142: How a Self-Teaching AI Product Emerges | Dialogue with Zebra CPO Xiu Jiaming
- 01The Two-Year Battle: Balancing AI Creativity with Educational Structure
- 02Education Data vs. Experience: Why Human Expertise Remains Irreplaceable
- 03The Unique Oral Language Learning Gap: A Problem Only AI Can Solve at Scale
1. Key Themes
The Two-Year Battle: Balancing AI Creativity with Educational Structure
The development of Banma's AI spoken English product required an intense two-year effort focused on what Xiu Jiaming calls "wrestling with AI models." The core challenge was preserving AI's creative capabilities while maintaining strict educational boundaries. As Xiu explains: "In our view, so-called hallucinations aren't necessarily a negative term. Sometimes they represent the AI model's creativity. This is actually a major advantage of AI models" [00:00:04]. However, this creativity must be balanced: "We need to ensure AI's expression is limited within the boundaries we've set for it. It has a range" [00:08:58]. The team spent enormous effort training models with adversarial techniques, reward-based learning, and continuous monitoring to achieve this balance.
Education Data vs. Experience: Why Human Expertise Remains Irreplaceable
A profound insight Xiu shares is that in education—unlike other fields—data cannot replace experience. "Regarding education, data doesn't have more persuasive power than experience. No matter where the data comes from—the internet, major enterprises, or schools—you can't apply it when facing this specific child and deciding how to teach them" [00:28:30]. This explains why creating each 25-minute lesson requires a year of development [00:55:31] and involves a 200-person team [01:13:19]. The complexity means even training an experienced English teacher to work on one aspect of the product takes two months [01:03:48].
The Unique Oral Language Learning Gap: A Problem Only AI Can Solve at Scale
Xiu identifies oral language learning as uniquely suited for AI intervention because it's the only English skill that absolutely requires a conversational partner. "Other abilities like listening, reading, vocabulary, and grammar can actually be learned through self-study or taught by one teacher. But only speaking—this one ability—requires a real person or counterpart for authentic communication" [00:15:10]. Traditional solutions (foreign teachers, language partners) are prohibitively expensive or unavailable for most families, creating a persistent plateau effect: "Students might reach KET level through various methods, but then stay at that level all the way through high school and university" [00:16:48].
2. Contrarian Perspectives
AI Tutors Don't Need to Be Human-Like—They Need Their Own "Personhood"
Counter to the prevailing view that AI should maximize human similarity, Xiu argues for AI establishing its own distinct identity: "We never try to make the product X% like a human, because we didn't design it to be like a human from the start... Just like we never say a dog is X% like a human. A dog is a dog. Dogs have their own position in the world" [00:45:17]. Children recognize and accept AI tutors as a new category of being with consistent, predictable behaviors—neither human teacher nor inanimate tool. This distinct "personhood" actually reduces performance anxiety: "Compared to speaking with real foreigners, this is already much more relaxed. This is also a major advantage of AI—it lowers the affective filter barrier" [00:41:01].
Product Success Requires Doing What Teachers Cannot, Not What They Can
Most AI education products aim to match excellent teachers. Xiu sets a higher bar: "Currently, all AI-combined education products we observe on the market are at a level that can replace humans... But the point we can grasp with Banma is that teachers cannot replace us—no teacher you bring can complete the oral training that an AI tutor can accomplish" [01:11:08]. This philosophy guides their product selection: "If teachers can do it, if humans can do it, we won't do it. We'll roughly pursue this product direction" [01:11:37].
Higher Pricing Would Be Justified—They Deliberately Priced Lower
Against the typical startup wisdom of pricing low to gain market share, Xiu reveals: "We actually think we priced it slightly too low" [01:04:56]. At 3,600 RMB per level (one year), they consciously chose accessibility over revenue maximization, noting this price point removes barriers while being well below parent expectations for oral English instruction. During beta testing at 70% discount (1,500-2,000 RMB), "there was basically no one who wouldn't accept it" [01:05:27].
3. Companies Identified
Banma (斑马)
Description: Educational technology company under Yuanli Technology (猿辅导), specializing in children's digital content and education for 6-12 year-olds.
Why Mentioned: Creator of the featured product—Banma Spoken English, the first fully AI-native educational product from their parent company. The company has maintained an AI lab since 2014 and accumulated extensive children's voice data, giving them industry-leading child speech recognition capabilities [00:04:44].
Notable Quote: "Banma Spoken English is the first—or most advanced—product in our entire group that's completely built from the ground up with AI. Internally we call it 'full-stack,' meaning from front-end to back-end, from backend to UI, at every level, we designed it considering AI integration" [00:06:05].
OpenAI
Description: Leading AI research company developing large language models.
Why Mentioned: Referenced as example of how rapid model capability improvements could potentially disrupt products built on earlier model generations. Their product releases (like new model announcements) can obsolete features that required significant engineering effort [00:59:58].
DeepSeek
Description: AI model provider whose capabilities are integrated into Banma's product ecosystem.
Why Mentioned: Used selectively in Banma's product, particularly for post-session tasks like organizing student information and generating reports. "We use some DeepSeek capabilities, but not during the conversation process. For example, after class is finished, when organizing the child's information for these tasks... because DeepSeek has some advantages in these areas" [00:26:36].
4. People Identified
Steve Jobs
Description: Co-founder of Apple Inc.
Why Mentioned: Raised the famous question about why education hadn't benefited from technology progress during his final meeting with Bill Gates in 2011, setting up the podcast's central theme.
Quote (paraphrased by host): "The education field seems not to have benefited from technology's tremendous progress" [00:00:48].
Bill Gates
Description: Co-founder of Microsoft and philanthropist focused on education.
Why Mentioned: Provided the answer to Jobs' question that foreshadows current AI developments: "Technology can only fundamentally remake education when it can provide more personalized courses and enlightening feedback" [00:01:01].
5. Operating Insights
The 25-Minute High-Intensity Design: Maximizing Effective Output Through Micro-Segmentation
Rather than longer, relaxed sessions, Banma designed 25-minute classes requiring 100+ verbal outputs. "A lesson is 25 minutes, and during this time the child will speak effectively over 100 times. That's four utterances per minute completing expressions. This intensity is quite significant" [00:53:11]. Each session is fragmented into 10+ micro-segments of 2-3 minutes with crystal-clear objectives, ensuring AI maintains instructional control while adapting to student responses. This structure prevents drift while maintaining engagement—children are "in a quite tense state learning English" [00:41:09] but productively so.
The Proprietary "Oral Ability" Metric: Building Objective Progress Measurement
Banma developed a composite "oral ability index" measuring three dimensions (accuracy, fluency, richness), each with two sub-dimensions, aggregated through weighted formulas [00:52:12]. This will soon be released publicly, allowing parents to see concrete progress. Critically: "Because class materials are substantial—25 minutes with 100 utterances—one lesson can give a rough estimate. After four lessons, we can estimate quite accurately" [00:52:51]. This eliminates traditional exam dependency for oral skills.
The Parallel Development-Iteration Pipeline: Continuous Quality Improvement
Xiu describes two teams working simultaneously: "It's like two teams doing this work. One team is in front doing testing, and after testing is complete, another team does iteration. The iterated version goes live" [01:13:10]. Additionally, they score every AI teaching session using the same rubrics previously used for human teachers: "We would rate teachers, so we use the same dimensions to rate every AI model lesson. If a lesson scores low in certain dimensions, we iterate the course" [01:13:01].
Strategic Material Selection: Start With Hardest Product, Easiest Content
Contrary to typical go-to-market strategies, Banma launched with their most challenging product form but simplest content tier: "First and second grades are harder to make products for, but higher grades are easier for products though the content becomes more difficult. Fewer people might need it—mostly those who came up through our system. So we started with lower grades first, then expand in both directions" [01:14:11]. This builds the hardest capabilities first while serving the largest addressable market.
The Emotion-First Pedagogical Layer: Why "Super Warm" Personality Is Strategic
The AI tutor's notably warm, encouraging personality isn't about likability—it's essential pedagogy. "Especially for foreign language learning in oral communication scenarios, the language itself is already unfamiliar, causing high stress. So we try every method to keep them in a very comfortable, relaxed, pressure-free environment where they're willing to actively express and produce output" [00:42:00]. The team programmed scenario-specific responses (child tired, frustrated, happy, confused) with strict rules like "must reference specific content the child said" rather than generic praise [00:25:31].
6. Overlooked Insights
The "Scaffolding Strategy" Granularity: Real-Time Difficulty Calibration
Buried in the technical discussion is a profound teaching methodology insight. When students struggle with complex sentences, AI employs "scaffolding strategies"—but the decision tree is remarkably sophisticated: "Should you ask a general question or a choice question? This is based on your judgment of the child's level... If they still can't answer, you give hints. After hints, you have them repeat. After repeating, you gradually return to where they should learn this control point" [00:30:00]. This dynamic difficulty adjustment happens in real-time during conversation, representing an educational breakthrough that's technically invisible but pedagogically transformative.
The Teacher's Day Card Insight: Emotional Intelligence Through External Strategy, Not Model Training
A seemingly minor anecdote reveals sophisticated product design. When a child brought a homemade card to celebrate Teacher's Day with the AI tutor, "the AI's response was correct—smiling and saying thank you—but not enthusiastic enough" [00:48:56]. The solution wasn't more model training but adding external calendar-aware strategy layers for special occasions (Teacher's Day, birthdays, holidays) [00:49:38]. This demonstrates that product-level orchestration, not just model capability, creates emotional resonance. It's a template for how to systematically address edge cases that could otherwise undermine the AI's established "personhood."
Key Timestamps:
- Product overview: [00:02:33]
- Two-year development journey: [00:04:35]
- Technical architecture: [00:22:14]
- Hallucination management: [00:08:16]
- Education vs. data debate: [00:28:30]
- Commercial strategy: [01:04:09]
- Competition analysis: [01:08:08]
- Team composition: [01:11:49]