# LLMs Are Not Tools — Session Notes Working notes from conversation with Claudian, 10 March 2026. For resuming work on the Substack essay. --- ## The Key Reframing The essay should open with the **proper function puzzle**, not with Hoel directly. **Core insight:** LLMs are unusual in not having a clear proper function. This isn't just a puzzle to note — it's *the* important thing. Because without knowing the proper function, we don't know how to use them. And that means: the current slop might reflect our ignorance, not the system's ceiling. --- ## On Hoel's Argument Hoel is arguing **against AI hype**, not making a positive claim about LLMs being "tools for writing." His argument: 1. Writing is where LLMs should have excelled ("words are its womb, its mother, its literal atoms") 2. Writing hasn't been transformed (slop, no "move 37" moment) 3. Therefore LLMs are tools, not transformative intelligence His slogan "bits in, bits out" is about a **quality ceiling**: you can't get more quality out than you put in. B-tier bloggers become A-tier (raised ceiling), but A-tier bloggers don't improve (no surplus at the top). **Key correction:** Hoel doesn't explicitly say "tools for writing" — he uses writing as the *test case*. The attribution needs to be framed as interpretive: "Hoel's focus on writing suggests..." not "Hoel says..." --- ## Two Explanations for Slop Same evidence, different diagnoses: 1. **Hoel's explanation:** LLMs are limited tools. This is what they do. Bits in, bits out. 2. **Alternative:** We haven't learned the practice yet. "Bits in, bits out" assumes we're already putting in what we can. Maybe we're not — maybe better practice would extract more. If the second is right, improvement is possible. --- ## The Proper Function Section Needs development with examples: - Hammer → hammering (clear proper function) - Google → searching (clear) - Hammer holding down papers → accidental function - Googlewhacking → accidental function of search - Spork / Swiss Army knife → multi-functional but still has proper functions Then: what's ChatGPT for? - Predicting tokens → mechanism, not use - Chatting → fits some uses badly (code, translation, philosophy) - Writing → too blunt (a pen is for writing too) No clear answer. That's the puzzle. --- ## Frippertronics — Repositioned Not "do exactly what I do" but an illustration of a **general structure**: - Material goes through the system - Comes back altered - User responds to the return - Loop continues **Crucial point:** Still entirely user-originated. Fripp's playing is always the source. The system transforms but doesn't replace. The thinking remains yours. The Frippertronics analogy needs to be placed as illustrating something more general — a possible way of engaging, not the prescription. --- ## Medium and Recalcitrance "Medium" in art theory isn't just "what it goes through." It's about: - Resistances and possibilities - Dealing with the medium's difficulties - Achievement and appreciation **The tracing analogy:** If you trace a drawing, you've sidestepped the medium's recalcitrance. The result might look okay but it's hollow — you haven't actually dealt with drawing as a medium. For LLMs: if you just accept what comes back, you've sidestepped the recalcitrance. The generic tendencies have won. Good use requires wrestling with the return — resisting the generic, sharpening against it. --- ## Why Quality Might Emerge The loop forces engagement: - The return offers material you wouldn't reach alone (even if wrong or flat, it's something to push against) - Accepting generic returns is like tracing - Quality comes from wrestling with the return, resisting it, sharpening against it "Thoughts articulated through words" vs. "pseudo-thoughts articulated through words" — the medium can produce both. Good use distinguishes them. --- ## Structure Sketch (Tentative) 1. **Opening puzzle:** LLMs don't have a clear proper function. (Develop with examples.) 2. **Implication:** If we don't know the proper function, we don't know how to use them. We can't extract what's potentially there. 3. **Two explanations for slop:** Hoel's (ceiling) vs. alternative (haven't learned the practice). Explore the second. 4. **The loop structure:** Frippertronics as illustration. Material goes through, comes back, user responds. Still user-originated. The thinking remains yours. 5. **Why this might produce quality:** Wrestling with recalcitrance. Resisting generic tendencies. Using the return as material, not as answer. 6. **Slop explained:** Generic tendencies dominate when we don't engage the medium properly. Slop is a failure of practice, not a fixed outcome. 7. **Against "bits in, bits out":** The ceiling might be higher than we're reaching because we haven't learned the practice yet. --- ## Outstanding Questions - How to transition from "exploring a possibility" to "here's what I do" without being lame? - What's the textual equivalent of "Fripp couldn't have made those sounds without the loop"? What thinking becomes possible that couldn't happen otherwise? - Should the essay give an example of good use, or just describe the structure? --- ## Links - Current draft: [[Writing/substack/LLMs Are Not Tools - Codex Revision]] - Hoel's essay: [[Clippings/Bits In, Bits Out]] - Related published work: [[Growing the Image]]