#substack ## [[Full Plan]] v5 (Four Sections) ### I. Introduction/Problem — Drop-in opening + plan LLMs have a deserved reputation for producing pap. At time of writing the arteries of the internet are not completely clogged, but they’re getting greasier. But—and I would [[say this]] is especially true since the beginning of the year—it’s getting easier and easier for users to get these things to produce decent, non‑pap text: text that makes it hard to tell, even for an expert in the field, whether it is a student or an LLM that has written an answer, or an essay, or a thesis, or a book. This is, clearly, going to lead to a profound shift in the way that students are taught and evaluated. **I don’t have a solution for this, but I do have something that is hopefully wafting in the direction that a solution is.** The starting point is this: LLMs aren’t going anywhere. Students of all ages and shapes and types are going to use them, so we need to think about the best ways of using them. ### II. Understanding vs Knowledge: What LLMs Really Are — Drop-in + plan This is the way I recommend we shift our thinking: > LLMs are not machines for knowing > LLMs are machines for understanding We often address LLMs as if they were oracles: in normal use, they aim to provide honest, truthful answers about widely agreed-upon facts. While there are uncertainties and exceptions, let's not quibble about this stuff right now. Indulge me. Indeed, they are getting better at getting things right. [[Reasoning models]], and tool use, for example have helped LLMs get very high scores at various benchmarks %%the benchmarks example is a bit weak, I need to think of some other options to back up my claim in this paragraph%%. I think it is hard to deny now that these systems are becoming more intelligent, in some broad, problem related sense at the least. If you like to throw your essay assignments into an LLM and hand in the results, the future is bright, your going to get to get better and better answers [[over time]]; if you studiously avoid LLMs in your writing so far, I think you are in a tough spot. Imagine every who does cheat, writing better, richer, *more* assignments than you. That will not be a very fair outcome, but I think it's a likely one, with the other most likely being that [[essay writing]] assignments are dropped, pretty much immediately as ways of assessing students. A student who feeds his [[essay question]] into ChatGPT is treating it like an oracle. Even if that [[essay question]] is complex and relies on reference to tens of thousands of words of supplementary information, it is still a question that the student wants a plausible answer to. And by 'plausible' I mean something like 'plausibly written by somebody knowledgeable about the subject'. I don't think this sort of behaviour can be stopped at this point. So I'm not going to try and put forward a solution. What I'm instead going to do is try to describe a particular approach, an attitude to LLMs that we might want to cultivate. While I'm not going to deny that there are legitimate times we can use an LLM as an oracle, I think there's maybe another way to approach these machines. And that is as machines not for knowing but for understanding. Let's try and talk about the difference between knowing something and understanding something. It seems to me, [yes, I am evoking the seems-to-be operator] one way of making the difference is in terms of isolated facts versus networks of dependence. I can know “A rainbow appears when sunlight passes through raindrops.” But I understand it when I can say: light is bent and reflected inside spherical drops, and because different wavelengths bend differently, the colours separate into an arc (red outer, violet inner), the sky looks brighter inside it, and a fainter, reversed secondary arc can appear. Of course this means that understanding comes in degrees – %%Insert example here.%% In the way I'm thinking of it, knowing seems slightly cheaper, epistemologically, than understanding. It seems that I can know some isolated fact, while understanding it very poorly, or at least comparitively poorly to those around you. A galavanting abroad, you might might be sufficiently confident in your knowledge of the meaning of some foreign word as you utter it, and even recognise it is said back, but if you hear it as simply a component in an otherwise impentrible torrent of foreign words (I am not going anywhere bad with this), can we really say that you understand it to any great degree? %%Possible extra sentence here saying, of course there's plenty one could say about this but you know, bear with me, why not, see where I go.%% I now want to talk about three things. How understanding might be thought of as an [[aesthetic experience]]. Why this is a good thing. And that LLMs are best approached as engines for enhancing human understanding. %%Here it would be good to follow the [[intellectual pursuits]] paper quite closely.%% Think about what it's like to really get something. *To grok* means to understand something so thoroughly that you grasp it intuitively and completely, almost as if becoming one with it; it implies deep, holistic comprehension rather than surface-level understanding. The term was coined by Robert A. Heinlein in his 1961 science-fiction novel “Stranger in a Strange Land,” where in the book’s Martian language “grok” literally means “to drink,” and figuratively “to take something in so fully that it becomes part of you.” %%need to check this%% From the late 1960s onward it was adopted in tech and hacker subcultures, where it continues to denote deep mastery or intuitive understanding of a system or concept. And grokking something feels nice. Think of something you know about or a good at, something you could bore on about for *ages*. When you truly *get* something, thinking thoughts about it is pleasing, something we like to do. %%work in ideas form Schellekens paper here, making sure that credit is properly attributed.%% %%Mention why it is [[aesthetic pleasure]], make reference to the aesthetic qualities she thinks are involve in noetic [[aesthetic pleasure]]; speak of the benefits of appreciating [[intellectual pursuits]], in schellekens opion, by which I mean epistemic benefits as well the other one%% The heart of this proposal is the recommendation that we do not treat LLMs as machines for knowledge but as machines for understanding. A lot of people use LLMs as something like an oracle. They ask it questions. What should I do or [[what does]] this mean or what is this? These sorts of questions can often be thought of as just requests for knowledge. And they're fine on their own. I'm certainly not recommending that people stop asking LLMs questions. But what I am interested in is... I'm interested in listening to this. I'm interested in getting them from a long-term view online. I don't know if they [[look like]] LLMs, but we can see where it is in the room. If you want to see a video, you can see the video. - Knowing is the brute fact. Paris is the capital of France; the derivative of \(x^2\) is \(2x\). Understanding is the feeling of grokking—when the relations start to show themselves, when you can see why this bit must follow from that bit, and what else would have to be true if the picture really fits. - Ask an LLM, “Write 800 words on the categorical imperative.” Out comes something competent, [[true enough]] in outline. You read it; for as long as you remember those sentences, you’ve acquired a little parcel of knowledge. This is exactly how the chat‑essay trick works: type the prompt, get the answer back. - It’s both not what we want students to do and, more to the point, an inapt use of these systems. Not an oracle, then, but—if you’ll allow the phrase—a comprehension stretcher: a way of working that pulls your sense‑making wider and tighter at the same time. - Why this works: they are infinitely patient and instantly malleable; they aren’t persons, so there’s no social friction in asking “silly” questions; and their words give just enough resistance to push against. That resistance is cognitive traction: like tyre on road, it lets thought move. - What stretching looks like: ask for the same idea at different levels until the invariant skeleton shows; ask for a near‑miss that almost fits and say exactly where it breaks; ask for two competing unifications and then a minimal one that holds them together. In the back‑and‑forth, you begin to feel fit—coherence, unity, even a kind of elegance—as the shape of the thing becomes graspable. - Bridge: That felt fit matters. The more you sense coherence, unity, and elegance in the way an explanation hangs together, the more you want to keep going. The next section is about that appetite. • Plan notes: ▪ Do not mention Schellekens here; just surface the predicates (coherence, unity, elegance) and the “felt fit.” ▪ Emphasize instrumentality (non‑personhood), malleability, and patience. ▪ Keep tone descriptive and essayistic, not manual-like. ### III. The Aesthetics of Understanding: Drawing on Schellekens and Chiang — plan + optional drop‑ins • Purpose: ▪ Give the analytic account of why the “felt fit” in II is aesthetically charged and why that motivates sustained inquiry. • Moves: ▪ Schellekens (analytic summary): coming‑to‑know as a genuinely aesthetic process in intellectual pursuits; key predicates—harmony, coherence, unity, elegance; “epistemic inventiveness” as the way aesthetic appreciation opens new paths of thought. ▪ Map directly to II’s practice: ◦ While typing and reformulating: sensing which relations cohere and which don’t. ◦ While reading outputs: perceiving an elegant minimal explanation or a unifying frame. ▪ Personal shade (compact): a single moment where iterative re‑expression produced a cleaner unifying account and the felt shift followed (keep it terse). ▪ Chiang (two sentences, restrained): employ as vivification of appreciating one’s own cognitive processes (“patterns within patterns,” a modest gestalt), scaled to everyday work—no lyricism. ▪ Clarify scope: we are not claiming all understanding is inherently aesthetic; rather, this mode readily affords aesthetic appreciation. • Optional drop‑in sentence (use once): ▪ “What keeps you at the desk isn’t the promise of another fact but the moment when the explanation tightens—when harmony and coherence arrive together and the whole feels lighter.” ### IV. Implications: How to Cultivate Aesthetic Appreciation — plan + optional drop‑ins • Purpose: ▪ Translate the reframing into practice without drifting into policy; complement contemplation and reading. • Moves: ▪ From using to engaging aesthetically: attend to coherence, unity, elegance as you work; treat the model as an instrument that affords traction, not an oracle to obey. ▪ Practical modes (described in prose within the article): ◦ “Shooting the shit”: exploratory, low‑stakes back‑and‑forth to surface latent connections. ◦ Prompt‑craft as thinking: restatements across abstraction levels; alternative framings; part–whole decompositions; charitable steelmanning. ◦ Comparative unification: ask for multiple unifiers, then for the most elegant minimal account with reasons. ◦ Micro‑log reflection (not assessment): note which prompt revision increased traction and why. ▪ Educational gesture (light): if essays are harder to evaluate, we can still cultivate the desire to understand by foregrounding the aesthetic rewards of seeing how ideas hang together. ▪ Complementarity: this mode supplements contemplation and reading; it’s distinct, legitimate, and worth practicing. • Optional drop‑in cadence line: ▪ “Treat the outputs as a surface to push against. Elegance shows up when the pushing is in the right direction.” ### Voice & Phrase Capture Log (running) • “deserved reputation for producing pap” • “arteries of the internet… getting greasier” • “fatal for the essay” • “wafts us in the direction of a solution” • “LLMs aren’t going anywhere” • “students of all ages and shapes and types” • “best ways of using them” • “the feeling of grokking” • “type in an essay prompt and get the answer back” • “LLM as oracle” • “an inapt use of LLMs” • “comprehension stretcher” (alt: “comprension stretcher”) • “cognitive traction” / “tyre on road” • “coherence, unity, elegance” • “patterns within patterns” (Chiang) ### N. Final Answer Here is the full plan, end to end, with Section I beginning almost verbatim from your new opening and Section II reoriented exactly as requested. Section I keeps the essayistic voice—pap, greasier arteries, likely fatal for the essay—and ends by tilting toward “best ways of using them.” Section II then opens with a descriptive contrast between brute knowledge and the feeling of grokking, gives the oracle-mode example that mirrors the chat‑essay pattern, and pivots to your framing of LLMs as “comprehension stretchers,” explaining why that orientation works via malleability, infinite patience, non‑personhood, and cognitive traction. It closes by naming the felt predicates—coherence, unity, elegance—to hand off cleanly to Section III. Section III provides the analytic account of the aesthetics of understanding through Schellekens (harmony, coherence, unity, elegance; epistemic inventiveness) and uses a restrained nod to Chiang solely to vivify the beauty of one’s own cognitive processes as patterns begin to fit, keeping the tone precise rather than lyrical. Section IV turns that into practice without becoming a policy program: exploratory “shooting the shit,” prompt‑craft as thinking, comparative unification, and brief reflective notes, while emphasizing complementarity with contemplation and reading. I’ve also kept a running log of your turns of phrase and folded them into the plan where they belong. Going forward, I will always return the complete four‑section plan in each iteration and continue harvesting voicey lines so we retain them for drafting.