Summary

Talkit’s core bet is that large models are moving human-computer interaction from the “hands” to the “mouth,” and language learning should move from tapping answers to speaking through real-world tasks. 刘夜 calls Duolingo “a company focused on training the hands,” while he wants to build “a company focused on training the mouth”; Talkit is trying to create a learner-friendly virtual English world where users learn to express themselves through tasks such as ordering food, interviewing, renting an apartment, filing a claim and bargaining—as if they had “immigrated to the US.”
The moat is not just the conversational model, but a single Gen World Engine that combines TBLT pedagogy, 3D characters, task generation and precise difficulty control. 刘夜 says Transformers have enabled systems to judge for the first time whether “hi bro” or “hi there” achieves a learning objective, rather than accepting only enumerated answers such as “how are you / fine thank you and you”; the system must also deliver n+1 difficulty, spaced repetition, interest matching and personalized assessment—“learn by speaking, and learn by completing tasks.”
Talkit is not trying to fight Duolingo head-on for beginners; it is targeting users who have practiced for hundreds of days and now want to move from training their “hands” to training their “mouth.” 刘夜’s market framework is roughly 2B language learners worldwide, including 1.5B English learners, with 90% classified as weak-motivation users; a survey of thousands of users found that many leave Duolingo after 300, 500 or even 1,000 days because they “seek to improve their spoken language.”
This is still an early product-validation bet: a $10M seed round and a 40-plus-person team, but the full version launched only a week before the interview, with no meaningful revenue or profit data yet. The team is tracking target users’ conversation time and retention—“whether they show up, and then whether they keep talking”; the beta had plenty of bugs, and although 刘夜 says some users already spend 1-2 hours a day on it, he rates the current product only “30 points.”
The competitive variable 刘夜 cares about most is effortless, not perfect effectiveness from day one. The host argues that Duolingo may be selling not only learning outcomes but also the feeling of continuous progress; 刘夜 agrees with Duolingo’s effort to make language learning effortless and defines language learning as a lifestyle and entertainment experience. “You can never teach a student who doesn’t show up,” so the first requirement is to keep people coming back and willing to speak.
Double Reduction did not leave 刘夜 with simple anger; it produced a mix of relief and regret that had to be converted into rebuilding. Looking back at the post-2018–2019 battles over traffic, renewals, advertising and ROI, he believes the industry had become “limited value, internal competition and a betrayal of the original mission”; the regret could only be resolved by building “a bigger company, a more successful company, a more valuable company,” which is why he rejected several easy-to-monetize directions that did not feel like his own. He also believes enormous commercial potential and value are necessary but insufficient conditions: willingness must exceed resistance.
For investors, the more important founder signal is his willingness to endure prolonged non-consensus around a product judgment, rather than rushing into a “revenge startup.” Over 3 years he examined more than 30 industries, once drank roughly 6 cups of coffee a day for 6 months and visited hundreds of stores, only to stop the project after the building was up and funding was about to arrive; at Talkit, he spent 24 months testing more than 100 versions and overturned the core architecture 3 times because he believes “loneliness means scarcity,” while acknowledging that loneliness may also mean being wrong—“What if they’re all wrong?”

Deep dive

1. Talkit Enters Genuine Product Validation After a $10M Seed Round

The host introduces 刘夜, who founded Homework Box in 2014, a company that served roughly 100M users. After Double Reduction, he launched Vision Flow and re-entered AI education with Talkit, raising a $10M seed round from investors including 理想 of Li Auto, Alibaba partner 曾明, 雨嫣, Zoom angel investor HP, Canadian VC Celtic House and investors in Perplexity. The host calls it the largest seed round in the AI language-learning space.
The new company has more than 40 employees. After 2 years of development, Talkit’s full version was released only the weekend before the interview; 刘夜 gave no revenue or profit figures, saying only that the company was beginning to see encouraging user-behavior data and would have plans for future rounds.
刘夜 does not define the company as an “AI speaking-practice companion,” but as “building a virtual world where everyone can learn a language as if they had immigrated to an English-speaking country, as if they had immigrated to the US.” The vision is also more explicit than it was at Homework Box: to change the basic way humanity learns languages over the coming decades.

2. Large Models Changed Interaction, but Language Learning Is Still Stuck on the Fingers

刘夜 describes himself as a longtime early adopter: one of Gmail’s first 5,000 global users, among the first 1,000 beta users of World of Warcraft, one of roughly the first 100 Google AdWords accounts, and a user of nearly 70% of the games on Vision Pro and Quest—after taking “several boxes of motion-sickness pills.”
ChatGPT gave him a key insight: “Humans have spent more time talking to machines than talking to people for the first time in history.” Almost overnight, the world began speaking to chat windows. Human-computer interaction shifted from the hands to the mouth, but most people still learn English with their eyes, hands and multiple-choice questions.
When US investors pressed him on Talkit’s difference from Duolingo, his answer was deliberately extreme: “Duolingo is a company focused on training the hands. I’m focused on training the mouth.” Talkit therefore committed to All in Talk—engaging the eyes, ears and mouth as people do when learning a native language, rather than mistaking click-through accuracy for language ability.

3. Course, Explore and Apply Form a Loop from Input to Live Conversation

Talkit’s first pillar is Course. Beginners still need to use their hands to build a foundation in words, chunks, vocabulary and basic grammar; internally, the team calls it “A Better Duolingo,” with the main difference being immersion rather than a rejection of structured practice.
The second pillar, Explore, provides rehearsal. Its interface resembles a short-video feed, with recommendation algorithms continuously serving up living scenarios: users may need to complete an interview, ask for someone’s WeChat, persuade another person, bargain, resolve a dispute, order food, buy a plane ticket, check out, file a claim or rent an apartment.
Relationships in these scenarios persist. Users can turn characters they meet while handling tasks into friends, continue to make friends and receive emotional responses from them. 刘夜’s language-learning formula is: “You have to study, use speaking in life, and make lots of good friends”; without all 3, it is difficult to approach native-level proficiency.
The final pillar, Apply, lets users connect directly with real people in English-speaking countries inside the app. Course input, AI-world rehearsal and live conversation form the loop 刘夜 calls “study—life—make friends—real-world application.”

4. TBLT Waited More Than 40 Years for Models That Could Understand Open-Ended Speech

刘夜 believes people have learned their native language inside a 3D world from day one, rather than by “watching TV and learning.” Talkit therefore adopts TBLT, or Task-Based Language Teaching, allowing users to “learn by talk, talk to finish the real scenarios, the real task.”
He says the theory emerged in the 1980s and later became mainstream in offline classrooms, but remained difficult to productize. Traditional systems could only define “how are you / fine thank you and you” as the correct answer; if a user said “hi bro” or “hi there,” systems based on methods such as BERT could not determine whether the communication task had been completed just as effectively.
刘夜’s view is that only after Transformers pushed natural-language understanding to what he calls “99.99%” could systems flexibly assess whether a learning objective had been met and generate personalized evaluation, reinforcement and next-step instruction. Models make it possible to productize teaching methods that were previously difficult to scale.

5. Gen World Engine Must Generate Souls, Scenarios and Teachable Tasks at the Same Time

Gen Souls Engine handles “soul generation”: every character has a 3D Avatar and can also have a profession, appearance, personality, movements and emotional reactions. 刘夜 says one intern can use AIGC to generate roughly 1,000 characters a month; when users speak well, characters respond with gestures, facial expressions and actions, rather than simply playing “good job” and fireworks.
He compares Talkit with Roblox. Roblox grew through an editor and user-generated content into what he calls a “$100B metaverse company,” while Talkit wants models to lower the cost of producing worlds. Gen Scenario Engine has not yet been fully solved: the current backgrounds are mainly 2D images that look 3D. He expects true 3D-generation models to mature rapidly “over the next year or two.”
Gen Task Engine is one of the central challenges. 刘夜 says roughly 20% of ChatGPT’s DAU use it to learn foreign languages, but free-form conversation is difficult to sustain and not accurate enough to generate consistently. Talkit therefore uses model infra, prompts, routines, serial and parallel calls, agents and conventional engineering to constrain dialogue.
The constraints include n+1—ideally no more than 1 or 2 new words per sentence—along with spaced repetition, interests, learning style, language style and session length. Native English speakers in the real world may speed up or deliberately make things difficult for learners; Talkit wants to build a “language-learning-friendly virtual world.”

6. Duolingo Proved That Weak-Motivation Markets Sell Sustained Participation Before Learning Outcomes

刘夜’s market breakdown is roughly 2B language learners worldwide, including 1.5B English learners and roughly 500M learners of Spanish, French, German, Italian, Russian, Japanese, Korean, Chinese and other languages. Only about 10% study for exams; the other 90% are weak-motivation learners.
He rejects the use of “hard demand” to describe this market, calling it “stale and rotten vocabulary” inherited from the transition from agrarian to industrial society. Entrepreneurship, podcasts and personal growth are not needs that people die without; much of what consumers buy is lifestyle, experience, achievement and social identity.
The host’s inference is that the underlying reason people pay for Duolingo may not be “I learned it,” but “I’m making progress, and I like myself more.” 刘夜 strongly endorses Duolingo’s value in sustained participation, positive feedback and lifestyle, summarizing it as: “let language learning effortless, then effectiveness.” Entertainment is the goal; learning may be the valuable byproduct.
His best example is a friend who maintained an Arabic streak for 1,000 consecutive days without ever meeting an Arab person. That does not mean the product failed: it delivered showing off, companionship and small but frequent positive feedback. “When life disappoints you, go memorize words”—because the feeling of progress arrives quickly enough.

7. Talkit Is More Likely to Capture Duolingo Churners Looking to Train Their “Mouth”—But Speaking Must First Stop Hurting

According to data cited by 刘夜, Duolingo has accumulated roughly 1B users, 150M MAU and 50M DAU, with about 12M expected to maintain 365-day streaks; meanwhile, roughly 850M registered users have churned. “You can never teach a student who doesn’t show up,” so continuous learning must come before effectiveness.
Talkit conducted a survey of thousands of users, and the team summarized a common need among a group of churners as “seek to improve their spoken language.” 刘夜 believes that after 300, 500 or 1,000 days of streaks, users have mastered fill-in-the-blank exercises and vocabulary and will eventually say: “I can’t play fill-in-the-blank games to train my fingers for the rest of my life. I need to start training my mouth.”
The host’s key objection is that memorizing words produces easy positive feedback, while speaking repeatedly frustrates beginners. 刘夜 does not answer with a single feature; he breaks the solution into engage: characters must be interesting and emotional, remember the user, keep difficulty reasonable, and give tasks plot, challenge, interaction and rewards.
He uses his own experience to explain the power of immersion. He failed middle-school geography and history, yet playing Uncharted Waters taught him the ports of the world, while Romance of the Three Kingdoms helped him remember roughly 2,000 generals. The content did not suddenly become easier; the learning experience changed. Large models supply the “engine” that makes that change possible.

8. 2 Years, More Than 100 Versions and 3 Rebuilds Produced a Product Worth Only “30 Points”

Over the past 2 years, the team tested more than 100 versions and overturned the product architecture 3 times. 刘夜 did not accept his first interview until after the full release because “if the product isn’t ready, I simply don’t release it.” He compares the 24-month development cycle with 哪吒’s birth: ordinary people take 10 months, while 哪吒 took 24.
He remains extremely restrained about the result: “This product is nowhere near successful.” For now, it is merely “30 points, good enough to show.” His more vivid standard is that the child is at least “born with all four limbs intact, appears able to breathe and can breathe on its own.”
The 2 years of repeated validation centered on 3 layers of philosophy: user value should be effortless first, then effectiveness and enjoyable; pedagogy should use TBLT like native-language learning; and the product must create “an emotional immersive world for language learning.”
The beta still had many bugs, but 刘夜 says some users already spend 1-2 hours a day on it. At this stage, the team is focused on target users’ conversation time and retention: “whether they show up, and then whether they keep talking.”

9. The One-Day Decision After ChatGPT Was Built on 3 Years of Searching Without Direction

During the pandemic, 刘夜 examined more than 30 industries, including healthcare, biopharma, new-energy vehicles, consumer electronics, chain coffee, recruitment platforms and blue-collar labor, but never found a direction that made him pull the trigger. When ChatGPT appeared, he “decided in a single day to pursue this direction.”
He recalls that GPT-3.5 had not yet launched. Working in an offline environment, he spent a day writing the key prompts and generation flows for several engines and simulating the product through a conversation stream. He then spoke with 李想, repeated the pitch at a class reunion a week later, and quickly closed the $10M seed round—there was “not enough to go around.”
After fundraising, he did not immediately manufacture progress through publicity. 刘夜 says he spent almost all his time at the company, traveling to the US only once a quarter; unlike at Homework Box, when competitors’ frequent launches would trigger an urge to follow, this time he was able to stay quiet for 2 years.
The pace also addresses the host’s concern about a “revenge startup.” An unfulfilled ambition may provide motivation, but 刘夜’s evidence was not a verbal denial: he was willing to postpone the urge to prove himself until his product philosophy had a usable vehicle.

10. The Language Barrier Is Ultimately a Problem of Allocating Capability

刘夜 uses the roughly 8M Chinese people in the US as a concrete backdrop. Many Chinese technical workers feel that “we do all the work,” while their reporting line may run to an Indian colleague. He describes a friend with a US computer-science Ph.D., an early Meta employee and current Snapchat employee who can discuss technology but struggles to discuss management and small talk.
The friend summarized the 2 sentences he says most often as “Yes, I can do it” and “Can you speak again?” 刘夜 concludes that the language barrier fails to “unleash everyone’s potential”: effort, intelligence and diligence do not automatically translate into the opportunities people deserve.
Against the host’s observation that many jobs are done by Chinese workers who report to Indians, 刘夜 offers a response worth preserving: many Chinese people who are best at expressing themselves remain in Zhongguancun, Wangjing and domestic tech giants, while Indian talent travels globally more naturally because of India’s economic conditions and English advantage. The gap should not be reduced to Chinese people being unable to express themselves.
His own spoken English was so weak 2 years ago that entering the US was difficult, and Americans asked, “How can you provide English training if your English isn’t good?” His answer was not to evade the gap but to say, “I started learning.” For him, a language product is both a market opportunity and a question of whether Chinese people can participate more freely in global competition.

11. When Double Reduction Arrived, His First Feeling Was Not the End of the World but Complicated Relief

Before Double Reduction formally arrived in 2021, the leading companies already knew what was coming and had little time left. On the night it was implemented, 刘夜 called his co-founder in the US. When asked how he felt, he answered, “Pretty good,” and the co-founder said, “Pretty good.”
The relief came from changes that had begun much earlier. 刘夜 sees 2018–2019 as the key transition from traffic and growth to commercialization, followed by battles over dual-teacher large classes, advertising, talent and sales conversion. The team discussed renewals, scripts, traffic and ROI every day; one teacher even endorsed 3 companies at the same time.
His retrospective was “limited value, internal competition and a betrayal of the original mission.” Most leading online-education founders still cared about education, but could only “numb themselves with hard work.” He often asked whether he would let his own child attend this kind of tutoring; the answer was “no, in one second.”
The host asks how there could be no pain when the investment was wiped out. 刘夜 answers: “When you know it’s the end of the world, you don’t have the mood, because all the problems stop being problems.” The regret remained, but only building “a bigger, more successful, more valuable company” could truly dissolve it.

12. What He Rejected Was Not Money, but Projects That Could Not Answer “Why Must I Be the One to Do This?”

刘夜 studied finance in college but spent roughly 90% of his time on computer science and programming competitions. After graduation, he did not look for a job and went straight into entrepreneurship. His first company spent 10 years in Chinese SaaS and low-code; when he switched into education, he gave all his shares to those who stayed, taking only 3 people with him. He eventually retained 10% as a keepsake after colleagues insisted.
He also says that in 2013 he had an opportunity to use “nearly 100,000 BTC” to build an exchange, but ultimately walked away and chose Homework Box, which raised only $1M and was barely believed capable of challenging TAL Education Group. For him, the choice between financial gain and education had already been tested once.
His post-Double-Reduction coffee exploration came closest to becoming real. He drank roughly 6 cups a day for 6 months, visited hundreds of stores in Beijing and Shanghai, interviewed brand executives, staff, architects and planners, and studied site selection, equipment and production. Aranya had even approved land and built a building, while investors’ money was about to arrive.
At the final step, he still stopped the project because he believed he could only “open a bunch of coffee shops and make money,” without creating a sufficiently large, unique value. His distinction was: “Are you an arbitrage merchant, or are you an entrepreneur?” Water and air are highly valuable, but their value alone does not make them his entrepreneurial opportunity.

13. After Starting From Zero, the Real Danger Was Not a Lack of Opportunities but a Lack of Purpose and the Inability to Be Needed

The host initially interpreted the 3-year gap as finally having time to do enjoyable things. 刘夜 corrected him immediately: “That’s not accurate.” He actually wanted to find an opportunity quickly; age pressure and the inertia of more than 20 years without a real break left him unsure how to spend his time, deeply panicked, distressed and lost.
He admits that he was unusually enthusiastic then. Whenever a friend asked, he would “help you 10 times more,” because he was bored and “really wanted to be needed.” Many projects appeared capable of raising money quickly and making a lot of money, and investors even pressured him to invest, but he could never pull the trigger: “It’s not something I want to do.”
This was not his first period without a goal. While looking for a new direction in 2013, he went to a psychiatric hospital, was diagnosed with depression and was told to try medication only if psychotherapy failed. The doctor asked whether he could speak and whether he could exercise. After he answered yes to both, the doctor told him to talk to people and run intensely, saying everything on his mind out loud.
刘夜 followed the advice for 1-2 months and says his symptoms “all went away.” He did not become depressed again during the second gap for 2 reasons: Homework Box’s results had shown him that opportunities still existed, and the pandemic paused everyone together, reducing horizontal comparison. But feeling “not as bad” was not the same as feeling relaxed; he simply had not yet concluded that “this life is over.”

14. Double Reduction Left Three Lessons: Value, Cycles and Globalization

The first lesson remains value. Whether a company is at a high or a low, it ultimately has to return to the question of whether it creates unique value for users; “everything else is bullshit.” Doing something valuable but easy is not scarce. What deserves long-term investment is something “difficult enough and valuable enough,” ideally something “that nobody will do if you don’t.”
The second lesson is cycles. At the best of times, you need to know conditions may deteriorate; at the worst of times, the turn may already be approaching. Young founders often cling on as conditions move from good to bad, then give up just before they move from bad to good. What looks like impatience or a lack of persistence may fundamentally be a failure to read the rhythm.
The third lesson is perspective. 刘夜 observes that the online-education founders with the means to do so “almost without exception” turned toward globalization. He answers the question of persistence through continuity: “I have three decades, first decade, second, then is the third.” The first decade was SaaS, the second education, and now he has entered the third. He also explains why weak English need not be a barrier with “willingness greater than karma, then greater than ability”: the key is whether willingness exceeds resistance.
His slogan, “Swim until the seawater turns blue,” comes from an image written by 余华 and the view that the sea near Zhejiang is yellow while the blue ocean can be clearly imagined in the distance: “The blue sea is far away, so the sea of suffering is not bitter.” His conclusion for young founders follows: amid an explosion of opportunities, many are false; spend 2 or 3 years looking for the long-term right thing. Once you find it, do not fear loneliness, because “loneliness means scarcity,” and do not fear mistakes—“Only by making mistakes can you live with yourself… What if they’re all wrong?”

General Catalyst CEO, Hemant Taneja: Lessons Scaling GC to $40BN in AUM Organizational Capability Is the Real Moat for AI Companies | Conversation with Palona AI Co-founder 任川