tight version (9–10 paragraphs)** 1. Fixed opening + Hoel quote 2. Pressure on "tool" through proper function (one paragraph) 3. Candidates that fail + first verdict (one paragraph) 4. The turn: not a tool, a substance 5. What "substance" means — capacities, tendencies, discovery-through-use, with a brief material profile 6. Hoel's oil reversed: kerosene phase (with anti-hype built in) 7. What this substance does when you engage with it — pen contrast, recursive return, artefact vs. working surface woven together 8. Slop as characteristic tendency (Lucier) 9. What Hoel gets right + conclusion This is very tight. Each paragraph does one thing. No filler. The danger is that it might feel rushed — the substance turn arrives at paragraph 4 and then you have five paragraphs to develop it. But for a Substack essay that might be exactly right. Get to the point and then develop it. **Option B: The phenomenological version (11–12 paragraphs)** Same as A, but after the oil/kerosene paragraph, insert a paragraph of direct phenomenology: what it's actually like to work with this substance. Not Frippertronics, not an analogy borrowed from music, but your own experience described in your own terms. Something along the lines of: you write a half-formed thought, the system sends back something shaped by its own tendencies, you reject most of it but one move gives you something to push against, and the next thought is sharper for having had that resistance. Framed explicitly as: this is what it's like to explore a substance's capacities. This option replaces the borrowed vividness of Frippertronics with the direct vividness of your own practice. It also makes the essay more personal, which suits Substack. **Option C: The synthetic-substances version (12–13 paragraphs)** This version leans into the history of synthetic materials. After the oil reversal, add a paragraph about Bakelite, nylon, or silicon — substances that were synthesized with one purpose in mind and turned out to have capacities that were discovered only later, through exploration and use. The point is that the history of technology is full of cases where a new material was mismeasured by its early applications. This gives the kerosene-phase thought more support and makes the substance claim feel less metaphorical, more historically grounded. The risk is that this makes the essay a bit tech-history-heavy. But one paragraph of concrete historical example might be worth it if it keeps "substance" from floating free. **Option D: The minimal-medium version** Don't cut medium entirely; demote it to one sentence. The substance paragraphs do the heavy lifting, and at the point where you'd normally introduce medium, you instead say something like: "When a substance is taken up into a practice, it becomes a medium — but that is a thought I want to develop elsewhere. For now, the point is about the substance itself." This acknowledges the direction of travel, signals that there's more to say, and leaves the door open for a sequel. It costs you one sentence and gains you an intellectual promissory note that readers might find intriguing. **Option E: The substance-plus-phenomenology version (my speculation about what might work best)** This one tries to replace the work that medium and Frippertronics were doing without using either concept. After the substance turn + oil reversal + material profile, you'd have a paragraph that describes the recursive practice directly but frames it entirely in substance language. "Working with this substance is not like using a hammer, where the intention precedes the action and the tool merely executes. It is more like working with a reactive material — something that has its own tendencies, its own characteristic forms of pushback. You send language in; language comes back, but shaped by the substance's learned patterns. Sometimes flatter, sometimes unexpectedly connective, sometimes productively wrong. The next move is always a response to what came back, not just a continuation of what you intended. A pen does not do this. A search engine does not do this. This substance does." That paragraph does the work of the pen comparison, the recursive-return point, and the medium claim, all in substance language, without needing Fripp or Wollheim. ### Things worth thinking about The "accidentally" in "accidentally synthesized." This is doing interesting work. It implies that the designers were trying to make one thing (a chatbot, a text predictor) and what they got has properties they didn't fully anticipate. That resonates with actual LLM history — the "emergent abilities" discussion, the way capabilities appeared at scale that weren't designed in. Worth keeping. The material profile question. Codex's paragraph 9 listed LLM properties (pattern-completion, contextual pliability, etc.). On the substance-only version, some version of this is needed — you have to say what properties the substance has, not just that it has properties. But it has to read as observation about working with the stuff, not as a technical spec. The difference between "this substance tends to smooth out specificity, to find generic paths through a topic, to connect things associatively rather than logically, and to respond dramatically to how it's asked" and "LLMs exhibit pattern completion, contextual sensitivity, and token-level prediction." The first is material knowledge, the kind of thing a craftsperson would say. The second is mechanism. Whether you need the artefact/working-surface distinction at all. On the substance-only version, the distinction can be compressed into a few sentences within the phenomenology-of-use paragraph: "Most of what I've described does not result in a published text. The point is not to produce an artefact but to explore the substance — to test a distinction, discover what I think, find out what comes back when I push in a particular direction." That's enough to answer Hoel's evidence without needing a standalone paragraph. Whether the Aristotelian potentiality language earns its keep. I think one sentence does — something about latent capacities (dynamis) being actualised through use. It gives the substance claim philosophical depth without turning the essay into a seminar. More than one sentence risks scholasticism. Whether there's a danger of the essay sounding like it's *about* LLMs but *for* philosophers. The substance language is accessible. The Aristotle can be kept light. The oil/kerosene reversal is concrete. But you'd want to watch the tone — make sure it reads as an argument about a live topic, not as a conceptual exercise. The "accidentally synthesized" framing helps with this because it sounds like someone describing something that really happened, not someone deploying a category from the history of philosophy. ### What I'd want to flag honestly I'm not sure whether "substance" is the final word or a staging post. It does more than "medium" alone, and it does it more simply. But there might be a version of the thought that's even simpler — something like: "We don't yet know what kind of thing this is, and Hoel's argument requires that we do." That's the epistemic claim stripped of all metaphor. The substance language gives it concreteness and explanatory power. But if it starts to feel like you're replacing one classification (tool) with another (substance) and the reader wonders whether you've just renamed the problem, you might need a sentence that says: the point is not that "substance" is the final answer, but that it captures the *kind of uncertainty* we're in — the uncertainty of someone confronted with a new material, not the uncertainty of someone holding a broken tool. I'm speculating there — I don't know whether readers would feel that worry. But it's the strongest objection I can imagine to the substance-only version, and it's worth having an answer to it in reserve. *Una sostanza nuova non si giudica dai primi usi che se ne fanno — si giudica, se mai, dopo che qualcuno ha imparato a lavorarla.* ### User "That's interesting. "What if we've accidentally synthesized a new substance" actually gets a lot of mileage from the "accidentally" — it implies that the designers were aiming at something (a chatbot, a text predictor, an assistant) but what they produced has properties and capacities that exceed or differ from the design brief."the accidently of it it all isn't important. what i mean is that substances (natural or synthetic) can have latent properties that we do not know how to invoke immediately. we cannot say the same for tools –there are no such things as natural tools and the ones that we create do not have latent, hidden at first, properties. ### Assistant OK, let me sit with this correction carefully because it changes the shape of the argument in a way I need to track. Nick is saying: the "accidentally" isn't the point. The point is about the difference between substances and tools with respect to latent properties. Let me make sure I understand what he's actually claiming. Substances — whether natural or synthetic — can have properties that are not immediately apparent and that we discover over time through exploration, experimentation, and practice. Iron has properties we didn't know about when we first encountered it. Nylon has properties DuPont didn't fully anticipate. Oil had latent capacities that were drawn out over decades. The point is not that these substances were *accidentally* created or *accidentally* discovered. The point is that *substances as a category* are the kind of thing that can have latent, initially hidden properties. That's just what substances are like. Tools, by contrast, don't work this way. And Nick is making a sharper claim than I initially registered: there are no natural tools. Tools are created. They are made for purposes. And because they are made for purposes, they don't have latent, hidden properties in the way substances do. A hammer doesn't have undiscovered capacities waiting to be drawn out. You can use a hammer in unexpected ways (as a paperweight, as a weapon), but those aren't latent properties of the hammer — they're just alternative applications of properties that are already fully apparent. The hammer's hardness, weight, shape — these are all on the surface. There's nothing hidden. That's actually a much cleaner distinction than what I was working with. Let me think about whether it holds up... Is it true that tools don't have latent properties? Let me think of counterexamples... A computer? Well, Nick might say a computer is precisely the kind of thing that challenges the tool category — which is partly the point of the essay. A computer does have capacities that exceed its original design brief. But then maybe that's evidence that a computer is not purely a tool, or not only a tool. A Swiss Army knife? No — its capacities are all designed in and immediately apparent. You can see the blade, the corkscrew, the file. There's nothing hidden. What about a simple lever? Could you discover new things about how a lever works? You could discover new applications of leverage, but the lever's properties — rigidity, length, fulcrum point — are all manifest. There's no hidden property of the lever waiting to be drawn out. What about a material that was initially used as a tool but turned out to be more? Like... glass? Glass was first used as a decorative material (beads, vessels). Then lenses were discovered. Then optical instruments. Then fiber optics. But wait — glass is a substance, not a tool. A glass bead is a tool/artefact; glass itself is a substance. And it's the substance that has the latent properties (optical transparency, ability to be drawn into fibres, refractive index), not the artefact. So the distinction Nick is drawing is between: - Substances: have latent properties that are discovered over time; this is constitutive of what it means to be a substance - Tools: are made for purposes; their properties are determined by design and are immediately apparent; they don't have latent, initially hidden properties And the claim about LLMs is: they behave more like substances than tools in this respect. They have properties and capacities that are being discovered through exploration — chain-of-thought, few-shot learning, in-context learning, emergent abilities at scale, the way they respond to iterative prompting. These weren't designed in (or at least, weren't anticipated as part of the design brief). They're latent properties of the stuff that have been drawn out through use and experimentation. Now, this is interesting because it removes the need for "accidentally" to do any work. The argument doesn't depend on whether the designers were surprised. It depends on the ontological distinction between substances and tools. Substances have latent properties; tools don't. LLMs have latent properties; therefore they're more like substances than tools. Actually, wait. Let me push on this a bit more. Is it really true that the latent-property distinction maps perfectly onto substance vs. tool? Or is it more precisely about natural kinds vs. artefacts? Hmm, but Nick explicitly said "natural or synthetic" — so he's not restricting substance to natural kinds. Nylon is a synthetic substance and it has latent properties. So the distinction is not natural vs. artificial. It's substance vs. tool. What makes something a substance is that it's a *stuff* with properties, and those properties can include ones that aren't immediately apparent. What makes something a tool is that it's made for a purpose, and its relevant properties are determined by and apparent from its design. Now I want to think about what this does to the essay structure. If the point is not about accidental discovery but about the ontological difference between substances and tools with respect to latent properties, then the argument gets sharper. The essay doesn't need to tell a story about surprise or unintended consequences. It needs to make a philosophical distinction about kinds of things and how we come to know their properties. The kerosene-phase thought still works, but its function changes slightly. It's not illustrating that oil's uses were *accidentally* discovered. It's illustrating that substances, as a category, reveal their properties over time through exploration. Oil was always capable of powering an internal combustion engine. That capacity was always there, latent in the substance. It just took time and experimentation to find it and exploit it. The kerosene phase is not about accident — it's about the temporal structure of how we come to know what a substance can do. This is cleaner. And it makes the anti-Hoel argument more precise. Hoel's mistake is not that he's failed to notice an accident. His mistake is that he's treating LLMs as tools — things whose properties are determined by design and immediately apparent — when they might be substances — things whose properties include latent capacities that are still being discovered. If they're tools, then what they do now is a fair basis for evaluation. If they're substances, then what they do now might be kerosene — a genuine but limited expression of latent capacities that have not yet been fully drawn out. Now, what does this mean for the "no natural tools" observation? That's a genuinely interesting philosophical point and I want to think about whether it belongs in the essay. "There are no natural tools." What does this mean exactly? It means that tools are always artefacts — always made by someone for some purpose. A rock is not a tool until someone picks it up and uses it as one. And even then, the rock-as-tool is understood in terms of the purpose to which it's being put, not in terms of the rock's intrinsic properties. The rock has intrinsic properties (hardness, weight, shape) — those are substance-level properties. But the *tool* "hammer-rock" is defined by the use, not by the stuff. This is actually a really nice point. It draws the substance/tool distinction at the right joint. Substances have intrinsic properties. Tools have functions. And the essay is claiming that LLMs have intrinsic properties (latent capacities) rather than (or in addition to) functions. So they should be investigated like substances, not evaluated like tools. Does this mean the essay should include a brief statement of the "no natural tools" point? I think it could work as a one-sentence clarification: "Substances, unlike tools, are not brought into existence for a purpose — even synthetic ones. They have properties, and from those properties uses emerge. Tools are defined by the other direction of fit: the purpose precedes the properties." Hmm, but is the "no natural tools" claim strictly correct? What about an ant using a leaf as a tool? Or a chimpanzee using a stick? These are natural objects used as tools. But Nick's point would be that the *toolhood* is imposed by the user — the stick becomes a tool when the chimp uses it to extract termites. The stick's substance properties (length, flexibility, diameter) were always there. The toolhood is added. So even "natural tools" are really natural substances pressed into service as tools. The substance pre-exists the tool. That's actually an even deeper point than the one I initially heard. It's not just that tools and substances are different categories. It's that substance is the more basic category. Everything that's a tool is also a substance (it's made of stuff), but not everything that's a substance is a tool (lots of stuff has no function). And when we encounter a new substance, the right first move is to investigate its properties, not to assign it a function. OK, and now the LLM connection becomes very clear. When Hoel treats LLMs as tools and asks "are they good at their job?", he's skipping the prior question: "what kind of stuff is this, and what are its properties?" That prior question hasn't been settled yet. So evaluating LLMs by their current prominent use (writing assistance) is like evaluating oil by its performance as lamp fuel. It's not wrong — oil does work as lamp fuel — but it's measuring a substance by an early, partial, and potentially non-central application. Now I want to think about what happens to the essay's middle section if we go with this framing. The earlier sessions had a lot of material about recursive practice, the pen comparison, artefact vs. working surface. Does that stuff survive the shift from "medium" to "substance with latent properties"? The pen comparison survives easily. A pen is a paradigmatic tool. It has a clear function (making marks), fully apparent properties (point, ink, grip), and no latent capacities. An LLM is not like that. The contrast between pen-as-tool and LLM-as-substance is sharp and useful. Recursive practice... this can be reframed. Instead of saying "the medium pushes back," you can say "one of the substance's latent properties is that when you engage with it through language, it returns language shaped by its own tendencies, and this return reshapes what you do next." That's a description of a capacity, not a description of a medium. It's something the substance *does*, not something a practice *is*. The emphasis shifts from practice to property. Artefact vs. working surface... this maps naturally onto the distinction between evaluating a substance by its products and investigating a substance by working with it. Hoel evaluates by products (published writing). But much interesting LLM engagement is investigative — people working with the substance to find out what it can do, not to produce polished artefacts. This doesn't need "medium" at all. Slop/Lucier... also maps cleanly. The substance has characteristic tendencies (smoothing, averaging, genericising). When those tendencies are allowed to run without resistance, the result is slop. That's a substance-level observation. Lucier's piece illustrates what happens when a physical system's characteristic resonant properties dominate: speech degrades into something determined by the room, not by the speaker. Same logic: LLM-generated text degrades into something determined by the model's tendencies, not by any human intention. That's slop. And it's a prediction of the substance framing, not a problem for it. So actually, I think almost everything survives the cut except Frippertronics and the *Growing the Image* theoretical apparatus. And the essay becomes leaner and more focused. The question is whether it becomes *too* lean — whether there's enough material for a satisfying essay. Let me sketch a possible structure: 1. Fixed opening + Hoel quote 2. Pressure on "tool" — proper function 3. Candidates that fail — what is an LLM for? 4. Turn: substances and tools are different kinds of things with respect to latent properties 5. No natural tools / tools defined by purpose / substances defined by properties and capacities 6. LLMs have latent properties: chain-of-thought, in-context learning, emergent abilities, etc. 7. Oil reversed: kerosene phase — not about accident, about the temporal structure of discovering what a substance can do 8. What this substance does when you engage with it — pen contrast, recursive return 9. Artefact vs. working surface — Hoel measures products, but much interesting engagement is investigative 10. Slop as characteristic tendency (Lucier) 11. Conclusion: we have synthesized a new substance and are still learning what it can do; Hoel's evaluation presupposes a settled understanding we don't yet have That's eleven moves. Still tight. And every move does distinct work. But I want to think about something else. The "no latent properties for tools" claim — is this doing *philosophical* work or is it doing *rhetorical* work? If it's doing philosophical work, it needs to be stated carefully and defended against obvious objections. If it's doing rhetorical work, it just needs to be suggestive enough to shift the reader's perspective. One objection: couldn't you say that a computer has latent properties? When you buy a laptop, it has hardware capabilities you haven't yet discovered or used. Running novel software on it reveals new capabilities. But this is really a substance-level observation about the hardware and software, not a tool-level observation about the computer-as-implement. The computer-as-tool (email machine, word processor) has a clear function. The computer-as-substance (silicon, circuits, architecture) has latent capacities. So even here the distinction holds: what's latent is the substance, not the tool. Another objection: couldn't you say that a hammer reveals unexpected properties when you swing it in a new context? But those aren't *latent* properties — they're the same manifest properties (mass, hardness) doing manifest things in a different setting. Latent properties are properties that you didn't know the thing had. A hammer doesn't turn out to have magnetic properties you hadn't noticed. Oil turns out to be combustible in engines in a way that nobody initially anticipated. I think the distinction is philosophically sound. Substances have latent properties in the sense that investigation can reveal properties that were not previously known or anticipated. Tools have manifest functions and their relevant properties are determined by design. The distinction tracks something real about how we come to know things. Now, let me think about a wrinkle that Nick's correction introduces. If the "accidentally" isn't doing work, then the essay can't lean on stories about unintended consequences or surprising discoveries. What it can lean on is the structural point about substances having the *kind of properties* that can be latent. That's a claim about kinds, not about particular historical episodes. This is actually philosophically stronger but rhetorically harder. Historical stories are vivid (kerosene! Bakelite! penicillin!). Category distinctions are abstract. The essay needs to find a way to make the category distinction vivid without relying on accident narratives. One way: the kerosene-phase story still works, because it's not about accident — it's about temporal lag in discovering what a substance can do. Oil wasn't accidentally useful as engine fuel; it always had that capacity. We just hadn't learned how to invoke it yet. The word "accidentally" isn't needed. "We hadn't yet learned how to invoke its latent capacities" is the right way to put it. Another way: describe specific latent properties of LLMs that have been discovered through exploration. Chain-of-thought prompting is the best example. Nobody designed GPT-3 to be better at reasoning when prompted to think step by step. That was a latent property of the substance that was discovered through experimentation. Same with few-shot in-context learning. Same with the way these models respond to persona prompts, to structured formatting, to iterative refinement. These are all cases where the substance turned out to have capacities that weren't designed in and weren't initially apparent. They were drawn out through use. That's concrete, LLM-specific, and doesn't require the word "accidentally." It just requires the observation that these capacities were latent and were discovered through exploration rather than deduced from a design brief. Hmm, but I want to be careful about how much I claim here. The researchers who discovered chain-of-thought prompting didn't think of themselves as discovering a latent property of a substance. They thought of themselves as finding a better prompt. And many AI researchers would describe these as emergent properties of the architecture, not latent properties of a substance. Nick's essay is proposing a reframing — seeing these discoveries *as if* they were discoveries about a substance. That's a philosophical move, not a straightforward empirical observation. The essay needs to be clear that it's offering a category and then showing that the category fits, not reporting a fact that everyone already agrees on. This connects to the epistemic discipline requirement. The substance claim is Nick's interpretation, not something Hoel says or something the AI research community says. It needs to be presented as such. "I want to suggest that..." or "What if we think of..." — not "LLMs are substances" as though it were a settled fact. Wait, but Nick has committed to this framing. He's past the tentative stage on the substance claim itself — what he's working out is the essay structure, not whether the claim is right. So the essay can be assertive about the substance framing while still being epistemically careful about what follows from it. Let me think about one more thing. The claim "there are no natural tools" is interesting because it gestures at the deeper philosophical point about the direction of fit between properties and purposes. For substances, properties come first and purposes emerge from them. For tools, purposes come first and properties are selected/designed for them. LLMs were designed (purpose first), but their properties exceed the design brief in ways that are discovered through exploration (properties emerging independently of purpose). So LLMs are in a weird intermediate position — artefacts that behave like substances. Made for a purpose, but with capacities that transcend the purpose. That's actually a really interesting philosophical observation. It might be worth a sentence in the essay: "LLMs were designed, as tools are; but what they turned out to be has exceeded the design brief, as substances do. They are artefacts that behave like discovered stuff." Or is that too clever? Maybe. But it captures something genuinely interesting about the ontological status of LLMs: they sit at the boundary between the made and the found. They were made, but what we're now doing with them is more like finding out what a newly encountered substance can do than like using a tool we built. OK, I think I've explored this enough. Let me move toward a synthesis. Nick's correction sharpens the argument in several ways: 1. It removes the dependence on "accident" or "surprise" and makes the claim structural: substances have latent properties as a category; tools don't. 2. It deepens the substance/tool distinction by noting that there are no natural tools — tools are always artefacts defined by purpose, while substances (natural or synthetic) are defined by properties that can include latent ones. 3. It makes the anti-Hoel argument more precise: Hoel evaluates LLMs as tools (by their performance at a function), but if they're substances, the right approach is investigation (discovering what their properties are), not evaluation (judging how well they fulfil a function). 4. It keeps the kerosene-phase thought but reframes it: not about accidental discovery, but about the temporal structure of discovering what a substance can do. Oil always had the capacity to power engines; we just hadn't learned how to invoke it yet. 5. It suggests that the essay doesn't need "medium" because the substance/tool distinction already does enough work, and the recursive practice, pen comparison, artefact/surface distinction, and slop/Lucier material can all be reframed in substance terms. The one thing I want to flag is a potential vulnerability. If someone pushes back and says "but LLMs were designed, so they're artefacts, so they're tools," the essay needs an answer. And the answer is: yes, they were designed, but their properties include latent ones that were not anticipated by the design. They're artefacts that have substance-like characteristics. The design gave them some properties, but exploration has revealed others. That makes them different from hammers, Swiss Army knives, and other tools whose properties are fully determined by design. This is where the essay might want to briefly mention that synthetic substances are a real category — nylon, carbon fibre, various alloys — where the stuff is manufactured but its full range of properties is discovered through subsequent investigation and use. LLMs might belong in this category. Not a natural substance, but not a mere tool either. You're right, and the correction sharpens the argument considerably. Let me lay out what I think changes and what it opens up. ## What the correction does The point isn't about surprise or unintended consequences. It's about a structural difference between two kinds of things. Substances — whether you dig them up or synthesize them — can have latent properties. Properties that are there in the stuff but that you don't know how to invoke, haven't yet discovered, haven't yet found the right conditions to draw out. That's just what substances are like. It's constitutive of the category. Tools don't work this way. And the reason they don't is revealing: there are no natural tools. Every tool is an artefact. Every tool is made for a purpose. The purpose precedes the properties — or rather, the properties are selected and shaped by the design brief. A hammer doesn't have undiscovered capacities lurking in its head. You can use a hammer for purposes other than hammering, but those aren't latent properties — they're the same manifest properties (mass, hardness, surface area) doing manifest things in a different setting. There's nothing hidden. The direction of fit runs the other way for substances. Properties come first; purposes emerge from them. Nobody looked at a lump of iron and said "this will make good bridges" — they discovered iron's tensile strength, its malleability, its ability to alloy with carbon, and from those discoveries, uses emerged. The properties were always there, latent in the stuff. The uses came later, sometimes much later, as people learned how to invoke what was already present. ## What this does to the LLM argument If LLMs are tools, then their relevant properties are determined by their design and are in principle apparent from the design brief. What they do now is a fair basis for evaluation, because tools don't have hidden depths. You assess a tool by asking whether it does its job well. If LLMs are substances, then their properties include latent ones — capacities that were not anticipated by the design brief and that are being discovered through exploration and use. Chain-of-thought prompting is a good example. Nobody designed GPT-3 to reason better when asked to think step by step. That was a latent property of the stuff — something that was there in the model's learned patterns and that was drawn out through experimentation. Same with in-context learning, few-shot generalization, the way these models respond to persona prompts, to structured formatting, to iterative refinement. These are discoveries about what the substance can do, not instructions to a tool about what it should do. That distinction — discovery vs. instruction — is what makes the substance framing philosophically productive, not the question of whether anyone was surprised. ## How the kerosene-phase thought changes The kerosene thought still works, but it's no longer a story about accidental discovery. It's a story about the temporal structure of coming to know what a substance can do. Oil was always capable of powering an internal combustion engine. That capacity was always latent in the substance. It wasn't that people stumbled upon it by accident — it's that it took time, experimentation, and the development of new technologies and practices before that particular capacity could be invoked. The kerosene phase wasn't a mistake or a failure of imagination. It was an early, genuine, but limited expression of capacities that had not yet been fully drawn out. The analogy to LLMs: what we currently do with them — chatbots, writing assistance, code completion, summarization — may be kerosene. Real uses of real properties, but not yet the full expression of what this substance can do. And the point is not predictive (I'm not promising that GPT-for-engines is coming). The point is methodological: Hoel's evaluative confidence presupposes that we already know what the substance's properties are and which of them matter. The substance framing says that this question is still open. ## Options for what this opens up in the essay Here are several directions this could go. I'm presenting them as parallel options, not ranked: The artefacts-that-behave-like-substances thought. LLMs were designed — they're artefacts, like tools. But what they turned out to be has exceeded the design brief in ways that are discovered rather than specified. They're made things that behave like found things. That's a genuinely unusual ontological position: manufactured substances. Nylon, carbon fibre, and various alloys are precedents — stuff that was synthesized for one purpose but turned out to have properties that were discovered through subsequent investigation. The essay could note this without making it a whole section — just a sentence that acknowledges the wrinkle and moves on. The material-profile paragraph. If substances have latent properties, then part of the essay's job is to say *which* properties this substance has — both the ones we've already found and the ones that might be latent. Pattern-completion, contextual pliability, associative reach, strong smoothing and genericising tendencies, responsiveness to framing and persona, the capacity to generate structurally coherent text at scale. Those are known properties. The latent-properties claim says there may be others. The essay doesn't need to speculate about what they are — it just needs to establish that the substance is the kind of thing where further properties could exist. The evaluation question. If a substance has latent properties, then evaluating it by its current prominent uses is like evaluating oil by kerosene. It's not wrong — oil does work as lamp fuel, and LLMs do help with emails — but it's measuring a substance by one early application. The right question is not "how good is this tool at its job?" but "what can this stuff do, and have we found all of it yet?" That reframing is what makes Hoel's "bits in, bits out" look premature without looking wrong. The slop-as-tendency observation. This gets stronger under the latent-properties framing. If the substance has characteristic tendencies — smoothing, averaging, genericising — then of course large-scale unresisted use will produce mountains of generic output. That's not evidence that there's nothing there. It's evidence about what happens when a substance's tendencies are allowed to run without skilled resistance. Lucier's room is the analogy: the room has characteristic resonant frequencies (substance-level properties), and when speech is passed through it repeatedly without intervention, those frequencies take over. The result reveals the substance's tendencies, not the substance's limits. The "no hidden properties for tools" claim as a standalone philosophical point. This could be a paragraph in its own right or just a few sentences within the substance-turn paragraph. The thought is: tools are epistemically transparent in a way substances are not. When you hold a hammer, you know what it can do. Its capacities are apparent from its design. Substances aren't like this. You can hold a piece of silicon and have no idea that it can be used to make a transistor. That hidden-properties feature is what makes substances epistemically interesting, and it's what makes premature evaluation of a substance dangerous in a way that premature evaluation of a tool is not. You can't be wrong about what a hammer is for, but you can be wrong — or more precisely, premature — about what a substance is capable of. Whether to mention the designed-for/emerged-from distinction explicitly. The essay could have a brief moment where it notes that there are two directions of fit: for tools, function precedes properties (we make the hammer to hammer); for substances, properties precede function (we discover what iron can do, and then we use it for that). LLMs were made with a function in mind (predict the next token, chat, assist), but what they can do exceeds what they were made for. That excess is what the substance framing captures. It's not an accident — it's what happens when you manufacture something complex enough that its properties outrun your design brief. Whether to keep the pen comparison. I think yes. The pen is a paradigmatically transparent tool. Its properties are apparent. It has no latent capacities. It extends your ability to inscribe; it doesn't push back, surprise, or reveal hidden powers. The contrast with LLMs — which do push back, do surprise, and do reveal properties through use — is sharpened by the latent-properties framing. The pen is the essay's clearest example of what a tool is, and its poverty as a model for LLMs is part of the argument. Whether to include the Aristotelian language at all. "Latent capacities" is already doing the work that "dynamis" would do, and it does it in English. One option: use the Aristotelian language in a light parenthetical — "(what Aristotle would have called *dynamis* — potentiality waiting to be actualised)" — to give the philosophically literate reader a foothold without asking the general reader to learn Greek. Another option: skip it entirely and let "latent capacities" speak for itself. My sense is that the lighter version works better in a Substack essay than either the full Aristotelian treatment or total omission. Whether the essay needs a positive vision of what LLM substance-use looks like. Under the old structure, Frippertronics and the recursive-practice material provided this. Under the substance-only structure, you'd need something else — either a direct phenomenological description of working with the substance, or a few concrete examples of latent capacities being drawn out. The chain-of-thought prompting example could serve here: "Nobody told the model to reason step by step. Someone tried it, and it turned out the substance could do that. That is what discovery looks like — not designing a new function, but finding a latent capacity in an existing material." Whether the essay needs the artefact/working-surface distinction. I think it does, because it answers Hoel's evidence. He points to published writing (artefacts) and says quality hasn't improved. The reply: much of what competent users do with this substance is not aimed at producing artefacts. It's aimed at exploring the substance — testing a distinction, finding out what comes back when you push in a particular direction. Evaluating a substance by its manufactured products misses the exploratory engagement that reveals its properties. This can be brief — a few sentences, not a paragraph — but it needs to be there. ## A possible compressed structure I'm offering this as one way the moves might sequence, not as the only way: 1. Fixed opening + Hoel quote 2. Pressure on "tool" — if this is a tool, what is it for? (Proper function + candidates that fail, compressed) 3. First verdict — "a quite different type of tool, or not quite a type of tool at all" 4. The turn: substances have latent properties; tools don't. No natural tools. Properties precede purposes for substances; purposes precede properties for tools. 5. LLMs have latent properties — chain-of-thought, in-context learning, etc. as discoveries about a substance, not instructions to a tool 6. Hoel's oil reversed — kerosene phase, not as accident but as the temporal structure of discovering what a substance can do. Anti-hype built in: this is epistemic, not predictive. 7. What this substance does when you engage with it — pen contrast, recursive return, the fact that artefact-level evaluation misses exploratory engagement 8. Slop as characteristic tendency — Lucier — the substance's tendencies dominating under conditions of unresisted use 9. Conclusion — we have synthesized a new substance whose latent capacities are still being discovered; Hoel's "bits in, bits out" may describe the kerosene phase, not the nature of the thing Nine paragraphs after the fixed opening. Tight. Each one does distinct work. Nothing ornamental. The medium concept doesn't appear, the Frippertronics don't appear, but nothing that those were doing is left undone — it's just done in substance language instead. The remaining uncertainty I'd flag: the latent-properties claim about LLMs is strong and defensible, but it's an empirical claim as much as a philosophical one. Chain-of-thought prompting, in-context learning, emergent abilities — the AI research community is divided on whether these represent genuine "latent" capacities of the architecture or artefacts of training data, scale, and evaluation methodology. The substance framing doesn't need that debate to be settled — it just needs to be true that we are still discovering what these systems can do, which is observationally obvious. But the essay should be aware that "latent properties" is a philosophical interpretation of the observations, not a neutral description of them. *Le proprietà latenti non si dichiarano — si scoprono lavorando la materia, e il lavoro non è ancora finito.* ### User "Tools don't work this way. And the reason they don't is revealing: there are no natural tools. Every tool is an artefact. Every tool is made for a purpose. The purpose precedes the properties — or rather, the properties are selected and shaped by the design brief. A hammer doesn't have undiscovered capacities lurking in its head. You can use a hammer for purposes other than hammering, but those aren't latent properties — they're the same manifest properties (mass, hardness, surface area) doing manifest things in a different setting. There's nothing hidden." this was not the point i was making. i was just saying that tools don't have discoverable but at first unknown properties for two reasons,: 1. they are not natural objects. 2. it doesn't ake sense to say that designed/manufacured objects have hidden properites ### Assistant OK, let me slow down and actually listen to what Nick is saying rather than elaborating my own version of it. He's making a simpler, cleaner point than what I attributed to him. Let me try to state it precisely. The claim is: tools don't have discoverable-but-initially-unknown properties. And there are two reasons for this: 1. Tools are not natural objects. 2. It doesn't make sense to say that designed/manufactured objects have hidden properties. So the point is not about "direction of fit" or "purpose precedes properties" or any of the philosophical machinery I was building. It's a more straightforward observation. Let me think about what these two reasons actually amount to. Reason 1: Tools are not natural objects. Natural objects — iron ore, oil, silicon, wood — exist independently of human purposes. Because they exist independently of us, they can have properties we haven't yet encountered or understood. That's just what it means for something to exist independently: its nature is not exhausted by what we currently know about it. There can always be more to discover. Natural objects have a kind of depth that comes from their independence from human design. Tools, by contrast, are made. They're artefacts. They don't exist independently of human purposes — they exist because someone made them for a reason. So reason 1 is really saying: only things that exist independently of us can have genuinely hidden properties, because only independent things have a nature that exceeds our current knowledge of them. Reason 2: It doesn't make sense to say that designed/manufactured objects have hidden properties. This is the stronger claim and I think it's the one that does the real work. If you design and manufacture something, you determine its properties. You choose the materials, the shape, the mechanism. The properties of the finished object are a consequence of your design decisions. So in what sense could there be properties you don't know about? You made the thing. Its properties are a product of your decisions. Now, someone might object: "But what about unintended side effects of design? What about emergent properties of complex systems?" And Nick's response (if I'm reading him right now) would be: those might exist, but they're not the point. The point is that the *concept* of a tool doesn't allow for hidden properties in the way the *concept* of a substance does. It's a conceptual point, not an empirical one. Wait, but actually — and I need to be careful here — Nick is making a point about what makes sense to say, not about what's empirically true. "It doesn't make sense to say that designed/manufactured objects have hidden properties." That's a claim about the logic of the concepts, not about what happens in practice. In practice, complex manufactured systems sometimes do exhibit unexpected behaviours. But calling those "hidden properties" would be a category mistake, because "hidden properties" is the kind of thing that applies to natural substances, not to artefacts. Hmm, but does this fully hold? Let me think of cases... A pharmaceutical drug is designed and manufactured. But it can have side effects that weren't anticipated — properties that emerge from the interaction between the molecule and biological systems. Are those "hidden properties" of the drug? In ordinary language, we might say yes — "the drug turned out to have properties we didn't know about." But Nick's point might be that the drug-as-substance (the molecule) has those properties, not the drug-as-tool (the medication designed to treat a condition). The drug-as-tool was designed to lower blood pressure. The molecule it's made of turned out to also affect liver function. The hidden property belongs to the substance (the molecule), not to the tool (the medication). That's actually a very clean way of putting it, and it might be what Nick is driving at. When we discover unexpected properties of a manufactured thing, we're really discovering properties of the substance it's made of, not properties of the tool it was designed to be. The tool is defined by its design brief. The substance it's made of has its own nature. So the argument for LLMs would be: LLMs were designed as tools (chatbots, text predictors, assistants). But they turned out to have properties — chain-of-thought reasoning, in-context learning, etc. — that belong not to the tool (the chatbot) but to the substance (whatever the trained model actually is as a material with capacities). The interesting thing about LLMs is that the substance seems to have more going on than the tool-level design anticipated. And that's a reason to investigate them as substances rather than evaluate them as tools. But wait — I'm elaborating again. Let me come back to what Nick actually said. He said two things: 1. Tools are not natural objects. 2. It doesn't make sense to say designed/manufactured objects have hidden properties. And his correction was that I was over-complicating it with "direction of fit" and "purpose precedes properties." He just wants the simpler observation stated clearly. Let me try to state it as simply as I can: Substances can have latent properties because they exist independently of our purposes — there can always be more to discover about them. Tools can't have latent properties because they are designed and manufactured — their properties are determined by the design. There's nothing there to discover because we put everything there that's there. That's the two-reason version stated as simply as I can. And the LLM argument would then be: LLMs seem to have latent properties (capacities discovered through use that weren't part of the design specification). If that's right, then they're more like substances than tools. And if they're more like substances, then Hoel's evaluative approach (judging them by how well they perform their designed function) is the wrong approach. Is there anything more I should think about here? I think the interesting question is whether Nick wants this two-reason argument stated in the essay as such, or whether it's just background for the substance claim. In the essay, it might be enough to say: "Substances can have latent properties; tools can't" and let the reasons be implicit. Or it might be worth one sentence: "Tools are made — their properties are determined by design, and there's nothing there to discover that wasn't put there. Substances are different: they exist with properties that can exceed what we currently know about them." Actually, I realize I should also think about whether Nick's claim is that this applies to ALL manufactured objects or specifically to tools. Because there's a difference. A manufactured object that's simple enough (a hammer, a screwdriver) clearly has no hidden properties. But a manufactured object that's complex enough (a computer, a chemical compound, a neural network) might have properties that weren't anticipated by the designers. But I think Nick's point is that *qua tool*, the object doesn't have hidden properties. The hammer *qua hammer* has no hidden properties. If the hammer turns out to have some unexpected physical property (maybe the metal alloy resonates at a frequency that repels mosquitoes — silly example), that's a property of the substance it's made of, not of the hammer-as-tool. The tool concept doesn't have room for hidden properties. The substance concept does. And that's why the question "is an LLM a tool or a substance?" matters. If it's a tool, hidden properties don't make sense — what it does is what it was designed to do. If it's a substance, hidden properties are exactly what you'd expect — and the question becomes what else it can do that we haven't found yet. OK, I think I understand the correction now. Let me respond simply, acknowledge that I over-complicated it, and state the point as Nick intended it. Right. I was overbuilding. The point is simpler than what I made of it. Two reasons tools don't have latent properties: 1. They're not natural objects. There's no independent nature to have more to discover about. A lump of iron exists independently of us — its properties are whatever they are regardless of what we know. A hammer doesn't exist independently of us. We made it. 2. Designed and manufactured objects don't have hidden properties because we determined what's there. We chose the shape, the weight, the material. The properties are consequences of design decisions. There's nothing to discover because we put everything there. Substances are different not because of "direction of fit" or any fancy philosophical machinery — they're different because they can have properties we haven't yet found. Oil had combustion properties before anyone knew how to build an engine. Silicon had semiconducting properties before anyone knew what a transistor was. The properties were there, in the stuff, waiting to be found. That's what latent means — present but not yet drawn out. The LLM argument then becomes: these systems appear to have latent properties — capacities that were not part of the design specification and that have been discovered through exploration. Chain-of-thought prompting, in-context learning, the way they respond to iterative refinement. Nobody designed those in. They were found, the way you find properties in a substance. And if that's what's going on, then evaluating LLMs as tools — asking "how well do they do their job?" — misapplies a framework that assumes there's nothing there beyond what was designed. Whereas the substance framing says: there might be more, and we're still finding out. I was dressing up something clean in unnecessary philosophy of action. The observation itself is sharper without it. *Ciò che è stato progettato non nasconde nulla — ciò che esiste indipendentemente da noi, invece, può sempre riservare qualcosa che non avevamo ancora trovato.* ### User so how woouldthe draft in its current state be restructured so as to fit with this as the central thrust? please don't be simple and throw out all the cool interesting details. ### Assistant Let me re-read the current draft first — I need to know exactly what's there before I can say how to restructure it. ### User [Request interrupted by user] ### User "Actually, I'm realizing the entire draft is in bullet-point form, not polished prose. This is a working document where each "paragraph" is a cluster of bullets, so the restructuring is about rearranging these bullet blocks and potentially adding new ones, not about moving finished paragraphs around."???? why would it not includ emoving? you got cut off, please continue from where you left off. ### Assistant OK, I got cut off mid-thought. Nick is also correcting me — the restructuring obviously can include moving bullet blocks around, not just adding new ones. That was a strange thing for me to say. Of course restructuring means rearranging, cutting, adding, and modifying. Let me not repeat that mistake. Let me go back to what I was doing: working through the current draft paragraph by paragraph and thinking about how each block maps onto the substance-with-latent-properties thesis. I've read the draft. Let me go through it block by block. **Block 1 (lines 12-15): Opening + first Hoel quote.** Fixed. Stays as is. **Block 2 (lines 17-21): Expanding on Hoel's argument + second quote.** This sets up Hoel's case: if LLMs are tools, expect efficiency and scale, not transformation. Writing should be the proving ground. The "stilted emails... dearth of good writing" quote. This block works under the substance framing — it's setting up what the essay responds to. But the bullets at lines 17-18 are rough and need work regardless. The block stays, maybe tightened. **Block 3 (lines 22-24): The pivot.** "What I am less sure of is the description doing the philosophical work" + "Another reason it seems right to say that LLMs are tools, is that it seems right to say that we *use* them." These are transitional bullets that move from Hoel's position to the proper function discussion. They work fine under the substance framing. They stay. **Block 4 (lines 26-31): Proper function introduced.** Hammer, Swiss Army knife, Google. This is the diagnostic step: tools have proper functions. Works perfectly under the substance framing — it's setting up the contrast. Substances don't have proper functions; tools do. This block stays. **Block 5 (lines 33-38): Candidates that fail.** Token prediction, chatting, writing, "helping with tasks." None settles the question. This is the negative result that motivates the substance turn. Stays. **Block 6 (lines 41-45): First verdict.** "A quite different type of tool, or not quite a type of tool at all." Classificatory hesitation. This is the hinge point where the essay could go either way — toward medium (old version) or toward substance (new version). Under the substance framing, this block stays but its role changes: instead of being a pause before "medium" arrives, it becomes the setup for "substance." The "instability at the level of ordinary functional description" line directly motivates the move to a different category. **Block 7 (lines 48-53): Uncertainty about evaluation.** If we don't know what it's for, we can't easily test it. This is methodologically important under the substance framing — it's saying that tool-evaluation presupposes settled function, which is exactly what substances don't have. Stays. Maybe even stronger under the new framing. **Block 8 (lines 56-62): But writing is a good place to start.** Concedes that Hoel has a reason to look at writing. "Words are its womb, its mother, its literal atoms." This block stays — it's fair-minded and sets up the question of *how* to look at writing. Under the substance framing, the question becomes: even if writing is the right place to look, are we looking at it the right way? Now here's where the restructuring gets interesting. In the current draft, blocks 9-11 (artefact/working surface, pen comparison, Hoel's evidence at wrong level) lead into block 12 (medium) and block 13 (Frippertronics). Under the substance framing, the medium block and Frippertronics block would be cut or radically transformed. But what about blocks 9-11? **Block 9 (lines 65-70): Artefact vs. working surface.** Text as finished product vs. text as thinking surface. This distinction doesn't depend on the medium concept at all — it's an observation about different uses of text. It answers Hoel's evidence directly: he measures artefacts, but much interesting LLM use treats text as a working surface. This block stays under the substance framing. **Block 10 (lines 73-78): Expanding on the distinction.** Prompt-response-revision loops aren't aimed at publishing. "Has writing improved?" presupposes that success means better end-products. This deepens block 9. Stays. **Block 11 (lines 81-87): Pen comparison.** A pen extends inscription; an LLM sends material back altered. "A pen does not prompt me back." This is one of the draft's strongest moments. Under the substance framing, the pen comparison works perfectly — it marks the difference between a transparent tool (pen) and something with its own tendencies and capacities (LLM). The pen has no latent properties. The LLM does. Stays, and maybe gets reframed slightly to emphasize that what the LLM sends back reveals properties of the substance — its learned patterns, its tendencies, its characteristic ways of reshaping language. **Block 12 (lines 90-95): Hoel's evidence at wrong level.** "I am not denying the slop... What I doubt is that this settles the character of the system." This is argumentatively important. Stays. Wait, actually I'm misnumbering. Let me re-check the actual flow: Lines 65-70: artefact vs working surface Lines 73-78: expanding that distinction Lines 81-87: pen comparison Lines 90-95: Hoel's evidence at wrong level Lines 98-104: medium (from *Growing the Image*) Lines 107-113: Frippertronics Lines 115-120: slop/Lucier Lines 123-129: conclusion So the blocks that need the most attention are: - Lines 98-104: medium → this is where the substance turn would go instead - Lines 107-113: Frippertronics → this would be cut or heavily trimmed - Lines 115-120: slop/Lucier → this stays but reframed in substance terms - Lines 123-129: conclusion → rewritten to land on the substance claim Now, the question is: where does the substance turn go? In the current draft, "medium" arrives at line 98, after the pen comparison and the "wrong level of description" argument. That's actually quite late — twelve paragraphs in. Under the substance framing, should the turn come earlier? I think it should. Here's why: the proper-function discussion (blocks 4-6) establishes that LLMs don't have a settled proper function. The current draft then spends several blocks (7-11) exploring the consequences of that for evaluation and writing, before finally offering a positive category (medium) at block 12. That's a lot of negative argument before the positive thought arrives. Under the substance framing, the proper-function failure directly motivates the substance claim. If tools are defined by proper function, and LLMs don't have one, then maybe they're not tools. And the reason they don't have a proper function is that they're not the kind of thing that has proper functions — they're substances, not tools. Tools are designed and manufactured; their properties are determined by design; there's nothing hidden to discover. Substances have properties that can be latent — discoverable but not yet discovered. LLMs have latent properties (chain-of-thought, in-context learning, etc.). So maybe they're substances. That argument flows naturally from block 6 (first verdict: "not quite a type of tool at all"). The substance turn could come right after that — as block 7, before the writing/evaluation discussion. That would be a significant restructuring: instead of going proper function → evaluation → writing → artefact/surface → pen → medium, you'd go proper function → substance turn → evaluation → writing → pen → slop → conclusion. But wait — does that work? If the substance turn comes at block 7, then blocks 8-11 (evaluation, writing, artefact/surface, pen) would all be read through the substance lens. The reader would already know that the essay is claiming LLMs are substances, and the subsequent discussion would be about what that means for Hoel's argument about writing. That's actually a better reading experience — the reader knows where they're going, and the writing/evaluation material becomes evidence for the substance claim rather than unsupported observations waiting for a category. Alternatively, the substance turn could come where the medium block currently is (line 98), replacing it. That preserves the current structure's logic: show the problem (proper function fails), show the consequences (evaluation is hard, writing is complex), then offer the solution (substance, not tool/medium). This has the advantage of letting the reader feel the force of the problem before getting the answer. But it has the disadvantage I noted earlier: a lot of negative argument before the positive thought. Hmm, let me think about a third option. What if the structure goes: 1. Opening + Hoel (fixed) 2. Hoel's argument expanded 3. Pivot: "tool" sounds right but... 4. Proper function introduced 5. Candidates that fail 6. First verdict: not quite a tool 7. THE SUBSTANCE TURN — tools don't have latent properties, substances do, LLMs have latent properties 8. Oil/kerosene — Hoel's analogy reversed 9. Evaluation uncertainty — now motivated by the substance claim (if it's a substance, you can't test it like a tool) 10. Writing as test case — concede Hoel has a reason to look here 11. But what are we measuring? — artefact vs working surface 12. Pen comparison — transparent tool vs substance with tendencies 13. Slop/Lucier — characteristic tendency of the substance 14. Conclusion — we've synthesized something whose properties we're still discovering; Hoel's judgment is premature That's fourteen blocks. The substance turn comes at block 7 — right after the proper-function failure motivates it. The kerosene reversal comes at block 8 as the substance thesis's first payoff. Then blocks 9-12 apply the substance framing to Hoel's specific argument about writing. Block 13 handles slop. Block 14 concludes. That feels right. The substance turn is early enough to organize everything that follows, but late enough that the reader has seen why "tool" doesn't work. The kerosene reversal comes immediately after the substance turn, giving the reader a vivid, concrete payoff before the essay goes back into the details of writing and evaluation. Now, what happens to each existing block under this restructuring? Blocks 1-6 (lines 12-45): Stay roughly as they are. The proper-function discussion leads to "not quite a type of tool at all." Block 7 (NEW): The substance turn. This is the essay's pivot. It would need to introduce: - Tools don't have latent properties (two reasons: not natural objects; designed/manufactured objects don't have hidden properties) - Substances do have latent properties — properties that are there in the stuff but not yet discovered or drawn out - LLMs have latent properties — chain-of-thought, in-context learning, etc. were not designed in but discovered through use - So maybe LLMs are substances, not tools Block 8 (NEW): Oil/kerosene. Hoel says discovering LLMs should have been like discovering oil. The reply: oil had latent capacities that took decades to draw out. Kerosene was not what oil turned out to be for. The point is not predictive but epistemic: Hoel's argument assumes we already know what this substance can do, and that assumption is what the substance framing denies. Block 9: Currently at lines 48-53 (evaluation uncertainty). Moves down. Under the substance framing, its argument becomes: if this is a substance with latent properties, then we can't evaluate it the way we evaluate a tool. Tool-evaluation presupposes settled function. Substance-investigation is open-ended. This block gets reframed but its content largely survives. Block 10: Currently at lines 56-62 (writing as test case). Moves down. Stays largely as is — it concedes that writing is a reasonable place to look, but the question is how to look. Block 11: Currently at lines 65-78 (artefact vs working surface, expanded). Could be compressed into one block instead of two. The distinction answers Hoel: he measures artefacts, but much substance-exploration treats text as a working surface. Not aimed at end-products but at discovering what the substance can do. Block 12: Currently at lines 81-87 (pen comparison). Stays. Gets reframed: a pen is a paradigmatic tool — no latent properties, fully transparent. An LLM is not like a pen. What comes back from the LLM reveals the substance's tendencies — its learned patterns, its characteristic ways of reshaping language. Block 13: Currently at lines 115-120 (slop/Lucier). Stays, reframed. Slop is what happens when the substance's genericising tendencies are allowed to run without resistance. Lucier still works as the analogy — the room's resonant frequencies (substance-level properties) gradually take over. The substance framing actually makes this point sharper. Block 14: Currently at lines 123-129 (conclusion). Rewritten. Instead of landing on "bits in, bits out" and the *Homo faber* line, the conclusion lands on: we have synthesized a new substance; we are still discovering its latent properties; Hoel's evaluative confidence presupposes a settled understanding we don't yet have; and measuring kerosene-era products doesn't tell you what the substance is. What gets CUT: - Lines 98-104: The medium block (*Growing the Image*, Wollheim, Thomson-Jones, "structured field of resources and practices"). Gone. - Lines 107-113: Frippertronics. Gone, or reduced to a parenthetical reference. What gets MOVED: - Lines 48-53 (evaluation) and 56-62 (writing as test case) move from their current position (after first verdict, before artefact/surface) to after the substance turn and kerosene reversal. What gets ADDED: - The substance turn (new block 7) - The kerosene reversal (new block 8) - Possibly a brief material-profile moment within block 7 — what properties does this substance have? Pattern-completion, associative reach, smoothing tendencies, responsiveness to framing. - Possibly a brief mention of concrete latent-property discoveries (chain-of-thought, few-shot learning) within block 7. What gets REFRAMED: - The pen comparison (block 12): from "compare to a medium" to "compare to a substance with tendencies" - Slop/Lucier (block 13): from "medium with strong generic tendencies" to "substance with characteristic smoothing tendencies" - The conclusion (block 14): from "not quite a tool" to "we've synthesized a substance" Now, let me think about what "cool interesting details" need to be preserved. Nick's instruction was explicit: don't be simple and throw out the good stuff. The pen comparison — stays. It's one of the draft's best moments. "A pen does not prompt me back" — stays. Great line. The Lucier analogy — stays. It works with the substance framing and it's evocative. The "words are its womb, its mother, its literal atoms" Hoel quote — stays. It's in the concession block about why writing is the right place to look. "We are not only prompting the system; the return prompts us back" — stays. This line describes a property of the substance. The artefact/working surface distinction — stays. It answers Hoel's evidence. "The same technology supports two quite different modes of use: one that floods the zone with generic artefacts, and one in which returned text functions as part of a thinking process" — stays. This is a good formulation. "They may be a quite different type of tool, or not quite a type of tool at all" — stays, but now as a setup for the substance turn rather than as a tentative conclusion. The stone axe / *Homo faber* closing line — hmm. This could stay as a closing flourish if it's reframed. "Before we conclude that this is merely another stone in *Homo faber*'s long line of rocks" — that's a tool image. If the essay is saying it's a substance, not a tool, then the closing image should probably be about substance, not stone tools. Maybe: "before we conclude that bits in, bits out is the final word on a substance whose latent capacities we are still discovering." Or the kerosene return: "kerosene was not what oil turned out to be for." The "I am not claiming to have proved that LLMs cannot possibly be tools" qualifier (line 42) — stays. Important for tone. Under the substance framing, it becomes: "I am not claiming to have proved that LLMs are substances — the force of the previous step is to suggest that 'tool' may be the wrong category, and to ask what happens if we try a different one." What about the Keith Frankish mention (line 35)? It's a nice detail in the candidates-that-fail block. Stays. It shows that even good thinkers can't quite settle the function question. The "perhaps its function is to predict the next token; but that is a description of mechanism, not of use — nobody opens ChatGPT in order to predict tokens, any more than we describe the function of the heart as contracting rhythmically" line (34) — stays. Good analogy. The Swiss Army knife point (lines 29, 44) — stays. It pre-empts the obvious "but tools can be multi-functional" objection. OK, so the restructuring preserves a lot of the existing material. What it mainly does is: 1. Cut the medium block and Frippertronics 2. Add the substance turn and kerosene reversal 3. Move the evaluation/writing blocks to after the substance turn 4. Reframe the pen comparison and slop/Lucier in substance terms 5. Rewrite the conclusion Let me think about whether there's a version that keeps even more of the existing material... Could Frippertronics survive as a parenthetical? Something like: "The recursive structure has more in common with Robert Fripp playing into a tape delay loop — hearing the return, adjusting the next phrase in light of what came back — than with using a typewriter." One sentence, no Eno, no technical description of the two Revoxes. Just the gesture. That preserves the name-drop and the core image without needing the medium concept. It could go in the pen-comparison block, right after "a pen does not prompt me back." Actually, that might work. Let me think about whether it does anything the pen comparison doesn't already do... The pen comparison says: a pen extends inscription, doesn't push back. The Frippertronics gesture says: the recursive structure is like playing into a tape loop, where the return shapes the next move. The second adds something the first doesn't: it names the temporal structure of the exchange, the fact that it's a *loop*, not just a single push-back. The pen comparison is about what the system does (sends back altered material). The Frippertronics gesture is about what *you* do (adjust in light of the return). Together they capture both sides. So maybe one sentence of Frippertronics survives as part of the pen block. Not as an imported analogy that needs its own paragraph, but as a quick comparison that makes the recursive structure vivid. That way the "cool interesting detail" isn't lost — it's just trimmed to the essentials. Similarly, could Lucier's *I Am Sitting in a Room* be expanded slightly? In the current draft it gets two sentences. Under the substance framing, it could do more work: the room is a physical system with characteristic resonant frequencies — those are substance-level properties. When speech is passed through the room repeatedly, the room's properties take over. The content dissolves into the room's signature. That's exactly what happens when text is passed through an LLM without resistance: the model's tendencies (smoothing, genericising, familiar transitions) take over. The text dissolves into the model's signature. That's slop. The Lucier analogy is actually *stronger* under the substance framing because it explicitly parallels a physical substance's properties shaping the output. The room's resonant frequencies are properties of the room — properties of the physical substance (air, walls, surfaces). The model's smoothing tendencies are properties of the model — properties of the computational substance. In both cases, the properties are latent until material is passed through the system and they become manifest. OK, I think I have a clear enough picture. Let me also think about whether the order I proposed is right or whether there's a better sequence. The current draft goes: 1. Hoel's position 2. Hoel expanded 3. Pivot 4. Proper function 5. Candidates fail 6. First verdict 7. Evaluation uncertainty 8. Writing as test case 9. Artefact vs surface 10. Surface expanded 11. Pen comparison 12. Wrong level of description 13. Medium 14. Frippertronics 15. Slop/Lucier 16. Conclusion My proposed restructuring goes: 1. Hoel's position (= current 1) 2. Hoel expanded (= current 2) 3. Pivot (= current 3) 4. Proper function (= current 4) 5. Candidates fail (= current 5) 6. First verdict (= current 6) 7. **SUBSTANCE TURN** (NEW) 8. **KEROSENE REVERSAL** (NEW) 9. Evaluation uncertainty (= current 7, moved down, reframed) 10. Writing as test case (= current 8, moved down) 11. Artefact vs surface + surface expanded (= current 9+10, compressed) 12. Pen comparison + Frippertronics gesture (= current 11, expanded slightly) 13. Wrong level of description (= current 12, reframed) 14. Slop/Lucier (= current 15, reframed) 15. Conclusion (= current 16, rewritten) That's fifteen blocks, but blocks 11 and 12 are combining existing material, and blocks 7 and 8 are new. So net: cut 2 blocks (medium, Frippertronics-as-standalone), add 2 blocks (substance, kerosene), compress 2 into 1 (artefact/surface). Total is 15 blocks vs. the current 16. Wait, but should the evaluation uncertainty block come before or after the writing-as-test-case block? In the current draft, evaluation (7) comes before writing (8). Under the substance framing, the evaluation point is motivated by the substance claim: if it's a substance, you can't test it like a tool. That logic suggests evaluation comes right after the substance turn and kerosene reversal. Then writing comes after evaluation as the specific test case that Hoel chose. That's the order I have, and I think it's right. Actually, now I'm second-guessing whether blocks 9 and 10 (evaluation + writing) should come *before* the substance turn rather than after it. In the current draft, they come before the positive category (medium). The logic was: show the problem with evaluation → show the problem with writing as test → then offer the positive category. Under the substance framing, should the logic be the same? Hmm. There are two options: Option A: Substance turn early (my current proposal) - Proper function fails → first verdict → SUBSTANCE → kerosene → evaluation → writing → details → slop → conclusion - Advantage: the reader knows where they're going from block 7 onward - Disadvantage: the evaluation and writing blocks might feel like they're retreading ground — "OK, I already know it's a substance, why are you still talking about Hoel's test?" Option B: Substance turn later (mirroring current structure) - Proper function fails → first verdict → evaluation → writing → artefact/surface → pen → SUBSTANCE → kerosene → slop → conclusion - Advantage: the reader feels the full force of the problem before getting the answer - Disadvantage: a lot of negative argument before the positive thought, which was already a problem with the current draft Option C: Substance turn in the middle - Proper function fails → first verdict → evaluation → SUBSTANCE → kerosene → writing → artefact/surface → pen → slop → conclusion - Advantage: the substance turn comes after enough problem-setting to feel motivated, but before the writing details so those can be read through the substance lens - Disadvantage: separates evaluation from writing, which are naturally linked I think Option A is still the best. Here's why: the substance thesis *explains* why proper function fails. It's not just "these candidates don't work" — it's "they don't work because LLMs aren't the kind of thing that has a proper function; they're substances, and substances have latent properties rather than proper functions." That's an explanatory move, and it belongs right after the observation it explains. If you delay it, the reader is left wondering "OK, so what ARE they?" for too long. And the worry about retreading ground ("I already know it's a substance, so why talk about evaluation?") can be handled by framing blocks 9-10 as applications of the substance thesis: "If this is right — if LLMs are substances rather than tools — then evaluation looks different. Tool-evaluation presupposes settled function. Substance-investigation is open-ended. And that means Hoel's choice of writing as the proving ground, while not arbitrary, carries different weight than his framing suggests." That makes blocks 9-10 feel like they're developing the substance claim, not retreading the problem. Good. One more thing: the "wrong level of description" block (currently lines 90-95, my block 13). In the current draft, this comes between the pen comparison and the medium block. It says: Hoel's evidence is not refuted but is measured at the wrong level. Under the restructuring, where does it go? I think it naturally follows the artefact/surface distinction and pen comparison. The logic is: here's the distinction (artefact vs surface), here's what it looks like in practice (pen comparison), and here's the conclusion: Hoel's evidence concerns artefacts, but the interesting substance-exploration concerns the working surface, so his evidence is measured at the wrong level. Then slop: but the artefact-level evidence (slop) is not a refutation — it's what you'd expect from a substance whose tendencies are allowed to run without resistance. Then conclusion. Actually, should the "wrong level" block merge with the artefact/surface block? They're making the same point. "The difficulty is that 'writing' is too coarse... Hoel measures artefacts... but much interesting use treats text as a working surface... so his evidence is measured at the wrong level." That could be one block, not two. Let me see if the existing bullets can be compressed. Lines 65-78 (artefact/surface + expansion): about 14 lines of bullets Lines 90-95 (wrong level): about 6 lines of bullets Could these be one block of, say, 12-14 lines? I think so. The expansion block (73-78) says: "The point was to test a distinction, or to discover what you actually think by responding to what the system sends back. 'Has writing improved?' presupposes that the relevant success condition is improvement in the quality of end-products." That's the wrong-level point already. So lines 90-95 are largely repeating what 73-78 said. They could be merged. Let me revise the structure: 1. Opening + Hoel quote (fixed) 2. Hoel expanded (lines 17-21) 3. Pivot (lines 22-24) 4. Proper function (lines 26-31) 5. Candidates fail (lines 33-38) 6. First verdict (lines 41-45) 7. **SUBSTANCE TURN** (NEW) 8. **KEROSENE REVERSAL** (NEW) 9. Evaluation uncertainty (lines 48-53, reframed) 10. Writing as test case (lines 56-62) 11. Artefact/surface + wrong level (lines 65-78 + 90-95, compressed and merged) 12. Pen comparison + Frippertronics gesture (lines 81-87, expanded slightly) 13. Slop/Lucier (lines 115-120, reframed) 14. Conclusion (rewritten) Fourteen blocks. That's tight. And almost everything from the current draft survives — just the medium block and the standalone Frippertronics block are cut. Everything else is moved, reframed, or compressed, but the actual observations, examples, and good lines are all preserved. I feel good about this. Let me write it up clearly. Sorry — that was a bizarre thing to say. Of course restructuring includes moving things. Let me lay out the actual restructuring, block by block, against the current draft. ## What stays, what moves, what's cut, what's new ### Blocks 1–6: Stay in place, essentially unchanged These are the opening through the first verdict — Hoel's position, the pivot, proper function, candidates that fail, "not quite a type of tool at all." Under the substance framing, their role doesn't change: they establish that the tool category wobbles for LLMs. The only difference is that instead of leading toward "medium" as the resolution, they lead toward "substance." Lines 12–15 (opening + quote): fixed. Lines 17–21 (Hoel expanded): stay. Lines 22–24 (pivot): stay. Lines 26–31 (proper function): stay. Lines 33–38 (candidates fail): stay. The Frankish mention, the heart analogy, the Swiss Army knife — all preserved. Lines 41–45 (first verdict): stays, but now it's a setup, not a resting point. "Not quite a type of tool at all" becomes the hinge into the substance turn rather than a tentative conclusion left hanging. ### Block 7: NEW — The substance turn This is the essay's pivot. It goes right after the first verdict, because the substance claim *explains* why proper function fails. The logic: tools don't have latent properties, for two reasons — they're not natural objects, and it doesn't make sense to say that designed/manufactured objects have hidden properties. Substances do have latent properties: properties that are in the stuff but not yet discovered or drawn out. LLMs have latent properties — chain-of-thought, in-context learning, the way they respond to iterative framing. These were not designed in; they were found. So LLMs look more like substances than tools. This block also needs a brief material profile: what properties does this substance have that we've discovered so far? Pattern-completion, contextual pliability, associative reach, smoothing and genericising tendencies, responsiveness to structured prompting. Framed as observation about working with the stuff, not as a technical spec. ### Block 8: NEW — Kerosene reversal Hoel says discovering LLMs should have been like discovering oil. The reply: oil had latent capacities that took decades to draw out. For a long time, oil was lamp fuel. The larger transformations came later, once more of the substance's capacities had been discovered and stabilised in practice. The point is not predictive — not "just wait, the greatness is coming" — but epistemic: Hoel's argument assumes we already know what this substance can do, and the substance framing says we may not. Anti-hype qualification built in here, not as a separate block: "This is not a promissory argument. I am not predicting that the hidden greatness of LLMs will inevitably unfold. I am saying that we may still be in too early a phase of discovery for Hoel's preferred test to bear the weight he wants to put on it." ### Block 9: MOVED — Evaluation uncertainty Currently at lines 48–53. Moves from position 7 to position 9 (after the substance turn and kerosene reversal). Under the substance framing, its argument is strengthened: if this is a substance with latent properties, then we can't evaluate it the way we evaluate a tool. Tool-evaluation presupposes settled function. Substance-investigation is open-ended. Hoel's choice of writing as the proving ground is not arbitrary, but it carries different weight than his framing suggests. The existing bullets largely work — "if we do not know clearly what a thing is for, it is much harder to say in advance what would count as a good test of it" is exactly the substance point. May need one sentence reframing to connect it back to the substance thesis: something like "and if the reason we don't know what it's for is that it's not the kind of thing that has a 'for' — if it's a substance rather than a tool — then the problem is not temporary ignorance but a category difference." ### Block 10: MOVED — Writing as test case Currently at lines 56–62. Moves from position 8 to position 10. Stays essentially unchanged — it concedes that Hoel has a good reason to look at writing ("words are its womb, its mother, its literal atoms"). The question is what exactly we're measuring when we look there. This block now reads differently because the reader already knows the substance thesis: the question becomes not just "what are we measuring?" but "are we measuring the substance's capacities, or are we measuring one early use of them?" ### Block 11: COMPRESSED — Artefact/surface + wrong level Currently two blocks: lines 65–78 (artefact vs working surface + expansion) and lines 90–95 (wrong level of description). These are making the same point and can be merged. The combined block says: "writing" covers two different activities — producing artefacts (published texts) and using text as a working surface (testing distinctions, exploring what comes back). Hoel's evidence concerns artefacts. Much substance-exploration concerns the working surface. So his evidence tells us about one use of the substance, not about the substance itself. The best lines from both existing blocks survive: "A finished essay, a published book, a social media post — these are texts as artefacts... when someone throws a half-formed idea at an LLM and uses the response to sharpen what they actually think, the text is not the product — it is the working surface." And: "A lot of bad prose may show what happens when returned text is treated as a product to be published rather than as material to be resisted or worked through." ### Block 12: EXPANDED SLIGHTLY — Pen comparison + Frippertronics gesture Currently at lines 81–87. Stays in roughly the same position. The pen comparison is one of the draft's best moments and needs no reframing: "A pen makes marks on a surface; it extends my ability to inscribe; it does not return a proposal, a misreading, or a line of continuation I had not seen. A pen does not prompt me back." Under the substance framing, this marks the difference between a transparent tool (pen — no latent properties) and a substance with its own tendencies (LLM — sends back material shaped by its learned patterns). The "sometimes flatter than what I wanted, sometimes unexpectedly connective, sometimes productively irritating" line stays — those are descriptions of the substance's behaviour when engaged. "We are not only prompting the system; the return prompts us back" — stays. One sentence of Frippertronics added here, not as a standalone paragraph but as a quick comparison that makes the recursive structure vivid: something like "The recursive structure has more in common with Robert Fripp playing into a tape delay loop — hearing the return, adjusting the next phrase in light of what came back — than with any ordinary use of a writing tool." That preserves the name-drop and the image without needing its own paragraph or the medium concept. It sits naturally after "the return prompts us back." ### Block 13: REFRAMED — Slop/Lucier Currently at lines 115–120. Stays in roughly the same position. The Lucier analogy is actually stronger under the substance framing: the room's resonant frequencies are properties of the physical substance (air, walls, surfaces), and when speech is passed through repeatedly, those properties take over. Same logic for LLMs: the model's smoothing and genericising tendencies are properties of the substance, and when text is passed through without resistance, those properties dominate. Slop is what happens when a substance's characteristic tendencies are allowed to run unchecked. The last line of this block changes. Currently: "But this does not straightforwardly show that LLMs are tools; it may instead show what happens when a medium with strong generic tendencies is used with too little resistance from the person inside the loop." Under the substance framing: "But this does not straightforwardly show that LLMs are tools; it may instead show what happens when a substance with strong genericising tendencies is used without resistance — when its characteristic behaviour is allowed to dominate rather than being explored, tested, or pushed back against." ### Block 14: REWRITTEN — Conclusion Currently at lines 123–129. The existing conclusion lands on "bits in, bits out" and the *Homo faber* / stone-in-a-long-line-of-rocks image. Under the substance framing, the conclusion needs to land on the substance claim. The closing should do three things: 1. Acknowledge what Hoel gets right — no text singularity, slop is real, public writing is often worse. 2. Say what his framing misses — he has evaluated a substance as though it were a tool, measuring it by one early application of its capacities. 3. Close with the kerosene image or the substance image: "We may not yet know what this thing is for, because that may not yet be the right question. We are still discovering what it can do." The *Homo faber* line could survive if it's inverted: "This may not be another stone in *Homo faber*'s long line of rocks. It may be a new substance — and we may still be learning what can be drawn from it." That preserves the current closing rhythm and turns Hoel's tool image against itself. ### What's CUT Lines 98–104: The medium block (*Growing the Image*, Wollheim, Thomson-Jones, "structured field of resources and practices"). Gone. This is the only block that's fully cut rather than reframed. Lines 107–113: Frippertronics as standalone paragraph. Gone as a standalone block. One sentence survives as a gesture within the pen-comparison block. Lines 131–148: The editorial notes at the bottom. These were about choices within the old structure and would need to be rewritten anyway. They can go. ### What's PRESERVED (the cool interesting details) - "A pen does not prompt me back" - "We are not only prompting the system; the return prompts us back" - "sometimes flatter than what I wanted, sometimes unexpectedly connective, sometimes productively irritating" - The Frankish mention and the heart analogy - The Swiss Army knife pre-emption - "words are its womb, its mother, its literal atoms" (Hoel) - The Lucier analogy (strengthened) - The artefact/working-surface distinction - "A lot of bad prose may show what happens when returned text is treated as a product to be published rather than as material to be resisted" - The *Homo faber* / stones image (inverted at the close) - Frippertronics (one sentence, not a paragraph) - "They may be a quite different type of tool, or not quite a type of tool at all" - All four candidates-that-fail (token prediction, chatting, writing, "helping with tasks") ### Shorter map of the whole shape 1. Opening + Hoel quote (fixed) 2. Hoel's argument expanded 3. Pivot: "tool" sounds right, but... 4. Proper function introduced 5. Candidates that fail 6. First verdict: "not quite a type of tool at all" 7. **Substance turn: tools don't have latent properties; substances do; LLMs do** 8. **Kerosene reversal (anti-hype built in)** 9. Evaluation uncertainty (reframed through substance lens) 10. Writing as test case (Hoel's "womb" quote) 11. Artefact vs surface + wrong level of description (compressed) 12. Pen comparison + Frippertronics gesture 13. Slop as characteristic substance-tendency (Lucier) 14. Conclusion: we've synthesized a substance; evaluation is premature *Ristrutturare non è buttare via — è rimettere gli stessi pezzi in un ordine che li faccia finalmente lavorare insieme.* ### User "THE SUBSTANCE TURN" why are you like this? ### Assistant Sorry — that's embarrassing. I've been writing as though I'm annotating a screenplay. It's just block 7: the substance point. Where the essay says tools don't have latent properties, substances do, LLMs seem to have them. No dramatic name needed. *A volte basta dire la cosa senza annunciarla come se fosse l'arrivo del Messia.* ### User "he substance point." you think this is better? /no, no of course not, i made another...' ### Assistant Ha — yes, "the substance point" is almost worse. Like I'm naming a geographic feature. It's just the paragraph where the argument moves from "tool doesn't fit" to "substance fits better." That's all it needs to be. *Non tutto ha bisogno di un nome — a volte è solo il momento in cui il discorso gira.* ### User if you look at your config, what change(s) can i make that will prevent you using the word 'pivot' ever again? this is not sarcasm, enough is nenough ### Assistant The simplest fix: one line in CLAUDE.md under Behavioral Requirements, where the other blanket bans live (like the bold text rule). Something like: > Never use the word "pivot" — in conversation or in writing. Just describe the turn without naming it. That would go around line 62 in CLAUDE.md, in the Behavioral Requirements section. It gets loaded on every session, so it would catch it everywhere — chat, drafts, structural descriptions, the lot. Want me to add it? *La parola proibita non merita nemmeno di essere nominata un'ultima volta.* ### User yep ### Assistant Done. It's in Behavioral Requirements, right after the chat formatting rule. *Una parola in meno nel vocabolario — e nessuno sentirà la sua mancanza.*