# I would like to continue this chat with you. I didn't think that sonnet was giving me good enough... ## Skills Used - epistemic-discipline - smart-note ## Retrieval Notes - Session id: `76277479-d95b-47c1-a4c2-25d461431a69` - Last activity: `2026-03-09T20:34:15.765Z` - Files touched: `2` ## Artifacts **Created:** - [[Notes/LLMs Are Not Tools]] **Modified:** - [[Daily Notes/2026-03-09]] ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User I would like to continue this chat with you. I didn't think that sonnet was giving me good enough answers. 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. tell me about what frippertronics is. i sort of half remember. 20:11 Identified Fripp's pioneering tape loop innovation system Identified Fripp's pioneering tape loop innovation system Frippertronics is the name given to Robert Fripp's live tape-loop performance system, developed in the late 1970s in collaboration with Brian Eno. The basic setup involves two reel-to-reel tape recorders (originally Revox machines) positioned some distance apart, with the tape running between them in a long loop. The guitarist plays into the first machine, which records the signal; by the time the tape travels to the second machine for playback, a delay has elapsed — the length of which is determined by the physical distance between the machines and the tape speed. The output of the second machine is fed back into the first, so successive layers accumulate and sustain indefinitely, each pass slightly degrading and blending into the texture. Fripp began using the system as a solo performance format after the dissolution of the original King Crimson in 1974. He'd set up in unconventional venues — record shops, foyers, small halls — and build up these slowly evolving, shimmering drones and melodic fragments over a single guitar. The name "Frippertronics" came later, once the solo loop performances became a recognised thing in their own right. The Eno connection is worth noting: Fripp and Eno had developed the tape-loop approach together on the albums No Pussyfooting (1973) and Evening Star (1975), which are sort of the ur-texts of the whole ambient/loop aesthetic. Frippertronics is essentially Fripp taking that collaborative system and making it a solo live practice. It's a direct ancestor of the looping pedal setups that became ubiquitous later — though the tape system has a distinct sonic character, warmer and more subject to drift and decay. Bits In, Bits Out - by Erik Hoel Six years of AI and the world got stupider theintrinsicperspective.com Substack 12 mins 8:17 pm More Remove from Later Move to Shortlist Archive AI Art And Artists: What They Are, What They Could Be, What They Should Be [No summary] PDF Dominic McIver Lopes 33 mins pasted i want to talk about this tet as well as frippertronics. my main focus is llms though. can you see connections? There is one particulart as pect pfthe copied in article I am interested in. I wonder if you can work it out. 20:18 Excavated connections between feedback systems and LLM limitations, hypothesizing particularity versus genericism Excavated connections between feedback systems and LLM limitations, hypothesizing particularity versus genericism The user wants me to connect Frippertronics with the Hoel article about LLMs. Let me think about what the connections might be, and what particular aspect of the article they're most interested in. The most obvious connection is the "bits in, bits out" framing - Hoel's central thesis is that LLMs are tools, not intelligence. Frippertronics is literally a feedback/loop system where what comes out is a function of what goes in. But the more interesting connection is probably about the nature of the system and what it produces. Frippertronics works by layering and degrading - each pass through the tape loop slightly degrades the signal. What's striking is that this feedback process actually creates something genuinely novel - the tape loop transforms its input through physical degradation and accumulation rather than just replaying it. That's the real tension with Hoel's argument: while he claims LLMs are mere imitators constrained by their inputs, Frippertronics demonstrates how a feedback system can generate emergent properties that weren't present in the original material. The key question becomes whether Hoel's "bits in, bits out" framework actually captures what's happening in these systems, or if there's something about the process itself that produces genuine novelty. 3. The children's book example 4. The "views from nowhere" characterisation of LLMs 5. The book quality data 6. The Madeline image attempt The "views from nowhere" phrase is interesting in context of Nick's work - he's a philosopher of perception working on LLM aesthetics. Actually, thinking about Frippertronics specifically: the key feature is that Fripp's input is transformed and accumulated - the loop doesn't just replay; it builds layers, each degraded but contributing to a texture that couldn't be predicted from the input alone. There's emergence from the process. But Hoel's argument is essentially the opposite: with LLMs, there's no such emergence. The "loop" of training just encodes statistical regularities. The tape degradation through each pass - noise, warping, loss - is what gives Frippertronics its sound, which mirrors how LLMs trained on LLM-generated content degrade through recursive processing. Yet Frippertronics actually produces genuinely novel aesthetic experiences from that accumulation, something unpredictable from Fripp's isolated inputs alone. That tension between mere statistical layering and genuine emergence might be what Nick is really after. The Heidegger connection through Homo faber is worth exploring too - whether LLMs can achieve Zuhandenheit (the transparency of a tool in use) or remain Vorhandenheit (objects of conscious attention). Can they become extensions of human cognition, or do they always stay present-at-hand? I should probably lay out several possible connections and let Nick clarify which one resonates with him. Looking at the feedback loops angle: Frippertronics is literally a feedback accumulation system, and LLMs trained on internet text increasingly containing LLM output are in a similar loop - though a degrading one where output quality declines. There's also the "bits in, bits out" principle - Fripp's signal gets transformed but not transcended, just like Hoel argues LLMs transform input without transcending it. And then the core question of whether these are tools extending human agency or generating something more autonomous, where Hoel's position is clearly that LLMs remain tools. But I think the real insight Nick might be pointing toward is the perspective problem. Frippertronics is saturated with Fripp's particular sensibility - his note choices, timing, musical voice - the loop amplifies and extends that specificity rather than averaging it away. LLMs, by contrast, are views from nowhere, trained on the aggregate of all voices. That distinction between a tool that deepens particularity versus one that flattens it into generality might be what he's getting at. Now I'm thinking about how this connects to his environmental aesthetics work - Frippertronics as an aesthetic object invites a specific mode of appreciation, one that values the regularities and constraints of the system itself rather than just the surface output. I should lay out these connections and make an educated guess about which aspect he's really interested in. Show more Done Several connections suggest themselves, but let me try to identify the one you might have in mind. The most obvious structural connection is that Frippertronics is a bits-in-bits-out system, quite literally. Signal goes in, the loop accumulates and returns it, what comes out is a function of what went in — Fripp's note choices, timing, dynamics. And yet the outputs feel generative, genuinely surprising even to Fripp. So it's immediately a test case for Hoel's thesis: does a feedback system that only processes its own prior outputs constitute something more than a tool? The connection that strikes me as most philosophically pointed — and possibly the one you're circling — is Hoel's phrase "views from nowhere." He deploys it almost in passing, but it carries a lot of weight. His claim is that LLMs produce generic outputs because they have no genuine perspective; they're statistical centroids, not voices. Frippertronics is almost the precise inversion of this. The loop doesn't average across guitarists — it amplifies and extends a particular perspective, a specific set of choices made by one player in one moment. The "nowhere" problem doesn't arise because the system is saturated with Fripp's particularity. The tool extends a perspective rather than dissolving one. That seems directly relevant to your LLM aesthetics work — it raises the question of whether the "view from nowhere" problem is intrinsic to how LLMs are trained, or whether it's contingent on training regime, fine-tuning, prompting strategies. Is that the angle you had in mind? Or is it something else — the Heidegger passage, perhaps, or the degradation/slop-feeding-slop feedback problem? miss! but it was a tough job. I just wanted you to have a think about things. let's start with this question: do you think frippertronic music is created using is *a tool*? 20:21 Interrogated whether feedback loops transcend mere tool status Interrogated whether feedback loops transcend mere tool status It's genuinely not obvious, and I think that's the point. In the banal sense, yes — it's physical apparatus that Fripp operates. But Hoel's "bits in, bits out" framing implicitly assumes a fairly linear input-output relationship. What Frippertronics does is structurally different: the output feeds back into the input. The loop is the instrument in a meaningful sense, and the accumulating layers start to shape what Fripp plays next. He's responding to the system as much as driving it. There's a feedback coupling between performer and apparatus that makes the "in/out" directionality blurry. So it seems resistant to clean tool-classification. It's not a hammer, or even a guitar. The system has a kind of temporal memory and a generative dynamic that partly escapes Fripp's intentions — which is precisely what he valued about it. The degradation, the drift, the unexpected resonances between layers. You could say: a tool is something you use on material. But Fripp is, in a meaningful sense, inside the system, in a loop with it. Which — if I'm still not finding the connection you want — might suggest the question isn't whether Frippertronics is a tool, but whether Hoel's tool/non-tool binary is actually the right frame for thinking about LLMs either. The feedback dynamics might be the key. Getting warmer? article from substack: Bits In, Bits Out - by Erik Hoel Six years of AI and the world got stupider theintrinsicperspective.com Substack 12 mins 8:17 pm More Remove from Later Move to Shortlist Archive AI Art And Artists: What They Are, What They Could Be, What They Should Be [No summary] PDF Dominic McIver Lopes 33 mins 7:26 pm Green youth surge driven by progressivism, not economics A popular narrative on both the Left and the Right is that the rise of the Green Party reflects young Britons’ struggles to buy a house and manage the cost of living. But this phenomenon is explained less by rent than it is by progressive social views. In the wake of the Greens’ victory in [...]Read More... unherd.com Eric Kaufmann 4 mins 3:35 pm Philosophy's Future All rights reserved. Book Russell Blackford,Damien Broderick 4hr 0min left Mar 6th OpenAI’s “compromise” with the Pentagon is what Anthropic feared Anthropic pushed for moral boundaries. OpenAI settled for softer legal ones, and now it stands to benefit as the Pentagon rushes out a politicized AI strategy during strikes on Iran. technologyreview.com James O'Donnell 5 mins Mar 6th ‘Star Trek’: 100 Greatest Episodes To celebrate the 50th anniversary of the airing of the original Star Trek, THR counted down the best 100 episodes across all 6 series. hollywoodreporter.com Graeme McMillan 44 mins Mar 5th In Defense of Effeminate Gay Boys Ben Appel shares his experience growing up as an effeminate gay boy and argues that such boys are naturally male, not girls. He warns against rushing medical treatments that change children's bodies based on gender confusion. He believes society should accept gender-nonconforming boys and let them be themselves without pressure to change. theatlantic.com Ben Appel 8 mins Mar 4th The “machinal bypass” and how we’re using AI to avoid ourselves | PNAS Researchers coin “machinal bypass” to describe using generative AI to replace our own presence and avoid hard human work. They warn this can erode essential human aspects of research, care, and creativity. AI is fine for chores, but not for tasks needing lived, messy, personal engagement. pnas.org pnas.org 11min left Mar 3rd We should challenge our complacency around disability The author shares her struggles caring for her disabled brother after their father's death and challenges simple ideas about kindness and disability seen in films. She highlights how real-life situations are more complex than stories that suggest being nice solves everything. The article calls for more understanding and less judgment about the challenges faced by disabled people and their families. thecritic.co.uk Victoria Smith 6 mins Mar 2nd Book Review: The Geography Of Madness ... astralcodexten.com Scott Alexander 31 mins Mar 1st On malevolence When a diabolical disability meets an inveterate narcissist katrosenfield.substack.com Kat Rosenfield 5 mins Feb 26th 30 Years On: Autechre’s Tri Repetae Revisited Autechre’s album Tri Repetae stood out in 1995 for its unique and thoughtful electronic sound amid a sea of generic dance music. The album was not a commercial hit but gained respect for its melody and experimental style. Over time, Autechre moved toward more abstract and challenging music, leaving behind the relatable human touch of Tri Repetae. thequietus.com Gary Suarez 6 mins Feb 26th To inbox To previous document To next document toggleLeftPanelIcon Appearance Remove from Later Move to Shortlist Archive More toggleRightPanelIcon SearchNavIcon Find in document... Clear Previous result Next result Done The Intrinsic Perspective Bits In, Bits Out Six years of AI and the world got stupider When I was ten years old I visited the ruins in Cornwall where King Arthur had been conceived and born, at least in legend. There at Tintagel Castle, surrounded by the ocean air and the jagged rocks, I separated from my mother and sister and made my way down to the beach. And on the beach of Tintagel, right near Merlin’s cave, I spotted it in the sand. A stone. But not just any stone—a stone ax head. It could have been nothing else. It was shaped just like an ax, being unnaturally thick at the head, which was a smoothed blunt blade, and with near right angles it tapered to a point at the back. There was a notched cleft down the middle, to tie it to a shaft. As a 10-year-old boy in a place already dreamy with legend, I pocketed it, for I felt the Neolithic ax had come to me specifically, as if a Lady of the Lake had tossed it ashore. I am looking at it on my desk now. Later I learned that such Neolithic axes are not so rare—much like ancient Roman coins, they were mass-produced, and you can buy them on eBay cheaply due to finds like mine. This one from Tintagel beach is a perfect specimen, although the ocean likely washed away its provenance. But on that day, even amid all the other sandy stones, it stood out to me immediately. I knew it was a tool instinctively, the way a baby knows the nipple. The philosopher Henri Bergson wrote that: We should say not Homo sapiens, but Homo faber. Homo faber means “man the maker.” For if anything defines humans, it is tool use. I know it is now standard, in our rush to dethrone humanity, to play up that other animals also sometimes use tools. But unlike other animals, tools are our evolutionary niche. We have been making stone tools for at least 3.3 million years. We co-evolved with tools. At first, we made them from wood and bone and stone; later we began to craft abstract tools too. Language is a tool. Math is a tool. All of our vaunted cognition is, in some sense, a tool for a more protean mental firmament, which probably is consciousness itself. Heidegger’s term for this aspect of our consciousness was Zuhandenheit: “readiness to hand.” Now, we live in an age of tools that can talk back to us. When ChatGPT and the other LLMs appeared on the scene, and I first typed into a chat window, I experienced amazement. It was the legendary Turing Test, and I was living it! REPENT, THE SINGULARITY IS NIGH! Do not doubt that we are at the absolute peak of the AI hype cycle. In monetary terms, investment cannot actually go on increasing at these rates without leading to basically impossible numbers. A major falling out between the DoD and Anthropic dominates the news (even amid war). In the months leading up to all this, a bunch of commentators have jumped aboard the bandwagon of AI hype: The Wall Street Journal advises to “Brace yourself for the AI Tsunami” and The New York Timesis saying that AI might completely change the fundamentals of human existence. Popular bloggers are writing that “AI can already do social science research better than most professors,” and that “the humanities are about to be automated” and that “superintelligence is already here.” Meanwhile, fictional doomsday reports from the year 2028 go hyper-viral and impact the stock market as the authors imagine human unemployment leading to an endless great depression. The infamous METR graph continues to accelerate (and everyone ignores that METR tasks are a tiny sample size for a slim number of domain-specific programming tasks, all based on a dubious analogy to human-time-spent-on-task, and some of the authors have tried to downplay it because their graph has been so misused). There are usually many charts involved in the Great AI Debate. I could show you charts too! Like that, despite improvements on benchmarks, the actual reliability of AI is on an almost-flat trajectory for many tasks. source But others can then show other charts, and so on, ad infinitum. I don’t think you can decide the future by a couple of charts. No, there’s a better form of argument about AI, one which I am finally comfortable making: the argument from experience. There simply has been enough time now to see clearly how LLMs transformed the intellectual work of writing, and how this reflects their fundamental nature. My proposal is that we simply extrapolate what has happened to text production to all the other intellectual domains LLMs will ever touch. For if everything that anyone can do on a computer is soon to be automated (as Andrew Yang is now preaching will happen in the next 12-18 months), then this process should have started with writing years ago. Yet, beyond mass-producing stilted emails and stilted social media posts and stilted essays, the impact of LLMs on writing itself has not really been to improve or accelerate good writing overall. We are not in a glut of good writing. We are in a dearth of it. This is surprising and counterintuitive, because for an LLM, words are its womb, its mother, its literal atoms—yet their impact on writing as a whole has been mostly to generate mountains of slop, while, on the positive side, helping with efficiency and research and editing and feedback, all things that only marginally improve already-good pieces. There are no signs of a burgeoning “text singularity” seen in the words output by our civilization, and words are the most sensitive weathervane to AI capabilities. If LLMs were a true source of intelligence to rival humans, then discovering them should be like discovering oil. And if we were climbing the curve of an intelligence explosion their surplus intellect would be improving our civilization’s text as a whole in noticeable ways. If LLMs are tools, then we should expect their impacts to be a mirror of us, and concern efficiency and scale, rather than quality, and depend strongly on how people use them. So let me ask you: if you took an observer from 2016 and teleported them a decade ahead to our time, and then showed them your social media feed or your emails and other media in general, what would their main response be? Would it be “Wow, everything is more intelligent now!” Or would it be “Why is everyone writing like a pod person now?” It’s been six years since GPT-3, and there has been no “move 37” moment for writing (as there was for AlphaGo’s creative play of Go). Not even close. HECK, LET’S LOWER THE BAR If you ask a leading AI to write a children’s book, you’ll see that AI has not demonstrated exponential improvements at whatever amorphous and hard-to-define (but very real) skill “children’s book authorship” is. Indeed, the upper bound of the needle hasn’t moved much for writing in general, as people forget that older models like GPT-3 were already extremely good at short sprints of text when they kept it together. LLMs have been able to write a passable approximation of a children’s book for almost half a decade… and yet, the real lesson from this is that approximation is not automation. Below is Anthropic’s latest model (arguably the smartest AI in existence) trying to write a good children’s book, by itself, without guidance and hand-holding. The result, which it called “beautiful,” was exactly what a smiling alien would write if it had never interacted with a child before. It was a “blurry jpeg” of a children’s book. Here’s the sappy ending: gag me with a spoon People love to onanistically declare that the “it’s just autocomplete” or the “stochastic parrot” criticism is soooo outdated and soooo stupid. And yet… isn’t “You shine by being yourself” basically autocomplete for a children’s book? Can we admit that? Or do we have to pretend these insipid outputs are from a machine on the verge of artificial superintelligence (coming in literal months)? I simply refuse to play along. If you actually interact with them, and ask them to do things they aren’t directly trained on, these models remain spectacularly intellectually shallow and incompetent when there’s no detailed human prompting to give them clues and hints and guideposts. Charged with writing children’s literature independently, based on their own ideas and own attempts at style, an AI like Anthropic’s Claude will always draw from the same small well. The outputs still feel like an LLM to anyone with an ear for language, or an eye for content. The entire life of the artist, indeed, the life of the mind in general, is defined by resistance to slop. If you want to write an actually good children’s book you’ve got to put your own perspective into it, and LLMs are “views from nowhere.” Consider how deep and complex good children’s books actually are. The Giving Tree is about unrequited parental sacrifice. The Velveteen Rabbit is about ontology. Madeline is about Paris-maxing. The Rainbow Fish? Communism. The Very Hungry Caterpillar? Metamorphosis. In Where the Wild Things Are, at the end of the book, the Wild Things try to eat Max. That’s a good children’s book. (I was going to show a beautiful picture from Madeline here, as my daughter is obsessed with the “Pooh-pooh to the tiger in the zoo” scene. I took a photo of the book and asked the smartest-AI-in-existence-with-extended-thinking-on-at-the-highest-subscription-tier-available to make it look better. Presented are the results, in triptych.) Consider AI video generation. “We’re going to automate Hollywood!” Okay, well, what happened to automating publishing? That’s a much simpler task, one that you’ve had well over half a decade to do. How’d that go? Hmm? Everyone on social media and at the companies simply declared victory (“Wow, this fan fiction looks good enough to me!”) and moved on to more capabilities, stuffing more reinforcement learning into the models, building more elaborate scaffolds, without actually getting it across the finish line for producing non-trivial and non-annoying and non-fluff text unless the model is being spoon-fed by a human. Like, you think people are going to craft highly-detailed prompts for their own personalized movies? Again, we can look at writing, where everything already played out. You can already prompt your way to personalized books! It just sucks and no one does it, because when you ask the LLM to operate independently, its ideas are mostly slop. This is why when a company like Anthropic says they’ve automated their code production, when really they mean they are still writing code, just at a slightly higher level of abstraction (and that’s why they still have over 100 open software developer positions, and why Boris Cherny, creator of Claude Code, said that “Engineering is changing and great engineers are more important than ever.”) HAVE LLMS MADE BOOKS BETTER? A paper looking at large-scale Amazon data quietly appeared earlier this year, asking “Have LLMs Boosted the Creation of Valuable Books?” The answer to this question is that in the post-LLM era (which they date as after 2022), at least based on numbers of Amazon ratings, the average book got worse. E.g., you occasionally hear hype of the AI-assisted author mass-producing books, and yet, investigation usually reveals that they have sold more like zero books and that the story is a scam. The actual effects of LLMs on publishing were that: (a) the average book got worse, (b) the top 1,000 books in each category improved somewhat, and (c) the top 100 books in each category didn’t change in quality. How to read: Books with high reader ratings are on the right side of the x-axis here. So you can see that 2023-2025 books are mostly slop and the best don’t change much. Also, this may be hyping up the effects of AI for various reasons, e.g., the researchers do adjustments (it is not the case that only 2020-22 books are uniquely so low rated on the left-hand side, that’s an artifact of adjustments) among other issues. Now, the average book a reader encounters is more likely to be in the top 1,000 than below that, so the average consumer experiences an increase in quality (mostly due, it appears, to more “shots on goal” rather than actual improvements). But this beneficial effect for consumers seems to be driven by often-very-low-quality-anyways categories like “Travel” and “Outdoors,” rather than, say, “Science” and “Literature.” So do these effects look like a new source of surplus alien intelligence? Or does it look like tool use? Consider that the authors who were already successful pre-LLMs had the most efficiency gains, supporting the “tools” theory; meanwhile, new post-LLM era debut authors produced much worse work. After failing to automate publishing books and writing in general, the claim is now that AI will go on to automate science, math, all of academia, and finally humanity itself? But why would the chart for scientific papers (or anything else) look different from the chart for books above? LLMs can now “write a scientific paper” or “write a mathematical paper” in the exact same sense that they’ve been able to “write a book” or “write a short story” or “write an essay” for several years, all to some effect, but overall the results have been objectively mediocre given the hype, and the world is somewhat stupider, rather than smarter, at least on average. WELCOME TO WRITER HELL, MATHEMATICIANS Looking into the crystal ball that the last half-decade represents for writers reveals that, more likely than superintelligence, we are going to enter a world of immense, overwhelming, scientific and philosophical and mathematical slop. Will there be some good outcomes as well? Yes! Just as there have been for writing. I am not entirely an AI pessimist. I am an AI realist—there are indeed positives to the technology, and I’m trying to find them myself (like for research, or, e.g., my attempt at making the Madeline image better above, or the fewer spelling mistakes I make now, or sometimes I ask LLMs to double-check something, etc.) Yes, some B-tier bloggers have suddenly and mysteriously transformed into A-tier bloggers, who all kind of sound the same in their A-tier-ness. Previously A-tier bloggers have gotten a lot less use from the technology, and are not noticeably better than they were years ago. LLMs have broadly failed to automate text generation in general for the precise reasons I laid out all the way back in my 2022 essay “AI Art Isn’t Art.” AI-art isn't art They struggle because they are fundamentally imitators, and when not told who or what to mimic they are intellectually shallow. You put more bits in, you get better bits out. Fine. That’s a tool. A computer or a piano is like that too. This is not something that will trigger a singularity of self-improvement or take the jobs of all of humanity or create “machines of loving grace” or any of the stuff people are now regularly promising is literally going to happen in like… a year. E.g., humans can edit their text, making the block of marble look ever more like the statue inside. When it comes flying solo, first drafts by LLMs are often better than last. LLMs can’t even recursively improve a five paragraph essay, let alone themselves. So we will experience a long march by AI across intellectual disciplines (most lately, mathematics) where the hype reawakens with each new expansion, and in the wake of the long march some things do change but ultimately the world is not reconfigured as has been promised and the actual experienced intelligence level of the world, especially the top where it matters most, remains mostly unchanged, because it’s still just humans using tools. Bits in, bits out. That’s been the effect of LLMs on text production (in both book publishing and on social media), and it seems very likely to be their effect on almost everything else too. In this conservative view, by 2030, the top 100 math papers of the year won’t look spectacularly different from the top 100 math papers of 2020. The top 1,000 papers? Maybe they’ll be marginally improved, like with books or blogs. And at the backend, the entire field of mathematics will be buried, absolutely buried, in slop. Companies will continually say their systems accomplish things “autonomously” but that word is hard to define for LLMs, where a prompt is an injection of human intelligence, and a scaffold too is an injection of human intelligence (but the advantage of scaffolds is that you can put a ton of domain-specific knowledge and tips and tricks and guides in the scaffolds, keep them private because it’s proprietary, and then say the models “solved it autonomously!”). At some point you have to use your capacity as Homo faber and call it: LLMs have behaved precisely as we would expect tools to behave when it comes to changing the nature of first-impacted and frontline intellectual disciplines like writing. The best users gain efficiencies and expand, to some degree, their capability range, especially for the mid-list of intellectual output. The worst users flood the zone. So at least when it comes to the near-term future and the foreseeable scaling of the current technology, “merely” the exact same thing that has happened to writing will happen to every subject on Earth. But that’s not replacing humanity. That’s not the singularity. You’re just confused about what we are. We are Homo faber, and we have been doing this for 3.3 million years, and our rocks have gotten very complex—so complex you’re forgiven for not thinking they’re rocks. toggleRightPanelIcon Copy Copy Help ghostreader ghost bodyghost glasses ### Assistant Let me first load the required skills and read some context before diving into this deep contemplation. ### Assistant Now let me read Nick's research context and check for relevant session notes on LLM aesthetics work. ### Assistant Good — I now have the context on Nick's research. He's working on "Generating Philosophy with AI" (arguing LLMs can produce genuine philosophical contributions) and the Substack "The Dereliction of Thought" which explores whether human practice with LLMs counts as "really thinking/writing." His position is NOT about whether LLMs think, but about whether human practice with LLMs counts as genuine thinking/writing. He also has the gardening metaphor for AI generation. Let me also check for the Lopes paper that was mentioned in the pasted content. ### Assistant Good. The Lopes "AI Art and Artists" paper is in Readwise. Let me read it since it was referenced in the pasted content. Let me also read the previous session conversations to understand the full context of what's been discussed. ### User [Request interrupted by user] ### User It is not necessarily about the aesthetic work, i think we could just think of it as all text. with me iterating on what comes out each time. Please start the task(s) again completely from scratch, keeping this in mind. ### Assistant Okay, let me start completely from scratch. Nick wants me to think about Frippertronics, the Hoel article, and LLMs — but the connection he's interested in is not specifically about aesthetics. It's about all text. And he's emphasised the iterative aspect — "with me iterating on what comes out each time." So let me re-read the situation. The previous Sonnet conversation went like this: 1. Nick asked about Frippertronics. Got a decent summary. 2. Nick pasted the Hoel article and asked about connections to Frippertronics and LLMs. Sonnet guessed "views from nowhere" — miss. 3. Nick asked: "do you think frippertronic music is created using a tool?" Sonnet gave a good answer about feedback coupling, the blurriness of in/out directionality, Fripp being inside the system. Nick said "getting warmer." So what's the thread? Let me think carefully. The question Nick posed — "is Frippertronics a tool?" — and the answer that was "getting warmer" involved the idea that Fripp is *inside* the feedback loop. The system has temporal memory. The output shapes the next input. Fripp responds to what the system gives back. The in/out directionality is blurry. Now Nick says: think of it as all text, not just aesthetics. "With me iterating on what comes out each time." That phrase is telling. "Me iterating on what comes out each time." That's... that's literally describing how Nick uses LLMs. He puts text in. Gets text out. Reads it. Responds to it. The LLM responds to his response. And so on. The conversation *is* the writing process — he's said this explicitly in his context ("Sessions with Claude *are* the writing process"). So the Frippertronics analogy isn't about aesthetics or views from nowhere or degradation. It's about the *mode of production*. It's about what happens when a human enters a feedback loop with a system. Fripp plays → the loop captures and returns → Fripp hears it back → this shapes what he plays next → the loop captures that too → and so on. The music that emerges isn't "Fripp's output" in a simple sense. It's the product of an iterative feedback process between Fripp and the system. Neither party alone could have produced what results. Nick writes/prompts → the LLM responds → Nick reads it, reacts, iterates → the LLM responds to *that* → and so on. The text that emerges isn't "Nick's output" or "the LLM's output" in a simple sense. It's the product of an iterative feedback process. Now here's where Hoel comes in. Hoel's argument is: LLMs are tools. "Bits in, bits out." They're just very complex rocks. Homo faber uses tools. The tool doesn't change the fundamental nature of what's happening — it's still human intelligence driving everything. But is that right for the iterative case? When Hoel says "bits in, bits out," he's implicitly describing a linear, one-shot process. You put a prompt in, you get output out. That's the children's book example — ask Claude to write a children's book "by itself, without guidance and hand-holding," and you get slop. The "views from nowhere" criticism applies to one-shot generation. But that's not how Frippertronics works, and it's not how Nick works with LLMs. The iterative loop changes the nature of the process. Each pass through the system is shaped by all previous passes. The human is responding to the system's output, and the system is responding to the human's response to its output. Hoel's strongest evidence is all about one-shot or low-iteration use: the children's book written "independently," the Amazon data about average books getting worse (mass-produced slop is definitionally low-iteration), the "views from nowhere" criticism. His test case is asking an LLM to do something alone. But Frippertronics shows that the same basic mechanism — bits in, bits out, through a feedback system — can produce genuinely novel, genuinely valuable output when the human is *in the loop*. The loop is what matters, not the one-shot. Wait, but Hoel might say: "Yes, that's exactly my point! The human is what makes it work. The tool is just a tool. Fripp makes Frippertronics great because Fripp is great. The LLM is just the tape machine." That's the "tools" reading. And there's something to it. But here's where it gets interesting. Can you cleanly separate Fripp's contribution from the tape loop's contribution? The music that results from Frippertronics isn't music that Fripp could have played on a guitar alone. It's not that Fripp composed something and the tape loop is just a playback mechanism. The tape loop *generates* the layering, the accumulation, the degradation, the unexpected resonances. Fripp provides the input, but the system transforms it in ways Fripp doesn't fully control and responds to. The resulting music is a *joint product* that neither Fripp alone nor the tape loop alone could produce. Similarly — and this is Nick's point, I think — the text that emerges from iterated LLM conversation isn't text that Nick could have written alone (or at least, not in this form, at this pace, with these particular articulations). And it's not text the LLM produces alone (that's Hoel's children's book slop). It's a joint product of the iterative loop. So the connection Nick is pointing at — let me try to be precise — is something like: Hoel's "bits in, bits out" framing treats LLMs as a one-shot input-output mechanism (like a hammer or even a piano). But Frippertronics demonstrates that a bits-in-bits-out system, when used iteratively with a human in the loop, produces something that can't be adequately described as "just tool use." The feedback loop generates something that transcends the simple tool model. And this is directly relevant to Nick's own practice. He uses LLM conversations as his writing process. He iterates. The text he produces is the product of this feedback loop. Hoel would presumably say "that's still just you using a tool" — but is that description adequate? Is it adequate for Frippertronics? Actually, let me think about what specific aspect of the Hoel article Nick said he was interested in. He said "there is one particular aspect of the copied in article I am interested in." Sonnet guessed "views from nowhere" — miss. Then Nick redirected to the question of whether Frippertronics uses "a tool." Let me re-read the article more carefully for what connects to the iterative/feedback dimension... Hoel says: "You put more bits in, you get better bits out. Fine. That's a tool." He also says: "humans can edit their text, making the block of marble look ever more like the statue inside. When it comes flying solo, first drafts by LLMs are often better than last. LLMs can't even recursively improve a five paragraph essay, let alone themselves." That's interesting! He's claiming that LLMs can't improve through iteration — that recursive self-improvement doesn't work for text. But that's only true for the LLM iterating *alone*. What about the LLM iterating with a human? That's the Frippertronics setup. There's also: "when you ask the LLM to operate independently, its ideas are mostly slop" — the emphasis on *independently* versus in conversation. And: "there has been no 'move 37' moment for writing" — no creative leap. But could the iterative human-LLM loop be where such moments happen, not in the LLM's solo output? And: "if you actually interact with them, and ask them to do things they aren't directly trained on, these models remain spectacularly intellectually shallow and incompetent when there's no detailed human prompting to give them clues and hints and guideposts." Hoel seems to view "detailed human prompting" as the human doing the real work and the LLM just executing. But in the Frippertronics model, the prompting isn't separate from the output — it's part of the loop. Fripp's playing *is* the prompt, and what comes back shapes the next prompt, and the music is the whole thing. Let me also think about what Nick's established positions are. He argues in his "Generating Philosophy" paper that LLMs can produce genuine philosophical contributions. He frames AI generation through gardening/cultivation (environmental aesthetics). His Substack position is about whether human practice with LLMs counts as "really thinking/writing" — and his answer seems to be yes, at least the "thinking with" version where judgement remains his. So the Frippertronics connection supports Nick's position against Hoel's: the iterative loop IS a form of genuine creative production, not just tool use. The system amplifies and transforms the human's input in ways that make the result genuinely novel and not merely derivative. But wait — Nick said the previous answer was a "miss" when Sonnet focused on views from nowhere and the perspective/particularity angle. And he said "getting warmer" when the answer was about feedback coupling and being inside the system. Then he said: "think of it as all text. with me iterating on what comes out each time." So Nick wants me to think about text production specifically. Not music. Not aesthetics in general. The Frippertronics analogy applied to how text gets produced through iterated human-LLM interaction. Let me think about what Hoel's article specifically claims about the *process* of text production with LLMs... Hoel says: "on the positive side, helping with efficiency and research and editing and feedback, all things that only marginally improve already-good pieces." He characterises LLM contributions as marginal improvements. Not transformative. Not generative of new quality. He says: "Previously A-tier bloggers have gotten a lot less use from the technology, and are not noticeably better than they were years ago." This is interesting. If A-tier writers don't benefit, that suggests the LLM adds nothing to genuine quality. But is Hoel right about this? Or is he not considering the iterative, conversational mode at all? He also says: "Boris Cherny, creator of Claude Code, said that 'Engineering is changing and great engineers are more important than ever.'" This actually supports the iterative model — the human in the loop matters more, not less. Then there's: "a prompt is an injection of human intelligence, and a scaffold too is an injection of human intelligence." Hoel frames prompts as injections. One-directional. Human → machine. But in a conversation, the prompt is a *response to* the machine's last output. It's bidirectional. It's a loop. Let me think about what the "one particular aspect" might be... Actually, I think I need to stop trying to guess the specific aspect and instead do what Nick asked: think deeply about the connections, give lots of options and ideas, explain my reasoning. He wants maximum feedback. So let me think about ALL the connections I can see between Frippertronics, the Hoel article, and LLMs (specifically regarding text, not aesthetics): 1. THE ITERATIVE LOOP VS. ONE-SHOT MODEL Hoel's entire argument is built on one-shot examples. Children's book written alone. Amazon books (mass-produced, presumably low-iteration). Social media slop. He never considers the case of sustained, iterative human-LLM dialogue producing text. Frippertronics is the counterexample: a bits-in-bits-out system that, through iteration and feedback, produces genuinely novel output. The question: does Hoel's argument survive the shift from one-shot to iterative? 2. THE TOOL QUESTION Hoel: LLMs are tools. Homo faber uses tools. But as the Sonnet conversation established, Frippertronics resists clean tool classification because the human is inside the system. Similarly, when Nick writes philosophy through extended LLM conversations, he's not using a tool on material — he's in a feedback loop with a system that responds to him and that he responds to. The in/out directionality is blurry. "Bits in, bits out" doesn't capture the recursive, reflexive structure. 3. DEGRADATION VS. ACCUMULATION In Frippertronics, each pass through the tape loop degrades the signal slightly — but this degradation is productive. It creates texture, warmth, complexity. In LLM iteration, each pass through the conversation accumulates context, sharpens ideas, produces refinements. Hoel's claim that "first drafts by LLMs are often better than last" — that LLMs degrade rather than improve through iteration — is true for LLMs iterating alone. But in the human-LLM loop, the human provides the evaluative correction that keeps the iteration productive, just as Fripp's musical choices keep the tape loop from becoming noise. 4. AUTHORSHIP AND THE JOINT PRODUCT If the music produced by Frippertronics is genuinely a joint product of Fripp and the system — not just Fripp using a tool — then what about text produced through iterated human-LLM conversation? Is it the human's text? The LLM's text? A joint product? Hoel doesn't consider this question because he treats LLMs as straightforwardly tools. But the Frippertronics analogy makes the question harder to dismiss. 5. "BITS IN, BITS OUT" AS SELF-REFUTING Here's a thought: isn't ALL cognition "bits in, bits out"? Hoel himself notes that "Language is a tool. Math is a tool. All of our vaunted cognition is, in some sense, a tool." If the criterion is "bits in, bits out," then human cognition is also just a loop. The brain receives sensory input, processes it, produces output. The question isn't whether something is bits-in-bits-out — everything is — but whether the process generates genuine novelty. Frippertronics does, despite being a simple feedback loop. Does the human-LLM conversation loop? 6. HOMO FABER AND THE NATURE OF TOOL USE Hoel invokes Homo faber — man the maker — and claims we've been using tools for 3.3 million years and LLMs are just the latest. But Frippertronics might show that some "tools" fundamentally change the nature of the maker's activity. A hammer extends the arm. A guitar extends the voice. But Frippertronics creates a system in which the maker is a component rather than a master. Is the human-LLM conversational loop similarly transformative of the activity itself? 7. THE SCAFFOLDING PROBLEM Hoel says: "you can put a ton of domain-specific knowledge and tips and tricks and guides in the scaffolds, keep them private because it's proprietary, and then say the models 'solved it autonomously!'" He means this dismissively — the scaffolding is human intelligence pretending to be machine intelligence. But isn't the conversational loop itself a kind of scaffolding? And isn't the question whether the scaffolding transforms the process rather than just enabling it? 8. THE WRITING QUESTION SPECIFICALLY Nick's Substack position: the question isn't whether LLMs think, but whether human practice with LLMs counts as "really thinking/writing." The distinction is "thinking with" (judgement remains his) vs. "outsourcing" (ceding judgement). Frippertronics maps onto this perfectly. Fripp is "playing with" the system — his musical judgement remains his, he decides what to play next, he responds to what comes back. But the resulting music isn't "just Fripp playing guitar." It's something different, something new, something that only emerges from the loop. Nick writing philosophy through conversation with Claude is structurally identical. His philosophical judgement remains his. He decides what to pursue, what to reject, what to develop. But the resulting text isn't "just Nick writing" — it's something that emerges from the iterative loop. Hoel would presumably say: "Fine, but you're still doing the thinking. The LLM is just a sophisticated tape recorder." But is that adequate? Is Fripp "just playing guitar" when he's performing Frippertronics? 9. THE SCALE OF INTERVENTION In Frippertronics, Fripp's input can be minimal — a single note, held and allowed to accumulate and transform through the loop. The system does enormous amounts of the "work." But no one would say the music is "just" the tape loop. Fripp's minimal input, precisely because of the loop, generates maximal output. Similarly, a brief prompt or a nudge in conversation with an LLM can generate extensive, detailed text. The human's input can be minimal in quantity but transformative in direction. Hoel seems to assume that if the LLM generates most of the text, the text is "the LLM's" — but Frippertronics shows that who generates the volume isn't the same as who generates the value. 10. THE SPECIFIC PASSAGE NICK MIGHT BE INTERESTED IN OK, let me think again about what specific part of Hoel's article connects most directly. Nick said there was "one particular aspect." Given that he then redirected to the tool question and said "think of it as all text, with me iterating on what comes out each time"... Could it be this passage? "humans can edit their text, making the block of marble look ever more like the statue inside. When it comes flying solo, first drafts by LLMs are often better than last. LLMs can't even recursively improve a five paragraph essay, let alone themselves." This is Hoel's claim that LLMs can't recursively improve. That iteration doesn't work for them. But iteration DOES work for Frippertronics — and it does work for human-LLM conversation. The difference is the human in the loop. The LLM alone degrades through iteration (like a tape loop with no one listening and playing back into it — just noise). But the LLM with a human iterating? That's Frippertronics. That's where the productive iteration happens. Or could it be the Homo faber / Heidegger passage? "All of our vaunted cognition is, in some sense, a tool for a more protean mental firmament, which probably is consciousness itself. Heidegger's term for this aspect of our consciousness was Zuhandenheit: 'readiness to hand.'" Zuhandenheit — ready-to-hand. The tool disappears in use. Does the LLM disappear in conversational use? Does the Frippertronics apparatus disappear for Fripp during performance? If so, then the relationship is phenomenologically different from "using a tool" in the reflective, deliberate sense. Actually, another possibility: "So let me ask you: if you took an observer from 2016 and teleported them a decade ahead to our time, and then showed them your social media feed or your emails and other media in general, what would their main response be?" Hoel's point is that the *average* quality of text has gone down. But Frippertronics wasn't about the average. It was about one specific musician using a feedback system to produce specific, high-quality work. The average person with a tape loop doesn't make Frippertronics. Fripp does. Similarly, the average person with ChatGPT produces slop. But does a skilled practitioner — someone who knows how to be in the loop, who has good judgement about what to iterate on — produce something genuinely new? Let me think about this differently. What does Nick actually DO when he writes philosophy with me (or with Claude in general)? He: 1. Raises an idea or question 2. I respond with analysis, connections, possible developments 3. He reads my response, identifies what's interesting, pushes back on what's wrong 4. I respond to his pushback 5. He iterates further 6. Eventually, text crystallises that neither of us would have produced alone This is structurally Frippertronics. Nick plays a note (raises an idea). The system (LLM) captures it, transforms it, plays it back. Nick hears it back, responds to what the system generated, plays a new note shaped by what came back. The layers accumulate. Something emerges. Hoel's framework can't describe this because it insists on a clean input/output separation: bits in, bits out. But in the iterative loop, the output becomes the next input. There is no clean separation. The process is circular, not linear. Now, is this actually a GOOD argument against Hoel? Let me steelman Hoel's position. Hoel might say: "Fine, you're in a feedback loop. But your judgement — your ability to evaluate, select, redirect — is doing all the real cognitive work. The LLM is like a very elaborate mirror: it reflects your ideas back at you in slightly different forms, and you use your intelligence to pick the good reflections and discard the bad ones. You're still doing the thinking. The loop is just a method of externalising your thought process." And there's something to this. When Nick says "that's interesting, pull on that thread," he's exercising philosophical judgement. The LLM isn't exercising judgement. So maybe Hoel is right that it's still "just" tool use, just a more sophisticated feedback tool. But here's where Frippertronics pushes back. Fripp would say (and has said) that the loop produces things he wouldn't have thought to play. The accumulation of layers creates textures and harmonies that surprise him, that he responds to in ways he wouldn't have if he were playing guitar alone. The loop isn't just reflecting — it's TRANSFORMING. The degradation, the accumulation, the interaction between layers — these are productive processes that generate novelty. Similarly, when an LLM responds to a philosophical idea, it often makes connections, raises objections, or articulates implications that the human hadn't considered. Not because the LLM is "thinking" but because the pattern-matching across vast corpora surfaces relevant material that the human's individual experience wouldn't have produced. The human then evaluates this, and the cycle continues. So the question becomes: is that productive transformation enough to make the process more than tool use? Or is it still "just" tool use, with the tool being very good at its job? I don't think this question has a clean answer. And maybe that's the point. The Frippertronics analogy shows that the tool/non-tool binary is too crude for describing these systems. Some forms of "tool use" are so deeply interactive, so thoroughly feedback-coupled, that the distinction between user and tool starts to dissolve. Not because the tool is intelligent, but because the process is iterative. Let me also think about what's at stake for Nick's specific projects. For "Generating Philosophy with AI": If LLM-assisted philosophical text is produced through an iterative, Frippertronics-like process, then calling the LLM "just a tool" undersells what's happening. The process generates genuine philosophical contributions not because the LLM is a philosopher, but because the feedback loop between human philosophical judgement and LLM pattern-matching produces text that neither could produce alone. For the Substack "The Dereliction of Thought": The Frippertronics analogy is a powerful way to frame the "thinking with" vs "outsourcing" distinction. "Thinking with" = Frippertronics (Fripp is in the loop, exercising judgement, but the system contributes something irreducible). "Outsourcing" = asking the LLM to write something solo (Hoel's children's book example — no loop, no feedback, no human judgement, and the result is slop). Hmm, let me think about one more thing. The "particular aspect" of the Hoel article. Let me re-read key sentences... "You put more bits in, you get better bits out. Fine. That's a tool." "LLMs can't even recursively improve a five paragraph essay, let alone themselves." "a prompt is an injection of human intelligence, and a scaffold too is an injection of human intelligence" "So we will experience a long march by AI across intellectual disciplines... and in the wake of the long march some things do change but ultimately the world is not reconfigured as has been promised and the actual experienced intelligence level of the world, especially the top where it matters most, remains mostly unchanged, because it's still just humans using tools." "the best users gain efficiencies and expand, to some degree, their capability range, especially for the mid-list of intellectual output." That last point is interesting. Hoel admits that the best users expand their capability range. But he frames this as "efficiency" rather than genuine expansion. The Frippertronics counter: Fripp's capability range with the loop is genuinely expanded beyond what he could do without it. It's not that he plays the same music more efficiently. He plays DIFFERENT music. Music that only exists because of the loop. OK wait. Let me go back to what Nick actually said. He said: "think of it as all text. with me iterating on what comes out each time." He's telling me something specific. The key is: iteration on text, by him, each time. Not one-shot generation. Not the LLM operating independently. But HIM, ITERATING, on WHAT COMES OUT, EACH TIME. That's... that's literally the Frippertronics process described in textual terms. Fripp plays → output comes out → Fripp responds to the output → new output comes out → Fripp responds again → and so on. Nick writes/prompts → text comes out → Nick iterates on it → new text comes out → Nick iterates on that → and so on. The "particular aspect" of the Hoel article might be the recursive improvement claim: "LLMs can't even recursively improve a five paragraph essay, let alone themselves." Because Frippertronics shows that recursive processing through a feedback loop WITH a human operator CAN produce genuine improvement — not degradation. The degradation happens when the loop runs without human intervention. The improvement happens when the human is inside the loop, making choices. Or it might be the "bits in, bits out" framing itself as inadequate for describing iterative processes. Or it might be something about the distinction between what AI does "independently" versus what it does in conversation. Hoel keeps testing AI on its solo performance. But solo performance is the wrong test if the actual mode of use is collaborative iteration. You know what, I think I should present all of this to Nick rather than trying to pin down the one specific thing. He asked for a long, deep, detailed answer with lots of options and ideas. Let me give him the full landscape of connections I can see, with careful reasoning for each. One more thought: there's a temporal dimension here that's worth noting. Frippertronics is inherently temporal — the music unfolds over time, each moment shaped by what came before. The tape loop has a specific delay time, and the relationship between input and output is temporal, not instantaneous. Similarly, a conversation with an LLM unfolds over time. The context accumulates. Ideas develop across multiple turns. This temporality is completely missing from Hoel's analysis, which treats LLM use as essentially instantaneous: prompt in, output out. And one more: Hoel says there's been "no move 37 moment for writing." Move 37 was AlphaGo's creative, surprising play that no human would have made. But move 37 didn't happen in isolation — it happened within the context of a game, a sequence of moves and responses. It was an emergent product of an iterative process. If we're looking for the equivalent in writing, we shouldn't look at one-shot text generation. We should look at what emerges from sustained human-LLM conversation. Has anything genuinely surprising, genuinely novel, emerged from such conversations? Nick's own work might be evidence that it has. OK, I think I have enough threads to give a thorough answer. Let me compose it. ## Frippertronics, Hoel, and Iterated Text Let me lay out the full landscape of connections I can see, with reasoning. You asked for maximum depth, so here's everything, presented as parallel threads rather than ranked. ### The structural analogy Frippertronics is an iterative feedback loop: Fripp plays → the system captures, transforms, returns → Fripp hears what comes back → this shapes what he plays next → the system captures *that* → layers accumulate. The music is a product of the whole loop, not of either component in isolation. Your practice with LLMs is structurally identical: you raise an idea → the LLM responds with analysis, connections, possible developments → you read it, identify what's interesting, push back on what's wrong → the LLM responds to your pushback → you iterate further. Text crystallises that neither you nor the LLM would have produced alone. Hoel's "bits in, bits out" framework describes a linear, one-directional process. Prompt goes in, output comes out, done. But both Frippertronics and iterated LLM conversation are *circular*. The output becomes the next input. There is no clean separation between "bits in" and "bits out" because the process recurses. ### Where Hoel's evidence fails Here's what strikes me: every piece of evidence Hoel marshals is about *solo, one-shot, or low-iteration* LLM use. - The children's book: written "by itself, without guidance and hand-holding" - The Amazon book data: mass-produced books are definitionally low-iteration - Social media slop: generated and posted, not iterated - The Madeline image: one prompt, one output - "Views from nowhere": a property of text generated without a particular human's judgement in the loop He never tests the case that actually corresponds to Frippertronics: a skilled human in sustained, iterative exchange with the system. His argument is essentially: "When I ask the tape machine to play music by itself, it just produces noise. Therefore, tape machines are just tools." But nobody ever claimed the tape machine alone produces music. The claim — and the interesting case — is about what happens when Fripp is in the loop. ### The recursive improvement claim This specific passage in Hoel feels like it might be the one you're interested in: "Humans can edit their text, making the block of marble look ever more like the statue inside. When it comes flying solo, first drafts by LLMs are often better than last. LLMs can't even recursively improve a five paragraph essay, let alone themselves." He's claiming recursive improvement doesn't work for LLMs. And he's right — *for the LLM alone*. An LLM iterating on its own output does tend to degrade. This mirrors Frippertronics with no one at the guitar: the tape loop running by itself eventually decays into noise. Degradation without a human. But the whole point of Frippertronics is that Fripp is there, providing the evaluative intelligence that keeps the iteration productive. He listens to what comes back, decides what to play next, steers the accumulation. The loop improves because of his presence in it — not because the tape machine improves independently. Similarly, when you iterate on LLM output — reading, evaluating, redirecting, pushing, discarding — the text does recursively improve. Not because the LLM is improving itself, but because your judgement in the loop keeps the iteration productive. Hoel's claim that recursive improvement is impossible for LLMs is only true for the case nobody cares about. ### The tool question, revisited When Sonnet asked "is Frippertronics made with a tool?" and you said "getting warmer," the productive answer was: the feedback coupling between Fripp and the system makes the "tool" framing inadequate. Fripp isn't using the tape loop *on* material the way you use a hammer on a nail. He's *inside* the system. Apply this to text. When you're several turns deep into a philosophical conversation with an LLM — the context has accumulated, you're responding to something it articulated that you hadn't quite thought of, which it articulated in response to your pushback on something it said earlier, which was itself a response to your initial question — at that point, are you "using a tool"? Or are you inside a process? Hoel's Homo faber framing insists on a clean separation: the maker and the made, the tool-user and the tool. But the Frippertronics structure dissolves this. The maker is a component of the system, and the system is partly constituted by the maker's responses to what the system produces. Now, Hoel could respond: "Fine, but your *judgement* is doing all the real work. The LLM is a mirror — it reflects your ideas back in slightly different forms, and you pick the good ones. You're still the maker." But this doesn't work for Frippertronics either. Fripp's judgement is doing evaluative work, yes. But the tape loop isn't just reflecting — it's *transforming*. The accumulation, the degradation, the harmonic interactions between layers — these produce things Fripp didn't intend and couldn't have played on guitar alone. The loop generates material that Fripp then responds to. His judgement alone doesn't explain the output. Similarly, when an LLM responds to a philosophical idea, it surfaces connections across a vast corpus that your individual reading history wouldn't have produced. It articulates implications you hadn't considered, raises objections from angles you hadn't taken. Not because it "understands," but because pattern-matching across that much text produces genuine novelty in combination. You evaluate this, and the loop continues. Your judgement alone doesn't explain the output either. ### Temporality Something Hoel completely ignores: both Frippertronics and LLM conversation are inherently *temporal* processes. The music unfolds over time — each moment shaped by what came before. The tape loop has a specific delay, and the relationship between what Fripp plays and what comes back is stretched across a duration. Similarly, a conversation with an LLM unfolds over turns. Context accumulates. Ideas develop across exchanges. You come back to something said five turns ago. The LLM holds the thread. Hoel treats LLM use as essentially instantaneous: prompt in, output out, evaluate, done. But the temporal dimension is where the interesting stuff happens. The accumulation of context over a long conversation is what makes iterated text production different from one-shot prompting — just as the accumulation of layers over time is what makes Frippertronics different from playing guitar through a delay pedal. ### "Move 37" Hoel says there's been no "move 37" moment for writing — no creative breakthrough produced by AI that rivals AlphaGo's famous play. But move 37 didn't happen in a vacuum. It happened within a game, an iterative sequence of moves and responses. It was an emergent product of a process, not a one-shot output. If we're looking for the equivalent in text, we shouldn't look at children's books written by Claude "independently." We should look at what emerges from sustained, iterated human-LLM exchange. The test isn't "can the LLM write a surprising essay from a single prompt?" The test is "does the conversation loop produce ideas, articulations, or connections that surprise the human participant?" And I'd guess your experience says it does — not every time, not reliably, but sometimes. Whether that constitutes a "move 37" is another question, but Hoel isn't even looking in the right place. ### What this might mean for your work Without imposing hierarchy on your various projects, I can note how the Frippertronics analogy touches several threads you're exploring: - For the argument that LLMs can produce genuine philosophical contributions: the iterative loop provides the mechanism. Neither you nor the LLM alone produces the contribution. The loop does. The question of whether this "counts" is a question about whether the loop is a legitimate mode of philosophical production. - For the "thinking with" vs "outsourcing" distinction: Frippertronics gives you a precise structural analogy. "Thinking with" = being in the loop, exercising judgement on what comes back, steering the iteration. "Outsourcing" = asking the system to run independently (Hoel's children's book example). The distinction isn't about whether you use the system. It's about whether you're *in* the loop. - Against Hoel specifically: his argument is compelling against one-shot and low-iteration use. It's not obviously wrong about that. Average books did get worse. Social media is more slop-filled. But his argument doesn't address the case that matters to you: the iterative, feedback-coupled mode of text production. Frippertronics is the counterexample that shows why "bits in, bits out" is an inadequate description of feedback systems, even though it's literally true of them. ### Open questions I'm not sure about A few things I think are genuinely unresolved: Does the Frippertronics analogy *help* philosophically, or does it just make the same point you could make without it? The structural parallel is clear, but analogies can obscure as much as illuminate. Is there a disanalogy that matters? Frippertronics involves continuous, real-time feedback. LLM conversation involves discrete turns with reflection between them. Does the discreteness change the analysis? Hoel might accept the Frippertronics point and say: "Sure, the best users in iterative mode produce good work. That's what I said — the best users expand their capability range. But that's still tool use, and it doesn't change the top 100." Is the Frippertronics analogy strong enough to push past that response? And the hardest question: does the analogy actually show that the LLM contributes something irreducible, or just that the *process* of externalising your thinking (into any responsive medium) is productive? Would you get the same benefit talking to a colleague? Writing in a journal? Is the LLM's specific contribution — its pattern-matching, its corpus-derived connections — what matters, or is it just the loop structure itself? *Quando Fripp entra nel circuito, non suona più la chitarra — abita il sistema, e il sistema abita lui.* ### User btw, I have never once shown any interesting in the bits in bits out thing. I can't even remember what that is, but i don't think it is very interesting. i am interested in two, potentially but I am not sure related questions: I am interested in the possibility that LLMs are not tools, indeed this might be thee title of the substack article i am considering writing. 1. I am interested in the idea that LLMs are NOT tools. this is for two reasons. tools have a proper function. LLMs don't have an obvious proper function. relatedly (still the same point), they are unpredictable (note, i do not mean unreliable) in a way that tools are not. 2. I am interested in the idea I like the idea of a self loop as modelled on frippertronics. if you know me, you'll know where i am going with this......... 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 check Nick's existing work on this to make sure I'm grounding my thinking properly. ### Assistant Good. Now let me check for Nick's "Growing the Image" paper — the gardening metaphor — since that's directly about how AI generation works, and also search for anything about "inner speech" or "coupling" which is another thread he has. ### Assistant Now let me also check for Nick's "Growing the Image" paper, since the gardening metaphor for AI is his published position and directly relevant to the "LLMs are not tools" question. ### Assistant Now I have a very clear picture. The "Growing the Image" paper explicitly argues that Midjourney is "not an agent or a tool, but a new type of artistic medium." And the specific argument against the tool classification is exactly what Nick is now pointing me toward: tools have a proper function, and an unpredictable tool is a malfunctioning one ("if the outcome of a traditional machine becomes unpredictable, we do not think that it is creative or original---we think that it is broken"), whereas LLMs/Midjourney are *reliably unpredictable* — "this is a feature, not a bug." Now I understand the full picture. Let me compose the contemplation. OK. Let me start from scratch and think clearly about what Nick is actually telling me. Two claims, potentially related: 1. LLMs are NOT tools. Two reasons: (a) tools have a proper function and LLMs don't have an obvious one; (b) LLMs are unpredictable (not unreliable) in a way tools are not. 2. The self-loop modelled on Frippertronics is interesting. And he says: "if you know me, you'll know where I am going with this." OK, so what do I know about Nick's existing positions? I just read the "Growing the Image" paper. In it, Nick and Terrone argue that Midjourney is not an agent, not a tool, but a *medium* — specifically, a dynamically recalcitrant medium analogous to gardening. The argument against the tool classification relies on exactly the two points Nick just mentioned: - Tools have a proper function. An unpredictable tool is malfunctioning. Esposito (2022): "If the outcome of a traditional machine becomes unpredictable, we do not think that it is creative or original — we think that it is broken." - Midjourney is *reliably* unpredictable. This is a feature, not a bug. So it doesn't fit the tool category. And the positive proposal is: medium. Specifically, a medium with *dynamic recalcitrance* — like gardening, where the gardener iterates on natura naturans, coaxing it but never fully controlling it. Now, Nick says "I am interested in the idea that LLMs are NOT tools" — this directly echoes the argument from his own published paper. But that paper was about Midjourney and images. Nick is now saying "think of it as all text." He's extending the argument from image generation to text generation, to LLMs more broadly. And the Frippertronics "self-loop" — this connects to the gardening iteration model. The gardener acts → waits → sees what natura naturans does → acts again. Fripp plays → the loop captures → Fripp hears what comes back → plays again. Nick prompts → the LLM responds → Nick reads → prompts again. So what is Nick going toward? Let me think... He has the "LLMs are not tools" argument already established for images. He has the "medium" concept as the positive alternative. He has Frippertronics as a model for the iterative self-loop. And he's considering a Substack article titled something like "LLMs Are Not Tools." So the Substack version would take the philosophical argument from "Growing the Image" and: 1. Apply it to text/LLMs rather than images/Midjourney 2. Make it accessible for a non-specialist audience 3. Frame it against Hoel's claim that LLMs ARE tools (the Homo faber argument) The Hoel article becomes the foil. Hoel insists: LLMs are tools. Homo faber. Bits in, bits out. Nick's response: no, they're not tools, and here's why — they don't have a proper function, they're unpredictably productive (not malfunctioning), and the mode of engagement is iterative, like gardening, or like Frippertronics. And Frippertronics gives him a much more vivid, accessible analogy than gardening for a Substack audience. The gardening metaphor works in a philosophy journal, but Frippertronics gives you something immediately graspable: the musician inside the feedback loop, responding to the system, the system responding back, neither fully in control. Actually wait. Let me think about whether Frippertronics is really the same as the gardening metaphor or whether it adds something different. Gardening emphasises: dynamic recalcitrance, iteration, coaxing, natura naturans. The gardener plants, waits, prunes, waits, and so on. The medium has its own generative dynamics. Frippertronics emphasises: the feedback loop, the temporal accumulation, the human being *inside* the system rather than acting on it from outside. There's a difference. The gardener is somewhat outside the garden — she acts on it, waits, acts again. There's a temporal gap. Fripp is more tightly coupled — he's hearing the output in real time, his playing is shaped by what comes back, and what comes back is shaped by his playing. It's a tighter loop. For LLM text production, which model fits better? In some ways, Frippertronics fits better than gardening does. You prompt, you get a response *immediately* (no waiting like a gardener), you respond to it, the LLM responds to your response, and so on. The loop is tight, real-time (or near-real-time). You're *inside* the conversation, not acting on something from outside and waiting. So Frippertronics might give Nick a better model for *text* than gardening does for *images* (or even for text, where the gardening metaphor was developed for images specifically). But wait — is there also something about the proper function point that Frippertronics illuminates? What is the proper function of a Frippertronics tape loop? As a piece of equipment — two Revox tape machines — each machine has a proper function (recording, playback). But the *system* — the loop configuration — doesn't have a proper function in the same way. Fripp repurposed tape machines into something they weren't designed for. The system is, in Hoel's terms, "just" tape machines (tools). But the configuration creates something that exceeds the tool category. Hmm, but that might be overcomplicating it. Let me refocus. What does Nick mean by "the self-loop modelled on Frippertronics"? He called it a "self-loop." That's interesting. Not just "a loop" but a *self* loop. In Frippertronics, the output feeds back into the input — the system loops on itself. But with a human in the loop. In the LLM case: you write → the LLM responds → you respond to the LLM's response → the LLM responds to your response to its response → and so on. Is this a "self" loop? In what sense? One possibility: it's a loop of *your own thinking*. The LLM is reflecting, transforming, and returning your own ideas. You're not getting input from another mind (there is no other mind). You're getting your own thinking processed through a system and returned to you in transformed form. Like Fripp hearing his own playing come back altered. It's a self-loop because it's YOUR cognition looping through an external system. This connects to Nick's inner speech work! Inner speech is already a self-loop — you produce language, hear it, respond to it, produce more. The LLM adds a transformative step in the middle of that loop. You externalise your thinking → it gets processed → it comes back transformed → you respond. It's still *your* thinking, but augmented by the transformation. And this is precisely NOT what a tool does. A tool extends your capacity in a predictable way — a hammer amplifies force, a telescope amplifies vision. A tool doesn't *transform your input unpredictably and feed it back to you*. What the Frippertronics loop does — and what the LLM conversation does — is something different from tool use. It's a cognitive self-loop mediated by an unpredictable, responsive system. Now: where is Nick going with this? He said "if you know me, you'll know where I'm going." Let me think about what I know about his commitments and projects: 1. He has the "Growing the Image" paper arguing Midjourney is a medium, not a tool. 2. He's working on "Generating Philosophy with AI" — arguing LLMs can produce genuine philosophy. 3. He has the inner speech / LLM coupling thread — hybrid inner speech, cognitive traction, the double-extension argument. 4. He's planning the Substack "The Dereliction of Thought" — about whether using LLMs counts as really thinking/writing. 5. His position is that it IS really thinking/writing — "thinking with" rather than "outsourcing." So the connections I can see: The "LLMs are not tools" argument, applied to text, would position against Hoel's essay. Frippertronics provides the model for what they *are* instead (or at least for what the interaction *is*). And this feeds directly into the Substack. But what specifically is the positive proposal? In "Growing the Image," the positive proposal was: medium. LLMs are a medium with dynamic recalcitrance. Is Nick going to say the same thing for text? That LLMs are a *medium* for philosophical writing, with dynamic recalcitrance? Or is the positive proposal something different for text? Something more like: the LLM is part of a cognitive self-loop. Not a medium you work *in* (as you work in marble or oil paint), but a system you loop *through* — like Frippertronics. Actually, I think there might be a tension here. The gardening/medium metaphor emphasises the creative product — the image or text that results. The Frippertronics/self-loop metaphor emphasises the *process* — the thinking that happens in the loop. For images, the product is what matters (you're making an artwork). For philosophy, the process might matter more (you're trying to think, and the text is both the means and the product of thinking). So maybe: for images, LLMs are a medium (not a tool). For text/philosophy, LLMs are something else — a cognitive loop? A thinking partner? Part of an extended cognitive system? This connects to the inner speech thread. The "double-extension" argument: language is a cognitive technology → inner speech is internalized language → LLM coupling is re-externalized inner speech. The LLM isn't a tool you use ON text; it's a system you think THROUGH, the way you think through inner speech. Let me think about whether the "not a tool" argument works differently for text than for images... For images (from the paper): - Proper function: artistic tools have proper functions (paintbrush makes marks, camera captures light). Midjourney doesn't have a proper function in this clear sense. - Unpredictability: an unpredictable tool is broken. Midjourney is reliably unpredictable — feature, not bug. For text/LLMs: - Proper function: what's the proper function of ChatGPT? To answer questions? To generate text? To assist? It's genuinely unclear. You can use it for writing, coding, research, brainstorming, translation, analysis... It doesn't have a single, definite proper function the way a hammer or a calculator does. - Unpredictability: LLMs are reliably unpredictable in their text outputs. Same prompt, different responses. More importantly, they generate content you couldn't have predicted — and this is valuable, not a malfunction. The "not a tool" argument actually works *better* for LLMs-as-text-generators than for Midjourney-as-image-generators, because the proper function indeterminacy is even more extreme. Midjourney at least has a fairly definite purpose: generate images from prompts. ChatGPT/Claude? It does... everything? Nothing in particular? Actually, this is an interesting point. Hoel keeps calling LLMs "tools" but never says what their proper function IS. He just says "bits in, bits out" — which is a description of how they work, not what they're for. A hammer is FOR driving nails. A telescope is FOR magnifying distant objects. What is an LLM FOR? This is genuinely unclear, and that unclarity might be precisely what makes them NOT tools. OK so let me try to figure out where Nick is going. He's interested in: 1. LLMs are not tools (proper function problem + unpredictability) 2. The self-loop modelled on Frippertronics If I connect these: LLMs are not tools because tools have proper functions and are predictable. LLMs are something else — part of a self-loop, like Frippertronics, where the human iterates through an unpredictable system and the result is a joint product of the loop. For the Substack article: this would be a direct response to Hoel. Hoel says LLMs are tools, Homo faber, 3.3 million years of tool use. Nick says: no, they're not tools, and here's why (proper function, unpredictability). And what they actually are (or what the practice with them actually is) is better modelled as a self-loop, like Frippertronics. The Frippertronics analogy works particularly well for a Substack audience because: - It's vivid and concrete - It's cultural (music fans will know Fripp/Eno) - It captures the iterative, feedback-coupled nature of actual LLM use - It sidesteps the agent/tool binary entirely Now, where exactly is Nick "going" that I should recognise? I think he's heading toward saying that what you do with an LLM — the conversational, iterative, textual process — is *thinking*. Not tool-assisted thinking (which implies the tool is separate from the thinking). Not delegated thinking (which implies the LLM does the thinking). But thinking that happens in and through the loop, the way Fripp's music happens in and through the tape loop. The self-loop IS the thinking. Not: you think, then use a tool, then think again. Rather: the loop — you-LLM-you-LLM — is a single process of thought. Just as the Frippertronics loop is a single process of music-making, not "Fripp plays guitar + a tape machine records it." And this connects to the Substack's position: "thinking with" LLMs counts as really thinking. Not because the LLM thinks (it doesn't). Not because it's just a tool you're using to think (it's not a tool). But because the loop constituted by your engagement with it IS a form of thinking — extended, externalized, looped through a responsive, unpredictable medium. This also maps onto the inner speech thread: inner speech is already a self-loop (produce → perceive → respond → repeat). The LLM loop is a turbocharged version of this. And the "double-extension" argument: language → internalized → re-externalized through LLM. And it maps onto the "Growing the Image" argument, but shifted from product (artwork/medium) to process (thinking/self-loop). The "medium" metaphor works for creative products. The "self-loop" metaphor works for cognitive processes. Let me also think about what this means for Hoel specifically. Hoel's argument is: LLMs are tools. Tools don't change the nature of human activity. Therefore LLMs won't change the nature of intellectual work. QED. Nick's counter: LLMs are not tools. Here's why (proper function, unpredictability). What they are is better understood through the self-loop model. And the self-loop DOES change the nature of intellectual work — it constitutes a new form of thinking, just as Frippertronics constitutes a new form of music-making. Hoel's evidence (bad children's books, Amazon slop, social media deterioration) shows what happens when you treat LLMs as tools — give them a job and expect output. That IS how most people use them, and the results ARE bad. But that's because tool-use is the wrong model. When you engage with them as a self-loop — iterating, responding, developing ideas through the feedback — the results are different. The difference isn't just in the skill of the user (Hoel's "best users gain efficiencies"). It's in the *mode of engagement* — loop vs. one-shot, process vs. product. Hmm, but I should be careful not to present this as too tidy. Let me think about where the argument might have difficulties... One worry: is the "proper function" argument really that strong for LLMs vs tools? You could say: the proper function of ChatGPT is "to generate helpful text in response to prompts." That's vague, but lots of tools have vague functions. A Swiss Army knife does many things — is it therefore not a tool? Nick might respond: a Swiss Army knife has multiple specific functions — blade for cutting, screwdriver for screwing, etc. Each implement has a proper function. The knife as a whole is a multi-tool. But ChatGPT doesn't have identifiable sub-functions that each have a proper function. It's not that it has many tools combined; it's that what it does isn't tool-like at all. The Swiss Army knife is predictable — the blade will cut, the screwdriver will turn. ChatGPT is not. Another worry: is "unpredictable (not unreliable)" really enough to distinguish LLMs from tools? Lots of tools have unpredictable elements. A kiln produces unpredictable results in glazing. A camera in certain conditions produces unpredictable depth-of-field effects. These are still tools. Nick addressed this in the paper for images: the watercolour case. The watercolour painter can't control exactly how paint distributes atom-for-atom, but she controls what depicted objects go where. The unpredictability is in fine details, not in the substance of the output. With LLMs, the unpredictability is substantive — you don't know what ideas, arguments, connections the system will produce. The "content" is unpredictable, not just the "surface." For text, this is even stronger than for images. When you prompt an LLM with a philosophical question, you genuinely don't know what argument it will make, what connections it will draw, what objections it will raise. This isn't fine-grained unpredictability (typos, formatting). It's content-level unpredictability. And this content-level unpredictability is precisely what makes the self-loop productive — you couldn't iterate on predictable output. Actually, that's an important point about the self-loop. The loop REQUIRES unpredictability to be productive. If the system just returned exactly what you expected, there'd be nothing to respond to, nothing to iterate on. The unpredictability is what makes the loop generative. This connects unpredictability (point 1) to the self-loop (point 2) — they're not just "potentially related," they're necessarily connected. The self-loop works BECAUSE LLMs are unpredictable. In Frippertronics too: the tape loop is productive precisely because the accumulation and degradation are unpredictable. If the loop just played back exactly what Fripp played, note-for-note, it would be useless. The unexpected resonances, the harmonic interactions between layers — these are what make Fripp want to respond, what keep the loop generative. So the two points are deeply connected: LLMs are not tools (because they're unpredictable in a non-malfunctioning way), and the way to understand what they ARE is through the self-loop model (which works precisely because of that productive unpredictability). Let me think about the Substack framing specifically. Nick's Substack is "The Dereliction of Thought" and the angle is: using LLMs counts as really thinking. "LLMs Are Not Tools" as a Substack article would: 1. Present Hoel's position as foil (LLMs are tools, Homo faber, etc.) 2. Argue: no, they're not tools (proper function problem + unpredictability-not-unreliability) 3. Propose the self-loop model via Frippertronics 4. Connect to the thesis: the self-loop IS thinking, not tool-assisted thinking This fits the Substack's positioning perfectly: anti-anti-AI, not cheerleading but arguing against dismissive framings. And it's a direct engagement with a specific, well-known piece (Hoel's article), which is good for Substack readership. Let me also think about whether I'm missing anything about the Frippertronics → self-loop connection. Nick specifically said "self-loop." In Frippertronics, the signal loops back into itself. The output of the second tape machine feeds back into the first. It's a *self*-referential system — the output becomes the input. In the LLM case: your thinking goes out, gets transformed, comes back, shapes your next thinking, goes out again. Your thinking is looping through an external system. But it's YOUR thinking that loops — hence "self" loop. The LLM is the medium through which the loop passes, but the cognitive content is yours (modified, transformed, augmented, but yours in the sense that your judgement, evaluation, and direction remain yours). This connects to the "thinking with" vs "outsourcing" distinction. The self-loop is "thinking with" — your thinking passes through the system and comes back, and you evaluate and redirect. It's not outsourcing because you're in the loop, exercising judgement at every stage. And the Frippertronics analogy makes this vivid: nobody would say Fripp "outsourced" his music to the tape loop. He's in the loop. The music is his. But it's also not him "just playing guitar" — the loop contributes something. The music is a product of the coupled system, and the coupling is what Fripp values. I think this is where Nick is going. Let me try to articulate it as clearly as I can... The thesis for the Substack article would be something like: Hoel says LLMs are tools. But they're not tools, because they lack a proper function and they're substantively unpredictable (not unreliable). What they are — or rather, what the practice of using them for thinking is — is better modelled as a self-loop, analogous to Frippertronics. In the self-loop, your thinking passes through a responsive, unpredictable system and comes back transformed. You respond to the transformation, and the loop continues. This loop IS thinking — not tool-assisted thinking, not delegated thinking, but a new mode of thinking that is genuinely productive precisely because the system is unpredictable and you are inside the loop. One more thought. Nick said "this might be the title of the substack article." He's thinking of writing this as a Substack piece. Given his Substack planning (first-person, philosophical, accessible, anti-anti-AI, drawing on personal experience), the article would likely: - Open with Hoel's article as the provocation - Personal experience of LLM use as evidence - The proper function / unpredictability argument (from his own published work, made accessible) - Frippertronics as the positive model - The claim: what I do with LLMs is thinking, and the Frippertronics model explains why This connects back to the earlier Sonnet conversation: Nick was using the Hoel article as a jumping-off point for thinking about his own work, and now he's identifying what specifically interested him and where he wants to go with it. Let me now think about what ideas, options, and feedback I can give Nick for this "kicking ideas around" stage... Some questions and ideas worth raising: 1. How does the medium concept from "Growing the Image" relate to the self-loop concept? Are these the same argument (medium = self-loop), different arguments (medium for images, self-loop for text), or is the self-loop a development/refinement of the medium concept? 2. The Frippertronics analogy is vivid and works well for Substack. But is it a perfect analogy? Potential disanalogies to think about: - Frippertronics involves continuous audio feedback; LLM conversation involves discrete text exchanges. Does the discreteness matter? - Frippertronics degrades the signal (each pass through the tape adds noise); LLM conversation doesn't obviously degrade (but does it accumulate in a similar way?). - In Frippertronics, the system is "dumb" — it's just tape machines. The tape doesn't "understand" the music. LLMs... also don't understand, but they produce semantically structured responses. Is this a relevant difference or does it actually strengthen the analogy (even something that doesn't understand can be a productive loop partner)? 3. The proper function argument: how far does Nick want to push it? Two versions: - Weak: LLMs don't have a CLEAR proper function (but maybe they have one that's hard to articulate) - Strong: LLMs don't have a proper function AT ALL (they're not the kind of thing that has proper functions) The weak version is easier to defend but less interesting. The strong version is provocative but harder. What would it mean for something to NOT be the kind of thing that has a proper function? 4. The relationship between unpredictability and the self-loop: I suggested these are necessarily connected (the loop requires unpredictability to be productive). Is this right? Or could you have a productive self-loop with a predictable system? (Maybe: a calculator is predictable, and you can loop through it — compute, check, compute again — but nobody would describe this as a self-loop in the interesting sense. The UNPREDICTABILITY is what makes the loop genuinely generative rather than merely confirmatory.) 5. What about the Heidegger angle from Hoel's article? Hoel invokes Zuhandenheit (readiness-to-hand). When a tool is working well, it disappears — you don't think about the hammer, you think about the nail. Does an LLM disappear in use? Nick's "Life in the Chat" note suggests it does, in some ways — the conversation becomes transparent, you're thinking about philosophy, not about the LLM. But it also doesn't — the surprising outputs, the need to evaluate and redirect, keep it present. This ambiguity might further support "not a tool": tools properly so called achieve Zuhandenheit; LLMs oscillate between transparency and opacity. 6. The "machinal bypass" article from PNAS (mentioned in Nick's Readwise feed): "using generative AI to replace our own presence and avoid hard human work." This is the "outsourcing" side — the anti-pole of the self-loop. The self-loop is precisely NOT machinal bypass, because you're present in the loop, doing the hard cognitive work of evaluation and direction. Frippertronics is precisely NOT Fripp leaving the room and letting the tape run. 7. How does this relate to Nick's gardening metaphor? One option: the gardening metaphor applies to creative production (making images, making text-artifacts). The self-loop applies to thinking (the cognitive process). They're the same phenomenon described at different levels — product vs process. In "Growing the Image," Nick described the product-level. In "LLMs Are Not Tools," he'd describe the process-level. 8. What about other feedback-loop artistic practices that might support the analogy? Generative music systems (Brian Eno's later work, generative algorithms). Improvisation in jazz (soloist responds to rhythm section, rhythm section responds to soloist). Even conversation itself — if LLMs aren't tools, and conversation isn't tool-use, then maybe the LLM interaction is a form of conversation? But Nick has been careful about this — LLMs are "subject-shaped surfaces," "hollow," "articulate non-subjects." The self-loop framing avoids the conversation trap by emphasising that it's YOUR thinking looping, not a dialogue between two thinkers. 9. The Esposito quote from the paper is excellent for the Substack: "If the outcome of a traditional machine becomes unpredictable, we do not think that it is creative or original — we think that it is broken." This is such a clean formulation of the point. LLMs are reliably unpredictable → not broken → therefore not machines/tools in the traditional sense. One sentence does a lot of work. 10. Potential objection Nick should consider: "Sure, but the unpredictability comes from the training data and the stochastic sampling process. It's engineered unpredictability, not genuine creativity." Response: the tape loop's unpredictability is also engineered (or at least, is a consequence of physical properties of tape). But that doesn't make it less productive. The source of the unpredictability doesn't determine its cognitive/creative value. Actually, let me also think about whether Nick is heading somewhere even more specific that I haven't considered yet... "If you know me, you'll know where I'm going with this." He has the extended mind / inner speech thread. The "double extension" argument. Language → inner speech (internalization) → LLM coupling (re-externalization). The self-loop modelled on Frippertronics would be: your inner speech loop, externalized through an LLM. The LLM takes the place of the autonomous (Type 1) processes that generate content in normal inner speech, but with greater capacity for novelty. Could the destination be: LLMs are not tools; they're extensions of inner speech? The self-loop is the cognitive self-loop of thinking, externalized and augmented? That would connect: - "Not tools" → the negative argument (proper function, unpredictability) - Self-loop → the positive proposal (what they ARE: extensions of the cognitive self-loop of thinking) - Frippertronics → the vivid analogy (makes the self-loop concept concrete and graspable) - Inner speech → the theoretical foundation (already developed in his research notes) And the Substack article "LLMs Are Not Tools" would be the accessible, first-person version of this theoretical structure. Yes, I think that's where he's going. The full chain: Hoel: LLMs are tools → Nick: No, because [proper function, unpredictability] → Nick: What they are is part of a self-loop (Frippertronics model) → Nick: This self-loop is a form of thinking (inner speech extended) → Nick: Therefore, using LLMs IS thinking, contra deBoer and other skeptics → Nick: And this is the thing I do all day, as a philosopher, and it works. The Substack article wouldn't need all the inner speech apparatus (that's for the paper). It would need: Hoel as foil, the not-a-tool argument, Frippertronics as the model, and the claim that the loop IS thinking. First-person, concrete, philosophical but accessible. Let me also consider: what's the relationship to "Growing the Image"? In that paper, the positive alternative to "tool" was "medium." For the Substack, the positive alternative to "tool" would be... self-loop? Medium? Something else? I think "medium" and "self-loop" capture different aspects. "Medium" emphasises the material you work with/through — its recalcitrance, its unpredictability, its resistance. "Self-loop" emphasises the process — the feedback, the iteration, the circular structure. For images, "medium" is natural. For thinking, "self-loop" is natural. But they're compatible — the LLM is a medium through which the self-loop passes. One more thought: for the Substack specifically, Frippertronics has another advantage over gardening. Gardening is slow — you plant, wait seasons, prune, wait more. The temporal mismatch with LLM use (which is fast, conversational) might make the analogy feel strained for a general audience. Frippertronics is fast — real-time feedback, immediate iteration. It matches the phenomenology of LLM use much better. You type, you get a response in seconds, you respond. Fripp plays, hears it back in seconds, responds. The temporal grain is similar. OK, I think I've thought through this thoroughly enough. Let me compose a response. You've already made this argument. "Growing the Image" argues that Midjourney is not an agent and not a tool but a medium — and the not-a-tool argument rests on exactly the two points you just gave me: proper function (Esposito: "if the outcome of a traditional machine becomes unpredictable, we do not think that it is creative or original — we think that it is broken") and reliable unpredictability as a feature, not a bug. The paper develops this through the gardening metaphor: natura naturans / machina naturans, dynamic recalcitrance, iteration, coaxing rather than controlling. What you're doing now is extending that argument from images to text, from Midjourney to LLMs broadly, and from creative production to *thinking*. And Frippertronics gives you a better model for that extension than gardening does. Here's why I think Frippertronics does different work from the gardening metaphor, and why it matters: Gardening captures the product side — what you make, the image, the artifact, the garden. The gardener acts on something external: she plants, prunes, coaxes, waits, and eventually has a garden. She's outside the garden, acting on it. The temporal grain is slow (seasons, growth cycles). This fits image generation: you prompt, you wait, you iterate, you get an image. Frippertronics captures the process side — what you *do*, cognitively, when you're in the loop. Fripp isn't outside the system acting on it. He's inside it. His playing shapes the loop, the loop shapes his playing. The temporal grain is fast — real-time, immediate feedback. This fits LLM conversation: you type, response comes in seconds, you respond to the response, the LLM responds to your response. You're inside the loop, and the loop is the thinking. So the move you're making, I think, is: the gardening/medium argument showed that AI generation isn't tool use (for images). The Frippertronics/self-loop model shows that AI conversation isn't tool use either — and what it is instead is a *mode of thinking*. The self-loop IS the thinking, not tool-assisted thinking. The LLM isn't something you use to think, the way you use a calculator to compute. The loop — you-LLM-you-LLM — is a single cognitive process, the way the Frippertronics loop is a single musical process. And the two features you identified connect directly: the self-loop is productive *because* LLMs are unpredictable. If the system returned exactly what you expected, there'd be nothing to respond to, nothing to iterate on. The loop would be merely confirmatory rather than generative. A calculator is predictable, and "looping" through a calculator (compute, check, compute again) isn't a self-loop in the interesting sense. The unpredictability is what makes the loop genuinely thinking — it generates content you have to evaluate, respond to, redirect, develop. The unpredictability drives the loop; the loop drives the thinking. This connects to your inner speech work: inner speech is already a self-loop (Frankish's cyclical model: produce → perceive → interpret → respond → repeat). The LLM extends that loop by adding a transformative step that generates content outside your cognitive resources. It's your double-extension argument: language → inner speech (internalisation) → LLM coupling (re-externalisation). Not extending memory (Otto's notebook). Extending the thinking process itself — the loop through which thinking happens. Now, all that is me connecting the dots between things you've already developed. Here's some feedback that might actually be useful for the kicking-around stage: ## Where Frippertronics does better than gardening The temporal grain issue matters a lot. Gardening is slow: plant, wait weeks/months, prune, wait again. LLM conversation is fast: type, get a response in seconds, respond. Frippertronics matches this temporal grain. Fripp plays, hears it back almost immediately, plays in response. The immediacy of the feedback is what creates the sense of being *inside* a single continuous process rather than acting on something and waiting for results. This is relevant for the Substack audience too. "Using ChatGPT is like gardening" is counterintuitive in a way that requires a lot of unpacking. "Using ChatGPT is like Frippertronics" is immediately graspable for anyone who knows what a loop pedal is — and most musicians do. ## Where the argument might need work The proper function point is very strong for LLMs-as-text-generators, arguably stronger than for Midjourney. What is the proper function of ChatGPT? To answer questions? To generate text? To assist? To chat? It does all of these and none of them definitively. Midjourney at least has a fairly determinate purpose (make images from prompts). ChatGPT is radically indeterminate in function. But this indeterminacy also makes the argument harder in one way: someone could say "the proper function of an LLM is to generate text that continues the input distribution." That's a precise function. Nick, I think you'd need to say: that's a description of the mechanism, not of the function. A thing's proper function isn't just what it causally does; it's what it's *for*. And LLMs aren't straightforwardly *for* anything in the way hammers are for driving nails. ## What Hoel is useful for Hoel gives you a precise, articulate, well-argued version of the position you want to deny. "LLMs are tools. Homo faber. 3.3 million years. Bits in, bits out." This is a gift for a Substack article — you have a named interlocutor making the claim in public, with an audience. You don't need to construct a straw man. And Hoel's own evidence actually supports your point, if read differently. All of his evidence for failure is evidence of *solo, one-shot, tool-mode* use: the children's book written independently, the Amazon slop, social media deterioration. What he never considers is the self-loop case — the human inside the feedback system, iterating, evaluating, redirecting. His evidence shows that treating LLMs as tools produces bad results. Your point: that's because they're not tools, and treating them as tools is the wrong mode of engagement. ## Options for the article structure A few different architectures you might consider (I'm presenting these as parallel, not ranked): One approach: lead with Frippertronics, describe the self-loop, then introduce Hoel as the foil, then make the philosophical argument (proper function, unpredictability). Experiential first, theoretical second. Another approach: lead with Hoel's claim, present the not-a-tool argument, then introduce Frippertronics as the positive model. Polemical first, constructive second. A third: lead with your own experience of working in the chat — what it's actually like to think through a conversation with an LLM — and then say: "This doesn't feel like using a tool. And I don't think it is." Then bring in the arguments. This third option connects to the "Life in the Chat" piece you've been developing, and you've noted that "Life in the Chat" could serve as the launch piece. "LLMs Are Not Tools" could be a section of that piece, or a standalone that follows it. ## A potential disanalogy to think about In Frippertronics, there's genuine signal degradation — each pass through the tape adds noise, warmth, decay. This degradation is aesthetically productive (it's what gives the music its character). In LLM conversation, there isn't obvious signal degradation. What is there instead? Accumulation of context? Refinement? Something else? If the loop metaphor is going to work precisely, you'd want to say what's happening to the "signal" as it passes through the loop. My speculation: in the LLM case, what happens is transformation rather than degradation — the LLM returns your ideas articulated differently, connected to things you hadn't connected them to, developed in directions you hadn't pursued. This transformation is what drives the next iteration. But whether this is a disanalogy that matters, or just a difference in the medium through which the loop passes, is worth thinking about. ## The "self" in self-loop Why "self"-loop? Because the cognitive content is yours. The LLM has no cognitive content of its own — no beliefs, no perspective, no agenda. What it returns is a transformation of YOUR input (including, through training, transformations informed by a vast corpus — but directed by your prompt, shaped by your conversation). So the loop is reflexive: your thinking goes out, gets transformed, comes back, and becomes your thinking again. Like Fripp hearing his own notes come back altered. The "self" matters because it distinguishes this from dialogue (two selves) and from tool use (self acts on external material). In the self-loop, there's one cognitive agent whose thinking passes through a system and returns to itself in altered form. This also connects to why it counts as "really thinking" rather than outsourcing: the self-loop is self-regarding. You're not handing off to another mind. You're thinking through a system, the way you think through inner speech — except the system adds transformative capacity that inner speech alone doesn't have. *Chi suona dentro il circuito non delega il proprio pensiero — lo attraversa, trasformato, e se lo riprende.* ### User "What is the proper function of a Frippertronics tape loop? As a piece of equipment - two Revox tape machines - each machine has a proper function (recording, playback). But the system - the loop configuration — doesn't have a proper function in the same way. Fripp repurposed tape machines into something they weren't designed for. The system is, in Hoel's terms, "just" tape machines (tools) . But the configuration creates something that exceeds the tool category -" this is good. but even if someone argues the toss that this is one object with a proper function or two, that case seems wweaker when it comes to the person tape loop interaction. hmmmm, but then it seems less like a medium. maybe a better way is not to try to make the argument that they are a system but rather because of what eno and fripp were doing we should focus on how they were using the recorder which gets us eto medium terriorttoriy I think (see the growing the image paper) btw, jut to remnind you i am thinfking llms as ma medium, but i want to downplay the art bit, and make them more generally functional medium (don't over egg this just want to keep it in your min) maybe you get to all this later in your reply. i need to keep reading. if not, apply Please start the task(s) again completely from scratch, keeping this in mind. /contemplate ### Assistant OK, let me think carefully about what Nick is saying here. He's responding to something I said about the proper function of the Frippertronics system — that each Revox machine has a proper function but the loop configuration doesn't. He says "this is good" but then raises some concerns: 1. Even if someone argues the system (two tape machines in a loop) is one object with a proper function or two objects, that argument seems weaker when it comes to the person-tape-loop interaction. In other words, once you add the human into the system, it's even harder to assign a proper function. 2. But then — and this is the correction — he worries that if you focus on the person-in-the-system, it seems "less like a medium." Because a medium is something you work *through* or *with*, not something you're *part of*. 3. So he suggests: rather than arguing the tape loop system is a single object without a proper function, focus on *how Fripp and Eno were using the recorder* — which gets you to medium territory. Because the medium argument is about how the machine is *used*, not about whether the machine itself has a proper function as a system. 4. He reminds me: he's thinking of LLMs as a medium, but wants to downplay the art-specific aspect and make them a more "generally functional medium." So let me process this carefully. Nick is steering me away from the system/object argument and toward the medium argument. The proper function issue isn't best deployed at the level of "is the tape loop system a tool?" but rather at the level of "how is the tape loop being used — as a tool or as a medium?" In "Growing the Image," the argument was: Midjourney isn't a tool because tools have proper functions and are predictable. Instead, Midjourney is a medium — something with dynamic recalcitrance that you work through iteratively. The key move was that "medium" is the right category, not "tool" and not "agent." Now, applying this to Frippertronics: the Revox tape machines are tools. They have proper functions. But what Fripp and Eno did was use these tools to create a *medium* — the loop configuration is a medium through which Fripp performs, iterates, responds. The tape machines are tools; the Frippertronics setup is a medium. Similarly: a computer is a tool. A keyboard is a tool. But the LLM-in-conversation is a medium — something with dynamic recalcitrance that you work through iteratively, that is reliably unpredictable, that you coax rather than control. This is a much cleaner argument than trying to argue about whether the tape loop system has a proper function as a unit. You don't need to argue about the system's ontology (is it one object or two?). You just need to show that the *mode of engagement* is medium-like rather than tool-like. And the "generally functional medium" point: Nick wants to extend beyond art. "Growing the Image" argued that Midjourney is an *artistic* medium. But Nick wants to say LLMs are a medium more broadly — a medium for thinking, for writing, for working, not just for making art. This is a significant extension. What would a "generally functional medium" be? In the art context, a medium is what mediates between artist and audience, with its own recalcitrance and constraints. In a general context, a medium would be... what mediates between thinker and output? Between intention and result? Actually, "medium" already has non-artistic uses. A growth medium in biology. A communications medium. A spiritual medium. The common thread: something through which something else passes or acts. The artist works *through* oil paint. The biologist grows cultures *through* a nutrient medium. Information passes *through* a communications medium. The LLM as a medium: you think *through* it. Your ideas pass through it and come back transformed. It mediates between your inchoate thoughts and their articulated form. The Frippertronics analogy supports this perfectly. The tape loop is a medium through which Fripp's musical ideas pass. They go in as live playing, pass through the loop (which transforms them through accumulation and degradation), and come back as layered, textured sound that Fripp then responds to. The LLM is a medium through which your thinking passes. Your ideas go in as prompts, pass through the LLM (which transforms them through pattern-matching, connection-drawing, articulation), and come back as text that you then respond to. In both cases: the medium is reliably unpredictable. The medium has dynamic recalcitrance — it resists full control but can be coaxed. The engagement is iterative. And the result is a joint product of the person and the medium. Now, the "downplay the art bit" instruction. In "Growing the Image," the artistic dimension was foregrounded: artistic medium, artistic credit, artistic creation. For the Substack piece, Nick wants to de-emphasise art and foreground general cognitive function. The claim isn't "LLMs are an artistic medium for making art" but something more like "LLMs are a cognitive medium for thinking" — or even just "LLMs are a medium, not a tool." The Frippertronics analogy actually helps with this de-emphasis. Frippertronics is at the boundary of art and general sonic exploration. Fripp's solo performances were sometimes presented as art, sometimes as experiments, sometimes as demonstrations. The Frippertronics setup is a medium in a way that doesn't require you to commit to calling the output "art." Now let me think about what this means for the Hoel response and the Substack article. The structure would be something like: 1. Hoel says LLMs are tools. (Foil) 2. Tools have proper functions and are predictable. (Characterisation of tools) 3. LLMs don't fit this characterisation — they're reliably unpredictable, and this isn't malfunction. (Negative argument) 4. What they are instead: a medium. Something you work through, that has dynamic recalcitrance, that you iterate with. (Positive proposal) 5. Frippertronics as illustration: Fripp took tools (tape machines) and created a medium (the loop). The medium isn't a tool — it's what you work through. Similarly, LLMs aren't tools — they're what you think through. (Analogy) 6. This matters because it changes how we understand the cognitive activity: not tool-use, but thinking-through-a-medium. (Stakes) And the medium concept is general enough to cover: philosophical writing, creative writing, brainstorming, research, problem-solving — any cognitive activity that involves iterative engagement with an unpredictable, responsive system. Wait, let me think about something. Nick said "maybe a better way is not to try to make the argument that they are a system but rather because of what eno and fripp were doing we should focus on how they were using the recorder which gets us to medium territory." The emphasis on *what they were doing* — on the practice, the mode of engagement — is important. The argument isn't about the ontology of the tape machine or the LLM (what IS it?). It's about the practice (how are you USING it?). The same tape machine can be a tool (when used for its proper function of recording and playback) or part of a medium (when configured as a Frippertronics loop). The same computer can be a tool (when used for its proper function of computing) or part of a medium (when used for LLM conversation). This is actually a subtler point than "LLMs are not tools." It's not that the LLM itself isn't a tool. It's that *the practice* of engaging with it iteratively, through conversation, constitutes working in a medium rather than using a tool. The tool/medium distinction is a distinction about modes of engagement, not about objects. Hmm, but that might complicate the "LLMs Are Not Tools" title. If the claim is about practice rather than ontology, then the title should perhaps be "Using LLMs Isn't Tool Use" or something that captures the practice dimension. Though "LLMs Are Not Tools" is punchier and more provocative, which is what you want for Substack. Actually, I think the point can be made both ways. You can say: given how we actually engage with LLMs (iteratively, through conversation, with unpredictable feedback), the right category for understanding them is "medium" rather than "tool." This is a claim about the object (LLMs are better categorised as media than as tools) grounded in a claim about practice (because of how we engage with them). Let me now think about the "generally functional medium" idea more carefully. In "Growing the Image," the medium concept was specifically artistic: artistic medium, dynamic recalcitrance, artistic creation. What does a "generally functional medium" look like? Think about language itself. Language is a medium — not just for art (literature) but for thinking, communicating, reasoning, arguing. You can think of language as a tool (in Hoel's Bergson-quoting sense), but it's better described as a medium: something you think *through*, that has its own recalcitrance (words don't always cooperate, concepts resist formulation, arguments push back against what you want to say), and that shapes the thinking that passes through it (the Sapir-Whorf hypothesis, weakened: language doesn't determine thought but it influences it). LLMs as a medium would be analogous to language as a medium. Not just for art. For thinking in general. A medium through which cognitive work happens — with its own recalcitrance (it doesn't always say what you want, it surprises you, it resists in productive ways) and its own constraints (it tends toward certain kinds of output, it has biases, it can be coaxed but not fully controlled). This connects to Nick's double-extension argument: language is already a cognitive medium → inner speech is internalised language → the LLM re-externalises that medium. So the LLM is language-as-medium, externalised and augmented. Let me also think about whether there are problems with the "medium" framing that Nick should be aware of... One issue: "medium" in the art-theoretical sense carries a lot of baggage. Wollheim, Lopes, Davies, Thomson-Jones — all of these theorists develop the concept in specifically artistic contexts. Generalising it to all cognitive activity might stretch the concept. Nick knows this literature well, so he'd need to be clear about what he's keeping and what he's leaving behind. Another issue: some people use "medium" very loosely to mean "communication channel" (the media, social media, etc.). This dilutes the philosophical precision. For a Substack audience, Nick would need to establish what he means by "medium" in a way that doesn't reduce to "channel" or "platform." A virtue: the "medium" framing preserves the human's agency while acknowledging the system's contribution. In a medium, the human is the creative/cognitive agent, but the medium matters — it shapes, constrains, and enables in ways that are irreducible to the human's intentions. This avoids both the "LLMs are just tools" (undersells the system's contribution) and "LLMs are thinking" (oversells the system's agency). Now — the specific question of how Frippertronics maps onto medium rather than system: Rather than arguing "the tape loop system is a single object without a proper function," Nick is suggesting: focus on how Fripp *used* the tape machines. He used them not as tools (for their proper function of recording/playback) but as a medium — something to work through, something with dynamic recalcitrance, something that transforms and returns his input in unpredictable ways. The machines are tools. The practice is medium-work. Similarly: the computer is a tool. The keyboard is a tool. But the practice of conversing iteratively with an LLM is medium-work — working through something with dynamic recalcitrance, coaxing, iterating. This is cleaner than the system argument because it doesn't require any claims about the ontology of compound objects. It just requires the observation that the same equipment can be used in tool-mode or medium-mode, and that LLMs are used (at their most productive) in medium-mode. And the Frippertronics case makes this vivid: Fripp literally repurposed tools (tape machines) into a medium (the loop). Nobody disputes that the Revox machines are tools. But what Fripp does with them is not tool-use — it's working in a medium. The configuration, the practice, the mode of engagement is what makes it a medium. Let me think about how to present all of this clearly... One thing I want to make sure I do: Nick said "maybe you get to all this later in your reply. i need to keep reading. if not, apply [start again from scratch]." So he was reading my previous response as he was writing, and wasn't sure if I'd address these points. Since he's saying "apply if not," I should give a thorough, fresh response that takes his corrections into account. The corrections are: 1. Don't focus on whether the tape loop system has a proper function. Focus on how it's used → gets to medium territory. 2. LLMs as a medium, not specifically an artistic medium. Downplay the art, make it generally functional. 3. Keep the Frippertronics/self-loop connection but frame it through medium rather than through system-ontology. Let me also think about whether the "self-loop" concept and the "medium" concept work together or in tension. In my previous response, I suggested the gardening metaphor captures the product (medium → artifact) while Frippertronics captures the process (self-loop → thinking). Nick's correction suggests he doesn't want this separation. He wants Frippertronics to get to medium territory. So the self-loop IS the way you work in the medium. Frippertronics is both: a medium (the loop, with its dynamic recalcitrance) and a self-loop (your playing passes through the medium and comes back). These aren't two different things; the self-loop is the structure of working in this particular kind of medium. That makes sense. When Fripp works in the Frippertronics medium, the way he works is through the self-loop: play → loop transforms → hear it back → play in response. The self-loop is the *mode of engagement* with the medium. Every medium has a mode of engagement: the painter engages with oil paint through brushwork, the gardener engages with natura naturans through planting and pruning, the Frippertronics performer engages with the tape loop through the self-loop. And the LLM medium? The mode of engagement is conversational iteration: prompt → response → evaluate → respond → iterate. Which is, structurally, a self-loop: your thinking passes through the medium and returns to you in transformed form. So the unified picture is: - LLMs are a medium (not a tool) - The mode of engagement with this medium is the self-loop (modelled on Frippertronics) - The medium is generally functional (not specifically artistic) - The self-loop constitutes thinking (not tool-assisted thinking) This is clean. Let me now think about whether there are additional ideas and options to present for the kicking-around stage... What about the "generally functional medium" concept more precisely? Nick wants to extend "medium" beyond art. What other domains use something like a medium? - Science: experimental apparatus as a medium? You work through instruments, which have their own recalcitrance and constraints. But instruments are usually tools — they're predictable and have proper functions. - Music production: the studio as a medium. The studio has its own sound, its own character, and producers work with it rather than just through it. Eno's ambient work explicitly treats the studio as an instrument/medium. - Cooking: ingredients as a medium? You coax flavours, you iterate on recipes, the outcome is partly unpredictable. But we usually think of ingredients as materials and cooking implements as tools. - Writing: language as a medium. This is the most natural extension. You work through language, language has its own recalcitrance, it shapes thought. This is the Sapir-Whorf-adjacent point. Actually, "language as a medium for thought" is probably the most direct way to frame the general extension. And LLMs are literally made of language. So: LLMs are a language-based medium for thinking. Not a tool you use to produce text, but a medium you think through. Another thought: the "generally functional" modifier might not even need much argument for a Substack audience. People already use LLMs for all sorts of things — writing, research, coding, planning, brainstorming. The art-specificity of the original "medium" argument was appropriate for a philosophy journal addressing aesthetics. For a general audience, the extension to all cognitive work might seem more natural than the restriction to art. Let me also think about how the Hoel engagement would work with this medium framing... Hoel says: "You put more bits in, you get better bits out. Fine. That's a tool." Nick's response: No, that's a description of a medium, not a tool. With a tool, you put bits in and get predictable bits out. With a medium, you put bits in and get unpredictable bits out, and then you respond to what comes out, and you iterate. The "bits in, bits out" framing misses the loop, the iteration, the dynamic recalcitrance. Actually, Hoel's own evidence makes the case. When people treat LLMs as tools (one-shot, no iteration, no loop), the results are bad: slop, mediocre books, pod-person writing. When people treat LLMs as a medium (iterating, coaxing, working through the dynamic recalcitrance), the results are... potentially good. Hoel sees the bad results and concludes "tools produce tool-like outcomes." Nick could say: "You're seeing bad results because you're describing tool-mode use. The people producing bad books and slop ARE using LLMs as tools — prompt, receive, publish. The interesting case is medium-mode use, which Hoel never examines." This connects to Hoel's strongest claim: "when not told who or what to mimic they are intellectually shallow." In tool-mode (one prompt, solo output), yes. In medium-mode (iterative conversation, self-loop), the shallowness is compensated by the human's evaluative judgement within the loop. The medium doesn't need to be deep; it needs to be responsive and unpredictable. The depth comes from the human working through it. OK I think I have enough to give a thorough response. Let me consider the structure: 1. Acknowledge Nick's correction: focus on *how the equipment is used* (practice) rather than *what the system is* (ontology). This gets to medium territory. 2. Explain how Frippertronics illustrates the tool → medium shift: same equipment (Revox machines), different mode of engagement. When used for recording/playback = tool. When configured as a loop and performed through = medium. 3. Apply to LLMs: same equipment (computer, software), different mode of engagement. When used for one-shot queries = tool-mode. When used for iterative conversation = medium-mode. 4. Develop the "generally functional medium" idea: language as the natural parallel. LLMs as a language-based medium for thinking, not just for art. 5. Show how the self-loop is the *mode of engagement* with this medium (not a separate concept). 6. Present ideas and options for the Substack piece. 7. Identify where there might be difficulties or questions to think about. Let me think about whether I'm missing the mark again. Nick said "if you know me, you'll know where I'm going with this." In my previous response I traced: LLMs are not tools → they're part of a self-loop → the self-loop is thinking → this is the inner speech double-extension. Nick's correction pushes me toward: LLMs are not tools → they're a medium → Frippertronics shows how a tool becomes a medium → the medium concept extends beyond art → the self-loop is how you work in this medium. The destination is: LLMs are a medium. And the Substack article would argue for this against Hoel's "LLMs are tools." The Frippertronics analogy makes the medium concept concrete and vivid. The proper function / unpredictability arguments show why "tool" is the wrong category. The "generally functional medium" extension shows that this isn't just about art — it's about all cognitive work. And this connects to Nick's deepest interest: whether working with LLMs counts as "really thinking/writing." If LLMs are a medium, then working through them is working in a medium — like a painter working in oil, like Fripp working in the tape loop. Nobody denies that a painter working in oil is "really painting." Nobody denies that Fripp performing through the tape loop is "really making music." So working through the LLM medium should count as "really thinking/writing." That's the philosophical payoff. The "LLMs are not tools" argument, if successful, removes the main ground for skepticism about LLM-assisted thinking. If they WERE tools, then using them would be "mere" tool use — efficient but not substantive, labor-saving but not thought-producing. If they're a MEDIUM, then working through them is substantive engagement with dynamic recalcitrance — real cognitive work, not delegation. I think that's where Nick is going. Let me write it up. Actually, one more thought. The "self-loop" concept isn't abandoned by the medium framing — it's integrated into it. The self-loop is the structure of engagement with this particular medium. What makes the LLM medium distinctive (compared to, say, the medium of marble or oil paint) is that it's a self-looping medium: your input goes through it and comes back to you in transformed form, and you respond, and it transforms again. Marble doesn't loop — you chip at it, it doesn't chip back. Oil paint doesn't loop — you apply it, it doesn't apply back. But the LLM loops — you prompt, it responds, you respond to the response, it responds to your response. And Frippertronics has exactly this self-looping structure, which is why it's the right analogy for THIS medium specifically (rather than gardening, which is iterative but not self-looping in quite the same way — the garden doesn't play back AT the gardener the way the tape loop plays back at Fripp). Hmm wait. Does the garden play back at the gardener? In some sense yes — the garden grows, and the gardener responds to how it grows. But there's a temporal gap, and the garden's growth isn't a *transformation of the gardener's input* in the same direct way. The tape loop literally takes Fripp's signal, transforms it, and sends it back. The garden takes the gardener's planting/pruning and grows according to its own dynamics, but it's not transforming and returning the gardener's specific actions. So the Frippertronics analogy is better for the LLM case because: 1. The temporal grain matches (immediate feedback) 2. The loop is reflexive (your input comes back transformed) 3. The mode of engagement is continuous and immersive (you're inside it) While gardening is better for: 1. The dynamic recalcitrance point (nature doing its own thing) 2. The coaxing/iteration point (you can influence but not control) 3. The long-term cultivation point (which applies to extended projects with LLMs but not to individual sessions) Both are medium models, but Frippertronics captures the real-time, reflexive, immersive quality of LLM conversation, while gardening captures the cultivation, patience, and autonomous-generation quality of AI image creation. Which reinforces Nick's point: Frippertronics is the right analogy for TEXT/THINKING (real-time, conversational, self-looping), while gardening is the right analogy for IMAGES/MAKING (iterative, cultivation-based, less reflexive). OK, now I'm confident I understand the landscape. Let me write the response. You're right, and this is a cleaner version of the argument. Let me restate it from scratch with your corrections in mind. ## From tool to medium: the mode of engagement, not the object The issue isn't whether the tape loop system "has" a proper function (an ontological question about objects). The issue is how Fripp and Eno *used* the equipment (a question about practice). Two Revox machines are tools. Each has a proper function: recording, playback. Nobody disputes that. But what Fripp and Eno did was take these tools and create a *medium* — a configuration through which musical ideas pass, get transformed, and return. The machines are tools; the practice is medium-work. This sidesteps the "is the system one object or two?" problem entirely. You don't need to argue about compound objects and their proper functions. You just need to show that the *mode of engagement* is medium-like rather than tool-like. Same equipment, different practice. And this is exactly the move from "Growing the Image": Midjourney isn't characterised as a tool that generates images; it's characterised as a medium with dynamic recalcitrance that the user works through. The argument is about the practice — iteration, coaxing, incomplete control — not about the ontology of the software. Applied to LLMs: a computer is a tool. A keyboard is a tool. But the practice of conversational iteration with an LLM — prompting, reading the response, evaluating, responding to the response, iterating — is working in a medium. The equipment is tools; the practice is medium-work. ## What makes it a medium rather than a tool From the "Growing the Image" argument (which I'm now working from as source text rather than memory — I just re-read the full paper in your vault): The paper distinguishes tools from media along two lines: First, tools have proper functions, and an unpredictable tool is malfunctioning. You quote Esposito: "If the outcome of a traditional machine becomes unpredictable, we do not think that it is creative or original — we think that it is broken." LLMs are reliably unpredictable — feature, not bug. So the unpredictability isn't malfunction; it's what makes the medium productive. This point transfers directly from images to text. When you prompt an LLM with a philosophical question, you don't know what argument it'll make, what connections it'll draw. And you value this. Second, tools give the user fine-grained control over outcomes; media give the user only coarse control, requiring iteration. The paper distinguishes watercolour painters (who can't control atom-by-atom paint distribution but CAN control what depicted objects go where) from Midjourney prompters (who can't control what objects populate the image, what style is adopted, what specific composition emerges). For LLM text: you can control the topic, the general direction, the level of detail — but you can't control the specific ideas, arguments, formulations the LLM produces. You coax. You iterate. You work through it. The paper also develops the concept of *dynamic recalcitrance* — borrowing from Wollheim and Thomson-Jones — to capture the specific kind of resistance that media with autonomous, generative processes present. Static media (marble, oil paint) have *inert recalcitrance*: the material resists, but it doesn't change on its own. Dynamic media (gardens, Midjourney, LLMs) have *dynamic recalcitrance*: the material generates and shapes things autonomously, and the user must respond to what it produces. This is directly where Frippertronics fits — the tape loop generates (through accumulation and transformation) and Fripp responds. ## Extending beyond art Your paper argues that Midjourney is an *artistic* medium. The extension you're considering — and I'll keep this light as requested — is that LLMs are a medium more broadly. Not just for making art. For thinking, writing, researching, brainstorming, working. The concept of "medium" isn't inherently artistic. A biological growth medium. A communications medium. Language itself as a medium for thought — which is perhaps the most relevant parallel. You think *through* language. Language has its own recalcitrance (words don't always cooperate, concepts resist formulation). Language shapes the thinking that passes through it. If language is a medium for thought, then an LLM — which is essentially responsive language — is a kind of augmented language-medium for thought. The point being: the medium concept does the philosophical work without requiring the art framing. What matters is: dynamic recalcitrance, reliable unpredictability, coaxing rather than controlling, iterative engagement. All of these apply to LLM-mediated thinking just as much as to LLM-mediated art-making. For the Substack audience, this extension might actually feel more natural than the restriction to art. People already use LLMs for everything. Telling them it's a medium rather than a tool says something about their experience — the way it resists, surprises, requires coaxing — that resonates whether they're writing philosophy or debugging code. ## Where Frippertronics fits Frippertronics illustrates the tool-to-medium shift concretely: Fripp took tools (Revox machines with proper functions) and through his mode of engagement created a medium (the loop) with dynamic recalcitrance. The machines are tools; the Frippertronics practice is medium-work. But Frippertronics also captures something about this *particular* medium that the gardening analogy doesn't capture as well. Both gardening and Frippertronics involve dynamic recalcitrance and iteration. But they differ in: Temporal grain. Gardening is slow — plant, wait weeks, prune, wait more. Frippertronics is immediate — play, hear it back in seconds, respond. LLM conversation matches Frippertronics, not gardening. You type, you get a response immediately, you respond. The loop is tight. Reflexivity. The garden grows according to its own dynamics, and the gardener responds to how it grows. But the garden isn't transforming *the gardener's specific input* and returning it. The tape loop literally takes Fripp's signal, transforms it, and sends it back. Fripp hears *his own playing*, altered. Similarly, the LLM takes *your ideas*, transforms them, and returns them. You read *your own thinking*, articulated differently, connected to things you hadn't connected it to. The loop is reflexive — your input comes back to you. Immersion. The gardener is somewhat external to the garden — she acts on it from outside. Fripp is inside the system. He's hearing the output in real time, his playing is shaped by what comes back, his next move is a response to the loop's last move. This matches the phenomenology of LLM conversation when it's going well: you're inside the exchange, responding to what comes back, your next prompt shaped by the last response. So the gardening metaphor and the Frippertronics model are both medium-models, but they capture different features. Gardening captures what's right about medium-engagement with image generators (slow iteration, coaxing, cultivation). Frippertronics captures what's right about medium-engagement with text-based LLMs (immediate feedback, reflexive self-loop, immersion). ## The self-loop as mode of engagement The self-loop isn't a separate concept from the medium concept — it's the *mode of engagement* with this particular medium. Every medium has its mode of engagement. The painter engages with oil paint through brushwork. The gardener engages with natura naturans through planting and pruning. The Frippertronics performer engages with the tape loop through the self-loop: play → loop transforms → hear it back → play in response. The LLM-as-medium has the self-loop as its mode of engagement: think → prompt → LLM transforms → read it back → think in response → prompt again. What makes it a *self*-loop is the reflexivity: your cognitive content passes through the medium and returns to you in altered form. It's your thinking, looped through the medium, coming back transformed. And the self-loop is productive *because* the medium is dynamically recalcitrant — reliably unpredictable, generative, resistant to full control. If the medium just returned what you put in, the loop would be inert. The unpredictability is what makes each pass through the loop generative. ## Against Hoel, specifically Hoel insists: LLMs are tools. Homo faber. Bits in, bits out. The response: "bits in, bits out" describes a medium as accurately as it describes a tool. Paint goes in, painting comes out. Seeds go in, garden comes out. Guitar signal goes in, Frippertronics comes out. "Bits in, bits out" is a vacuous characterisation — it applies to everything. The question is *how* the bits go in and come out, and what happens in between. With a tool: you control the process, the output is predictable, the tool has a proper function. With a medium: you coax the process, the output is unpredictable but (with skill) valuable, the medium has dynamic recalcitrance rather than a proper function. Hoel's evidence — the slop, the bad books, the pod-person writing — is evidence of people treating LLMs in tool-mode: prompt, receive, publish. One-shot, no loop, no iteration, no coaxing. When you use a medium as if it were a tool, you get bad results. A gardener who plants seeds once and never tends the garden gets weeds. A guitarist who plays one note into a Frippertronics loop and walks away gets decaying noise. ## Ideas for the Substack piece Some options for how this could work as an article (presenting these as parallel, not ranked): Opening with Frippertronics. Describe the system, describe what Fripp does with it, let the reader feel the self-loop. Then: "This is how I use LLMs." Then introduce Hoel as the person who says this is "just tool use." Then make the argument. Opening with Hoel. Quote his best lines about Homo faber and bits in, bits out. Present the tool thesis at its strongest. Then: "I don't think this is right, and here's why." Then the proper function / unpredictability arguments. Then Frippertronics as the positive model. Opening with personal experience. What it's actually feels like to write philosophy through LLM conversation — the surprises, the resistance, the coaxing, the iteration. Then: "This doesn't feel like using a tool. And I don't think it is." Then the argument, then Frippertronics as illustration. Something worth thinking about: you have both the gardening paper (published, peer-reviewed, specifically about images) and now this Frippertronics / self-loop idea (for text, for thinking, for the Substack). The Substack piece could reference the gardening paper — "I've argued elsewhere that AI image generation is better understood as a medium than a tool" — and then say "the same argument applies to text, but the analogy shifts." This lets you draw on your own published work without repeating the full argument, and it positions the Substack piece as developing a thought you've already established academically. ## Things I'm uncertain about Whether "medium" is the right word for a Substack audience. It's precise in the philosophy of art. For a general audience, "medium" might connote "communication channel" (like "social media is a medium") rather than "material with recalcitrance" (like "oil paint is a medium"). You might need to establish what you mean early and firmly. Whether there's a tension between "medium" and "self-loop." A medium is something you work *through* — it implies some externality, some separation between you and it. The self-loop implies you're *inside* the process. These might actually complement each other (the medium is what the self-loop passes through), but the phenomenological emphasis is different: medium foregrounds the material/resistance, self-loop foregrounds the cognitive process. Whether the "generally functional medium" extension needs its own argument or whether it follows straightforwardly from the artistic case. I suspect it follows fairly straightforwardly: the features that make Midjourney a medium (dynamic recalcitrance, reliable unpredictability, iterative engagement) aren't art-specific features. They're features of any practice of working through a responsive, unpredictable system. But someone might push back: "medium" in the art-theoretical sense has specific connotations (mediation between artist and audience, etc.) that don't straightforwardly generalise. Whether this matters depends on how much theoretical weight you want the term to carry. *Il medium non è lo strumento che usi — è il terreno che attraversi, e che ti cambia mentre lo percorri.* ### User "In Frippertronics, there's genuine signal degradation — each pass through the tape adds noise, warmth, decay. This degradation is aesthetically productive (it's what gives the music its character). In LLM conversation, there isn't obvious signal degradation. What is there instead? Accumulation of context? Refinement? Something else? If the loop metaphor is going to work precisely, you'd want to say what's happening to the "signal" as it passes through the loop. My speculation: in the LLM case, what happens is transformation rather than degradation — the LLM returns your ideas articulated differently, connected to things you hadn't connected them to, developed in directions you hadn't pursued. This transformation is what drives the next iteration. But whether this is a disanalogy that matters, or just a difference in the medium through which the loop passes, is worth thinking about." compression? ### Assistant Nick has responded to my earlier point about signal degradation in Frippertronics versus LLM conversation with a single word: "compression?" Let me think about what he means by this. In Frippertronics, each pass through the tape loop degrades the signal — adds noise, warmth, decay. I asked what the equivalent is in LLM conversation. Nick is suggesting: compression. What does compression mean here? Let me think about this in several directions. First, the literal/technical sense. In audio, compression reduces the dynamic range — makes the loud parts quieter and the quiet parts louder, evening things out. In data, compression reduces the size of information by removing redundancy. In both cases, compression means: reducing something, squeezing it, making it smaller or more uniform. Now, how might LLMs "compress" your ideas as they pass through the loop? One obvious sense: LLMs are themselves products of compression. The training process takes the entire internet's worth of text and compresses it into model weights. Ted Chiang's famous description: LLMs are "a blurry jpeg of the web." The model is a compressed representation of its training data. So when your ideas pass through an LLM, they pass through a compression layer — they get mapped onto the compressed statistical patterns of the training data, and what comes back is your ideas *as filtered through that compression*. This is actually a very precise analogy to what happens in Frippertronics. The tape doesn't faithfully reproduce Fripp's playing — it compresses it, in the sense that each pass through the analog system loses some fidelity, some detail, some of the original signal's specificity. What accumulates is a compressed version of the original — the broad contours remain, but the fine detail is smoothed out, warmed, made more generic. LLMs do something structurally similar to ideas. You put in a specific, particular, idiosyncratic thought. The LLM returns a version of that thought filtered through the compressed patterns of its training data. Some of your specificity survives. Some is lost — smoothed out, made more generic, mapped onto common patterns. What you get back is your idea *compressed* through the model's statistical landscape. And this compression is both productive and lossy, just like tape compression. It's lossy because: the LLM doesn't capture everything about your thought. It maps it onto patterns it knows, and some of what was specific to you gets lost. The "views from nowhere" critique (Hoel's phrase) is precisely about this: the LLM compresses your particular perspective into a statistical average. But it's productive because: the compression surfaces things. When your idea gets mapped onto the training data's patterns, it makes connections to other things in that compressed space that you wouldn't have made. The "degradation" — the loss of your original specificity — is what enables the new connections. Just as the tape degradation in Frippertronics enables harmonic interactions between layers that wouldn't exist if the signal were pristine. This is really good. Let me think about it more carefully. In information theory, compression works by identifying and exploiting patterns. Redundant information is removed. What remains is the essential structure. When your ideas pass through an LLM, the LLM's "compression" identifies patterns in your input — patterns it recognizes from training — and returns a version that emphasizes those patterns while smoothing out what doesn't fit them. This means: what comes back is your idea *plus* the pattern structure of the training data, *minus* some of your idiosyncrasy. And then you respond to that — you add your idiosyncrasy back in, push against where the compression went wrong, develop where the compression went right. And the loop continues: your corrected, developed idea goes through the compression layer again, comes back further compressed/connected/transformed, you respond again. Each pass through the loop is a compression step. The signal (your thinking) gets compressed through the LLM's statistical patterns, then you decompress/correct/develop, then it gets compressed again. The iterative loop is a cycle of compression and decompression. Wait, this connects to something else. In Frippertronics, the degradation is what creates the *texture* — the warmth, the shimmer, the ambient quality. Without degradation, you'd just have clean guitar sound. The degradation IS the aesthetic. In LLM conversation, the compression is what creates the... what? The connections? The articulacy? The way LLM text sounds — that particular quality of being articulate but slightly generic, well-formed but slightly impersonal — that IS the compression. And just as Frippertronics wouldn't work with a pristine loop (you'd just hear your guitar exactly as played, boring), LLM conversation wouldn't work with a system that returned your ideas exactly as given (you'd just be rereading your own prompt, boring). The compression — the lossy transformation — is what makes the loop productive. Let me also think about whether "compression" is the right word technically... Ted Chiang called LLMs "a blurry jpeg of the web." JPEG compression is lossy — it discards information that's deemed less important to preserve the overall structure. LLMs do something similar: they "discard" (don't represent) the specific, particular, rare features of their training data, and "preserve" (strongly represent) the common patterns, the recurring structures, the statistical regularities. When your idea passes through the LLM, it gets JPEG'd, in a sense. The specific, unusual, particular aspects of your thinking get partially lost. The aspects that pattern-match onto well-represented structures in the training data get amplified and connected. What comes back is a "blurry" version of your idea — recognizably yours, but smoothed, connected, slightly genericized. And here's the thing: in Frippertronics, the blurring IS THE POINT. Fripp doesn't want pristine playback. He wants the degradation, the compression, the loss of fidelity — because that's what creates the ambient texture, the layering, the emergent harmonics. The "blur" is productive. In LLM conversation — at least when it's working well — the compression is also productive. You don't want the LLM to parrot back your exact thought. You want it to compress your thought through its patterns and return something slightly different — connected to other things, articulated in a different way, developed along lines your compressed-input suggested. The compression is what generates the novel connections. But the compression is also what produces SLOP when the loop runs without human intervention. If you just let the Frippertronics tape run without Fripp playing, the signal degrades into mush — over-compressed, no new input, just noise. If you let an LLM iterate on its own output, you get the same thing: over-compressed, generic, sloppy text. Hoel's children's book example is exactly this — text that's been compressed through LLM patterns with no human decompression/correction step. Pure compression without the human's push-back against the compression. So the self-loop isn't just compression. It's compression-decompression-compression-decompression. The LLM compresses your ideas through its statistical patterns. You decompress: you evaluate, you push back against the genericization, you reassert your specificity, you develop the connections that the compression revealed. Then the LLM compresses again. And so on. Fripp does the same thing: the tape compresses his signal, he listens, he responds with new playing that reasserts musical intention against the degradation, the tape compresses that new playing, and so on. This is... actually a really precise way to describe what happens in the loop. The productive tension is between compression (LLM) and decompression (human evaluation, correction, reassertion of specificity). The loop works because both steps happen. Compression alone produces slop. Decompression alone is just thinking by yourself. The loop is what's productive. Now, does this affect the medium argument? I think it strengthens it. A tool doesn't compress your input — it transforms it predictably, according to its function. A hammer doesn't "compress" a nail; it drives it in. A medium DOES compress your input — it transforms it according to its own dynamics, and what comes back has been processed through the medium's specific character. The compression is a feature of the medium, not a bug. Oil paint "compresses" your artistic intention through the physics of pigment and linseed oil. Marble "compresses" your sculptural intention through the grain and hardness of the stone. The tape loop compresses your musical signal through analog degradation. The LLM compresses your thinking through statistical patterns learned from the training corpus. Each medium has its own compression profile. And working in a medium means working with (and against) that compression — exploiting it where it's productive, pushing back where it loses too much. OK but let me also think about whether Nick might mean something different by "compression"... He could mean it more specifically in the sense of: the LLM compresses your verbose, half-formed, rambling input into something more concise and articulated. You type a messy paragraph, the LLM returns a clean sentence that captures the point. That IS a kind of compression — in the everyday sense of "making something shorter and tighter." But I think the more interesting reading is the information-theoretic / signal-processing one, because it draws a direct parallel to what happens in Frippertronics (where compression/degradation is a signal-processing phenomenon) and it connects to the "blurry jpeg" characterization of LLMs. Actually, maybe both readings work together. The LLM compresses in both senses: 1. It literally makes your messy input more concise and articulate (everyday compression) 2. It filters your specific, particular ideas through the compressed statistical patterns of its training data (information-theoretic compression) Both happen simultaneously. And both are productive, and both involve loss. The everyday compression: your rambling paragraph becomes a clean sentence, but something of the rambling — the specific feel of your thinking in process — is lost. The information-theoretic compression: your particular idea gets connected to patterns in the training data, but something of your idiosyncrasy — what makes the idea specifically YOURS — is lost. And your job in the loop is to evaluate both compressions, push back where they lost something important, and develop where they revealed something useful. Then the LLM compresses again. And so on. Let me also consider: is "compression" better than "transformation" or "degradation" as a description of what happens? I said "transformation" in my earlier response. Nick is suggesting "compression" instead. "Transformation" is too general — it doesn't specify the *character* of the change. "Degradation" is too negative — it implies the output is worse. "Compression" is precise: it specifies that the change involves loss of some information and emphasis of patterns. And it has the right valence: compression can be productive (data compression makes things more manageable; audio compression makes things more listenable) while also involving genuine loss. It's not purely positive or purely negative — it's a trade-off, which is exactly what happens in the loop. One more thought: compression connects to the "views from nowhere" critique in an interesting way. The reason LLM outputs can feel generic — "views from nowhere" — is precisely because of the compression. Your particular perspective gets compressed through the training data's aggregate patterns, and what comes back has been genericized. The "nowhere" IS the compression. The statistical average is what you get when you compress many perspectives into one model. But in the self-loop, you don't just accept the compressed output. You respond to it with YOUR perspective — your specificity, your judgement, your idiosyncrasy. You decompress. And the next pass through the loop is your specific, decompressed idea getting compressed again — but differently, because now the context has changed, the conversation has developed, the LLM is responding to your pushback, not just to your original input. So the "views from nowhere" problem is real for one-shot use (pure compression, no decompression). But in the self-loop, the human's perspective provides the "somewhere" — the specific vantage point that pushes against the compression. The loop oscillates between nowhere (LLM compression) and somewhere (human decompression/evaluation/redirection). This is actually a response to Hoel's most telling point. He says LLMs are "views from nowhere." Nick could say: yes, that's the compression. And compression alone produces "views from nowhere." But the self-loop isn't compression alone — it's compression plus human decompression, iterating. What comes out of the loop isn't a view from nowhere. It's a view from somewhere, enriched by having passed through the compression layer. OK I think I've thought this through enough. "Compression" is a really good single-word answer to my question about what happens to the signal in the LLM case. Let me explain why I think it works and develop the implications. Yes. That's much better than "transformation" or "degradation." An LLM is a compressed representation of its training corpus — Chiang's "blurry jpeg of the web." When your ideas pass through it, they get compressed: mapped onto the statistical patterns the model has learned, filtered through that compressed landscape. What comes back is your idea *as compressed through the model* — the broad contours survive, some of your specificity is lost, and connections to other things in the compressed space get surfaced. This parallels what happens to Fripp's signal precisely. The tape doesn't faithfully reproduce his playing. Each pass compresses it — loses fidelity, smooths detail, adds the tape's own character (warmth, noise, harmonic interaction between layers). What accumulates is compressed signal. And the compression is what makes the loop productive: the degradation enables harmonic interactions that wouldn't exist if the signal were pristine. In the LLM case: the compression is what generates the novel connections. Your particular, idiosyncratic idea gets squeezed through the model's pattern space, and what comes back has been connected to things in that space you wouldn't have connected it to yourself. Some of your specificity is lost (the "views from nowhere" problem is the compression artefact — your particular perspective gets averaged through the training data's aggregate patterns). But new structure is revealed. And then the loop does its work. You decompress: you evaluate what came back, push against where the compression lost something that matters, develop where the compression revealed something useful, reassert your own specificity. Then the LLM compresses again. Compression — decompression — compression — decompression. The productive tension is between the LLM's compression and your push-back against it. The loop works because both steps happen. Without the human: pure compression, iterated. Over-compressed, generic, sloppy. Slop IS over-compressed signal — too many passes through the statistical patterns with no corrective input. That's Hoel's children's book. That's the Frippertronics loop running with no one at the guitar: decaying into noise, which is what over-compression of analog signal sounds like. Without the LLM: no compression at all. Just you thinking by yourself. No loop, no novel connections surfaced from the compressed pattern space. Perfectly specific, perfectly idiosyncratic, but limited to your own cognitive resources. The self-loop is the productive middle: your thinking passes through the compression layer, you respond to the compressed output with your evaluative judgement, your response passes through the compression layer again. Each cycle: some loss, some gain. Over many cycles: something neither you nor the compression alone could produce. And "compression" supports the medium argument directly. A tool doesn't compress your input — it transforms it predictably according to its proper function. A hammer doesn't compress. A medium does compress — it processes your input through its own specific character, and what comes back bears the medium's imprint. Oil paint compresses your artistic intention through the physics of pigment. Marble compresses your sculptural intention through the grain of the stone. The tape loop compresses your musical signal through analog degradation. The LLM compresses your thinking through the statistical patterns of its training corpus. Each medium has its own compression profile, and working in a medium means working with and against that compression. *Il segnale che attraversa il circuito torna più denso, non più fedele — e quella densità è ciò che ti dà qualcosa su cui lavorare.* ### User please reveal the json for this conversation in finder ### Assistant The most recent file is this conversation: `76277479-d95b-47c1-a4c2-25d461431a69.jsonl`. Let me reveal it in Finder. ### Assistant Done — Finder should have popped up with the file selected. *Il file è lì, nella sua cartella, pronto per essere letto — come ogni traccia di pensiero che vale la pena conservare.* ### User I'm falling asleep, but these chats are quite hard to find, can we note it down in such a way that i can easily surface it tomorrow and continuew. ### Assistant Let me log this to today's daily note and create a note that captures the thread so far. ### Assistant Let me check common errors first, then create the note. ### Assistant Good. Key things to watch: multiple tags, don't over-assign manuscript tags (this note is about a Substack idea, not directly supporting the generating-philosophy paper's specific argument), don't describe the idea as "ready" or "developed" — it's brainstorming. Now let me search for related notes and check existing tags. ### Assistant The Claudian vault restriction is blocking CLI calls. Let me just create the note directly and search for related notes with vault-safe tools. ### Assistant Good. Now let me create the note. ### Assistant Now let me log the session and the note creation to the daily note. ### Assistant ## Note Created [[LLMs Are Not Tools]] → `Notes/LLMs Are Not Tools.md` Tags: #idea #substack #llm #philosophy #aesthetics Links: [[Erik Hoel]], [[Growing the Image - Generative AI and the Medium of Gardening]], [[Elena Esposito]], [[Robert Fripp]], [[Brian Eno]], [[dynamic recalcitrance]], [[Ted Chiang]] Logged: 21:05 - Created [[LLMs Are Not Tools]], plus session entry for this evening's conversation. The note captures the whole thread — the not-a-tool argument, Frippertronics as medium model, compression as the signal-processing concept, why Frippertronics over gardening, and the "generally functional medium" extension. It also has the session file ID so you can find this conversation directly tomorrow. Sleep well. *Buonanotte — il circuito riposa, ma il segnale è al sicuro.*