167: Onion Academy’s 杨临风: Using AI to Manufacture Shortcuts Is Killing Real Learning
Summary
杨临风’s core judgment is that AI can deliver an answer, but it cannot complete the process of actually learning something for a student. Education is not e-commerce: the outcome is not the whole product. Real delivery means having students experience understanding, questioning, thinking and application themselves; if AI merely indulges System 1 and manufactures shortcuts, it may amplify cognitive laziness. AI should instead help students get past sticking points and return to independent thought.
Onion Academy is betting not on general-purpose model capability, but on more than a decade of accumulated digital-native learning experience and education know-how. A 5–8-minute interactive animation costs about RMB100K on average, takes roughly 2 months to produce and involves more than 20 steps, backed by several hundred billion event tags tracking pause, exit and rewind rates by the second. 杨临风’s threshold argument is blunt: “Spend RMB50 and the output is still zero; only at RMB80 does the output start to work.”
This is a business model with extremely heavy upfront fixed investment that becomes roughly SaaS-like at scale, rather than one that piles up GMV through paid traffic and teacher labor. Onion has accumulated about 130M student registrations and 4M teacher registrations, relying mainly on word of mouth, organic growth and in-product conversion. It first broke even for the full year in 2022, with NPS around 50–60. For investors, the key attractions are reusable content and data, marginal costs that do not rise in line with users, and the compounding effect of “tomorrow’s Onion will definitely be better than today’s Onion.”
Pure foundation models still cannot teach complex knowledge from scratch to ordinary K–12 students, making bounded assistance and fallback support the most effective role for now. Adults can absorb highly logical, structured model output, while K–12 students’ ability to understand structured material is still developing, creating a structural gap. Onion therefore does not train a base model in-house; it focuses on evaluation, workflow decomposition and necessary post-training or fine-tuning, while mixing models from ByteDance, Alibaba, DeepSeek and others across workflows to balance speed, performance and cost.
杨临风 has low confidence in the idea of an “AI education giant,” arguing that education is not one market but a collection of segments built on mutually incompatible philosophies. Grading, exam marking, teacher-process standardization, learning-process standardization and narrowly defined learning experiences all have different requirements. Even if a general-purpose platform acquired a large number of vertical companies, it might not be able to connect independent methodologies with short, self-contained supply chains into one unified system. Once a specialized product reaches 95, a general-purpose entry point will struggle to replace it through traffic alone.
Onion did not chase the 2016–2020 waves of photo-based homework solving, large-class livestreaming or dual-teacher courses because those models neither reused its accumulated capabilities nor fit its inclusive human-machine interaction strategy. The team discussed them no more than twice before walking away, citing “not our strength, no reuse of our accumulation, and unhealthy business logic.” Livestream courses use people to teach people, so costs rise with scale; standardized products can give rural and urban students the same experience. Being “hard and slow” is therefore also the consequence of the path Onion chose.
AI’s biggest incremental contribution to education is not replacing schools, but moving the classroom from a teacher monologue to a three-way classroom involving teacher, student and machine. Onion provides deep, on-site service to more than 2,000 schools each year, making the machine an intelligent teaching assistant for teachers and an intelligent learning companion for students. When students use classroom white space to ask questions such as, “Molecular motion can make molecules diffuse—is there a way to make them gather?”, AI can answer first and the teacher can then showcase the question, shifting class time from transmitting conclusions to cultivating questioning, reflection and ownership of learning.
The company’s public-interest mission is not a donation program bolted on after profitability, but a constraint embedded in the product architecture from Day One. Onion has provided free support to about 60K rural teachers, covering more than 30K rural schools and roughly 10 county-wide programs. 杨临风 believes “what society lacks most is not money, but methods, ideas and people who can get things done,” so the company must first become a healthy, sustainable business, but “cannot be merely a business.” His example: if over the next 10 years Onion can help 5M to 10M children learn more independently and confidently, that would already be an extraordinary achievement.
Deep dive
1. Onion Started with the Urban-Rural Classroom Experience Gap, Not the Online-Education Boom
Around 2010, while still an undergraduate, 杨临风 began volunteering as a teacher in rural areas with several partners who would later become co-founders. In remote villages, they saw cases where “the only building in the entire village was the school.” Hardware had improved materially; what had not changed was the way lessons were delivered, teacher-student interaction and the student experience.
The team initially imagined putting interactive content and educational videos on tablets and donating them to rural schools. Inspired by EdTech products such as Khan Academy, Duolingo and Coursera that emerged around 2008–2012, they spent about 1.5 years preparing and launched their first pilot in Wuwei, Gansu, during the summer of 2011.
The company was not incorporated until the end of 2013 and has now operated for nearly 13 years. By 杨临风’s figures, Onion has accumulated about 130M student registrations and 4M teacher registrations. But there was no mature “online education industry” when it was founded; commercialization was simply the means of keeping the original education experiment alive.
2. Incorporation Solved Sustainability; Public Interest and Business Were Never an Either-Or
The most practical bottleneck during the public-interest phase was not equipment, but the need for large numbers of cross-disciplinary people in curriculum research, animation and technology. The operation could not depend on part-time volunteers indefinitely, and once the team reached scale, donations could not reliably cover payroll.
杨临风 further clarifies that the team’s core objective is neither public interest nor business for its own sake, but building a high-quality learning experience and reaching more children. Product R&D is the largest fixed cost; if paying urban users can cover that investment, the same product can remain free for rural students, teachers and schools that meet the relevant standards.
That principle has held since 2013: paying users and rural beneficiaries use the same product, not a stripped-down public-interest edition. 杨临风 later summarized it this way: “It wasn’t that the product happened to be suitable—we designed it this way from Day One.”
3. Independent Learning Is a System of Motivation, Capability, Tools and Belief
杨临风 rejects the idea that failure to understand can simply be blamed on students being “unmotivated, stupid or unwilling to learn.” For students of normal intelligence, most K–12 knowledge should be understandable. The more commonly overlooked question is whether knowledge is being transmitted in accordance with cognitive principles and whether students can build deep understanding step by step.
He breaks independent learning into 4 layers: first the willingness to learn, then the capabilities needed to execute, followed by suitable tools to reduce friction, and finally the belief that “I can do this independently.” When the next unfamiliar problem arrives, students will choose independent learning again rather than wait for someone to feed them the answer.
When 曼祺 asked which element was more fundamental, 杨临风 said the endpoint is indeed belief, but the chicken-and-egg problem cannot be ordered so simply. A child may lack self-belief because he lacks capability, while parents demand initiative before they are willing to let go. The system then remains stuck.
The “I can do it” belief is particularly important in the AI era: students should be able to use tools to work through lessons they did not understand, homework they cannot do, previewing, review, exam preparation and interest-driven exploration. Only when action starts first and produces successful experiences can habits and awareness “multiply by time” into a positive feedback loop.
4. System 1 Prefers the Easy Route; Tools Should Pull People Toward the High-Energy System 2
杨临风 uses the distinction between System 1 and System 2 to explain why learning runs against human nature. Humans are wired for reflection, abstraction and organizing logic, but deep thinking consumes a great deal of energy, and adults often avoid invoking System 2 as well. “When parents say their children don’t want to think independently, they should first ask themselves whether they are willing to think independently most of the time.”
Onion cannot directly rewire human consciousness. Its entry point is to lower the barrier to active thinking. The product first lets students experience that independent learning is not as difficult as they imagined, then gradually builds capability, habit and belief through repeated successes.
When a student gets stuck, the ideal fallback is not to hand over the answer immediately. It is first to explain that getting stuck is normal, then diagnose the blockage and reduce the problem to an understandable level. After crossing the hurdle, the student should revisit why it happened and discuss how to handle a similar problem independently next time, turning one help session into metacognitive training.
5. “Deep Understanding” Means Writing the Learning Objective into Every Explanation, Not Merely Getting the Conclusion Right
Onion uses frameworks including Understanding by Design (UBD), Bloom’s taxonomy and the Feynman technique. Understanding is judged not only by the answer, but by whether students can explain it in their own words, give examples and make comparisons; these verbs correspond to verifiable levels of understanding.
杨临风 stresses that “the teacher explaining it correctly” does not mean “the student understands it.” A course must reconstruct the causes and consequences behind the knowledge and ask, line by line: Why expand here? How will the student’s cognition change? Can this sentence connect to the previous one? Does the student have a chance to use independent thinking to verify understanding?
These methods are not wall-mounted slogans. Every curriculum-research teacher must master them as instructional scaffolding. The team also uses AI to organize and flag methodological principles, but the final work is still to embed them in course scripts, interaction points and product flows.
6. Five to Eight Minutes Is a Complete Achievement Loop, Not a Content Slice
K–12 students can typically sustain focused attention for less than 10 minutes at a time, so every 5–8-minute unit must take them through the loop of learning something, understanding it and successfully applying it. A verifiable success creates confidence, which drives the next learning cycle.
The design also reflects the unsupervised setting. Once students are home, games, short videos and other apps compete for their time. If content does not let them engage comfortably and receive continuous feedback, they will not voluntarily open Onion to work out what they failed to understand that day.
7. Mathematics Came First Because the Most Abstract Subject Was the Best Test of the Method
The team started with mathematics for 2 reasons: it generates the most questions from K–12 students and feels the most abstract, while resources were sufficient to advance only one subject at a time. If the approach could get children to study math independently, transferring it to other subjects should be easier. When the team later developed physics, visible experiments and real-world phenomena did indeed make the material somewhat easier to design than purely abstract math.
In teaching plane geometry, Onion first unifies different problems involving sides, angles and circles under the idea of “transforming relationships between sides and angles”: list what is known and unknown, observe how far apart they are on the diagram, then build a connection step by step like constructing a bridge, using congruence or shared angles.
The value of this general method is not merely solving one problem correctly. Every lesson asks “why are we thinking this way?”—training students to connect the unknown to the known and break complex tasks into traceable steps. This is what 杨临风 calls metacognition and logical thinking.
8. “Onion’s Flavor” Regulates Cognitive Load Through Purpose, Empathy and Relevance
A sense of purpose means students always know why they are listening to the current explanation. The course tells them in advance what they will do in the next minute and why, then returns them to the broader objective afterward. This avoids the situation where they understand every word but do not know why the five minutes mattered.
Empathy does not mean lowering the difficulty. It means standing beside the student, acknowledging that the slope is steep, then pointing to the first stone on the path: “Trust me. Step up, get your foot on it, and then we’ll talk about the next one.” This reverses the feedback of “It’s easy—how do you still not understand?”
Relevance in learning motivation is also built through characters and context. Characters such as Li Goudan and Wang Xiaochui have continued from the early videos, allowing students to recognize their own confusion in the characters’ struggles. Teachers in different subjects likewise have differentiated personalities, including a lover of classical poetry and a geography teacher who travels by motorcycle.
杨临风 sees these design choices as regulating “brain pressure”: there should be excitement and stimulation, but not enough tension to trigger withdrawal. For students moving at different speeds, Onion uses a complete knowledge map, difficulty tiers, recommendations and AI supplementary explanations rather than assuming one pace fits everyone.
9. Animation Visualizes Abstract Concepts, Making It Far More Expensive Than Filming Live Teachers
Onion chose animation for a straightforward reason: “There is no information on the teacher’s face; the information is in the image.” A limited screen should make the concept move rather than devote most of its space to a live person. A real teacher—the “physics guy”—appears only when a physics experiment requires a real-world phenomenon.
A 5–8-minute interactive lesson costs about RMB100K to produce on average, takes roughly 2 months and involves more than 20 steps. Videos include questions, pauses, choices and feedback, while scripts are repeatedly debated through collective curriculum research rather than recorded once by a star teacher from existing notes.
Those debates focus on student cognition: Why say this sentence? Can a child understand it? Does it connect to the previous sentence? If the team cannot convince itself that the entire logic is coherent, the content is not released directly to students.
10. High-Quality Learning Experiences Have a Nonlinear Threshold; Slowness Is the Price of Effectiveness
曼琪 noted that Onion produced only about 1,500 animations in its first 5 years, far slower than Khan Academy. 杨临风’s response was that quality below the threshold cannot trigger independent use: “Spend RMB50 and the output is still zero; only at RMB80 does the output start to work.”
Asked why Khan Academy has still been able to operate for so long, he cited 2 differences. Students who can use Khan Academy consistently are already a selected group, and Khan is a nonprofit that does not need to prove that ordinary families are willing to pay. Onion faces both greater knowledge difficulty and tougher commercial requirements.
Courses are not sealed away once completed. By 杨临风’s account, Onion’s videos have been played several hundred billion times, with several hundred billion data events recorded to track pause, exit and rewind rates by the second. When the system flags anomalous peaks, curriculum researchers revisit the material, revise the animation and relaunch it.
Long-running character IP also supports relevance and familiarity, but the core remains an iteratable methodology. That is why 杨临风 can say: “Tomorrow’s Onion will definitely be better than today’s Onion.”
11. The AI Tutor Is Embedded in a Defined Course Context; Its Job Is to Fill Gaps, Not Improvise Freely
Students can pause a video at any time and call up an AI tutor. The system combines the current course context, subject knowledge base, historical student data and memory to infer where the student may be stuck. It avoids suddenly invoking concepts the student has not yet learned or concepts from another subject or grade level, keeping the student on the original cognitive path.
“AI Photo-Problem Deep Learning” does more than generate a solution. It organizes the solution logic, summarizes the underlying thinking for that type of problem and assesses whether related knowledge points are weak. If remedial work is needed, the system prepares the relevant Onion content for one-click access.
杨临风 stresses that complete intervention should also include emotional encouragement and post hoc review. AI’s value is not to make the sticking point disappear, but to help the student understand why they got stuck after moving past it and what method they can use to solve a similar problem independently next time.
12. Onion Avoided the Livestream Boom Because “Using People to Teach People” Cannot Deliver Both Inclusion and Scale
Between 2016 and 2020, photo-based homework solving, dual-teacher courses and large-class livestreaming produced multiple $10B companies, but Onion pursued none of them. Human-machine interaction allows rural students with a phone and internet access to replicate the urban experience in full; once teacher services become essential, the result inevitably becomes “whoever can afford it gets to use it.”
曼琪 suggested that a dual-teacher model could also extend star-teacher coverage to large numbers of students, with local teaching assistants or AI handling questions. 杨临风’s observation was that remote broadcasts often leave local students behind, forcing assistants to constantly “translate.” Once the assistant has to interrupt the main instructor, the livestream gradually degenerates into a recording that can be paused.
Even if the recorded content is pedagogically sound, it still does not answer whether students have the mental energy to watch continuously for 10–40 minutes, whether they stay focused or whether the teacher’s words truly trigger thought. In the early days, the team put recordings by star teachers onto rural tablets and found that most students “simply could not get through them.”
Eye contact, feedback and a shared physical setting dynamically adjust explanations in a live classroom. Those factors are heavily degraded when a live teacher is compressed into a small screen. Onion’s conclusion: digital environments need digital-native learning formats, not a simple transfer of offline classrooms.
13. Duolingo Is Good at System 1 Practice, but Its Model Cannot Be Directly Extended to System 2 Deep Learning
杨临风 believes Duolingo is better suited to language practice than deep understanding. When a grammar module is pushed inside the app, user retention drops visibly. This type of experience relies more on low-energy System 1: fast feedback, short cycles and no requirement for a long chain of logic.
K–12 subjects such as mathematics and physics require students to think through 5 or 7 consecutive steps and actively invoke System 2. Copying the gamified surface of a practice product does not automatically solve how students bear cognitive load and form deep understanding.
Learning content in China is also generally harder than content abroad, so the same blackboard-style format can make students see “a huge, overwhelming mass” and retreat immediately. Khan Academy, Duolingo and Onion therefore cannot substitute for one another’s product paradigms; Khan and Onion are more closely aligned at the foundational level of lowering the barrier to independent learning.
14. A Structural Gap Created by Cognitive Maturity Separates General-Purpose Models from K–12 Students
杨临风’s current judgment is that even with excellent prompts and engineering, a pure foundation model still cannot explain System 2 knowledge from scratch to an ordinary K–12 student. “Assistance is no problem,” but most students cannot keep up when the model takes the lead directly.
Large models learn from highly structured material such as academic papers, and their chains of thought also aim to be more orderly. Adults and college students can digest this type of output; K–12 students’ ability to understand logic and structure is still developing. A model that resembles a mature, high-cognition adult may not be better at scaffolding a child starting from zero.
He does not rule out a fundamental change in training methods or the arrival of AGI that lets models converse fluidly with students and generate learning formats that most students can probably follow with one click. Onion’s existing work could “possibly” come under threat then, but until that point there remains substantial room for course design and context control.
15. Education’s Moat Is Process Design Because Models Cannot Understand on the Student’s Behalf
杨临风 distinguishes education from buying something. E-commerce only needs to deliver the correct product to the user; whether the order is placed through a foundation model or Taobao is irrelevant. Education “delivers a process”: how students absorb knowledge before the answer appears is the product itself.
Even if a model’s delivery capability becomes extremely strong, it still cannot complete the process of following the logic and forming understanding for the student. A vertical company’s know-how lies in controlling pace, identifying sticking points, regulating emotion and bringing students back to thinking.
曼琪 asked whether OpenAI could acquire Onion, obtain its large volume of interaction data and build the same product directly. 杨临风 acknowledged that it might do a good job in the segments Onion covers, but said that serving ordinary students with AIGC alone, without structured content, remains difficult for now.
16. Fixed Costs and Word-of-Mouth Growth Took Onion Through Its Operating Inflection Point in 2022
Onion’s pricing and GMV are below those of large-class livestream courses, but 杨临风 describes its commercial structure as roughly SaaS-like: the early phase requires sustained investment in course and product development, with costs weighted toward fixed items. Once the user base grows, R&D investment does not need to rise proportionally, allowing operating leverage to emerge.
Users come mainly through word of mouth and organic growth rather than advertising. After entering the product, they are retained and converted into paying users along the way, making the model relatively resilient. The team first broke even for a full year in 2022; the financing raised over the preceding years was mainly used to cross the fixed-R&D investment period.
Onion has not disclosed complete NPS data. 杨临风’s rough figure is 50–60, meaning the net score remains high after subtracting strong detractors from strong promoters. He believes sustained referrals and school repurchases measure real impact better than a sharp increase in users or revenue within a single year; over-stimulating commercial metrics can distort behavior.
17. Contrarian Positioning Creates Opportunity, but Requires Capital and Teams to Accept “Hard and Slow”
When the industry argued that “K–12 students cannot possibly learn independently,” 杨临风 saw an opportunity. Children are neither stupid nor unwilling to achieve good grades: “Inside every underperforming student lives someone who wants to become a top student.” Negative feedback simply turns lively primary-school children into dull middle-school students over time.
From the end of 2013 to the end of 2015, the team spent 2 years completing about 200 concept lessons for junior middle-school math. There were no users and no live product during that period—like finishing a feature film before testing the market. Looking back, 杨临风 admits it was “extremely irrational and risky,” but the market soon responded, and product data and user numbers began to grow.
In response to the livestream-course boom, the team held no more than 2 serious internal discussions before walking away: “If you don’t have the diamond, don’t take on the porcelain job.” The team lacked the capability, could not reuse its existing accumulation and doubted the health of burning cash for growth. Following the trend would only have turned the company into a different species.
Early supporters included the family foundation of Jerry Chen, founder of Morningside, and 李丰, who was pushing the project at the time, later founded Frees Fund and continued investing. 杨临风 does not speculate about their return motivations; he only says Frees Fund’s orientation toward “doing the right thing” is relatively close to Onion’s.
18. “Secret Weapons” in Test Prep Mainly Add Polish for Top Students and May Not Help Most Students, Even on Exams
曼琪 asked whether livestream-based tutoring was at least effective for exam preparation. 杨临风 still answered no, while explicitly limiting the claim to “most” students rather than making it absolute. Many courses teach short-term problem-solving tricks and “secret weapons,” but memory is interconnected in a network; the more fragmented shortcuts students accumulate, the harder long-term retention becomes.
Top students have already worked through the underlying logic and also possess strong memory, so exposure to more difficult problems and pattern summaries can indeed add polish. Most students need “snow in winter”—a clear explanation of the root cause. Mnemonics such as “add on the left, subtract on the right; add above, subtract below” or “the condition is sufficient, the conclusion is necessary” may instead become another memory burden.
The new middle-school and college entrance exams are also pushing questions toward contextualized, real-life and interdisciplinary formats. As contexts keep changing, traditional drilling is less likely to guess the test. 杨临风 sees the shift as a response against rote drilling and involution: if students still rely only on dead memorization, their exam scores may actually be lower.
19. Double Reduction Did Not Hit Onion Hard Because It Was Always Embedded in the National Curriculum and School Pace
When the Double Reduction policy took effect, Onion’s business saw limited direct impact. From Day One, all content had been built within the national curriculum framework and synchronized with textbook chapters and school schedules, rather than creating a separate after-school system for pushing top students.
The choice also originated in rural use cases. Teachers can only use a resource directly if they know which lesson it supports; if they first have to understand an independent curriculum system, a public-interest product becomes difficult to deploy as well.
杨临风 acknowledges that these constraints may have prevented the company from exploding, but they also made it “relatively stable and sustainable.” After the AI boom began in 2023, Onion likewise saw no obvious step-up in users; overall operations remained normal.
20. “AI Native” Is Not a Useful Classification; Demand-Side Education Has at Least 3 Transformation Paths
杨临风 believes every company is pursuing AI native, making it a supply-side commonality that cannot explain product differences. The meaningful classification should start with the problems schools, teachers, parents and students each need to solve.
The first path standardizes teacher workflows: handing as much of a tutoring company’s existing expertise, instruction or service as possible to AI to lower costs and increase delivery certainty. Companies with large-class livestream backgrounds, such as TAL, fit this path; products such as VIP may be similar.
The second uses technology to standardize peripheral learning workflows, such as question recommendation, homework grading and automated exam marking. These are genuine needs, but they are more often by-products or components of learning than the moment when a student actually forms understanding.
Onion chose the third path: the narrowly defined learning experience. It focuses on how a knowledge or teaching process, from initiation to student uptake, can become more vivid, flexible and engaging through AI. 杨临风 does not reject the first 2 paths; he emphasizes that different companies should stay within their own strengths and boundaries.
21. Education Is More Likely to Remain Fragmented; General-Purpose Giants Will Struggle to Absorb Every Niche
Asked whether AI will create an education giant, 杨临风 first stressed that he has “not very high confidence in the answer,” then offered his current judgment: no. The differences in philosophy and preference among schools, teachers, students and families are too large. Even if AI reaches its full potential, some people will still feel safer with an offline tutoring center.
Education is therefore “not one big market,” but many segments built on different foundational logics. If a vertical product delivers a local experience at 95, schools and families may choose it rather than accept a general-purpose platform that cannot reach 95 in that segment.
Even a well-funded platform acquiring 1,000 companies might not create chemical synergy. Education companies typically have short supply chains and can serve users end to end, reducing the need for upstream-downstream coordination; methodological and philosophical incompatibility is stronger than the potential for integration.
22. Impact Is Not Just Registrations; It Is Also Rural Reach and Whether the Weakest Students Rejoin the Learning Process
Onion has a full-time “Onion Teaching Assistant Initiative” team that proactively contacts rural schools, opens resources for free and provides training. It has supported about 60K rural teachers across more than 30K schools, with county-wide programs in roughly 10 counties, including Qumalai and Zhiduo in Qinghai’s Yushu Tibetan region, Dangchang in Gansu, Baoshan in Yunnan, Jinping in Guizhou and Xunwu in Jiangxi.
杨临风 sees this as a closed loop in the company’s founding path. The public-interest experiment showed that the required investment was far greater than expected, so the team “took the long way around” and formed a company. After the detour, the commercial product could still serve the people it initially wanted to help in the same form, confirming that the team had not lost its starting point.
The most tangible result came from an AI future-classroom school: during evening self-study, the student ranked second from the bottom voluntarily taught a problem to the student ranked last. In traditional environments, the 2 Grade 12 students had been the least confident and most likely to retreat into a corner; one dared to ask and the other dared to explain, and both treated it as entirely normal.
Whoever figured out a problem first would go to the front of the room to share it. What 杨临风 values is not an instant reversal in rankings, but students regaining the sense that “I am useful” through explanation and mutual help. That is what it looks like when the belief in independent learning is reignited.
23. The AI Future Classroom Makes Students Own the Learning; Schools Build the Environment and Habits
曼琪 cited Alpha School, where annual tuition is about $60K, students study for 2 hours each morning and spend afternoons on activities and AI projects. 杨临风 believes the key is not the shorter schedule, but that students “take responsibility for learning”: the time belongs to them, and if they have not understood something, they must keep asking, reviewing and genuinely working it out.
He is equally clear that schools “will exist 100%.” Most children cannot acquire habits, motivation and belief from zero on their own. The central responsibility of teachers and schools should be to create the environment and design activities that gradually teach students how to learn independently, rather than permanently carrying the responsibility for them.
Onion provides deep, on-site service to more than 2,000 schools nationwide each year. In the AI future classroom, the machine provides immediate support when students get stuck or generate new ideas, while teachers guide the class, design activities and build habits of questioning.
In a class on the disorderly motion of molecules, one student asked AI: “Molecular motion can make molecules diffuse—is there a way to make them gather?” Onion used Bloom’s questioning framework to provide scaffolding and left time for questions; AI answered first, then the teacher showcased the high-quality question so the entire class could develop a lasting sense of what a good question looks like.
24. AI Has Not Rewritten the Laws of Good Education; It Has Made the Old System’s Problems Impossible to Ignore
杨临风 believes the principles of genuinely effective education have not fundamentally changed for thousands of years. Disengagement from learning, high scores with weak capabilities, lack of purpose and poor innovative ability all existed long before AI. What changed in the pre-AI era was that the traditional exam route could still serve as a relatively stable path to employment and socialization.
AI is a catalyst because it is undermining the certainty of that path. ChatGPT appeared just over 3 years ago; multiply that by 3 and it is roughly 10 years—the point at which today’s seventh-graders will be graduating from college. By then, the professional world will no longer be reliably imaginable.
The one relative certainty is that children will face things 10 years from now that no one can imagine today. They therefore need the ability to learn new things and apply what they learn. The goal of education is not to predict the jobs of 10 years from now, but to build the capacity to handle unfamiliar tasks.
At the end of the episode, 曼琪 placed this alongside 马毅’s discussion in Episode 108 that “the essence of intelligence is learning.” Biological learning remains simpler and more powerful than mainstream AI, while industry is moving rapidly down the path of massive compute and massive data and has already achieved results in tasks such as coding. Whatever wall eventually appears, the consistently useful actions for most people remain “exercise and learn independently.”
25. Knowledge Will Not Disappear; It Will More Explicitly Serve as a Vehicle for Ways of Thinking
杨临风 uses ancient Greek geometry to explain that learning knowledge is not merely about building better temples; geometry was a vehicle for developing logical thinking. Mathematics teaches respect for logic and how to reason, while also training people to understand complex systems built layer by layer from axioms and theorems, then transfer that thinking to weather, finance, transportation, healthcare, economics and politics.
These ways of thinking are difficult to teach independently of knowledge. Telling a primary-school student in the abstract that “math is beautiful” only sounds absurd; a math major may agree. The difference comes from the cognition and values formed through long-term engagement with concrete knowledge.
The destructiveness of rote memorization lies in putting the cart before the horse: it preserves conclusions while severing knowledge’s role as a vehicle for thought. 杨临风 therefore believes that “as long as brain-computer interfaces have not arrived,” people still need knowledge to develop thinking, and designing the learning experience will only become more important.
The school model will move from a teacher monologue, through a two-way teacher-student classroom, toward a three-way classroom involving teacher, student and machine. The machine becomes the teacher’s intelligent assistant and the student’s intelligent learning companion. The most important K–12 objective is to make students self-reliant—with confidence, learning ability, discipline, planning skills, goal-setting ability and the metacognitive capacity to execute step by step.
26. AI Will Amplify Laziness and Inequality, but the Risk Ultimately Depends on Where the Product Leads People
曼琪 noted that large models, like recommendation algorithms, may respond by following user preferences and reinforce information silos. 杨临风 agrees: if AI merely indulges System 1’s laziness, students will become even less willing to think. But the problem existed before large models; the new possibility is that there is finally another route to “pull students outward.”
Intervention cannot be arrogant. If a product begins by declaring that all of a child’s existing choices are wrong, the child will simply say, “Sorry, bye.” The design principle should be: “Remember where you want to go, but don’t be so arrogant as to assume the student will definitely follow you,” balancing a clear objective with patient migration.
曼琪 raised a sharper social risk: active learners will be massively amplified by AI while everyone else is left further behind—“opportunity is fair, outcomes are more unfair.” 杨临风 believes this may happen. Education cannot stop AI from reshaping society; it can only try to keep the people most likely to fall behind from being left behind.
Onion will not train a base model independently, limiting itself to cost-controlled post-training or fine-tuning. The team continually benchmarks brands, versions and model sizes, while the AI tutor is split into multiple tiers that call different models. Model costs are not yet a large share, but are expected to grow with usage, and the token prices of the strongest models are indeed expensive.
27. Public Interest Is the Company’s Ballast; Over the Next 10 Years, Success Will Still Mean Helping People Take Charge of Learning
Responding to criticism that an elite background might make the plan overly idealistic, 杨临风 says the team became more, not less, respectful of learning’s difficulty precisely because it had seen students who could not understand and were barely staying awake. Helping people figure things out fascinates him in its own right; computer science, education technology and social innovation therefore converge in the same mission.
He defines his primary identity as CEO and only then as an education-technology product manager: first make the organization healthy and stable so the team can invest real money in innovation over the long term. But if the company stops creating learning experiences, its value is zero. “Onion must be a healthy, sustainable business, but it cannot be merely a business.”
As a child, 杨临风 believed in making a lot of money first and doing charity later. He later changed his mind: “What society lacks most is not money, but methods, ideas and people who can get things done.” If a company starts by building a different business on Day One and plans to pivot back to a social issue later, its success model, stakeholders and value network will usually lock it in place. He uses the ongoing tension between ideals and reality at companies such as OpenAI to illustrate that this need not come from deception; each local optimum can push an organization onto a faster vehicle that is harder to turn back from.
The public-interest mission is therefore not a composite KPI to be included in every commercial decision, but an inviolable underlying framework. Within that framework, the company must first become a successful product and business organization. It also serves as the team’s “ballast” through turbulence: entrepreneurship is not always enjoyable, but knowing that time is still advancing the issues they care about provides a sense of steadiness.
Onion’s mission is to “create learning experiences that inspire every student to take charge of independent learning,” while its vision is to be “learner-centered, empower educators and advance educational equity.” Over the next 10 years, 杨临风 says, helping 5M to 10M children become less resistant to independent learning—and perhaps genuinely independent and confident—would already be an extraordinary result. AI will increasingly act as a planner and strategy distributor, using memory, dialogue, mastery and emotional signals to choose the right next intervention for each child.