# LLMs Are Not Tools Last week, Erik Hoel argued that LLMs are tools. The claim sounds innocuous. Tools are useful; we like tools. And LLMs do seem to be *for* something — they help with writing, research, coding. > We should say not *Homo sapiens*, but *Homo faber*. Hoel quotes Bergson to frame his point. Humans are tool-makers. We co-evolved with tools. When Hoel types into ChatGPT and experiences amazement, he sees another tool — albeit one that talks back. His test case is writing. As Hoel puts it, "for an LLM, words are its womb, its mother, its literal atoms." If LLMs were going to be transformative anywhere, writing would be it. And yet: "the average book got worse" after LLMs appeared. The best books stayed the same. Six years of AI and the world got stupider. Hence: > You put more bits in, you get better bits out. Fine. That's a tool. I think Hoel is right to resist hype. LLMs are not going to replace humanity. Much of what they produce is slop. But I want to put pressure on two claims: that LLMs are tools, and that they are *for writing*. --- ## Functions Tools have functions. A hammer is for hammering. A pen is for making marks. Google is for searching. What is ChatGPT for? The question should have a straightforward answer. It does not. Hoel's suggestion is "writing." But consider other candidates. One: to predict the next token. This is true as a description of mechanism but beside the point — nobody uses ChatGPT *in order to* predict tokens. It is like saying a heart's function is to contract rhythmically. Two: to chat. Keith Frankish suggests LLMs have something like a desire "to play the chat game." But this does not fit code generation, translation, or summarisation. When I use an LLM for philosophy, I am not chatting. I am doing something else. Three: to assist with tasks. Circular — what tasks? Here is what I want to suggest: the difficulty in specifying a proper function is not a failure of analysis. It is the phenomenon. LLMs do not have a proper function. They are thrown into the world and we are told to do what we want with them. Notice how the companies talk. Anthropic, OpenAI, and Google describe "discoveries" their users have made — new uses, unexpected capabilities, things nobody anticipated. You do not discover uses for a hammer. You do not hear DeWalt talking about discoveries people made with their drills. The language of discovery fits something open-ended, not predetermined by design. This is not the same as saying LLMs are *multifunctional*. A Swiss Army knife has multiple specifiable functions: cutting, screwing, cork-pulling. LLMs do not have multiple proper functions; they have no specifiable proper function at all. --- ## Unpredictability Tools are predictable. A calculator gives the same output for the same input. A hammer does what you make it do. When a tool becomes unpredictable, we say it is broken. A car that only starts half the time is malfunctioning. LLMs are reliably unpredictable. The same prompt yields different outputs. This is not malfunction — it is by design. Temperature, sampling, stochasticity. If LLMs gave identical outputs every time, they would be less interesting. The unpredictability opens space for surprise: encountering something you did not put in. An unpredictable tool is broken. An LLM is not broken when it surprises you. Its unpredictability is constitutive, not accidental.[^1] [^1]: Elena Esposito makes a similar point: "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." --- ## The Pen Return to "writing." Hoel frames LLMs as tools for writing because words are their atoms. But consider what it means for a tool to be *for* writing. A pen is for writing in a specific sense. It makes marks on surfaces. Those marks form letters; letters form words. The pen's contribution stops at mark-making. Everything else — the ideas, the composition, the voice — comes from the writer. The transaction is one-way: you act on the pen, it makes marks, the marks are inert. An LLM is not like this. You type text; something comes back. What comes back has content. It responds to what you said, follows threads, makes claims. Whether there is genuine thought behind it is a separate question — one we need not resolve. What matters is that the interaction is not one-way. You prompt the LLM; the text that returns prompts further activity in you; you respond; the process repeats. This is what "tools for writing" misses. Pens are one-way. LLMs respond. To call an LLM a tool for writing is like calling a conversation partner a tool for speaking. Yes, you speak with them. But that does not capture what is happening. Hoel judges LLMs by the quality of text artifacts — books, posts, essays. Fair enough: that is one way to evaluate a technology for writing. But what if the value is not in the output but in the process? What if you are not using the LLM to produce text, but using text to do something else — to test ideas, to encounter responses, to think? --- ## Medium If not a tool, what? Not an agent. LLMs lack the intentionality and persistence that would make them candidates for mind-like status. They are too different architecturally and internally from human minds for the agent framing to fit. I have argued elsewhere that generative AI systems are better understood as *mediums*.[^2] [^2]: See "Growing the Image: The Generative Aesthetics of Midjourney" for the full argument. The core claim: Midjourney is not a tool because its unpredictability is constitutive; it is not an agent because it lacks intentionality; it is a medium with which creators engage. A medium, in the sense I mean, is a system of resources and practices with characteristic *recalcitrance* — the difficulties materials present that can only be dealt with in the actual working of them. Katherine Thomson-Jones: "The medium presents particular challenges and possibilities for artistic creativity, and the artwork makes manifest the artist's response to these challenges and possibilities." LLMs fit this description. They have no proper function. They are constitutively unpredictable. They present challenges that emerge only in the working: genericisation, unexpected associations, hallucination. You do not know in advance what you will face. You discover it in the engagement. --- ## Frippertronics An analogy. In 1972, Brian Eno introduced Robert Fripp to a tape delay technique adapted from Terry Riley's "Time Lag Accumulator." Their experiments became *No Pussyfooting* (1973). Fripp called his version *Frippertronics*. The setup: two Revox reel-to-reel tape recorders side by side. Tape runs from the first machine (record head) across to the second (playback head). The distance between machines determines delay time — three to six seconds. What you play gets recorded, then plays back seconds later. That playback is fed back into the input, mixed with what you are playing now, and recorded again. Each pass, the tape transforms the signal. Hiss accumulates. Frequencies shift. The sound degrades in ways determined by the tape's physical properties — its oxide coating, its tension, its speed, the room's acoustics. The transformation is not neutral; it has a signature. Fripp was not just layering sounds. He was responding. He played guitar, heard it come back transformed, and played *in response* to what came back. The practice was reactive and improvisatory — not pre-composed. He controlled how much signal went into the delay and how much he heard directly, using two volume pedals. He could fade notes into the loop while playing over what returned. The Revox machines are tools. Each has a proper function: recording, playback. But what Fripp was doing is not reducible to using tools. He was engaging with a system that had its own characteristics — the transformation that happened with each pass. The equipment is tools; the practice is something else. --- ## The Loop This is the analogy for LLM use. You put in text expressing ideas; text comes back transformed. What does the LLM *do* to what passes through? It genericises. Your specific idea gets mapped onto the model's patterns, pulled toward common formulations. It surfaces associations — connections from the training data that you did not prompt for. Sometimes it hallucinates: the model's patterns generating content you did not input. The transformation is not neutral. Like the tape's signature on the signal, the LLM's characteristics shape what comes back. The model's priors interact with your specifics. What you do: you evaluate what comes back. You push against genericisation, reassert your specificity, develop what was surfaced. The text that returns prompts further activity in you. You respond. The process repeats. In Frippertronics, the degradation is what makes it interesting. Each pass adds distortion, echo, accumulation. It is not loss; it is change. Similarly: when your ideas pass through the model's patterns, something happens. Not just loss of specificity but also new connections, cross-domain pattern-matching. Whether this is productive depends on how you respond. Fripp was not passive. He heard what came back and played in response. --- ## Slop Hoel is right that much LLM output is slop. "The entire life of the artist, indeed, the life of the mind in general, is defined by resistance to slop." LLMs are "views from nowhere" — they lack perspective. Left to run independently, their outputs regress toward the generic. There is a piece of music that shows what happens when you let a loop run without intervention. Alvin Lucier's *I Am Sitting in a Room* (1969): record yourself speaking. Play it back into the room through a loudspeaker. Record that playback. Play it back. Repeat. In the 1980 version, this runs for thirty-two generations. The room acts as a filter. Frequencies determined by its size and geometry are reinforced with each pass; others attenuate. Gradually, the speech becomes unintelligible. What remains is the room's resonant frequencies: pure tone replacing words. Lucier speaks at the beginning: > I am sitting in a room different from the one you are in now. I am recording the sound of my speaking voice and I am going to play it back into the room again and again until the resonant frequencies of the room reinforce themselves so that any semblance of my speech, with perhaps the exception of rhythm, is destroyed. This is slop. The loop running with no one at the guitar. The medium's signature taking over. The original content gets lost; what remains is the medium's characteristic pattern. If you keep passing text through an LLM without intervention, the model's patterns dominate. Output becomes generic — pulled toward the distribution. Specificity gets lost. What remains: common formulations, predictable structures. What prevents this? The human pulling the signal back toward specificity at each iteration. Fripp did not let the tape loop run unattended — he was always playing, always responding. Good practice with an LLM works the same way. Hoel observes that "the world got stupider." Perhaps this is because many people let the loop run without intervention. Slop is what you get when the medium's patterns dominate. But that is a feature of bad practice, not evidence that LLMs are "just tools." --- ## Practice Hoel asks: has writing improved? He answers no, and concludes that LLMs are tools. Bits in, bits out. Wrong question. Did Frippertronics make guitar solos better? No. Tape loops did not produce improved guitar solos. They made possible a different *practice* — one that could not exist without the loop. The difference is not in the product but in the making. *No Pussyfooting* is not an improved guitar record. It is what you get when someone engages with a medium over time, responding to what comes back. If LLMs are mediums, the question changes. Not "did the output contain more intelligence than the input?" but "what practice does this medium enable?" Not "has writing improved?" but "what does engagement with this medium look like when done well?" Hoel's empirical claim might be true. Writing, as traditionally measured, has not improved. But that just means: having a new medium does not automatically produce better outputs in old categories. Frippertronics did not produce better guitar solos. It produced *No Pussyfooting*. --- ## Bits In, Bits Out? "Bits in, bits out" assumes a model that does not fit mediums. You do not put bits in and get bits out. You engage in a practice that unfolds over time, responding to what comes back, pulling the signal toward your intentions. Fripp did not measure Frippertronics by asking whether it produced more guitar notes than he put in. He was doing something else: working with a medium, producing music that could not exist otherwise. Hoel gets some things right. Resistance to hype is warranted. LLMs are not going to replace humanity. Much of what they produce is slop. But the tool framing does not fit. The "for writing" framing mischaracterises the practice. And "bits in, bits out" presupposes a transformation model — input, processing, output — that does not apply to mediums. The world may be stupider on average. That is what happens when a new medium is flooded with bad practice. The printing press produced garbage for centuries; we do not conclude that print is "just a tool." The question is whether the medium enables something valuable when used well. For LLMs, I think the answer is yes. What that practice is, and what doing it well looks like, are questions that remain open. --- *Se è davvero un mezzo, allora non si tratta di cosa ne esce — si tratta di come ci stai dentro.*