Interview with Onion Academy’s Yang Linfeng: 12 Years to the AI Era

Summary

• Onion Academy rejects being classified as an online education company, defining itself instead as a “learning experience design company.” It does not provide “ride-hailing- or food-delivery-style” problem-solving services, but focuses on whether students enjoy learning and genuinely grow. With no brand advertising for 12 years, it has reached 110M cumulative registered users through student and teacher word of mouth, distributed “roughly evenly” across China’s tier-one to tier-five cities and counties—evidence, in Yang Linfeng’s view, that the product is genuinely low-threshold.
• Its business model is the inverse of conventional tutoring: subscription-based, with no offline delivery and no human service, priced at one-fifth to one-tenth of tutoring classes. One subject costs about RMB500 a year; seven subjects cost a little over RMB3,000 before additional discounts. “Usage comes before transaction”: students use it free first, and parents pay later once they realize its value. The company has raised no further financing since 2020 and has gradually reached self-sufficiency; nearly half of its 1,000-plus employees work in R&D.
• The costs and rewards of the productization path are unusually clear: the first 2 years were a “handcrafted workshop” that produced only 200 videos, each taking 8 weeks, with negligible user numbers. It was “like building an operating system or a chip—you only know whether it works once all the investment is on the table.” But once built, “Onion’s product is guaranteed to be better each day than the day before”—something a human-delivered services company “wouldn’t dare say.”
• His core judgment on this AI cycle is that “the bigger potential of large models is not AIGC, but their ability to dispatch and plan.” The model should act as a brain that orchestrates Onion’s proprietary content library, rather than generate explanations from scratch. The approach went live in 2024 and accelerated in 2025 after “DeepSeek equalized access to chain-of-thought technology,” producing the multi-agent “AI intelligent study companion.” His repeated anchor point: “A model being able to solve a problem and being able to make a student able to solve it are two different things.” No matter how strong the model is at problem-solving, it is useless if it cannot solve the student’s motivation problem.
• As more large-model companies enter the space, Yang ranks the moat as follows: “In the large-model era, the more data you collect on a single user, the greater your advantage.” Onion has accumulated 500B interaction records, including video-by-video playback heat maps, along with 12 years of research on students, privately owned copyrighted content, and a proprietary knowledge graph. It will not fine-tune its own model: “You keep that advantage for 2 or 3 weeks, and once a new foundation model comes out, it is no longer necessary.”
• The current industry landscape is hostile to investors. Learning-device growth slowed in Q3 this year; tutoring survives under new forms. Investors see education, healthcare, and finance as areas where AI may achieve commercial closure relatively easily, yet remain “stuck” on the question: “Once you invest in education, how do you exit?” There is no answer, only small checks. His review of the 2021 bubble is blunt: the huge financing volumes “were not used on the supply side; they were used to burn money on demand,” with advertising helping create the parental-anxiety “theater effect.”
• His 10-year outlook is a bet on autonomous learning as an asset. AI products will become readily available, and “all knowledge will immediately become obsolete.” A seventh-grader today will graduate and enter society 10 years from now, yet is still spending those years on photo-search problem solving and passive listening—“spending 12 years practicing an ability that is fundamentally unnecessary.” “It is the same logic as global warming: if we do nothing now, something major will definitely happen 10 years from now.”

Deep dive

1. Positioning: A Learning Experience Design Company, Not a Problem-Solving Delivery Service

• Li Xiang opened by admitting he was unsure how to classify the company. Yang Linfeng drew a clear boundary: by industry classification it belongs to online education or edtech, but its philosophy is different—“We would rather position ourselves as a learning experience design company,” focused on whether students enjoy learning, genuinely develop ability, and come to love learning because of Onion.
• The defining contrast was explicit: “We do not treat learning tasks as errands to be handled, or like ride-hailing and food delivery… We are not focused on the efficiency of solving a problem. The core question is whether the student genuinely grows through the process.” Every subsequent business choice follows from this distinction.

2. Origins: Harvard CS, Rural Teaching, and Cognitive Science as the Foundation of Learning

• Yang studied computer science at Harvard from 2007 to 2011. Most of the founding team were recent graduates who had done rural teaching and education work while in school. They discovered that “the underlying foundation of learning is actually cognitive science”: teach in a way that matches how people think and the results are twice as good; teach against it and learning becomes painful.
• His stock answer to whether learning is inherently against human nature: “It depends on whether the method follows human nature… If you can activate that layer of curiosity at the bottom of people’s minds, learning can become very interesting.” The team believes it holds the key and has a “duty-bound” responsibility to use it.
• The methodology is not based on intuition. It draws on taxonomies of educational objectives, educational psychology, and cognitive science, down to “how every sentence should be polished and what kind of example should be used.” The team also speaks with students online and offline and visits schools to collect frontline feedback.

3. Closing the Urban-Rural Gap: Optimizing Every Minute of the Learning Experience

• Li listed popular solutions such as donating tablets, consolidating schools, and training principals. Onion’s answer is to empower teachers and directly empower students: “What ultimately constrains a student’s learning experience is whether each minute of the learning itself is interesting and efficient.” That is the most basic element once the problem is broken down to its smallest unit.
• This led to 2 business lines: helping students self-study through to-C products, and improving in-class experiences through to-B products. The vision is captured in 3 phrases: “Put learners first, empower educators, and advance education equity.”

4. Cliffs and Slopes: The Best Metaphor in the Interview

• Why can a fellow student make something understandable in a few words while a teacher cannot? “A teacher has already made the climb. They stand on a mountaintop shouting at students, ‘Come on, climb up, it’s easy!’—but what students see is a cliff.”
• Onion’s job is to “build a gentle slope”: “Can we make it so that when most students look at knowledge, they see a slope rather than a cliff?” When Li asked whether top students could understand ordinary students’ pain, Yang returned to scientific methodology: “If Onion relied entirely on individual experience, this company could never have made it this far.”

5. 500B Interaction Records: Refining Heat Maps Like Livestream Commerce, With a Different Goal

• Onion tracks everything and has accumulated more than 500B student interaction records. It analyzes playback heat maps for every video: “The pause rate suddenly spikes at 3 and a half minutes, which means we explained something poorly here.” The script is rewritten, the video relaunched, and the result tested again; successful changes are incorporated into the methodology.
• Li compared the process with the completion-rate methodology of Douyin livestream commerce. Yang accepted the resemblance but stressed the difference: “We are not focused on getting people addicted to the activity. We care more about whether the process lets students move smoothly forward.” The deeper point is metacognitive: “When you learn knowledge, you are not learning the knowledge itself; you are learning to build logical and cognitive frameworks through knowledge.”

6. From Public Welfare to Company: R&D Was Too Heavy for Fundraising

• The timeline: Yang co-launched a public-welfare project in his junior year, 2010, and piloted it in 2 rural schools in Liangzhou District, Wuwei, Gansu, in 2011. After working at BCG for a year, he resigned and worked full-time on public welfare from mid-2012 through the end of 2013. The turning point was realizing that “having a public-welfare organization do R&D creates enormous pressure on fundraising,” so a company was established in parallel at the end of 2013—2 independent entities, “not a succession relationship,” but with a consistent mission.
• From the founding flag onward, and in every financing round, the company made its commitment clear to investors: “The product must be made freely available to schools and students in rural areas.” The public-welfare brand Onion Teaching Assistant Action has since reached more than 30,000 rural schools, more than 60,000 rural teachers, and about 4.7M students, all free of charge and supported by a dedicated full-time team.

7. Why It Had to Be a Product, Not People: Bringing SOTA Learning to Rural Students Without Dilution

• He takes the human-delivery model to its logical endpoint: when the cost is people, “your focus will inevitably be on those who can afford to pay,” leaving public welfare as little more than branding. Rural students receive the same experience as urban students, with no tiering: “We say the model is SOTA today, and we want to provide rural students with the best learning method currently available, without loss. After turning it over and over, there is only one way: use technology to build a product, not people.”
• The data validates the approach. Comparing Onion’s 110M cumulative registered users by county, Yang found that “the distribution was broadly similar to each county’s share of China’s total population; tiers one through five were essentially evenly distributed.” If a student has a phone and an internet connection, the product is usable.

8. Rural Areas Do Not Lack Buildings; They Lack High-Quality Experience

• China has invested heavily in education hardware and “should be one of the best-built countries in the world in this respect.” Donations to Hope Schools have declined as rural populations move to cities and village schools merge because they cannot fill their classrooms. “Overall, what is missing now is not buildings; what is missing is a high-quality education experience.”

9. The Handcrafted Workshop, 2013-15: 200 Videos in 2 Years, With Everything on the Line

• During the early fundraising rounds, “everyone thought we were crazy”—how could students possibly study independently? Yang’s rebuttal was a comparison between teachers: “Two history teachers explain the same knowledge. One is completely absorbed in it, the other is half-asleep. Is that a problem with the knowledge itself? Of course not.”
• In February 2014, Onion entered a high-quality Beijing public school and received its first wave of positive feedback. But after the Web version launched, “engagement was extremely high in the first month, then fell off a cliff.” The reason was simple: only the first 2 chapters had been built, so students reached chapter 3 and found nothing there. The lesson was that a systematic K12 product must first cover the entire academic year.
• The team then spent 2 years turning the core concepts of seventh- and eighth-grade mathematics into 5-to-8-minute animations. Each video took 8 weeks through the production pipeline, for about 200 videos in total. “For a company that had raised only one round of money to go 2 years without releasing any product, the risk was extraordinarily high.” In October 2015, the iOS and Android apps launched and received a homepage recommendation on the China App Store in the first week. Student word of mouth quickly pushed users past the first 1M—all free.
• The defense of this high-risk model was one of the interview’s most substantive points. Human delivery “starts fast, though it may not stay fast”; productization is “extremely slow upfront, with enormous investment,” but each iteration compounds inside the product. “We can say responsibly that Onion’s product is guaranteed to be better each day than the day before. If we were a human-services company, I would not dare say that.” Li’s comparison was accepted: it was like building an operating system or a chip—you only know whether it works once all the investment is on the table.

10. Extreme Testing and Subject Expansion, 2016-19; Self-Sufficiency After 2020

• The team did not simply work blindly for 2 years. At the end of 2013, it ran an “extreme test,” delivering the same demo simultaneously to Beijing’s most elite public schools and the most remote rural schools. “We tested both extremes, and then we could sleep easy.” Usage behavior on the consumer side showed no urban-rural difference.
• From 2016 to 2019, Onion expanded subjects and grade levels: from middle-school math to physics and chemistry, high-school math, and elementary math, eventually covering all 9 core subjects in middle and high school. It also dug deeply into use cases: previewing, getting unstuck on homework, same-day review, weekend consolidation, systematic winter and summer vacation review, and pre-exam preparation. “The solution for each narrow scenario is actually different.” Adding subjects directly increased reach and payment interest, though “every subject expansion required enormous effort.”
• 2020 onward became the monetization phase: the company built a complete business model and sustained revenue growth. “We have not raised another financing round since 2020.”

11. Business Model: Premium Subscription, With Payment Deferred

• The starting point was straightforward: setting competition aside, education remains “a field where parents have a strong habit of paying.” But students must be able to use the product without a barrier, and students’ ability to pay is constrained. So Onion “moved the payment barrier to the back,” adopting a premium model of free content plus paid value-added subscription packages, with deeper learning and difficult-problem explanations placed behind the paywall.
• With few direct benchmarks, the company spent a long time finding the right price. The industry norm was to sell tutoring through classes; Onion chose subscription, no offline delivery, and no human service. Its structure reflects the model: nearly half of its 1,000-plus employees work in R&D, split roughly 1:1 between technology and product on one side and course content on the other. The animation team is full-time rather than outsourced; the rest are in operations, sales, customer service, and a nationwide field-school service team of more than 100 people.

12. Usage Before Transaction: Students Are the First Users, and Teachers Start With Their Own Children

• His insight into the industry’s transaction structure deserves separate treatment: “For most education companies, the first customer they serve is not actually the student but the parent—the transaction happens before usage. We want usage to happen before the transaction.” Parents’ greatest fear is that their children lack motivation. “When they see their child become interested in a learning product, trust is established very easily.”
• Word of mouth works the same way among teachers. Onion has about 4M public-school teacher users, and online comments often say, “This was recommended by our teacher.” The most persuasive detail is that “when teachers first started using Onion, they would first choose to give it to their own children”—a level of conviction that makes the choice meaningful.

13. Zero Advertising Growth Versus the Distortions of an Ad War

• “We have never run a single brand advertisement—on elevators, the internet, television, or variety shows.” Acquisition is driven primarily by student word of mouth—“The teacher explained it in class, and I solved it on Onion in 5 minutes”—along with teacher referrals.
• Heavy advertising by peers still created real pressure. “Of 20 billboards at a bus stop, 15 were practically online education ads.” Parents were pushed toward large live classes, while products like Onion “never became a mainstream category in parents’ minds,” making trial harder to sell. After 2019, the company deliberately reduced its external communications: “The market was too overheated, and everyone instinctively benchmarked us against tutoring. It took twice the effort for half the result, so we would rather spend the time strengthening the fundamentals.”

14. Price Anchors and Learning Tablets: Hardware as a Supplementary Commercial Strategy

• It is difficult to establish a price anchor for a human-free digital course: “A Tencent Video membership also costs a little over RMB200 a year, so why should your course cost that much?” Bundling the course with hardware gave parents an intuitive reference point for “a learning machine costing several thousand yuan.” That is why Onion made a learning tablet. Today, parents generally own at least 1, and often 2, learning devices, and their understanding of digital-learning pricing has matured. Hardware “has always been a supplement to the commercial strategy, not the main business.”
• The initial price anchor was one-fifth to one-tenth of a tutoring class: about RMB500 per subject per year, and a little over RMB3,000 for 7 subjects. “The total price did not seem all that cheap,” so the company added further discounts for the full package. Li complained, “Everything can clearly be done on one pad, but there has to be dedicated hardware—I have never understood it.” Yang Linfeng replied, “Yes.” Li then added: “But parents are still anxious.”

15. Revisiting the 2021 Bubble: Money Went to Demand, While Students Practiced “Listening” for 12 Years

• His assessment of the boom is unsparing. Online education was once “China’s most money-burning industry”; absent the later disruption, the amount of capital deployed in 2022 “might have been several times higher than in 2021.” Yet most companies offered “astonishingly similar” services—find a teacher and teach, simply moved from offline to online. Li asked whether this had inflated the wages of Peking and Tsinghua graduates. The enormous financing volumes “were not used on the supply side; they were used to burn money on demand.” Advertising created a “theater effect”: “It seems everyone is taking classes. Do I need to enroll my child too, or else I am an irresponsible parent? That mindset is frightening.”
• The deeper criticism targets the learning method. Putting children in a physical classroom “can only imprison their bodies.” For 12 years, students practice one thing: listening. Yet they do not practice metacognitive abilities such as acquiring knowledge, planning steps, regulating pace, pausing, backtracking, investigating, and digging deeper—“the most important abilities after entering society.” They “spent 12 years practicing an ability they fundamentally do not need, which is putting the cart before the horse.”
• Did he ever consider pivoting? No. “We were never on that track in the first place, and we did not benefit from that track’s upside. So when something happened to that track, it did not have much to do with us.”

16. Industry Snapshot: Learning Devices Slow, Large Models Enter, Investors Worry About Exits

• The current picture is fragmented. Tutoring models continue under new forms, including live-teacher classes, livestream classes, and repackaged offline formats. Learning devices grew rapidly for several years but “started slowing in Q3 this year.” Smart teaching aids bundle textbooks with online courses and sell them at higher prices. Many large-model companies are also entering: “Photo-based problem solving is an excellent application scenario for a large model; you can charge forward with it,” making it a flagship feature for K12 acquisition. “The industry is still quite chaotic, with relatively intense competition.”
• Investor sentiment can be summarized as “conflicted.” After analyzing the market, investors conclude that “education, healthcare, and finance seem to be the areas where AI can most easily achieve a commercial closed loop,” but remain stuck on “once you invest in education, how do you exit?” There is no settled answer, only small checks. Well-known existing companies “should already be too expensive,” and face the same exit problem.

17. Lessons From Khan Academy and Duolingo: System 1, System 2, and America’s To-B Tradition

• Yang was in the United States from 2007 to 2011, during the rise of Khan Academy around 2007-08, Duolingo around 2008, and another product around 2009. More than 50% of U.S. K12 edtech products may be paid for by schools. The sector studied deeply how teachers teach and students learn, giving him the early conviction that “building an education product is absolutely not about pursuing something fast and easy.”
• Khan proved that “a student self-learning product without human participation” could work, but direct transplantation to China failed. Translation and subtitles did not work; a writing tablet with voice-over could not attract Chinese students. Chinese questions are harder and more varied. Duolingo and Khan also operate on different foundations: Duolingo mobilizes System 1, while Khan mobilizes System 2. Their processes and product-design logics are fundamentally different.

18. Efficiency-Completion Versus Capability-Building: The Long-Term Inefficiency of Shortcut Tricks

• He divides education products into 2 types. Efficiency-completion products treat problem solving like food delivery: “The transaction between us is complete and we are done,” with a focus on engagement, frequency, and retention. Capability-building products say, “I want to accompany you and change your underlying abilities.” Most Chinese online education companies belong to the first category, which is also easier to measure. That is Onion’s practical dilemma.
• The failure chain for efficiency-completion products is straightforward: provide an answer plus a mnemonic “trick” that works only for that particular problem → solve 100 to 1,000 problems and memorize 100 to 1,000 tricks → the exam changes and the trick no longer applies → “answers become too easy to search for, and students lose the motivation for deep thinking” → “apparent short-term efficiency improves, but exam performance does not.”
• Capability-building is more efficient over the long term. It teaches general methods, the known-unknown logic of parsing a problem, and reusable thinking frameworks, while explaining the scientific history behind concepts and how knowledge transfers. Tutoring institutions ask, “Why are you teaching all this?” But “more neurons are activated because the knowledge network is connected.” Solving 10 problems refreshes the chain of general problem-solving methods 10 times. “Once you introduce the time dimension, looking 2 weeks or 1 month ahead, capability-building learning is far more effective than efficiency-solving learning.”

19. Mobile Internet’s Real Gift: A Student’s First Device That Truly Belongs to Them

• Mobile internet “gave students autonomy over a learning device for the first time.” A computer sits in a parent’s study or living room, and when a student opens it, the parent’s first reaction is, “Are you going to play?” A phone is the student’s “first electronic device that genuinely belongs to them.” The industry once wondered whether students would study on phones; in hindsight, that concern was misplaced.
• Li asked the obvious question: Couldn’t textbooks support self-study? The answer was direct: “Textbooks are not organized primarily for students to teach themselves. We call them textbooks, not learning materials.” They are designed for teachers to teach from and are static. The user-side evidence is familiar: when a teacher assigns previewing page 167, the student’s actual response is, “It is not that I do not want to read it; I cannot understand it.” The next day is no different from having done no previewing at all.

20. Why MOOCs Did Not Take Off: Self-Study Is an Advanced Skill, and Parents Are Trapped in a Vicious Cycle

• On the claim that the internet transformed education with much thunder but little rain, Yang admits it is “not entirely an illusion.” Money went into livestream technology, operating teachers, and trial-course acquisition, “not into changing the way people learn.” The deeper reason Coursera and edX did not become mainstream learning methods is that “self-directed learning is without question a very advanced skill… It involves not only an ability but also a belief and a set of values.” It requires the right environment, habits, positive feedback, and the confidence that “I am someone who can understand things on my own.”
• The parental paradox is painful: parents sit beside children doing homework and ask, the moment the pen pauses, “You do not even know this problem?” “The less you let go, the less they can practice; the less they practice, the less they possess the ability; the less they possess it, the less willing you are to let go.” The cycle feeds itself.
• On the claim that education is against human nature, he says deep cognition does require the energy-consuming System 2. “But if learning is not a human instinct, how did humans evolve from apes? The instinct to learn is written into your genes; it has simply been covered by the brain. What kind of product can awaken that ability? That is the interesting question.”

21. AI’s Real Potential Is Dispatch, Not Generation: The Step Change Brought by DeepSeek

• The core judgment: “Many people focus on AIGC capabilities… but we believe its greater potential lies in its dispatch and planning capabilities—understanding what the learner needs, finding the right content from a professional library, and delivering it.” For example, a student photographs a textbook page about accelerating an electron in a magnetic field and says it is incomprehensible; the model automatically selects the most relevant clips from Onion’s proprietary video library and supplies the surrounding explanation.
• The timeline: because K12 products could not use overseas models, Onion began offering large-model products to users around the end of 2024. The acceleration in 2025 came because “DeepSeek equalized access to chain-of-thought technology, producing a qualitative leap in large models’ planning capabilities.” Combined with RAG, the model became the dispatching central brain, and “product iteration moved extremely fast this year.”
• The product is an “AI intelligent study companion.” “It is not one agent but the collaboration of multiple agents”: previewing, review, getting unstuck, wrong-answer training, exam preparation, learning motivation, and psychological support each have their own division of labor. They share a common infrastructure of course content, questions, the knowledge graph, interaction data, and teacher-student progress. There is also an AI planner: “If you want to raise a weak math score from 60 to 90 in 2 months and tell me you have time on Wednesdays and Fridays, I will schedule what to study each week and you can follow it.”

22. A Clear-Eyed View of the ChatGPT Moment: A Model Solving Problems Does Not Mean Students Can

• His first encounter with ChatGPT was “definitely shocking, as if I had seen a shadow of intelligence.” But he quickly saw the limit: purely generated textual explanations remained “a cliff” for students—“You do not read it; if you read it, you are not interested; if you cannot understand it, it is useless.” Many students also lack the ability to ask follow-up questions. If they do not dare ask in class when they are confused, they are even less likely to describe their confusion to a machine. His initial conclusion was that large models could solve the “small loop” of explaining the point where a student was stuck and serve as a Q&A assistant, but he had not yet figured out how to overturn the entire product flow—until chain-of-thought technology appeared.
• His anchor remains firm: “A model being able to solve a problem and being able to make a student able to solve it are two completely different things. No matter how strong the model is at solving problems, it is useless—you are largely solving the student’s motivation problem for them. We are thinking about how to connect such a powerful external tool into the process and activate the person.”

23. Moats Against Model Companies: Deep Single-User Data, Student Insight, and Proprietary Assets

• The most important industry insight is this: “In the big-data era, the larger your total data volume, the greater your advantage. In the large-model era, the more data you collect on a single user, the greater your advantage. Having 1 piece of data on each of 10,000 users is not very useful; having 10,000 pieces of data on one person lets you build an extremely precise profile.” Onion has the latter: students stay with it for years, and the data reveals latent personality traits—an impulsive student skips through videos, an overly cautious student rewinds twice after understanding, and a strategic student quickly skips what they already know.
• Second is 12 years of research into the “genetics” of students; in theory, Onion should understand this population better than a general-purpose model company. Third is proprietary assets: copyrighted in-house content, a proprietary knowledge graph, and data on student mastery.
• The company explicitly does not—and for now will not—train its own model. “Model training is expensive, and foundation models keep updating in an intensely competitive market. You fine-tune one model and the advantage lasts 2 or 3 weeks; once a new foundation model comes out, it is no longer necessary.” The better timing is to wait until model capabilities stabilize and the company has a clearer product logic. “This is not where we should be spending our time right now.”

24. Education Deals With People Who Change: Internet North-Star Metrics No Longer Work

• Some AI-plus-education startups mainly use the off-the-shelf capabilities of foundation models. Li mentioned several overseas products growing rapidly; Yang cited overseas university students using photo-based problem solving, essay correction, and English speaking dialogue.
• Yang’s education-science perspective was the interview’s most systematic critique: “Most internet products deal with an unchanging person. They do not consider your transformation; they quickly satisfy what you want right now and end. Education deals with a person who changes.”
• The chain of reasoning is clear. When a user rejects recommended content, an internet company changes the algorithm and pushes 10 more pieces. “The problem is not the algorithm. First, the user is not ready; their ability and cognitive level have not reached that point. Second, none of the 10 things you gave them are good.” Replacing the textbook with 10 others will not solve the problem either. “What you need to change is the format, and if the new format does not yet exist, your company has to build it itself.” Internet companies are accustomed to matching and distribution, not creating supply; that is too heavy.
• More fundamentally, the metrics collapse. System 2 does not easily generate dopamine, so traditional immersion-design methodologies fail. “Stickiness, retention, and interaction frequency are all unimportant in the face of learning.” Organizational inertia then pushes teams to keep adding retention features, taking them farther from the actual solution. He also makes the self-deprecating point: “Top students cannot understand ordinary people. Many executives are very used to mobilizing System 2 and cannot understand why most people are unwilling to. Which product can make students voluntarily contribute their System 2? That would already be a major breakthrough for humanity.”

25. The 10-Year Outlook and Organizational Self-Learning: A Global-Warming-Level Risk

• The arithmetic is concrete. ChatGPT has transformed the world in less than 3 years; the next 10 years are “3 three-year periods.” AI products will be readily available, “all your knowledge will immediately become obsolete,” and people will be forced to learn new fields every day. “The most important thing then will be autonomous learning ability, not this knowledge.” Yet “a child in seventh grade today will graduate from university in exactly 10 years after 6 years of secondary school and 4 years of college,” still using photo-based problem solving and passive listening. “To use an inexact analogy, it is the same as global warming: you do nothing now and it looks fine, but something major is bound to happen 10 years from now.”
• Asked about current difficulties, he gave a candid symmetrical answer: Onion itself must learn autonomously. “Onion is also a 12-year-old company, so it inevitably has a lot of accumulated experience and baggage.” The hardest part is the psychological threshold in human-machine division of labor. People can accept AI as a co-pilot, but “stop doing this task and give this role to AI” makes them uneasy. It is like a teacher being told, “Do not teach this content; let students self-study first”—the teacher feels as though they have lost their livelihood.
• What interests him most is not education companies but “the way companies generally design their AI organizational structures.” OpenAI is experimenting with having some departments managed entirely by AI. Altman may have said that by 2030, a company could become something related to AI operations and the CEO role. “The organization must first become capable of keeping up with the times before the product can keep up with the times. That may be the more fundamental issue. These experiences are extremely valuable to me; I spend a lot of time studying them every day.”