Manchi | 晚点聊 LateTalk Summary

Podcast: 晚点聊 LateTalk | Participants: Manchi (host), Yang Linfeng (CEO & Co-founder of Onion Academy)

1. Key Themes

The Architecture of Autonomous Learning: Willingness → Ability → Tools → Belief

Yang Linfeng argues that autonomous learning is not a single trait but a system of four interlocking components. Most parents and educators focus on willingness, but that is putting the cart before the horse. The real entry point is lowering the barrier through well-designed tools.

"If we summarize what autonomous learning's core is — 'wanting to learn proactively' is one element, the willingness. But willingness alone is not enough. There is something called ability. After having ability, the next step is whether I have suitable tools... The last thing built is a belief system called: autonomous learning is something I can complete independently." [00:09:08]

"Tools can help people lower the threshold of autonomous learning. This is what Onion can do first — we cannot change people's consciousness and cognition, but we can provide tools." [00:11:44]

Content Quality Has a Hard Threshold — Below It, Output is Zero

Onion spent two years making 200 animated videos before launching any product. Yang argues this was not reckless slowness but a recognition that educational content has a binary quality threshold: either it clears the bar and students engage, or it doesn't and the investment is entirely wasted.

"If this thing's quality doesn't reach the threshold, even if you spent 50 RMB, the final output is zero. Only when you spend to 80 RMB does this output possibly start to take effect." [00:25:45]

"The core is that we aren't doing training [peixun]. Training has a self-selection logic — you're willing to be tutored, I'll tutor you. But from day one, we are doing foundational education work [jiaoyu dixia gongjian], hoping to help rural students learn well." [00:47:36]

Digital-Native Learning Requires a Digital-Native Content Form

Yang makes a sharp distinction between transposing real teachers onto screens versus designing content native to a digital environment. The in-person "field" created by a teacher's physical presence disappears on a screen, making animated, purpose-built content more effective for solo digital learners.

"In the digital environment, there must be a digital-native learning form. This learning form in the current stage needs us to reconstruct it from scratch. So we started from zero to create a learning method suited to the digital environment." [00:34:34]

"The teacher's face has no information on it. The information is in the picture. So the so-called Onion animation is not a cartoon — we use animation to do abstract concept visualization." [00:22:17]

2. Contrarian Perspectives

Most Educational Tutoring is Ineffective Even for Exam Performance

Yang argues that the dominant Chinese tutoring model — teaching shortcuts, mnemonics, and quick problem-solving tricks — is not just suboptimal but actively harmful to memory and understanding, and is even counterproductive for the very exams it targets.

"Most of these teaching methods are teaching students big moves [dazao] — short-term problem-solving techniques. But these techniques, while seemingly effective for short-term problem-solving, are very unfriendly to human memory... mnemonics and shortcuts are naturally fragmented memory. They actually make it very hard for many people to remember." [00:49:36]

"Top students are particularly suited to that kind of tutoring. A top student seeing more difficult problem types, getting big-move patterns summarized for them — first, their mind is clear, they remember clearly, they have good memory. Second, their underlying logic is already clear to themselves; what they lack is seeing various problem types. So this method is exactly suited to these students as icing on the cake. But most students need help in the snow — this model actually isn't suitable." [00:50:05]

AI Large Models are Structurally Poor at Teaching K-12 Students — By Design

Yang makes a non-obvious technical argument: the very thing that makes LLMs powerful (highly structured, logical, academic training data) makes them mismatched to students whose logical and structural cognition is still developing. This is not a solvable prompt-engineering problem.

"Why are large models now especially useful for college students and adults? Because adults and college students have sufficiently strong cognitive ability... Middle and elementary school students' fundamental problem is that their logical and structural understanding ability is still growing, not yet that strong. Yet all of the large model's training data — including logic chains, thinking chains — are all designed to make the model think more systematically... This creates a natural gap when going to teach someone whose cognitive level hasn't yet matured." [00:37:26]

Companies That Start with a Business Goal and Plan to "Pivot Back" to Social Impact Almost Always Fail to Do So

Yang argues the product architecture and business structure must be designed from Day 1 around the social mission; retrofitting it after commercial success is structurally near-impossible.

"If an enterprise is aiming at a social issue, but its Day 1 solution is not aimed at this issue, but rather 'I'll do business first, earn money, then pivot to this solution' — I now think that will most likely not succeed... Most enterprises find it very hard to smash an already-successful business model." [00:131:22]

"Onion, precisely because from the very beginning it aimed at this social issue, we chose the human-machine interaction form, we chose the universal, inclusive approach — that's why today we can say we can accommodate both things. If Onion had started by thinking 'let's first make money in education, then go solve this problem,' we likely would have gone toward the money-making methods in between." [00:131:50]

Autonomous Learning is Not "Against Human Nature" — Parents Who Say Their Kids Won't Self-Learn Should Check Themselves First

Yang challenges the mainstream parental assumption that children are constitutionally resistant to autonomous learning by pointing to System 1/System 2 psychology and noting that adults have the same resistance.

"For parents, when you say your child is unwilling to think proactively about problems, first ask yourself — most of the time, aren't you also unwilling to think proactively? The key question is: how to guide someone into a state willing to think proactively." [00:11:14]

3. Companies Identified

Onion Academy (洋葱学园) EdTech company, 13 years old, focused on animated digital learning for K-12 students in China. Mentioned throughout as the subject company. Distinguished by: SaaS-like fixed-cost model, word-of-mouth growth with no advertising, 130 million cumulative registered students (1/3 of all Chinese K-12 students), 4 million registered teachers, 2,000+ deeply served schools annually, and a permanent policy of offering all products free to rural areas.

"We are a company that has chosen a non-consensus path, and moreover in education — a field that requires people and societal concepts to change before it will change." [00:121:41]

Khan Academy Referenced as the closest philosophical parallel to Onion's approach — belief in autonomous learning, knowledge board-focused rather than teacher-face-focused. Noted as a validation of the underlying thesis but operating in a different market context.

"Good learning methods can lower the threshold of autonomous learning — it's just that in the overseas environment versus the domestic environment there are some differences. One very important reason is that China's learning content is genuinely harder than overseas content." [00:034:58]

Duolingo Referenced as a distinct category: effective for language practice (System 1 learning) but explicitly not deep learning. When Duolingo attempted to add grammar modules, retention plummeted — cited as proof that gamified apps are not suited to System 2 cognitive tasks.

"Duolingo is not deep learning. It is particularly suited to practice like English... Duolingo's learning, we feel, is more biased toward System 1 learning, whereas Onion, including Khan Academy, and in fact most of the K-12 stage subject learning, is biased toward System 2 deep logic chain learning." [00:035:31]

New Oriental (新东方) Cited as a rare example of a Chinese education company that genuinely built content experience — specifically through crafting classroom atmosphere and teacher charisma in offline settings. Praised as an exception to the norm of content being undervalued.

"I think New Oriental is a company that was very content-strength-oriented... very early they realized that classroom atmosphere is a scene worth creating. I think they were also creating experience — just their experience was based on the offline, face-to-face field between people, the teacher's own charisma." [00:046:38]

Alpha School (US) A US private school (mentioned in passing) where students only attend formal instruction for 2 hours/day, with AI assistance and heavy active output in the afternoon. Cited as a real-world example of student-centered autonomous learning. Annual tuition: $60,000 USD — signals elite market validation of the model.

"This model, in fact, is not expensive. We are now cooperating with many public schools." [00:100:45]

4. People Identified

Yang Linfeng (杨临风) CEO and Co-founder of Onion Academy. Computer science graduate of Harvard. Started doing rural education charity work in 2010 before founding the company in 2013. Distinguished by: 13+ years of mission consistency, refusal to chase the online education boom of 2016-2020, deep grounding in learning science frameworks (Bloom's Taxonomy, Understanding by Design).

"Onion must be a healthy and sustainable business, but it cannot merely be a business." [00:134:45]

Li Feng (李丰) of Fengre Capital (丰瑞资本) Early-stage investor, formerly at IDG. Long-term supporter of Onion across multiple rounds. His fund's slogan is "do the right things," which aligned with Onion's non-consensus mission.

"Fengre has a slogan called 'do the right things.' I feel this is relatively close to Onion's philosophy... I think Feng Shu definitely didn't think we were the enterprise that could make him a lot of money. But I think Onion's philosophy might be more aligned with his cognition of education." [00:123:06]

Chen Lezong (陈乐宗) / Jara Chan of Morningside First investor in Onion Academy. Used a family foundation to provide the first check, before Li Feng's involvement. Notable for backing an early-stage, non-consensus education company through a philanthropic vehicle.

"Our first investor, strictly speaking, should be Chen Xinmorningside's Founder, Chen Lezong, Jara Chan — he used a family foundation to invest the first sum of money in us." [00:122:39]

5. Operating Insights

Use Granular Behavioral Data to Drive Iterative Content Improvement

Onion tracks pause rate, dropout rate, and rewind rate at the per-second level across hundreds of billions of learning interactions. Spikes in these metrics automatically flag problematic content segments, triggering a structured review and re-production cycle.

"We can see every second of every video — for example, student pause rate, including dropout rate, rewind rate. Based on these peaks, the system will alert teachers to pay attention, saying this part needs attention, this part may not have been taught well. Then we pull the curriculum teachers back together to re-discuss, then modify the animation and go live." [00:028:42]

Design Every Product Interaction Around the "Achievement Loop" — 10 Minutes Max

A product insight applicable beyond education: users need to complete a full cycle of learn → apply → verify within a single session (≤10 minutes for children, adaptable for adults) to generate the confidence that drives the next session. Without this closed loop, engagement collapses.

"Within 10 minutes, you need to form at least one such closed loop — then they have the motivation to carry on to the next cycle. This is why we controlled the video to 5-8 minutes." [00:015:00]

Target the User, Not the Payer — Especially in B2C Markets With Asymmetric Decision Power

Onion made a deliberate choice to design for students first, not parents (who pay). The insight: if the end user (child) has a genuinely good experience, parents discover the product through the child — not the reverse. This reduces customer acquisition cost and creates authentic word-of-mouth.

"All education or tutoring companies' first point of contact is entirely with parents, because parents are the ones paying. So from the very beginning, the elements they chose are all different. We were a completely different species from the start... our starting point is from rural charity and bottom-up universal education, and educational innovation — not from the tutoring perspective." [00:048:06]

6. Overlooked Insights

The "Bloom's Questioning Framework" in the Classroom May Be the Highest-Leverage AI Education Use Case — and It's Almost Never Discussed

Yang briefly describes an AI classroom experiment where students are given a structured framework for how to ask questions (based on Bloom's Taxonomy), and AI answers every student's question in real time. The result: the second-to-last-ranked student in the school spontaneously began tutoring the last-ranked student. This is a radically different use of AI in education than the prevailing "AI tutor explaining content" framing — it uses AI to democratize the right to ask questions and to train metacognitive skills. This has enormous implications for classroom design and EdTech product architecture that the host did not probe.

"Because we gave students time to ask questions and guided them with a framework for how to ask good questions, we find students in class actually asking very good questions... A student asked: 'molecular motion can cause molecular diffusion — is there a way to make them gather instead?' This is actually a very good question... In traditional settings, most teachers couldn't answer these questions, or didn't have time. But because there's AI, every question a student asks, AI can at least answer it first. And because of this classroom format, students' questions can be seen by the entire class." [01:03:39]

China's Rural Education Gap Is a Technology Distribution Problem That Has Already Been Largely Solved — But Almost No One Knows It

Yang mentions in passing that Onion has reached ~60,000 rural teachers across ~30,000 rural schools — covering roughly 60% of China's total rural teacher population — for free, with full product access. This is a massive, quiet redistribution of educational resources that receives almost no attention in the EdTech narrative, which focuses almost entirely on urban premium markets. For impact investors and policy researchers, this represents a template for how a commercial product architecture (SaaS fixed-cost, word-of-mouth growth) can simultaneously serve both premium and underserved markets without product dilution.

"The number of rural teachers we have cumulatively supported is now approximately 60,000+. These teachers are distributed across roughly 30,000 rural schools. China's total rural teacher count is approximately 100,000. So in fact our coverage is quite extensive." [00:125:58]