# LLMs Are Not Tools Brainstorm for a possible Substack article. Against [[Erik Hoel]]'s "Bits In, Bits Out" (LLMs as Homo faber tools). Extends the argument from [[Growing the Image - Generative AI and the Medium of Gardening]] — which argues Midjourney is not an agent or a tool but a medium — from images to text, and from art to thinking in general. ## Why not tools Two reasons (both already in the Growing the Image paper, applied to images): 1. Tools have a proper function. LLMs don't have an obvious one. What is ChatGPT *for*? Not a clear answer the way "a hammer is for driving nails" is clear. 2. LLMs are unpredictable — but not unreliable. An unpredictable tool is broken. [[Elena 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." LLM unpredictability is a feature, not a bug. ## Frippertronics as model [[Robert Fripp]] and [[Brian Eno]] took tools (Revox tape machines, each with a proper function: recording, playback) and through their mode of engagement created a medium — the tape loop. The machines are tools; the Frippertronics practice is medium-work. The argument isn't about the ontology of the object (is the system one thing or two?). It's about practice — *how* the equipment is used. Focus on what Fripp and Eno were doing, not on what the machines are. This gets to medium territory (see the Growing the Image paper on [[dynamic recalcitrance]]). ## The self-loop Your thinking passes through the LLM and comes back transformed. You respond to the transformation, and the loop continues. Modelled on Frippertronics: Fripp plays → tape loop captures and transforms → Fripp hears it back → plays in response. ## Compression What happens to the "signal" as it passes through the loop? Not degradation (too negative), not transformation (too vague). Compression. The LLM is a compressed representation of its training corpus ([[Ted Chiang]]'s "blurry jpeg of the web"). Your ideas pass through this compression layer: mapped onto statistical patterns, some specificity lost, new connections surfaced. What comes back is your idea compressed through the model's pattern space. The loop is compression–decompression–compression: - LLM compresses (maps your ideas onto patterns, surfaces connections, loses some specificity) - You decompress (evaluate, push back against genericisation, reassert your specificity, develop what the compression revealed) - LLM compresses again Slop = over-compressed signal with no human decompression step. The Frippertronics loop running with no one at the guitar: decaying into noise. ## Why Frippertronics, not gardening Both are medium-models. But they capture different features: - Gardening: slow iteration, coaxing, cultivation. Fits image generation (prompt, wait, iterate). - Frippertronics: immediate feedback, reflexive (your input comes back transformed), immersive (you're inside the loop). Fits LLM text conversation. The temporal grain matters. LLM conversation is fast, like Frippertronics. Not slow, like gardening. ## Generally functional medium The Growing the Image paper argues Midjourney is an *artistic* medium. The extension here: LLMs are a medium more broadly — for thinking, writing, researching, working. Not specifically artistic. The features that make something a medium (dynamic recalcitrance, reliable unpredictability, iterative engagement) aren't art-specific. --- Source conversation: Claudian session 9 Mar 2026 evening (file `76277479-d95b-47c1-a4c2-25d461431a69.jsonl`). Continue from here.