Every institution a student passes through, school, exams, university, the military, is a decade long bet on what skills will be worth having in the future. That bet has never been more expensive to place, or more likely to be wrong.

I'm a Korean teenager and student living inside every institution this essay is about. I also run a small AI lab and have published research.

I'm writing this in September 2026. Three major things have happened in the AI space in the first two weeks of the month. OpenAI shipped GPT-6 Astra and its president opened the launch with "Welcome to the AGI era". Whether or not that's true, it's the first time an AI lab has put the phrase out publicly. New York City became the first major government to draw an age line on AI, banning student use through eighth grade and restricting it for high schoolers. And on September 12, Dario Amodei published an essay called "We Must Pace the Frontier"¹, arguing that the labs themselves should slow down capability growth. Within a day, Sam Altman agreed and committed OpenAI to the same first step, and Elon Musk posted three words: "Dario is right." So in two weeks, a lab declared the AGI era, the three biggest names in the industry agreed to slow down, and a city banned the tool from classrooms. Everyone involved is reacting to the same fact that the technology is moving faster than the institutions around it.

I'm assuming powerful AI² arrives within a few years. By powerful I don't mean AGI or any other word labs use in launch posts. I mean something narrower and easier to check. A model that can do the work a fresh graduate is hired to do at an entry level for less than what the graduate costs, with no more supervision than what you'd give a new hire. Some of that is arguably already true today. I'm not going to argue for the premise here. I'm going to take it as a base and see where it leads. If you reject it, this essay isn't for you. If you accept even a decent chance of it, keep reading.

I'm confident about the direction and not about the timing. Models that can do entry level work might be everywhere by 2028 or might take until 2034. If the labs actually pace the frontier, the dates become later. Either way, the graduate of 2036 meets them. Section 3 is the part I'm least sure of, because that's a prediction about what humans are for and nobody has good data on that yet. I'll be clear about which claims I can back and which I can't throughout the whole essay.

Your life is mostly in a hardened path. You study your whole life until you're 18, take a big test which the rest of your life depends on. You get a degree at 23, discharged at 25, and you get your first real income at 26 or 27. That's the traditional life path for Korean men³. For a 17 year old today, their first paycheck is 2036, and every academy hour this week is a bet on that year.

The most concerning thing is that the adults with the most credibility in a teenager's life are pushing them harder into academics, using AI as "the ultimate threat to the job market" instead of taking the shift in the industry seriously. Most parents nowadays regard going all in on something other than academic stuff as "being risky" and not being on the right route. That's what I'm talking about as the "hardened path" for most people. Yet nobody is allowed to say it, or will. I'll come back to it later on. Some people do find their way off the hardened path and build something of their own. This essay is for them, and for the ones still deciding.

I will be talking about roughly five topics in this essay.

1. Risk costs: What the bet costs in hours, money and years, and what happens if it doesn't pay.
2. A Broad View on 2036: What the job market will look like for a graduate in 2036 if AI keeps advancing.
3. What Survives: Which skills are still worth spending a decade on, and which ones a model already has.
4. What teenagers can do inside institutions that won't change in time: What teenagers can do this year.
5. What Adults owe them: What adults owe teenagers, having pushed them toward a world the adults themselves have never lived in.

Before going further, let me be clear about what this essay is and isn't. I'm going to be hard on the education system, harder than most people my age are allowed to be in print. What I'm not going to do is blame the people executing it. Teachers, academy owners, and parents are responding to incentives the system hands them. The people who set those incentives are a different story. Section 5 is addressed to them. I'm also not telling anyone to drop out without any purpose; I'm still taking exams myself. And I'm not saying AI is safe, or that using it more is automatically good. The question here is, if the premise holds, is the bet teenagers are being told to place still a good bet, and what should they do if it isn't.

1. Risk costs

Most people around me think a stable life follows if you study your whole life and do what parents and teachers tell you. That was never true. Even before AI, if you had no goal or landing point, you didn't land anywhere. The advice back then was to pick a landing point, work toward it, and pick extracurriculars that got you there.

In my case the hours in school, academies, and commute add up to about 65 hours a week, 35 in school, 20 in academies, and about 10 commuting. Run that from the first year of middle school to Suneung, the national university entrance exam, and it's six years, roughly forty school weeks each, so somewhere around 15,000 hours. That's the number to keep in mind for the rest of this section. It's more hours than most people spend on their first three jobs combined, and it's spent before the first job starts.

Korea's total private education spending in 2025 was 27.5 trillion won (about $20 billion), with 75.7 percent of students participating and an average of 7.1 hours a week. Per participating student, the monthly average was 604,000 won (about $450). For high schoolers specifically, about 800,000 won (about $600) a month, and that figure went up even as the total fell.⁴

Before Suneung, there are twelve years if you count the years from elementary school, which every Korean parent does. Then the years nobody planned for. In the exam this November, 31.5 percent of registrants are graduates sitting it again, and counting everyone who isn't a current high school senior the share is 35.6 percent, the highest in 31 years.⁵ Then eighteen months of service for men. Add it up and the 17 year old reading this reaches their first real job around 2036, having spent something like my 65 hours a week for the better part of a decade to get there. That's the price of the ticket, but now, the question is what it buys.

Let me be fair to the bet, because as of now it isn't stupid. A degree from a top university in Korea still buys a few things. An employer who has never met you before assumes you're capable because you passed through. It buys a network: the people who passed through it with you end up running things and they pick up the phone when you call. It also buys a first look from large companies. Their hiring pipeline is built around where you studied. Whether that's good doesn't matter. As of 2026 it's all real.

The question is which of the three survives the premise. The employer's assumption breaks first. The whole point of a passing filter is that it's expensive to fake, and a model that can pass the exam makes the filter cheap to pass and therefore worthless as a signal. The first look from big companies weakens with it because those pipelines exist to find people who can do entry-level work, and entry-level work is exactly what the premise says the model does. The network doesn't break at all. People still hire people they know, which means the bet is now mostly a bet on the network, and nobody is pricing it like that. If the network is what you're actually buying, there are cheaper ways to buy it than 15,000 hours, and most of them involve doing something people want to talk to you about.

In the case that it doesn't pay, it's 2036. You're 27, you have the degree, you apparently did everything right. You apply for the job the degree was supposed to get you, and the company has already given that job to a model that costs less per day than your meal. You have no portfolio, because all the time was spent on the exam and there wasn't any left. You have the network, but the network can't hand you a skill. And the part that really sucks: the time is already gone. The 800,000 won a month, 65 hours a week, twelve plus years, none of it comes back. You can't un-spend a childhood, and that's the risk on the safe path. What frustrates me is that it's never the one that gets called a risk.

More people are taking what gets called "risks": startups, content creation, their own thing. The money already moved too. Funding for AI-native startups went up 218 percent from 2021 to 2025 while funding for the rest of tech fell 36 percent.⁶

I'm not going to pretend the other path is safe either. Most startups fail. Most content creators earn nothing. Most teenagers who "go all in" on projects produce a half built app and a lot of fake confidence. It isn't that one path is risky and the other isn't. It's that one path's risk is visible, so adults warn you about it, and the other path's risk is deferred to 2036 so nobody does. Both bets can certainly lose but only one of them is being called a bet.

2. A Broad View on 2036

I don't put much weight on terms like AGI or ASI. Nobody can say what qualifies, so nobody can say whether a model has reached it. I think models passed human ability at writing, coding, and some kinds of thinking a while ago. Yet this month OpenAI launched GPT-6 Astra and opened the event with "Welcome to the AGI era". Nobody knows if that's true, and that's the problem. Those words are marketing. They don't grade anything. Benchmarks are better, but once a benchmark is famous, every lab knows it's on the scoreboard and trains toward it, which defeats the point. The benchmark worth trusting is the one nobody was aiming at: an exam written for humans, new every year, that no lab strongly prepared for. I'll get to one in section 3.

I picked 2036 because it's ten years out. At some point real diffusion happens. Industries and policymakers figure out the tool is efficient when used properly and doesn't have to degrade the people using it.

A lot has to happen first for AI to be used routinely outside tech and for access to be anywhere near equal. Most of it comes down to cost: training, inference, and compute, which all reduce to efficiency.

There's a version of "academic results don't matter anymore" that's already half true, and I want to be careful with it. Big companies have been dropping degree requirements from job postings for years. Tim Cook said in 2019 that about half of Apple's US hires the previous year didn't have a four-year degree.⁷ IBM says about half of its US openings don't require a bachelor's, and Google and Accenture have removed the requirement for many roles. By 2026, about 70 percent of employers say they hire for skills and half have removed the degree from at least some postings.⁸ The catch is that postings and hires are different things. When Harvard Business School and the Burning Glass Institute went back and checked, 45 percent of the companies that dropped the requirement had changed nothing about who they hired, and across large firms fewer than one hire in 700 was affected.⁹ The announcement ran ahead of the behavior. My read is that they took the degree off the posting and kept it in the decision because they had nothing better to filter on.

I think the exponential continues for at least another three to five years. Two things drive it: hype and competition. You need both to get a run like the current one. Hype pulls in venture money and creates competition, because everyone wants a monopoly in the new field. Competition then pushes harder, because staying at the frontier is hard and everyone wants to.

Sam Altman, back when he was president of Y Combinator, gave a talk in 2018 that I think applies here. Real trends are the ones where early adopters use the new thing obsessively and tell their friends. When the iPhone came out it sold a million or two the first year and people dismissed it, but the people who had one used it for hours a day. His fake trend, "at least a fake trend as of August 2018," was VR: plenty of people bought headsets, and most owners he knew used them "never or very rarely."¹⁰ AI is on the iPhone side of that test. Look at how many people pay for a subscription and still hit the usage limit every week.

Once in a while a big technological shift happens, and during the early stages of it people divide themselves into mainly three groups: accepters, deniers, and neutrals. The same thing is happening with AI. The accepters use it for literally everything. The deniers mostly tried it once in 2023, had a bad experience, and never came back, and people rarely change a first impression. The neutrals have the apps installed and barely use them, which gets almost none of the benefit. I'll come back to the three groups in section 4, because the split appears to be mostly by age.

By 2036 I think the compute problem is mostly solved by three things together: more chips, more performance per chip, and models that need less compute for the same result. None of the three does it alone. Google is the current example. Gemini 3.8 Flash matches Claude Opus 5 on several coding benchmarks at about a sixth of the price,¹¹ and it got there on Google's own chips and its own efficiency work.

The other thing that's already moving is how the work gets done, not just how much it costs. Companies are putting agents in the loop: a model that runs a task end to end, with a person checking the output at the end instead of doing the steps. Right now that's more common in the software than in the org chart. About a third of enterprises have at least one agent in production, and fewer than one in ten have scaled one to the point of measurable value.¹² But the direction matters more than the current number. When one person can supervise ten agents, the company doesn't need one person per task anymore. It needs one person per ten tasks, and that person is the one who knows what the output should look like. That's the mechanism by which this lands on entry-level first. The senior person with judgment gets ten agents. The junior person who was going to spend three years earning that judgment gets nothing to earn it on.

So here's 2036 as I see it. It's less dramatic than the way it usually gets sold. The company still hires. It hires for different things.¹³ The model writes the code, drafts the report, and summarizes meetings before anyone has finished reading a single document. The company still needs a human. It needs someone to decide what to ask for, to catch the answer that is confident and wrong, and to sign off when it ships. Those are real jobs and there will be plenty of them, but none of them was on the exam.

3. What Survives

I'm not on the side of "make AI do everything." There are things a model can't do for me. And people have to keep learning, or they become the bottleneck on the tool. That's where most people fail. Breakthroughs still need someone who holds enough real structure in their head to notice a connection nobody asked for. Some people do that more naturally than others. Anyone can train it, and math and physics are where you train it. So let me be concrete about what a model couldn't do for me last month, and what that says about what stays.

I always review and refine design work a model gives me. First drafts are unstable; you refine step by step. If you have a design team, the better path is to have them design and the model replicate, then fix the details. Otherwise you get AI slop, and the second your product looks like AI slop, people lose trust in the brand.

Currently, most AI models are confidently wrong often enough that anything critical needs a second check. That's misinformation you feed yourself without noticing. If the task is critical, check it against a second model or a search.

Something smaller that I notice every week. A model solves whatever you hand it, and it's very good at that. It has no opinion about what you should have handed it though. Last month, I had five things I could build for Cura, and the model would happily have written all five. Picking the one that actually ships is the job. A model can rank them, but the ranking is a bet about what people want, and the person who eats the loss if the bet is wrong is the one who gets to place it. Choosing the problem is the part that stays.

Choosing between problems is the small version; a decision with someone else's life in it is the big one. The last decision I made that I couldn't hand off was bringing a friend onto Meredic. I asked a model, of course. It drafted the offer, listed what could go wrong, even guessed how it would go. All of that was useful, but none of it was the decision. The model doesn't work next to him every day. If it goes badly, the model doesn't lose a friend or a company but I do. That's the part no amount of capability changes. A decision is the thing that happens after, and someone has to be the one it happens to. Humans take responsibility because we can be fired, sued, and embarrassed. A model can't be any of those, so responsibility has nowhere to land on it, and that's why the last signature stays as the human's job and responsibility. That's what survives. Judging the output, choosing the problem, and owning the result.

Now what a model already has. Start with the subject schools spend the most years on. Peter Thiel opens Zero to One with a memorable line. Every moment in business happens only once. The next Bill Gates won't build an operating system, the next Larry Page won't build a search engine. If that's true, then history as schools teach it, a set of patterns to memorize and match, is close to worthless because the patterns don't repeat. What's left of history is smaller and stranger. The knowledge that nothing repeats, which you can learn from one paragraph, not twelve years. A model holds the rest. I'll be told judgment needs context. My bet is that judgment needs the habit of asking for context, not a decade of storing dates and names, and that habit is cheaper to build than a memory.

Here's a harsh statement of my own.¹⁴ I don't think history, especially the way schools teach it, is worth a teenager's years anymore. The last two hundred years or so is fine. That's the world we actually live in. Industrialization, the wars that drew the current borders, the companies and technologies that built the economy I'm about to enter. Everything before that though, the dynasties, the ancient empires, the BC dates, is a decade of useless memorization for context a model can hand me in ten seconds. In Korea that is most of what we spend history class on.

And it isn't only history. The exam itself is done as well. Korean students spend years on it and almost nobody gets it all right. In September a university student ran the major models through the full 2026 Suneung, every subject including the four hardest electives, through the API with no tools and no system prompt. GPT-6 Astra scored 450 out of 450, the first full-subject perfect score. Three other models were within 2.5 points.¹⁵ For comparison, five human test-takers scored perfectly on that same exam, out of more than 550,000.¹⁶ The exam that a Korean teenager spends twelve years and 800,000 won a month preparing for is now a solved benchmark.

There's a second thing that result means, and it's the one nobody in the education system has said out loud. A model that scores 450 on the exam can also teach the exam. Every academy in Korea sells one thing, which is a person who knows the exam better than you do. There are now four models that know it better than every person alive, and they cost less per month than a single class. I'm not saying academies disappear. I'm saying the thing they sell just became a commodity, and what's left to sell is the part a model can't do: sitting next to a specific kid and noticing where they're stuck. That's a much smaller business than the one that took 27.5 trillion won last year.

So if the exam is solved and the memorization is gone, what should a 15 year old actually spend the next three years on? Elon Musk's version is physics: AI is automating routine digital work like programming, so human value shifts to the laws of the physical world. Sam Altman's is coding: not for the syntax, but because it teaches you how to structure a hard problem.¹⁷ I agree with the basic idea. My bet is that math and physics stay valuable. You can only connect ideas you actually hold in your head, and a model can't do the holding for you.

Learn with AI. The dumbest thing I see my peers do is ask a model a question and copy the answer down without understanding a word of it. AI is a better tutor than most students ever get, at scale, and one on one. It breaks down hard topics, gives immediate feedback, adjusts the difficulty, and does it at midnight. The difference between using it as a tutor and using it as an answer key is what you ask for. Ask it to derive the thing instead of state it. Ask it why the step works. Ask it to give you a harder version of the problem you just solved. Ask it what you got wrong before it tells you the answer. None of that is a prompting trick. It's the difference between wanting to know and wanting to be done, and the model gives you whichever one you ask for.

Every other creator on YouTube or Instagram says "you need to learn how to use AI." I'm against the idea. The whole point of the tool is to extend what a person can do, and "learn to use AI" turns it into one more thing to study. People learn prompting to avoid learning the task. Every model generation makes prompting matter less, and I expect that to keep going until it isn't a skill at all. What you need is to get used to it, by using it, with minimal guidance and no guru. You don't learn to surf from a book.

Your time is fixed, so if physics and building go in, something has to come out. Here's what I'd pull. First, the memorization half of every subject. The dates in history, the vocabulary lists in English, the formula sheets in chemistry. A model holds all of it perfectly, forever, and Suneung will keep testing it anyway. Do the bare minimum that keeps your grade alive and not even one hour more. Second, the second and third academy on the same subject. One good teacher plus a model that answers at midnight beats three teachers who all cover the same textbook. I dropped my second science academy for exactly that reason, and my science grade didn't move. Third, the practice-test grind past the point of diminishing returns. The tenth mock exam teaches you less than the second. The hours between them are where a project could live. What stays is the part of each subject that's actually understanding, the derivation not the formula, the argument not the date, because that's the part you need in your head to connect anything to anything. And the hours you free up should go somewhere. Create more than you consume, or the model will have done both for you.

4. What teenagers can do inside institutions that won't change in time

Almost nobody I know at school has gone deep on anything, to the point where they'd be embarrassed to get it wrong. That should change. This section is about what a teenager can do this year, inside a school system that isn't going to change in time.

Start with how you hold the tool. Think of AI as a thinking partner, not a box that throws out answers. Used for learning it's overpowered, and most of my peers use it to get homework off their plate.

New York City drew its line on September 2. A one-year moratorium on student-facing generative AI from 2-K through eighth grade, covering nearly 600,000 students, two-thirds of the system. Companion chatbots banned in every grade, high school included. About 40 classroom tools with AI functions switched off. For high schoolers, two AI critical-thinking classes during the year and supervised pilot access for at most five classes per school, capped at about 50,000 students, or 5 percent of the district.¹⁸ The under-14 part might be right. I don't have data saying a 12 year old should be using a chatbot unsupervised, and neither does anyone else. The high school part is a mistake. A 17 year old is one year from a workplace where the tool is on every desk, and the plan for them is two classes about it and about an hour a week with it, in a few classrooms per school. Even the city's own tech industry group said the hour a week should grow. The city tried a ban on ChatGPT once before, in 2023, and reversed it within months. This one has a coalition and a report attached, which is better. It's still a policy that assumes next year looks like last year.

Back to the three groups from section 2. Teenagers are mostly accepters and adults are mostly deniers or neutrals, and that's partly fine. If I told you there was a new way to fly with no device, you wouldn't jump. A kid who grew up flying that way and never fell has no reason to be scared of it. That's the benefit of picking up a technology as a kid. You find out where it breaks before you have anything to lose.

Almost everything I actually learned as a kid happened outside a sanctioned channel. Cracked Minecraft, mods that could have wrecked the family computer, a Discord server we were all too young to be on, and eventually a Minecraft server with real players, where I learned to write plugins because the players kept asking for things. None of it was allowed and all of it happened. I'm not telling anyone to break rules. I'm saying a ban doesn't stop a 10 year old from doing the thing. They do it at home instead, with nobody watching and nobody teaching them to check what the model gives back. That's what the chatbot ban does to a 17 year old in New York.

There's a problem with section 3 that I have to answer. Choosing the problem, catching the error, owning the result. Nobody learns those in a classroom. People learn them by doing junior work badly for a few years under someone who corrects them. If companies hire fewer juniors, that ladder gets shorter, and the skills I'm telling you to aim for are the ones the ladder used to produce. So you have to build the rung yourself. A project with one real user is a junior job you gave yourself. You pick what to build, you ship it, it's wrong, someone complains, you fix it. The Minecraft server taught me that and no class did. It's also why a portfolio holds up where a degree doesn't. A degree proves you passed a test a model can also pass. A real user proves you shipped something and fixed it when it broke. The model can't do the second one for you.

The opposite of the Minecraft server is what people my age call LARPing. You post the plan, make the logo, write the Notion page, tell everyone about the startup. From the inside it feels like execution. The difference is that nobody is using anything. The server had players who complained when it crashed at 2 AM. A LARP has an audience and no users. I check myself with that now. If nobody would notice if I stopped, I'm performing. Wanting to be someone is fine, and most real projects probably start there. It just can't count as the work.

So, concretely, this year.

Ship one thing to one person who isn't your friend. Not a portfolio, one user. Friends will tell you it's great. A stranger will tell you it doesn't work on their phone, and that complaint is the start of the apprenticeship. The easiest way to find the stranger is to fix something you personally use and hate, then post it where people who use the same thing hang out. It doesn't have to be software. A spreadsheet that does something a teacher needs, a Discord bot for a community you're already in, a 3D printed part someone in a forum asked for. The bar is that one person you don't know uses it and comes back.

Take one subject past the syllabus with a model as the tutor. Pick the thing in physics or math you actually want to understand and go until you can derive it, not just recite it. The syllabus stops at the formula. Go to where the formula comes from, then one step past that, to the place where the textbook says "it can be shown that" and doesn't show it. That's usually a week of evenings with a model that will walk you through it as many times as you need. When you can explain it to someone else without notes, you have the thing the exam was pretending to test.

Build something physical. I 3D print. It can be anything, but a physical thing catches "confident and wrong" faster than anything else, because the part doesn't fit. A model will tell you the tolerance is fine and the print will tell you it isn't. Electronics work the same way. So does cooking, so does woodwork. The point is a loop where the world grades you instead of a person, because the world doesn't round up.

Drop the second academy. Section 3 already said why. Put the hours here. If your parents need a reason, the honest one is that the grade didn't move when I did it, and a model at midnight covers what the second teacher covered at 9 PM. If that isn't enough, try one subject for one term and compare the grade. That's a test they can check.

5. What Adults owe them

Adults pushed us toward a world they've never lived in. That's not an accusation. Nobody has lived in it. But it comes with obligations, and they belong to the people who run the system, the Ministry of Education, KICE, and the universities, not the teacher in the room. I count four.

The first is honesty. I said at the start that nobody is allowed to say it, and I meant allowed. A teacher's income depends on the exam. An academy sells exam preparation; it can't sell "this exam tests the wrong half." A parent who has paid 800,000 won a month for six years is the last person who can hear that it was mispriced. Nobody in that picture is lying. Everyone is responding to what the system pays them for, so the honesty has to come from above them. The people above them are the ones who could change the exam and have chosen not to, for a decade, while the technology that makes it pointless went from a demo to a perfect score. It would be one sentence said out loud in a classroom: this exam tests the part a model already does, and we're going to spend the hours we can spare on the part it doesn't.

The second is access instead of bans. Korea tried the opposite of New York and ended up in the same place. In March 2025 the Ministry of Education put AI digital textbooks into classrooms for third and fourth graders and first-year middle and high school students, in math, English, and informatics. By August the National Assembly had passed a bill reclassifying them from textbooks to supplementary materials, effective the next semester, after teachers and parents pushed back and after one earlier version of the same bill had been vetoed in January.¹⁹ By November the category had been deleted from the textbook regulations entirely. Pushed too hard, then pulled back, inside a single school year. Both mistakes, New York's and Korea's, come from the same thing: nobody in charge has a clear idea of what the tool is for. Section 3 already said what it's for. A tutor that answers at midnight, and a second opinion on anything critical. Teach that, and teach how to catch it being wrong. A ban doesn't stop a 17 year old from using a model. They use it at home instead, with no guidance. And a mandatory textbook nobody asked for doesn't teach anyone to use it either. It teaches them that the tool is something the government does to you.

The third is credit for work done outside school. I run a company and a research group and have a published paper. None of it can appear on my student record. Since the 2024 admissions cycle, under the Ministry's fairness reforms, awards, reading records, self-organized clubs, individual volunteer work, and certifications are either not recorded or not sent to universities, and writing a research paper is explicitly barred from the record.²⁰ What universities see is the regular curriculum and a teacher's notes on it. The kind of work section 3 says survives is exactly the kind the record can't see. A university that says it wants judgment and admits on a record that excludes every place judgment gets built is measuring attendance. I understand why the rule exists. Outside activities were being bought, and the reform was meant to stop parents with money from buying a record. It worked. It also means a kid who shipped something to a thousand users and a kid who didn't look identical on paper, and that's the paper the whole bet is placed on.

The fourth is the exam itself. If a model scores 450 out of 450, the exam isn't testing what a person can do anymore. It's testing what a person can do with the tool taken away. You don't grade a human by making them hammer a nail without a hammer. Every real workplace hands you the hammer. Put the model in the exam and grade what the student does with it: what they ask, what they catch, what they sign off on. That's the half nobody is testing. It isn't a strange idea. Open book exams have existed for a century, and the point of them was always the same, that the memorization was never the thing being measured. An open-model section of Suneung would be the same principle with a better book. KICE is already redesigning the exam; the 2028 version redraws the subject list, which is why this year's registrants spiked. As far as I can find, the redesign doesn't mention the tool once. That was the moment to do it, and it passed.

I'm not asking anyone to tear the system down. I'm asking the people who placed a bet on a teenager's behalf to say what it's on. The bet might still pay. But a bet nobody is allowed to name doesn't get updated, and this one hasn't been updated in a decade.