99% of Assignments Are Written by AI: What's Left of College?
Summary
AI has evolved from a student productivity tool into the default production layer for courses and assignments.饶杰武 estimates that “99%” of assignments and papers involve AI at a deep level, and says he “can no longer think of any assignment that did not involve AI”; 林童宇 moved from hand-writing code and asking ChatGPT to fix bugs to barely looking at code while letting an Agent execute end to end. For education systems, the debate over whether AI should be allowed is already behind the curve. The real question is how courses, exams, and hiring will measure individual contribution after AI enters the process.
The bottleneck in education’s value chain is shifting from teaching execution to screening, evaluation, and accountability.侯泰宇 defines a degree as a “credential of trust” that society issues to the market, and describes the work left to humans in the AI era as choosing when there is no standard answer. 林童宇 argues that the new system should stop asking whether students use AI and instead assess whether they can work with it, identify problems worth solving, and take responsibility for the results. Degrees will not disappear easily, but the pricing power of content delivery is falling; assessment, certification, and judgment training will matter more.
Personalized AI tutoring can lift grades and accelerate exploration, but it does not automatically supply understanding, taste, or quality-control skills.林童宇 describes his CS education as “30% coursework, 70% conversations with AI and project work,” and says AI-assisted review helped him earn his first 4.0 in fall 2023; 泓君 stresses that the brain still needs time to digest a deep principle, and that high-quality work in design, law, and code ultimately depends on a person’s ability to distinguish good from bad. 侯泰宇 believes AI can also support and accelerate parts of the understanding process by filtering information and providing personalized feedback. “If the tool makes a mistake and you cannot tell,” that is the most dangerous failure point for people with weak foundations.
College may retain its value, but increasingly as a venue for trust and relationships rather than a warehouse of knowledge.林童宇 believes knowledge-work skills are being eroded by cheaper, 24-hour AI, while what Tsinghua truly leaves students is a high-density peer group, mentors’ office hours, and people connections that build trust and persuade others. If he could return to his freshman year, he would “go all out socially,” meaning one-on-one deep conversations rather than generic networking. A liberal education, critical thinking, and the exploration of “who I am” are moving from the margins to the parts of college that are hardest to replace.
AI spending is concentrating in a handful of high-value workflows rather than being spread evenly across every chatbot.林童宇 went from finding $20-plus a month expensive as a student to spending more than $200 a month; 侯泰宇 bought Claude’s $200 Max Plan and said that after exhausting his Claude quota, he “didn’t even know what to do,” while several participants have already canceled ChatGPT. Claude/Claude Code handles execution, Manus handles Deep Research and design, and Gemini handles visual work—evidence that vertical delivery, memory, and workflow stickiness are more likely to create a paid moat than being able to “chat about everything.”
The strongest retention signal is also a risk signal: these users describe AI as a more powerful productivity addiction than short video.侯泰宇 says he is “crazy addicted,” able to go a day without short video but not a day without AI; 林童宇 attributes the dependence to less painful thinking and the exhilaration of creating above his level. This is less about entertainment or companionship than about losing the internet or a computer: once calendars, email, research, and coding are embedded in Agents, the cost of quitting has moved from a psychological habit to a production system shutdown.
AI does not inherently raise the productivity of senior knowledge workers, but it is rewriting the advantage of newcomers and the hiring bar.泓君 cites METR’s randomized experiment involving 16 senior engineers and roughly 200 real-world tasks: those using AI were actually 19% slower, even though the developers believed they were 20% faster—a gap of nearly 40 percentage points between perception and reality. The slowdown was greatest among experts who already had optimized workflows, while beginners benefited. The labor-market conclusion is therefore not simply that “AI replaces you,” but that “people who are better at using AI replace you,” even as companies still cannot explain how to certify that ability.
Deep dive
1. The First AI Natives Completed Their Entire College Years with Generative Models
ChatGPT launched on November 30, 2022. 林童宇 and 侯泰宇 had just entered college, while Jack was already studying at NYU; around the time this episode aired, all 3 were graduating in succession, making them part of the first cohort to treat generative AI as learning infrastructure.
Their majors were deliberately varied: 林童宇 studied law and minored in CS at Tsinghua, 侯泰宇 studied applied psychology at NYU, and 饶杰武 studied data science at NYU before pursuing management science and engineering at Columbia University. All 3 eventually moved toward AI programming, Agent practice, or entrepreneurship.
泓君’s central question was not whether students were cheating, but what remains of college when courses, assignments, research, and career choices all involve AI from the outset. The 3 heavy users offered an answer more complicated than “the future of college is just socializing.”
2. One “Definitely yes” Convinced a Law Student to Join AI If He Couldn’t Beat It
When applying to Tsinghua, 林童宇 wrote that he hoped to “lead the next wave—or, put another way, surf it.” But ChatGPT soon outperformed him on legal provisions, case analysis, and legal writing. He reasoned that in 2 or 3 years, when he graduated, AI would be stronger, while his own legal skills might not improve enough to compete.
In mid-2023, he asked Joseph, a partner at a US law firm with more than 10 years of M&A experience, whether ChatGPT would disrupt junior lawyers and paralegals. Joseph cut him off before he finished: “Definitely yes,” then explained how exposed entry-level legal work was.
Once an industry veteran confirmed his judgment, 林童宇 adopted an “if you can’t beat it, join it” strategy. He calls continuing to pay $20-plus a month for ChatGPT one of the best investments he made as an undergraduate—although at the time, it still felt to a student like “a completely unreasonable expense.”
3. A Personalized AI Tutor Rewrote the Learning Path as “30% Coursework, 70% Conversation and Projects”
林童宇 did not stop at an industry overview. He wanted to understand how AI was built from fundamental theory and why AI in 2017 and 2018 differed so sharply in user perception and real-world application from the products of 2022 and 2023. ChatGPT mapped out the path and explained the parts of Stanford and MIT’s open courses that he had not understood.
His comparison was blunt: many university professors are exceptional researchers but not necessarily effective teachers. ChatGPT, by contrast, was like “a teacher who knows you extremely well, knows your current level, adapts to your needs, and patiently guides you,” continuously changing its explanations around his individual gaps.
He estimates that roughly 30% of his gains came from courses, textbooks, and open classes covering Python syntax, computer science fundamentals, and disciplinary common sense. The other 70% came from progressively deeper conversations with AI and “learning by doing on projects, then reviewing everything after the project was finished.”
5 or 6 projects captured the jump in capability. On the first, he still wrote the code himself, with AI occasionally fixing a bug or filling in a section. Later, he barely looked at the code and let AI execute like a programmer. The models’ progress became visible directly in project completion time, quality, and economic value.
4. A “施天涛 Doppelgänger” Helped a Law Student with Weaknesses in the Subject Earn His First 4.0
In fall 2023, 林童宇 earned a perfect 4.0 for the first time. He fed all the course materials for Anglo-American case law into GPTs, had the models act as the professor and generate questions, then repeatedly queried them inside a folder containing the packaged context for targeted review.
For company law, he fed the work of 施天涛, both his professor and the textbook author, into AI and named the resulting GPT “施天涛’s Doppelgänger.” Even with a limited context window at the time, the conversations helped him grasp the professor’s academic views, priorities, and legal framework.
Law exams required students to write subjective answers by hand on the spot; they could not simply submit AI-generated work. What was actually tested was the ability to formulate a thesis, develop reasoning, cite legal principles, and connect them to statutory provisions. Looking back, the relevant knowledge is still among the easiest for him to retrieve, leading him to conclude that he “may actually have learned it pretty well.”
5. Banning AI First Spawned Evasion, Then Forced Departments to Rewrite Assessment
When 林童宇 first learned Python, instructors strongly opposed AI-generated code. In severe cases, using AI could be deemed academic fraud and even result in failure. Students were already using it anyway, so the new hidden assignment became making code “look less like AI”: deliberately adding typos or using AI-powered rewriting tools.
He saw Chinese faculty move over 2 years from indifference, conservatism, or outright hostility toward active adoption. Computer science departments now rarely impose outright bans, while law schools have not made the same clear shift. 饶杰武’s professors also moved from resistance to allowing students to “consult with LLMs,” provided they took responsibility for the final output.
林童宇 compares this with model training: a model can score extremely well on a benchmark yet perform “terribly” in an open environment. Students can likewise be good at exams without being able to solve real-world problems. Continuing to use old-school evaluation would amount to training students who score well on tests but perform poorly in open-ended environments.
His proposed metric is: “Whether you use AI is no longer the question; we’re not discussing that anymore.” Schools should assess whether students can use AI well, choose problems they genuinely want—or society needs—to solve, and deliver results in an open environment.
6. College’s Career-Skills Value Is Eroding First, While Self-Discovery Is Gaining Value
林童宇 divides college’s traditional role into 2 layers. The first answers “Who am I, where is my passion, and where do I want to go?” The second builds specialized knowledge and skills that can secure a job and support a family. Elite schools in practice emphasize the latter, while many students still cannot answer the first set of questions when they graduate.
The problem is that the information intake, processing, and output skills education systems treat as valuable are already performed better, more cheaply, and around the clock by AI. Career barriers for programmers, lawyers, finance professionals, and other knowledge workers are therefore starting to weaken.
He argues that college should once again become a place to explore instinct and potential, and should incorporate the ability to solve problems alongside AI into its evaluations. Career skills are not useless, but they can no longer be treated as the endpoint of education. Otherwise, schools will train exactly the kind of executor most vulnerable to automation.
7. Liberal Education Turns Critical Thinking from a Slogan into a Survival Skill
After encountering a liberal-education-style course at Peking University, 林童宇 deliberately avoided employment-oriented classes such as securities law practice and international commercial arbitration, which held little personal interest. He shifted his focus to the history of Western legal thought, jurisprudence, constitutional law, and modern Western philosophy.
New雅 Academy at Tsinghua spent an entire semester discussing Meditations on First Philosophy. After finishing the original text, he was briefly shaken by the thought: “The world you believe in may not be the world you believe in.” Doubt can deconstruct one’s convictions without necessarily leading to nihilism; it can instead force a search for what still holds.
The experience showed him a part of life that AI struggles to replace: choosing courses, books, and ways of engaging with the world from a critical perspective, then continuing to ask, “Why am I here? What do I like? What am I good at?” This kind of foundational cognition reduces the fear of losing one’s footing as AI advances.
8. The Scarcest Assets at an Elite School Are Peers, Mentors, and Opportunities to Build Trust
In high school, 林童宇 ranked 1st or 2nd in his school. After entering Tsinghua, he immediately fell into the bottom 50% and at one point felt “like a piece of trash.” Once his mindset changed, he realized that outstanding peers could create pressure while also serving as friends and teachers.
Teachers’ value is not limited to classroom content. Several important professors spoke with him one-on-one during office hours, using their disciplinary knowledge and life experience to help him see certain consequences in advance and avoid “unnecessary mistakes.” That kind of feedback is difficult to reproduce through a standard curriculum.
He does not believe anyone is truly irreplaceable. But as specific skills fade, people connections, the ability to build trust, and the ability to persuade others will become more important—and will determine whether someone can fully leverage the resources available in a new environment.
If he could return to freshman year with what he knows now, he would “go all out socially.” He does not mean partying, going to bars, or accumulating activities, but pursuing more one-on-one, substantive conversations.
9. Degrees Will Not Disappear Easily Because They Sell a Social Signal, Not a Knowledge Inventory
When applying to NYU, 侯泰宇 wrote that socializing was the most important reason to attend college. He initially considered the University of Toronto because Geoffrey Hinton taught there. After learning that Hinton did not teach often, he chose NYU, where Yann LeCun taught, and audited his Deep Learning course. Once enrolled, however, he often read papers or worked on products in class because standard courses were “preconfigured” rather than designed for him personally.
He admits that his goal in college was “very utilitarian”: obtaining a degree. A degree is not proof of “what someone has learned,” but a signal to the market that “this person can be trusted.” Education is not merely about convincing students that they have learned; it is about convincing society that they can be entrusted with responsibility.
On successful dropouts, he sees survivorship bias. From Sumerian tablet schools to vocational education in the industrial era, institutions that endure are consistently tied to screening, evaluation, and talent certification. In a highly uncertain future, a degree will remain a “credential of trust” that people need to possess.
10. Continuous Trial and Error Matters More Than Long-Term Planning—Along with Asking What Technology Is For
侯泰宇 has not set himself a specific endpoint in life. He is more interested in asking what virtue, beauty, morality, and “the good” really are. The answer may be as invisible as a Platonic ideal, but “that does not mean we should stop pursuing it. Though we may not reach it, our hearts yearn toward it.”
His warning is that technological optimism alone cannot guide technology. A broad humanities education should cultivate virtue, not just calculation. Only by knowing what is good and bad can people steer AI rather than equate what can be executed with what should be executed.
Design Your Life gave him a practical method: avoid making rigid assumptions about medium- and long-term goals, and instead run many short-term experiments. Startup products, internships, and conversations with AI are all forms of trial and error—ways to learn what is necessary in less time and preserve energy for interests and self-discovery.
11. The Real Faculty Divide Is Between Banning GPT and Training Judgment and Tool Use
On the day 侯泰宇 graduated, a professor in a required Applied Psychology course was still handing every student a sticker saying “Do not use GPT.” Another website course, by contrast, told students on day 1 to discard their old coding software and switch to Claude Code, Replit, and Lovable. The education system is already visibly split in 2.
Drawing on Joseph Weizenbaum’s distinction between deciding and choosing, he says deciding means deriving an answer from rules under given conditions. Chess, some diagnoses, and shortest-path problems can all be handled by machines, potentially better than by people.
Choosing means making trade-offs when no objective standard exists. It is rooted in values, experience, and sensitivity to other people’s pain. A machine can receive a case and the facts of a crime, but whether someone should go to prison—or how life and death should be weighed—still requires a human to bear responsibility for the judgment.
He therefore believes exams may shift from written tests to oral examinations, while the creative space may move from paper to the Chatbot input box: students may be assessed on the prompts they create rather than only on the final essay. 侯泰宇 mentions NYU’s new AI-focused school, while Stanford has converted a famous coding course into a web-coding course.
12. AI Expands Cognitive Breadth, but There Is a Fundamental Dispute over Whether It Can Compress Understanding Time
泓君 acknowledges that AI can filter information overload and determine “what to look at first and what to look at next,” but rejects equating faster retrieval with faster understanding. Forming a deep view of a specific question is still constrained by the brain’s processing capacity per unit of time; the history and origins of liberal-arts knowledge cannot be replaced by a summary.
侯泰宇 takes the opposite, top-down route. Traditional education builds from mathematics to linear algebra to machine learning, and students may forget midway why they are learning. AI can provide a real-world anchor first, letting learners see a top-tier answer and work backward to identify what they are missing.
While studying transformers, he had AI assemble a “panel of top experts,” including Geoffrey Hinton and Yann LeCun, to explain the overall architecture first. He then asked follow-up questions about every definition and example he did not understand. The route is no longer prescribed by the textbook designer; it is generated as the shortest path to a specific understanding.
He also acknowledges the top-down trap: being overwhelmed by the full body of knowledge at the summit. The solution is to identify the few “key distinctions” between the summit and the next level down. Answers should be verified through AI searches with links and then checked with friends who genuinely understand the field; the model’s output itself cannot be treated as the answer.
13. After “99% of Assignments Involve AI,” Competition Shifts to Decomposition, Judgment, and Quality Control
饶杰武 entered NYU in 2020 and first used ChatGPT in late 2022 while building a research website for an NYU professor. His data science courses did not teach complete front-end development, and GPT filled the gap. The toolset later expanded to Claude, Perplexity, Manus, Genspark, Claude Code, and Codex.
Asked how many assignments or papers had been helped by AI, he answered “99%.” The remaining 1% did not mean AI was completely absent; presentations still required him to appear and speak himself. Since late 2022 or early 2023, he has made a habit of “asking AI first” about any subject.
For assignments such as papers and websites that cannot be directly verified, he first provides the course materials and requirements, has the model break down the task, and then decides what comes first and second. For familiar machine-learning tasks, he opens Google Colab and configures the runtime himself; for unfamiliar courses, he lets AI take on more of the initial setup.
If a student understands the course and knows how to use the tools, AI output is likely already usable. Removing the AI “flavor” can produce an 80-, 90-, or higher-scoring result. If the student understands nothing, they cannot tell whether the model is hallucinating. When a model offers 3 or even 10 seemingly reasonable approaches, the final choice still belongs to the human.
14. By 2026, the Tool Stack Has Moved from a Single Chatbot to a Division of Labor Across Models
饶杰武 frequently uses Claude, Codex, and other Coding Agents. He also uses Variants to generate multiple animation concepts and pull the code directly; Typeless turns hesitant speech into structured prompts, while Dify is useful for taking an architecture from a paper and turning it into an experiment.
林童宇 divides his tools into Chatbots, productivity tools, and long-horizon agents. Gemini’s web app handles conversation; Typeless and Recast handle input and efficiency; OpenClaw handles frequent daily-life tasks and dispatches other long-horizon agents; Manus handles Deep Research and designing; Claude Code handles development.
The combination pushed his monthly AI spending from $20-plus early on to more than $200. His willingness to pay has a clear condition: the tools must generate more than $200 in monthly value, not merely satisfy the urge to try something new.
侯泰宇 remains most impressed by Claude Cowork, but also mentions Suno’s music capabilities, Gemini Dynamic View, and Inception Labs’ Mercury-2. The latter uses Diffusion LM and, according to the company’s own data, is roughly 10x faster than GPT-4o mini. His cautious conclusion is only that this architecture “might” break the trade-off between speed and quality.
15. A General Orchestration Layer Has Not Killed Vertical Agents; It Has Made Mature Products More Frequently Used
林童宇 continues to use Manus because it packages complex technology into a user-friendly product that requires no personal configuration. When he sees a well-designed website, he asks Manus to inspect how it was built in a browser rather than assembling an entire toolchain himself.
林童宇 acknowledges that OpenClaw still trails mature Agents in output quality and execution efficiency on specific tasks. But its instant IM interaction, memory, and ability to switch across tasks make it suitable as an orchestration layer. The result is not that it replaces Manus or Claude Code, but that it dispatches them more smoothly—and usage frequency rises.
侯泰宇 does not currently use OpenClaw. He understands each of his workflows and which model fits each one, so he does not need an end-to-end package. Claude’s API is also too expensive, and he does not trust the performance of other models enough. For him, fixed subscriptions are more controllable than letting a general-purpose Agent burn through large amounts of API usage.
16. The First Heavy Users Are Canceling ChatGPT and Handing High-Value Work to Claude and Manus
侯泰宇 has largely abandoned GPT and Gemini, moving entirely to Claude and buying the $200 Max Plan. After exhausting his quota one Saturday and having to wait until 7 p.m. on Sunday, Monday, and Tuesday for it to reset, he “completely lost trust” in alternatives such as Gemini CLI and did not even know what to do for 2 days.
He observes that Claude’s token-in and token-out costs continue to rise, while many competitors are using discounts to buy growth and users are still willing to pay Claude. In his view, that is a contrarian signal that tests whether the product is genuinely valued.
泓君 and 林童宇 have also canceled ChatGPT, which accompanied the first wave of AI adoption, while retaining Gemini for visual work and Codex for coding. 林童宇 maintains task.md or progress.md; when Claude Code pauses or runs out of quota, he hands the state files to Codex to take over.
Manus was once criticized as a “wrapper with no moat,” but a year later it had retained high-value tasks through user experience. 林童宇 particularly values its ability to browse Reddit and X through accounts logged into a local browser and generate Deep Research reports, as well as its Canvas design experience. ChatGPT’s app, by contrast, has fallen into a “three-way no-man’s-land” where it is not the strongest at conversation, Agents, or coding.
17. Memory Can Build a Personalized Model—and Create an Information Bubble
侯泰宇 assigns GPT, Claude, and Gemini different cognitive roles. GPT has memory enabled to understand him, brainstorm, search, and organize the first version of a prompt. Claude and Gemini have memory disabled so past preferences do not distort fresh judgments; Claude’s extended thinking then handles execution.
He does not let the GPT that understands him best make actual decisions because the model has been “contaminated” by his personal memories and lives inside an information bubble. 泓君 agrees: long-term memory not only makes answers increasingly accommodating, but can even assimilate English expression into Chinglish.
Gemini’s role is visual production, including Nano Banana Pro and Dynamic View. The outputs are then inserted into Claude’s workflow, creating a division of labor between brainstorming, execution, and visual production.
18. AI Amplifies Existing Capabilities and Exposes Weak Foundations at the Quality-Control Stage
侯泰宇 used applied psychology for a cross-disciplinary project examining whether avoidant, anxious, and secure attachment styles affect how people interact with AI. He used AI to assist with the research and extended theories originally developed for human relationships to human-machine relationships, concluding that both his exam performance and exploratory interests improved.
泓君 uses design to illustrate the underlying constraint: a good designer will generally deliver a better image whether drawing by hand or using AI, because layout, typography, and subtle distinctions require years of accumulated taste. Someone with no design training may still generate an image, but without aesthetic judgment, cannot guide the model toward a higher standard.
饶杰武 sees the same problem in code-heavy projects. Gemini coding may forget a bracket or append a new branch instead of replacing outdated code, then continue making baffling errors after running for dozens of hours. Clients care only about delivery; the operator must understand the critical points and correct the tool in time.
That is why the 3 participants experience AI interruption differently. 饶杰武 was already writing code in 2018 and 2019 and could fall back to Stack Overflow without Claude Code. 林童宇 says writing a paper after completely removing AI would be extremely painful. 侯泰宇, who had lived alongside AI from his first day of college, says bluntly: “Without AI, I can’t write code.”
19. AI Addiction Includes a “Productivity Placebo,” but the Youngest Users Reject the Premise of Going Back
侯泰宇 calls the dependence “insanely addictive—more addictive than short video”: he can go a day without scrolling short video but not a day without AI. Losing Claude would mean losing roughly 70% of his execution and 30% of his inspiration. 饶杰武’s dependence is more purely work-driven: compressing a 10-hour task into 30 minutes gives him a rush. 侯泰宇 also gets the learning and conversational gratification of waiting for complex answers.
饶杰武 says calendars, Gmail reply drafts, and software development are already embedded in AI and Agents. 林童宇 admits that AI has eroded some capabilities and worries that keeping a human in the loop may be “wishful thinking.” Drawing on Norbert Wiener’s The Human Use of Human Beings, he locates the human remainder in asking the right questions, making value judgments, and bearing the consequences.
泓君 cites METR’s randomized controlled experiment: 16 senior engineers with more than 5 years of experience completed roughly 200 real-world tasks. The AI group was actually 19% slower but believed it was 20% faster. The “productivity placebo” comes from immediate feedback masking verification, debugging, and rework; the “senior paradox” is that generalized suggestions disrupt expert workflows, while beginners benefit substantially from their lack of prior knowledge.
The labor-market message is therefore sharper: “You are not being replaced by AI. You are being replaced by people who are better at using AI than you are.” Competition is rising in law, legal services, and public-sector legal work, while some North American tech companies have frozen software roles and startups have become a fallback. Companies still do not know how to define “good at using AI.”
林童宇 expects the rush to continue for another 1 or 2 years. Individuals cannot extrapolate linearly and will have to rely on adaptability, conviction, and belief. His long-term direction is emancipation: he hopes technology will eventually free people from compelled work and provide basic living security, while acknowledging that institutions, economic structures, and technology are all far from mature.
After the episode was recorded, a Silicon Valley high-school student asked 泓君: “Why make that assumption? We’re not going to leave AI.” For digital natives, “What would you do without a phone?” was never a meaningful premise. A generation 3 or 4 years younger may already believe that anything AI can do no longer needs to be learned. College students are simply the first to sit for the exam; the paper asking what is useful and what is irreplaceable has actually been handed to everyone.