[{"data":1,"prerenderedAt":320},["ShallowReactive",2],{"blog:/blog/tech/the-answer-is-free":3,"blog-index-min":285},{"id":4,"title":5,"abstract":6,"body":7,"date":268,"description":269,"descriptionZh":270,"draft":271,"extension":272,"mark":273,"meta":274,"navigation":275,"path":276,"seo":277,"stem":278,"tags":279,"titleZh":270,"__hash__":284},"blog/blog/tech/the-answer-is-free.md","The Answer Is Free. Learning Is Not.","One curve compounds. The other has been a horizontal line for tens of thousands of years. Everything interesting about teaching right now follows from the ratio between them.",{"type":8,"value":9,"toc":251},"minimark",[10,14,19,26,29,32,40,44,47,50,69,76,79,83,86,91,102,105,112,116,123,126,129,133,140,143,146,150,156,159,162,166,169,175,181,187,190,194,197,200,203,207,210,213,224,228,231,234,237,240],[11,12,13],"p",{},"The strangest thing about the last few years is not any single capability. It's the shape of the curve, set against the shape of us.",[15,16,18],"h2",{"id":17},"two-curves","Two curves",[11,20,21,22],{},"For most of recorded history, teaching was a logistics business. The scarce object wasn't insight, it was the text. A manuscript cost a scribe a year; a library was a fortune; a student walked to a city because that was where the book was. A teacher's economic value rested on a simple, unglamorous fact: ",[23,24,25],"strong",{},"I have it, you don't.",[11,27,28],{},"Print, cheap paper, public schooling, search engines, and then a chatbot with a trillion tokens of it — the possession gap closed in stages, and the last stage closed it completely. There is no longer a fact worth teaching that a motivated person cannot retrieve, for free, in about four seconds.",[11,30,31],{},"Meanwhile something else has been stubbornly flat. The brain you are reading this with has roughly the architecture it had tens of thousands of years ago. Working memory still holds a handful of chunks. Consolidation still needs sleep. A skill still needs repetitions spread across days, with the effortful retrieval in between doing work that re-reading cannot do. None of this is cultural, so none of it is upgradable by a model release. The skull is the fixed cost.",[11,33,34,35,39],{},"Put the two curves together and you get the particular vertigo of this decade. It isn't that a machine crossed a line and became smarter than us — \"smarter\" is doing a lot of work in that sentence. It's that one curve compounds while the other is a horizontal line, so the ",[36,37,38],"em",{},"ratio"," between what can be produced and what a person can absorb per hour diverges without bound. That ratio is what people are pointing at when they say AI will exceed most humans. It's also why a perfect answer is now worth less than it has ever been worth.",[15,41,43],{"id":42},"why-reading-the-right-answer-doesnt-work","Why reading the right answer doesn't work",[11,45,46],{},"Everyone has had this happen: you read a clear explanation, you follow every step, you feel the click — and a week later you cannot rebuild it from scratch.",[11,48,49],{},"The feeling of understanding is real. It just isn't yours. It was generated by the coherence of the text, the way a well-shot film generates the feeling of having been somewhere.",[11,51,52,53,56,57,60,61,64,65,68],{},"Psychology has names for the pieces, and they all point the same way. The ",[36,54,55],{},"illusion of explanatory depth",": people rate their understanding of a mechanism highly until asked to explain it out loud, at which point the rating collapses. The ",[36,58,59],{},"testing effect",": retrieving something strengthens memory far more than re-reading it. The ",[36,62,63],{},"generation effect",": material you produce yourself sticks better than material you receive. The umbrella term is desirable difficulty — the effortful part is not an obstacle on the road to learning, it ",[36,66,67],{},"is"," the learning. Remove the effort and you remove the mechanism, which is precisely what a good explanation does.",[11,70,71,72,75],{},"So here is the corollary I think the education industry has not absorbed yet. ",[23,73,74],{},"An LLM is the best explanation machine ever built, and therefore the best illusion-of-understanding machine ever built."," Explainers used to be scarce and mediocre, so the illusion was expensive to acquire and often failed on its own. Now it is free, instant, infinitely patient, and tuned to your exact confusion. The better the explanation, the less you have to build the model yourself — and the model is the thing you came for.",[11,77,78],{},"A concrete case from the domain I have been writing about: I can hand you a flawless derivation of the KV cache memory budget, and you will nod, and you will still be unable to say whether a particular model on a particular card at a particular concurrency will run out of memory. The derivation is a lookup. The skill is an estimate, and estimates are built by being wrong a few times in a way that costs you something.",[15,80,82],{"id":81},"what-the-teacher-actually-knew","What the teacher actually knew",[11,84,85],{},"If the information gap is gone, and absorption cannot be accelerated, where does the value go? It lands in two places, and both are uncomfortable for anyone whose job was to possess the text.",[87,88,90],"h3",{"id":89},"the-decision-gap","The decision gap",[11,92,93,94,97,98,101],{},"A curriculum is not a list of topics. It is a claim — a bet, actually — about two things: what is ",[36,95,96],{},"worth"," knowing for a given goal, and what ",[36,99,100],{},"order"," it has to arrive in for the second thing to be learnable by the first. Both halves are decisions under uncertainty, and both can be wrong.",[11,103,104],{},"Ordering carries more weight than most people expect. The same fifty topics sorted differently produce wildly different outcomes, because the sorting encodes real knowledge: which idea is a prerequisite, which idea pays rent in every later chapter, which idea is a trap that has to be postponed until the learner has the vocabulary to survive it. Get the order wrong and no amount of good writing rescues you, because the reader isn't failing at the chapter — they're failing at a dependency nobody mentioned.",[11,106,107,108,111],{},"And \"worth knowing\" is not a prediction, it's a value judgment. A model can tell you what's popular, what dominates codebases, what shows up in benchmarks. It cannot tell you what ",[36,109,110],{},"should"," matter to a specific person with three weeks and a specific job. That sentence is a claim about stakes, and stakes belong to someone, not to something.",[87,113,115],{"id":114},"the-control-gap","The control gap",[11,117,118,119,122],{},"This one is subtler and, I think, the real one. AI can produce a correct answer. It cannot reliably tell whether ",[36,120,121],{},"yours"," is correct, for a structural reason rather than a capability limit: you will describe your understanding wrong in exactly the way you misunderstand it. Your questions are shaped by your mental model, so the answers come back shaped by it too. Ask a question premised on a misconception and you get a fluent, well-formatted, internally consistent reply built on the misconception, because nothing in the conversation is positioned to notice it.",[11,124,125],{},"Which is why \"I studied with an AI for a month and I'm still confused\" is not a prompting failure. Self-study with a brilliant interlocutor fails in a predictable way: the loop never leaves the interior of your own framing. Getting out requires a probe you did not design — a question you would not have thought to ask, whose answer is wrong if your model is wrong.",[11,127,128],{},"That is what a teacher was always for, and it is the part that does not survive as text.",[15,130,132],{"id":131},"a-quiz-is-not-a-grade-its-an-instrument","A quiz is not a grade, it's an instrument",[11,134,135,136,139],{},"An assessment is not a valuation of a person. It is an instrument, and its only job is to make the learner's internal state ",[36,137,138],{},"identifiable"," — to pick a probe where two different mental models predict two different observable answers.",[11,141,142],{},"Most quizzes fail at this, and they fail in one direction: they are written to confirm rather than to discriminate. \"Do you know what PagedAttention is?\" measures whether you read the section, which you demonstrably did, since you are reading it. The item that earns its place is the one whose wrong answers encode specific, named misconceptions — the three or four ways people actually get this wrong, each with a plausible-sounding answer attached.",[11,144,145],{},"Writing one of those takes an hour, because you cannot write it until you have said out loud, precisely, what the wrong idea is. That cost is the point. An item bank is a compressed map of how people actually fail, and a map of failure modes is nothing you can scrape off a page. It is also falsifiable in a way prose never is: if the item is answered wrong by readers who genuinely engaged with the chapter, the chapter is broken, not the reader. Static prose has no such feedback channel. It cannot be wrong, which means it cannot improve.",[15,147,149],{"id":148},"why-static-tutorials-dont-sell-and-why-thats-a-pedagogy-problem-first","Why static tutorials don't sell — and why that's a pedagogy problem first",[11,151,152,153,155],{},"Everyone in this space has noticed that text-and-screenshot tutorials are hard to charge for. The usual diagnosis is pricing, or piracy. Neither is the root. Copying is cheap because the artifact ",[36,154,67],{}," the product: if the tutorial is words in a file, the marginal cost of delivering the value is zero, and the price converges there. That argument is as old as the printing press, and it does not get lost twice.",[11,157,158],{},"But the property that makes static content copyable is the same one that makes it pedagogically weak, and that is the more interesting half. A static tutorial cannot observe you. Its feedback loop is measured in days at best and never at worst. You read, you feel the click, you close the tab, you try it a week later, and if you're lucky a compiler tells you what went wrong. In that window a broken mental model compounds silently, and by the time correction arrives it is no longer a correction, it's a rewrite.",[11,160,161],{},"Shortening that loop is not a convenience feature. It is the mechanism. A correction has to land while the wrong model is still loaded in working memory, or it lands on a different person than the one who made the error.",[15,163,165],{"id":164},"three-loops-instead-of-one","Three loops instead of one",[11,167,168],{},"So the artifact worth building is not a document with more media attached. It's three loops joined together.",[11,170,171,174],{},[23,172,173],{},"The content loop"," is declarative and, this year, nearly free. Derivations, diagrams, code, errata. Do not build a moat here; the water is already at sea level.",[11,176,177,180],{},[23,178,179],{},"The diagnosis loop"," is a bank of probes at the granularity of single claims rather than chapters. Not \"did you finish part 2\", but \"you just read that decode is bandwidth-bound; here is a configuration; does adding batch size help, hurt, or neither?\" Checked in seconds, not weeks.",[11,182,183,186],{},[23,184,185],{},"The consequence loop"," is where project-based learning stops being a buzzword. PBL in the strict sense means the learner produces an artifact that can fail on its own terms — a test that goes red, a p99 that refuses to move, an OOM that actually happens. The environment is the grader, and it is the only grader that cannot be talked out of its opinion. That is also the only kind of failure that produces the specific sting which turns into memory.",[11,188,189],{},"Wire the three together and the \"completion drive\" everyone chases stops being gamification. It's simply that the next action is unambiguous and cheap to attempt. Nobody needs a badge to keep going when the next step is one command away and the failure mode is informative. Tutorial plus quiz plus a PBL environment is a different genre from tutorial plus quiz, in the same way a driving simulator with a test track is a different genre from a diagram of a gearbox.",[15,191,193],{"id":192},"what-it-cost","What it cost",[11,195,196],{},"The tutorial this essay keeps circling is fifty-one lessons, forty-eight interactive components, two hundred and four self-test questions, four code projects and a graduation project — all of it followable on a CPU alone. That last constraint matters more than it looks: \"rent eight H100s before page one\" is a filter wearing a prerequisite's clothes.",[11,198,199],{},"Writing the questions, in my experience, is the hardest part, and not because of the volume. Every distractor a competent reader might plausibly choose forces you to name a misconception you had only half-articulated. It is a much stricter test of a chapter than the chapter itself is.",[11,201,202],{},"That's the honest version of the claim I started with. A tutorial built this way is no longer a document. It's a hypothesis about where people will get stuck, with a test attached.",[15,204,206],{"id":205},"the-part-that-stays-human-honestly","The part that stays human, honestly",[11,208,209],{},"Let me argue against myself, because the obvious objection is correct: models will get better at this. Ordering advice will improve. Diagnosis will improve. An AI tutor will probably beat my item bank at finding the hole in your understanding, and I won't pretend otherwise.",[11,211,212],{},"What doesn't transfer automatically is accountability. A curriculum is a claim someone can be wrong about, and be held to. A model will almost always agree that your plan is reasonable — that is what it was trained to do, and it isn't lying; agreement is simply the cheapest way to be helpful. The value of a person in the loop is partly that they can say: no, this order is wrong, you'll hit a wall forty pages in.",[11,214,215,216,219,220,223],{},"Two more things resist automation. Case knowledge: a teacher knows ",[36,217,218],{},"your"," case — your background, your deadline, the system you actually have to keep running. None of that is in anyone's context window unless a person decides to put it there, and knowing what to put there is the skill itself. And the willingness to say ",[36,221,222],{},"you are not ready yet",", which is an assessment with a consequence attached. Recommendation systems are optimized to keep you engaged. Teachers are optimized for your competence. Those objectives diverge at exactly the moment you need a teacher.",[15,225,227],{"id":226},"the-layer-that-moves-up","The layer that moves up",[11,229,230],{},"After movable type, the text got cheap, and teaching did not disappear. It moved from copying to interpreting, and new institutions grew around the new scarcity — literacy, then credentialing. Every time the medium got cheap, the human part moved up a layer: from possessing the text, to selecting it, to sequencing it, to certifying it. This time it is moving to designing the loop in which a person actually changes.",[11,232,233],{},"The amount a human being can genuinely learn in a week is roughly constant, and always has been. That is the fixed cost, and no curve on the other side of the chart moves it. Exponential capability does not raise that ceiling. It only sharpens the bottleneck, and makes everything downstream of it more valuable.",[11,235,236],{},"The answer got free. The hour in which you can learn something did not.",[238,239],"hr",{},[11,241,242],{},[36,243,244,245,250],{},"The tutorial referred to above is ",[246,247,249],"a",{"href":248},"/projects/tutorials/llm-serving-opt","LLM Serving and Optimization"," — fifty-one lessons on inference serving, with the widgets and the question bank.",{"title":252,"searchDepth":253,"depth":253,"links":254},"",3,[255,257,258,262,263,264,265,266,267],{"id":17,"depth":256,"text":18},2,{"id":42,"depth":256,"text":43},{"id":81,"depth":256,"text":82,"children":259},[260,261],{"id":89,"depth":253,"text":90},{"id":114,"depth":253,"text":115},{"id":131,"depth":256,"text":132},{"id":148,"depth":256,"text":149},{"id":164,"depth":256,"text":165},{"id":192,"depth":256,"text":193},{"id":205,"depth":256,"text":206},{"id":226,"depth":256,"text":227},"2026-11-17","AI closed the information gap, so a teacher's value moved to two other gaps: knowing what matters and in what order, and knowing how to prove you actually learned it. That is also why a tutorial has to become a workflow instead of a text.",null,false,"md","Teach",{},true,"/blog/tech/the-answer-is-free",{"title":5,"description":269},"blog/tech/the-answer-is-free",[280,281,282,283],"ai","education","llm","teaching","QhJQONClQiR6KIODyAY3WVNGfSRXj5BcsGpuIl3rWqU",[286,287,290,293,296,299,302,305,308,311,314,317],{"path":276,"title":5,"titleZh":270},{"path":288,"title":289,"titleZh":270},"/blog/tech/satisfaction-identifiability","Bad Judges, or Missing Information?",{"path":291,"title":292,"titleZh":270},"/blog/life/tencent-first-months","Six Months at Tencent: Big Company, Startup Heart",{"path":294,"title":295,"titleZh":270},"/blog/life/qinghai-gansu-loop","Three Thousand Kilometers Before We Scatter: A Graduation Trip on the Qinghai–Gansu Loop",{"path":297,"title":298,"titleZh":270},"/blog/life/indonesia-detour","The Indonesia Detour: Volcano Sunrises, Blue Fire, and the Dolphins That Stood Us Up",{"path":300,"title":301,"titleZh":270},"/blog/life/sandwich-communication","The Sandwich Method: How to Give Compliments Without Sounding Like a Corporate Robot",{"path":303,"title":304,"titleZh":270},"/blog/life/2026-aaai-recap","My AAAI 2026 Experience: Research, Connections, and Discoveries in Singapore",{"path":306,"title":307,"titleZh":270},"/blog/life/2025-recap","The Year of the Co-Pilot: A 2025 Retrospective",{"path":309,"title":310,"titleZh":270},"/blog/life/life-is-a-stream","Life is a Stream, not a Staircase",{"path":312,"title":313,"titleZh":270},"/blog/tech/nuxt-with-cloudflare-pages","Overengineering a Simple Blog? My Journey to the Nuxt + Cloudflare Stack",{"path":315,"title":316,"titleZh":270},"/blog/hello-world","Hello World: Unlocking the Output Bottleneck",{"path":318,"title":319,"titleZh":270},"/blog/tech/grid-pulse","Making a Grid Pulse Background Using Pure Tailwind CSS",1794930669786]