A Three-Hour Talk with Fellou Founder 谢扬 on Solitude and Productivity

Summary

• Fellou positions itself as “the world’s first Agentic browser”: a browser, workflow automation layer, and Agent combined, built to deliver tasks end to end. It supports local, local-virtualized, and cloud environments. Unlike the cloud-VM approach taken by Manus and OpenAI Operator, Fellou can log in locally with the user’s own accounts. 谢扬 says “98% of the data on the internet is non-public,” giving Fellou a far wider operating scope than cloud agents limited to public websites; in testing, a single task took 3.7 minutes versus 18.6 minutes for Manus. “I can fly—and then I can go into the ocean.”
• The core investment thesis is a data flywheel: the browser is “the Tesla inside the computer,” the best collection point for action data from the digital world—computer use. Inspired by Tesla’s FSD, 谢扬 sees every click, scroll, and copy-paste action online as training data. Fellou’s in-house Eko framework scores 91 on WebVoyager versus 89 for Browser Use and runs 2.83x faster. The framework is partly open-sourced to build a developer ecosystem, while the data-collection and experience-pool layers remain closed.
• The funding and model plan: at least $100M is needed to build 2 models—one GUI model using a post-training route, and another that he “can’t talk about yet”; Fellou will not build a language base model. Comparing Fellou with Dia, formerly Arc, which raised $50M, he says “AI-native browser” is too broad a definition and that “many of its capabilities are still primarily conversational.” The host adds a New York labor-cost comparison, arguing that $50M in the US may buy roughly what RMB50M buys in China.
• Authing was a complete lesson in the cycle: during the 2021 SaaS mini-boom, the company mistakenly scaled from 50-60 people to 200; after the Fed began hiking rates and US SaaS was hit—Okta’s market cap fell from more than $40B to more than $10B—it went through repeated layoffs but still delivered 20% and 60% growth, extended cash runway to 60 months in 2023, and reached breakeven in 2024. Its 700+ paying customers include EY, Lululemon, China Merchants Bank, and Haidilao; total funding was $30M. “I’m staying at the table” (“我就是要留在牌桌上”). “The dirty, difficult work is the moat.”
• He wants both paths at the fork in the road: “The business model may be Google; the technology model may be Microsoft.” Advertising is “the best business model—there may be nothing better than advertising,” while usage-based memory pricing and outcome-based pricing are also under consideration. 100M users is a “small target,” but he says there is “really no way to judge” how long it will take. The strategy is international first: one global product, deep localization, and a developer ecosystem. “C in overseas markets, B in China” may be the best model, in his view.
• He is not afraid of big-tech competition “because I think I’ll be fast enough”; he sees Quark as the most likely follower and ByteDance as the company that should follow: “Let’s do it together and see who outdoes whom.” He criticizes Manus for having 4 internal teams working on Manus, calling this kind of copycat competition “unhealthy.” Silicon Valley makes the table bigger: “If 4 people can sit at it, make it bigger so 8 can sit there, rather than pushing you off.” His defense is “attack forward, defend backward”: international To C on offense, and the To B browser market as the fallback.
• The person himself is the episode’s biggest alpha: a solitary geek driven by intuition and endowed with an exceptional tolerance for adversity. He grew up “nourished” by the internet from age 13 through the 22dm forum, and answered a launch-event question about the source of his confidence with: “I don’t know why either, but I just know.” Authing’s only real learning was that “all beautiful things are produced by deep relationships between people.” What prompted this communication-averse INTJ to actively build connections was AI itself: “AI needs wisdom… one person’s wisdom isn’t enough; I need the wisdom of a group.” The host once associated his experience with the “small-town youth” archetype, but that is not 谢扬’s own self-description.

Deep dive

1. The Opening: After Manus, Another Chinese Agent Product Bursts Onto the Scene

• Shijie sets the tone by calling Fellou “the world’s first Agentic browser”—something like “a group of AIs teaming up to work on your computer, while hiding in a shadow space that doesn’t interfere with your control.” Its founder, 谢扬, was born in 1996 and is a Forbes 30 Under 30 honoree; his identity-cloud company Authing is already a leading player in China’s SaaS identity market.
• The timeline: Fellou began in stealth in June 2024 and launched 10 months later. The recording took place the day after launch, after roughly 30 AI and technology media outlets had been invited; several outlets reported the next day that trying it had “changed the way people interact with a browser.”

2. Solitude as the Starting Point: No Gifts, Hand-Me-Downs, and a Childhood of Being Overlooked

• 谢扬’s own account: “I was relatively lonely as a child and didn’t have many friends, so the computer became a window through which I could see the world.” Seeing a different world, however, made him even more different from the people around him—and more isolated.
• The defining story: his uncle, a doctor at Peking University Health Science Center, brought New Year gifts for every child in the family except him. When he went to his mother crying, she said, “They forgot you.” He wore his older cousins’ hand-me-downs. “Being forgotten and ignored was the normal state of your childhood.”

3. Sent to Boarding School in Third Grade: When the Outside World Offers Nothing, Build Your Own Shield

• His father told him he would come back to pick him up that evening, but 谢扬 waited after school, got lost as soon as he left the gate, and was found by people at a barbershop; he was picked up 2 hours later—and sent to boarding school as usual the next day. “Just like the gift incident, maybe I simply wasn’t valued that much.”
• The experience produced a self-protective logic: stop asking the outside world for help, “because you know you won’t get anything by asking,” and instead ask, “How do I make myself stronger?” The work happens on 2 tracks—capability and mindset. It offers one clue to both his management style and his product philosophy.

4. Pascal’s Reed: The Awakening of Self-Awareness

• In high school, he read that “man is a thinking reed.” “That sentence opened something up for me.” He began simulating a version of himself in his mind and examining the person he was at present. “Man is great because of thought; man himself is very fragile.”
• He did not say a single word to a classmate during his first week of high school—not because he was antisocial, but because he had “no desire to communicate.” “It felt like there was nothing nourishing to talk about,” since his social life had already been completed online during 3 years of middle school.

5. The 22dm Forum: Debating Big Bad Wolf with Geniuses Online

• The entry point was almost accidental. While watching Pleasant Goat and Big Big Wolf, he noticed the website 22dm.com in the end credits—“22 Anime,” whose production team averaged 22 years old. He registered on July 21, 2009. “To be honest, that forum changed me.”
• His case for Big Big Wolf is worth preserving: persistence (“I’ll be back”), intelligence and invention, and responsibility toward his wife and children. “Pleasant Goat was always messing with him, and I really disliked Pleasant Goat.” The forum’s lurkers included a third-grader who had passed the TOEFL, a sixth-grader writing an operating system, and 15- or 16-year-olds admitted to Tsinghua or Peking University through informatics competitions—“genius users.”
• Another source of spiritual nourishment was SpongeBob’s optimism: “I’m ready, I’m ready.” His own summary: “Optimism is a method; pessimism is an attitude.”

6. “Child of the Internet”: He Shares Aaron Swartz’s Ideals, Not His Methods

• He calls himself a “child of the internet” for 2 reasons. He identifies with Aaron Swartz’s belief in freedom of information, but says, “Downloading huge databases and making them public—I think that violates my own values.”
• The deeper reason is biographical: “I was raised and nourished by the internet. Without the internet, I might have become a completely different person”—perhaps someone with “borderline personality.” The host then associated the experience with the “small-town youth” archetype; 谢扬 is from Ganyu in Lianyungang, Jiangsu. He makes a comparison to Zhang Yiming’s description of Toutiao and Douyin as windows onto the world: “For me, the computer was a window through which I saw the world.”

7. “I Don’t Know Why, but I Just Know”: Where His Software Intuition Comes From

• At the launch, a reporter asked what his strong conviction was based on. His answer: “I don’t know why either, but I just know.” The intuition dates back to age 13 or 14, when he would write programs while reverse-engineering how each piece of software worked—studying software the way a hardware blogger tears down a device.
• His teaching example for the host: clicking the close button on a Word window actually triggers several steps—checking whether to save, closing the window, and killing the process. That training pays off in Fellou’s privacy space: “When someone said adding RPA to a browser would interfere with the user’s screen, I instinctively knew virtualization could solve it. No one told me, but I just knew.”

8. A 13-Year-Old’s Portfolio: A Forum Flooder, Sword Art Online Interaction, and a SpongeBob Player

• His teenage résumé includes “a lot of things that might get me criticized if I talked about them now.” He wrote a forum-flooding tool that posted thousands of messages per minute—“going to war with Baidu Tieba, building a nuclear weapon inside Tieba.”
• The constructive side was more revealing. He recreated the full VR-style interaction system from Sword Art Online on his computer and posted it on Bilibili. He also built a personal SpongeBob player that remembered which episode he had last watched. “Features all the major video sites have today—I had already built them,” only locally at the time.

9. High-School Rebellion: A 6,000-Character Poster and 3 Years of Refusing Morning Exercises

• His parents were called in every semester. He wrote a 6,000-character wall poster and pasted it outside the grade director’s office, accusing the teachers of “formalism, failing to teach according to aptitude, and making empty, fake-sounding speeches.” When teachers went class to class asking who wrote it, he denied it. “I just wanted to tell him: you still won’t know it was me by the time you graduate high school.” He did, however, write about it on his blog.
• Refusing morning exercises was the deeper trauma. He repeatedly told a teacher, “I feel uncomfortable and miserable standing on the field,” but was criticized in front of 400-plus students. He took off his watch, smashed it on the ground, and walked off. The teacher tried to expel him that day, then reversed course at 10 p.m. “When I think about those 3 years now, I still want to cry. They were simply the darkest 3 years of my life.”
• His verdict on Chinese education: “It puts you into a mold. You have to pass through that mold before you can grow freely again.” The physical distress he felt looking at a mass of people gradually disappeared only after he entered university.

10. 3 Stages of Life: Build Companies, Make Films, Run a School—and Compress Compulsory Education to 4 Years

• His university life plan was to become an entrepreneur first—“I think I’ve done that now”—make 10 years of films at 40, and work in education at 50. “Too many emotions were suppressed in me when I was young. I want to turn those emotions into a film.”
• His education plan is specific: compress compulsory education from 9 years to 4; provide general education at ages 6-10, let students try different specialties and “find what they love” at 10-13, then allow them to choose technology, art, or education. “I can give them money, space, resources, and teachers to grow with them.” Shijie notes that this ambition does not conflict with commercial ambition. “Exactly. Right now I’m focused on productivity, using productivity to amplify commercial value.”

11. Being “Used” by an Upperclassman: The Productivity Obsession and the Origin of “Intellectual Parity”

• At university, an upperclassman recruited him under the pretense of building a product, then “sold me into doing labor for RMB800 a month.” The shock was not the money but the realization that “the world could actually operate this way.” It left him deeply hostile to outsourcing and determined to liberate productivity.
• His view of fairness is pragmatic. Perfect equality does not exist, but “raising productivity can raise the floor of fairness.” If a product can reduce an 8-hour workday to 2 hours, “you can do whatever you want with the other 6 hours.” IT products can create “intellectual parity” by giving people without know-how access to the workflows of the very best performers.
• Shijie’s challenge is worth retaining: class is fundamentally a question from Capital—how the means of production and wealth are distributed, not how labor or intelligence is allocated. 谢扬 partly concedes the point. He has read British industrial history and acknowledges that “from another angle, productivity is still a tool politicians use to control society. There’s nothing wrong with that.” But “solving real problems in reality is more practical than insisting on starting some kind of movement.”

12. University Entrepreneurship: Online Photoshop, a Tencent Silver Award, and an Online IDE

• His first project, an online Photoshop, was submitted to a 36Kr feature and took first place in both the daily and main rankings. It later won a silver award in a Tencent competition—the only undergraduate winner—with prize money of less than RMB100K. “It released all the emotions I had suppressed from middle school through high school. I felt seen.”
• He insists he was not a genius, just relentless. He completed compiler theory, databases, computer networks, and the AI “Dragon Book” in his first year, studying from 6 a.m. until 11 p.m. or midnight. “I was always the last person to leave the library.”
• He put all the prize money into his second project, an online IDE. The plan had 2 stages: cloud programming, including a Webflow-like visual editor, a mini-program and web-page editor, and a command-line file system inside webpages; then natural-language programming. “Those technologies later became useful in Authing and Fellou.” The second stage began in 2016, before Transformers existed, and the technology was not mature enough to finish.

13. The Pivot to NLP: A Daily Brief with 100,000 Global Users and a Securities-Analysis Tool

• After pivoting to NLP, he built 2 products. One was a daily briefing for machine-learning developers, delivering 5 items a day to roughly 100K users globally. The other was a To B tool for regulators analyzing listed-company filings: named-entity recognition for directors and executives, relationship mapping from public news, and positive/negative reports on share transfers. “That can be monetized.”
• His own assessment of the period: “Basically, I was still figuring out what I wanted to do.” He changed direction roughly every year.

14. His View of Money: RMB100K Bought Experience, and He Likes High-Odds Projects

• The IDE project raised several hundred thousand yuan and never made money, but he has no regrets. “Maybe what I wanted wasn’t RMB100K. I wanted experience, and I really valued that period.” He learned the full engineering stack; the IDE he later worked on at ByteDance was also an extension of that experience.
• His attitude toward money is not indifference. “I actually need money. Mainly to do big things—projects with high upfront costs, but also high ROI and high odds. Those are the projects I like.”

15. A 2-Month Gap After Graduation: “I Realized Society Was About to Leave Me Behind”

• He graduated in June but did not start work until August. Unable to decide what to do, he experienced an unexpected psychological shift. “The first month was fine, but by the second month I was having trouble sleeping.” He went from not wanting to join any group to feeling that he “actually needed society,” and decided to “quickly find somewhere to stay for now.”
• The ByteDance decision was characteristically 谢扬. He applied to ByteDance, Alibaba, and Baidu, with Alibaba originally his first choice. But ByteDance interviewed him from 10 a.m. to 2 p.m., and the interviewer took him to lunch. “The ByteDance cafeteria was especially good. I decided right then”—along with the appeal of its “Always Day 1” culture.

16. 10 Months at ByteDance: The Natural-Language Programming Idea That Was Rejected

• He worked on an internal IDE at ByteDance. “I wasn’t that happy. I felt constrained in a lot of ways.” The key disagreement came when he proposed: “Since ByteDance’s recommendation algorithms are so good, shouldn’t we build natural-language programming?” His manager shot it down: “You’re holding a hammer looking for a nail.” Shijie presses him: if ByteDance had pursued it then, perhaps it would have beaten Cursor.
• His assessment of ByteDance is balanced. Its technical innovation was not necessarily at the frontier; it was “extremely pragmatic, exploring only technologies that could immediately serve the business. There’s nothing wrong with that—but from my perspective, I was definitely constrained.” His entire interaction with Zhang Yiming: “I saw him in the bathroom once. We never spoke.”

17. Choosing a Direction in 2019: Identity Cloud as the “Door” Into IT Infrastructure

• He applied lessons from university when choosing the business. In 2019, with the internet era nearing its end, he wanted a field that was relatively vertical, could survive long enough, could make money, and fit the technology trend. Computer vision was nearing the end of its cycle; the remaining options were cloud-native and SaaS. He settled on identity cloud: “If you’re building or buying software, you can’t get around identity. It’s a bottleneck in enterprise IT infrastructure.”
• The metaphor became widely circulated. Enterprise construction is like building and decorating a house: “I’m the company making the doors. As long as the building grows, the number of doors grows.” He knew it was a small, attractive niche. “If confidence is 100 points, I can only give it 70. It will definitely make money, but its ultimate size is uncertain. This market simply isn’t like the US market.”
• The opportunity came from looking inward. When he was launching products at university, users kept failing to log in—verification codes never arrived, WeChat Chinese names conflicted with database character sets. “Identity isn’t easy to solve,” so he started there.

18. The First 3 Deals and “Seeing Through an Investor in 2 Minutes”

• With no To B network at all, he used product and content to attract customers with genuine pain points. The first deal was RMB8K, the second RMB50K, and the third RMB800K. He raised more than RMB7M in angel funding and budgeted it over 18 months.
• His verdict on capital is blunt: “A lot of Chinese capital doesn’t understand software. The practical consequence is that investors who don’t understand it can’t judge what a person is capable of building.” It is similar to a nontechnical investor lacking his instinct for the underlying system. He says it takes “maybe 2 minutes” to identify the right investor—“you can tell from their eyes.” With people who do not get it, he still finishes the pitch. “It’s a sign of respect.”

19. Authing’s Scorecard: 700+ Customers and “Outsider Dimensionality Reduction”

• Authing now has more than 700 domestic paying customers, including EY, Lululemon, Volkswagen, Porsche, Bentley, Toyota, China Merchants Bank, China Everbright Bank, Shanghai Foreign Service, Higher Education Press, Career International, Haidilao, Zhihu, LandSpace, and Yangtze Memory. He makes no attempt at false modesty: “The whole industry says it, so it must be the best.”
• He attributes the core advantage not to a particular product feature. “I wasn’t from this industry. I was an outsider. I used an outsider’s lack of knowledge about the industry and a completely new method to build a moat, and achieved dimensionality reduction against all my peers.”

20. Benchmarking Okta/Auth0: BuiltWith Research and One Sentence That Changed Okta’s CEO’s Expression

• The global competitive map: Okta is No. 1, with Microsoft Azure AD as its largest rival. Auth0 was a later entrant that quickly reached $100M in scale before Okta acquired it for $8B; ChatGPT uses Auth0 for login. 谢扬 was selected as one of the influential founders in the global identity field for the “Auth0 Identity 25,” which he says made him “the only Chinese person selected.” He also appeared on a NASA screen.
• His research method is reusable. He used BuiltWith.com to see how many websites referenced Auth0, then compared the results rigorously with each financing announcement. “Traffic jumped sharply after every funding round—that showed they really had market demand.” The list of sites itself became a market-entry target list: “Just go after them one by one.”
• At his first meeting with Okta’s CEO, he opened with: “EY and Roland Berger are our customers.” Those customers had moved over from Okta’s China business, he explained, and the CEO’s expression changed. The decisive advantage was localization: out-of-the-box support for the WeChat and DingTalk ecosystems.

21. The 2021 Expansion Mistake: A SaaS Mini-Boom and the Fed’s Hammer

• Fundraising went smoothly, reaching $30M in total from investors including Tiger, GGV, China Renaissance, and CDH. The final round closed at the 2021 SaaS peak. “We timed every financing round perfectly.” Then the company made the expansion mistake, growing from 50-60 people to 200.
• His cycle analysis centers on Fed tightening: “Deposits went back into banks, and investment money disappeared. SaaS companies in the US were all losing money and needed capital to take them out.” Okta’s market cap fell from more than $40B to more than $10B even though the company had generated $2.3B-$2.4B in revenue the prior year and traded at less than 5x sales. Even high-quality revenue could not protect the multiple.
• His self-diagnosis is unusually direct: “Everyone probably got carried away. Learning this lesson at 25 is a good thing. When I’m 30, I definitely won’t make a decision like that.” After the expansion, “revenue didn’t rise, and efficiency actually fell.”

22. Turning It Around in a Downturn: 60 Months of Runway, Ruthless Layoffs, and Staying at the Table

• He based his forecast on historical data. Comparing the recovery after the 2008 financial crisis with earlier cycles, he concluded that recoveries generally took 4-5 years and predicted conditions would not improve until the end of 2024. So even 20 months of runway did not feel safe; in 2023 he forced it up to 60 months, and Authing reached breakeven in 2024.
• He has no psychological barrier to layoffs. “Cut decisively in one stroke.” Does he feel guilty toward employees? “This is your job and your life. You have to solve it yourself; don’t expect anyone else to.” Even people with strong capabilities but no value inside the system had to leave, “return to the free market, and choose again.”
• When people advised him to “just burn through the money and shut the company down,” he refused. “I’m staying at the table” (“我就是要留在牌桌上”). “Even when the market is declining, I still want to grow.” Growth was 20% in the first year after layoffs and 60% in the second. “I did everything people said I couldn’t.” What kept him going was instinct: “There will be good luck ahead.” His motivational feed included Mao quotes pushed by Xiaohongshu, SpongeBob, and Naruto’s Naruto—“lonely and unpopular from childhood, but proving himself step by step and saving the entire village.”

23. 3 Product Generations and “What We Sell Is a Philosophy”

• The reason Authing kept spending on R&D was its 3-generation evolution: developer-centric identity, with SDKs and protocols for B2B2C login; full-scenario identity cloud, after discovering that domestic To C users had low value and shifting to internal enterprise employees, OIDC and other enterprise protocols, and pre-integration with more than 2,000 applications; then event-driven cloud-native infrastructure for customers such as Vanke with hundreds of thousands of employees, where onboarding, transfers, and departures can generate thousands of requests a day. “We spent another year-plus building it, and customers such as EY came over immediately once it was done.”
• His commercialization methodology: “First define the philosophy of what you’re doing in the industry. Take the high-level view. Customers aren’t buying the specific thing you built; they’re buying your philosophy.” JD.com’s philosophy is speed; Authing’s is “developer-friendly and event-driven.” “Good IT leaders all like abstraction.”
• He is also candid about the sales bottleneck. Some large customers with “very old-school, very state-owned-enterprise” cultures simply “don’t get the point, and won’t tell you the truth.” “If there are some accounts we can’t break through, let them go. Forcing a slick salesperson into the company would dilute the culture.”

24. Management and Human Nature: “I Grind Other People the Way I Grind Myself”

• Authing’s biggest challenge is people. “Every year I see human stories that make me more stunned than the year before.” One example was salespeople playing both sides and provoking conflict. “Maybe there was something in it for him. I don’t know why else he would do it.”
• He admits to his own style: “I’m stubborn. I grind other people the way I grind myself. I don’t care about their feelings, and I don’t care about my own. Once a target is set, it has to be completed, and that hurts a lot of people.” The most common feedback is that he has “no human touch.” Should he change? “No struggle. I won’t change. I want efficiency. The people I hire need to change.” A note posted in his office for 5 years reads: “Failed managers are always trying to persuade others that they are right.”
• He highlights only 1 learning: “All beautiful things are produced by deep relationships between people.” It came from watching many people leave while others stayed, including people he had strongly believed in and admired who chose to leave the company.

25. How AI Changed Him: From Draining Social Interaction to “I Need the Wisdom of a Group”

• Asked whether he now genuinely wants to communicate with colleagues, he answers candidly: “If I say no, I’ll sound hypocritical; if I say yes, I’ll be going against my instincts. I really don’t like it that much. Spending large amounts of time communicating with other people every day is draining for me.” Why do it anyway? “Because AI changed me. AI needs wisdom, and turns wisdom into data. One person’s wisdom isn’t enough; I need the wisdom of a group.”
• The evidence is Fellou’s own positioning. Marketing, product, research, algorithms, engineering, and investment—7 or 8 people—worked from 9 a.m. to 12:30 a.m. to develop it together. “It wasn’t a conclusion I reached alone.” He moved from assuming everyone else was stupid to believing that collisions between minds can generate insight.

26. The ChatGPT Moment: “I Thought I Was an Idiot”

• Asked whether he had ever doubted himself, he says: “Yes. When I saw ChatGPT, I thought I was an idiot. Why hadn’t I spotted this technology earlier and started researching the space?” He had vaguely heard that people in the US were building large models but had not paid attention.
• His emotions followed a curve. While domestic large models were taking off, he was still consumed by Authing’s revenue problems—“very anxious, but not quite that anxious.” He knew he could not catch up on foundation models, but after trying them, he thought, “They’re still far behind. Let them work on it a little longer while I organize my thoughts.” His conclusion: “It was still early, but it wasn’t too late. Last year I began to feel that if I didn’t start, it might become too late.”

27. Reading Papers During the 2024 New Year Holiday: 3 Directions Converge on Action-Space Data

• He read papers intensively during the Spring Festival and May Day holidays, narrowing his focus to 3 directions: reasoning, coding, and agents. They corresponded to improving model accuracy, creating tools inside computers, and enabling models to use tools like humans. The final conclusion: “Data and reinforcement learning are important,” and what was missing was “data from the action space, not just the language space.”
• Sampling the physical world is too expensive. Recording an action such as “pick up this cup and bring it to your mouth to drink” costs too much. So he focused on computer use in the digital world. Only after GPT-4o arrived did the combination of vision, behavior, and reasoning gradually become viable.

28. The Browser Is “the Tesla Inside the Computer”: an FSD-Style Data Flywheel

• The central analogy is the most important idea in the episode: “The browser is the car inside the computer. When you surf the web, you’re really driving through different networks.” Which button you click, what text you enter, whether you scroll up or down, and whether you copy and paste can all be recorded as foundational training data, allowing the data flywheel to compound like FSD.
• The browser controls both time spent and interaction. Of an 8-hour workday, “maybe 6 hours are spent in the browser.” It is also the container where humans and agents co-work. Unlike ChatGPT or Perplexity, which are “relatively independent individuals,” Fellou’s agent can automatically obtain the user’s context, like a colleague sharing the same office and context.
• The breakthrough came after he abandoned all 3 other directions. He had been deeply anxious. “One Sunday afternoon in June, the sun was quite nice. I was lying on my bed, and suddenly I figured out the browser.” He wrote the document that afternoon: how to build the data flywheel, what the browser needed to become, and what kind of team was required.

29. The Fourth Type of Browser: the Boundary with Manus Is “Operating Scope”

• His taxonomy has 3 existing categories: traditional browsers such as Chrome; conversational browsers such as Doubao and Arc; and retrieval-augmented browsers such as Perplexity’s newly launched Comet, which depends on deep retrieval. Fellou creates a fourth category, the Agentic browser, focused on end-to-end task delivery. “Find positive reviews of Xiaomi cars on Xiaohongshu” means it logs into Xiaohongshu and reads 20-30 webpages in batches until it finds the answer.
• The fundamental difference from Manus and OpenAI Operator is that “the operating scope is different.” A cloud virtual machine can generally run only public websites—“If you log into LinkedIn through Operator, your account gets banned for 3 days”—while Fellou runs locally with the user’s own account. “98% of the data on the internet is non-public; only 2% is public.” Extending the driving metaphor: “I can fly, and then I can go into the ocean.”

30. The Eko Framework: 91 Versus 89, 2.83x Faster, and Partial Open Source to Build an Ecosystem

• Fellou’s in-house browser-use framework Eko, a variation on “echo,” is meant to “turn human language into action.” It scored 91 on WebVoyager, above the 89 scored by Browser Use, the framework Manus wraps, and ran 2.83x faster. “That score came without further optimization.” In live testing, a single Manus task took 18.6 minutes, Fellou took 3.7 minutes, and OpenAI Deep Research averaged 11.5 minutes; the speed was independent of user volume.
• The open-source strategy is deliberate. Open-sourcing the framework “brings in more developer support; the ecosystem is a very important source of strength for us.” Competitors can use it, but “they may not use it well, because we know what can and cannot be open-sourced.” The data layer stays closed: market agents execute tasks in virtual environments, and humans label successful cases into an experience pool to prepare training data for future in-house models.
• His definition of a moat comes from Authing: “The dirty, difficult work is the moat. In theory, others can do it too, but when they start, they definitely won’t be as fast as you are now. Go ahead and do it—spend as much time as I did.” He compares it with DeepSeek’s engineering “tricks of the trade.” “Yes, technical ornamentation.”

31. Proactive Intelligence: 3 Data Streams Predict the Next Step; Privacy “Never Existed”

• Proactive intelligence uses 3 types of data to predict the next action: user behavior; environmental state, including form fields, cursor position, scrolling, and selections; and environmental events such as refreshes, redirects, and switches. Few-shot examples provide action templates. “It isn’t that accurate yet, but it can already do some simple work.” For example, when a user opens Baidu to search for a Forbidden City itinerary, Fellou might offer: “I found 5 high-quality sources. Want to open them all?” The instant the user clicks Reply in Gmail, it may already have drafted 3 replies.
• Shijie asks whether that means there is no privacy left in web browsing. His answer is blunt: “There was no privacy to begin with. If you have a Cookie and Session, you can be tracked.” He still emphasizes legal compliance and encrypted storage. Phone numbers, ID numbers, passwords, and bank-card numbers are never collected.

32. Shadow Space: Like the Shadow Nobita Cut Off with Scissors

• The clearest product explanation comes from Doraemon. In one episode, Nobita cuts off his own shadow with scissors, and the shadow goes to work for him. Shadow space is a virtual space reflected by a mirror; the agent works there without taking over the user’s screen. “Nobita sleeps here, the shadow sleeps there, and neither affects the other.” Data can still connect and be shared across the 2 spaces. 谢扬 believes Manus chose cloud virtual machines rather than a local space partly because it worried browser RPA would disrupt the user experience.
• The 3 environments correspond to 3 types of tasks: local for tasks dependent on local context, such as organizing chat histories and email; local virtualization, or shadow space, for long-running tasks that depend on local context, such as “throw the workbook into the mirror, have it finish the assignment, and bring it back”; and a cloud desktop for long-running tasks that do not depend on local context, such as planning a 30-day team-building trip. More than 700 tasks have been labeled by environment so the model can choose and switch by itself.

33. Who Takes Responsibility When It Makes a Mistake: Rollbacks, Action Logs, and “Don’t Touch Bank Accounts Yet”

• He does not evade the responsibility question: “The user bears it.” Fellou asks for confirmation before every operation; in the event of an accidental action, behavior logs answer the question of responsibility—“whoever clicked it is responsible.” Citing the paper Agentic Infra, he says any incident must support rollback. “If it deletes a file, it must be able to find that file and restore it,” similar to version recovery in cloud documents but more complicated.
• He is candid about the current boundary: “Delete all my tweets”—the agent can do it. But any deletion or modification requires caution. “For now, I won’t let it control my bank account.”

34. Deep Search and the Credibility Fight: “What Matters Is That People Think It’s Good”

• Deep Search predates OpenAI’s Deep Research. The browser’s essence is “browsing and reaching a conclusion.” Users have 2 different needs: those who want only the answer, for whom AI can search 20-30 webpages and produce a report; and those who want to learn, for whom the browser provides proactive reading assistance. “AI can’t browse Xiaohongshu for you.” Every conclusion has a source, and hovering over the data reveals the original text.
• Shijie’s hardest pushback concerns a claim that Fellou’s own test of 49 cases beat OpenAI, Anthropic, and Manus. Depth, breadth, and readability are subjective dimensions. “Is it credible?” He responds by turning the argument around: “Subjective dimensions explain more. Many answers are subjective. What is user experience? What matters is that people think it’s good.” The judges selected cases randomly and blindly. He also attacks benchmark gaming: “We reproduced Browser Use. It wasn’t as good as it claimed. It really has no shame, and we can’t compete by having no shame.”
• The use case he personally contributes most is recruiting. When searching for C++ and Ethereum engineers, “I have no interest in watching the retrieval process. I just want the result. Give me a list and tell me which of these 10 people I should call first.”
• For users who insist on checking everything themselves, the value is memory and efficiency: scattered information is consolidated into memory and recalled when needed. “You may not need it every moment, but whenever you do need it, it’s there.” Shijie’s summary: “A giant warm boyfriend.”

35. Generative Interfaces and the Case for Visualization

• Fellou wants to build Claude Artifacts-style generative interfaces. When booking a flight, it could combine a calendar itinerary with available flights in a new interface. While writing, relevant information could automatically appear on the right. The interface would be personalized to each user, and anyone who dislikes it could simply say, “Change it.”
• Shijie’s skepticism is preserved in full. She prefers modular text and points to the fleeting popularity of H5 pages as evidence that “visualization is not necessarily the most efficient way to present knowledge.” 谢扬 concedes with a qualification: “It depends on the scenario. Education and marketing require a quick, comprehensive grasp of information. If you’re doing research with a lot of text, OpenAI’s approach is more suitable. I agree with that.” But he is betting on habit: “Once you get used to a faster way to make decisions, you can’t go back.”
• Why does it have to be a browser? Accessing the internet and rendering a GUI both depend on it. “Web-based is currently the best rendering solution. It’s much better than writing C++.”

36. Big-Tech Competition: Not Afraid, “Because I’ll Be Fast Enough”

• On Dia, formerly Arc, which raised $50M: “AI-native is too broad. Many of its capabilities are still primarily conversational. It’s still missing something.” He respects the team for having the courage to discard its baggage and start over, and says it “could become a very formidable rival.” Fellou needs at least $100M, in his view, to build a GUI model through post-training rather than from scratch, plus another model he “can’t talk about yet.” It definitely will not build a language model. The host jokes that he too could raise more than Dia.
• Big-tech data advantages are less straightforward than they appear. “Having lots of data and users doesn’t mean the data belongs to the team building that product.” It would take someone at Zhang Yiming’s level to order the data and businesses of 100 teams into a single system. “The cost is actually very high. It isn’t easier than building this data business outside a big company.” Quark is the most likely follower because “the team is quite innovative”; ByteDance is the company that should follow. If ByteDance entered tomorrow: “Let’s do it together and see who outdoes whom.”
• He criticizes Manus for having 4 internal teams working on Manus, calling it “completely meaningless.” The goal should be to define the industry, not follow it. The host calls this form of competition “unhealthy.” Silicon Valley makes the table bigger: if 4 people can sit there, make it big enough for 8 rather than pushing someone off. Fellou’s defensive depth: “Attack forward, defend backward. The floor is high enough, and the ceiling is infinite.” It attacks international markets for 100-300M users and falls back to the To B browser market. “The floor is that people will use your browser. The ceiling is infinite.”

37. The Business Model: Google’s Money, Microsoft’s Technology, and 100M Users as a “Small Target”

• He is pursuing both branches. Advertising is a natural fit for Proactive recommendations: “Among all the business models where you can make money lying down, advertising is the best. There may be nothing better.” Usage-based memory pricing and outcome-based pricing are also options. Shijie adds that advertisers finding people who match a profile more quickly can itself count as a productivity gain. His end-state formulation: “The business model may be Google; the technology model may be Microsoft.” His blunt assessment of Microsoft: “First-rate marketing, first-rate technology, third-rate products.” His vision echoes Bill Gates: “I want every person to have a digital companion in every device.”
• 100M users is a “small target.” If every user spends $1-$2 a day, or subscribes for RMB10 or $10 a month, the numbers are already enormous. But on timing he says: “I have no way to judge it right now. Really no way to judge it. It depends on how mature the technology becomes.” The strategy is international first, with China alongside it: one global product, deep localization, and joint solutions with developers in exchange for users. “C overseas, B in China is the best model. There’s a very good chance I’ll do that.” The biggest variable in overseas expansion may be Google, which controls Chromium.
• The name comes from “fellow,” meaning partner. W was unavailable, so it became U: Fellou.ai. On engineering, he uses it 7-8 hours a day and keeps 20-30 webpages open; memory usage is 1.5-2GB, with a potential reduction to 400-500MB. Shijie links this to Windows 95’s role in popularizing PCs and Microsoft reaching 100M users in 1999, speculating that an agent boom could force upgrades in CPUs, GPUs, and storage.

38. Ethics, Job Loss, and “I’ll Share the Money and Support Them”: an Unembarrassed Idealist

• He focuses the ethics debate on responsibility. The user currently bears responsibility when an agent makes a mistake; what happens when execution is accurate but the decision itself is wrong? Agent-enabled spam and gray-market activity could become more severe. “A security agent is also a direction we urgently need next. The offense has already appeared—but who is going to build the defense?” He believes alignment can be addressed under regulation: regulators need to define what values an AI should have.
• Shijie punctures the slogan that people will work 8 hours and spend 6 hours idle. It will not happen: if AI can handle 6 hours of work, a 10-person team becomes 2 people. “Capitalists won’t let you idle for 6 hours.” His answer is put on the record by the host: “If this company makes a lot of money, I’ll distribute the money to them and support them.” The host says, “I’m bringing this quote back to beat you with 10 years from now.” 谢扬 replies: “Please do. If the profits are sufficient, I’m serious.”
• The closing arc: his lowest point was the first half of 2024, when he had not committed to a direction and felt “like I was about to be thrown into the trash by the era.” Now, after launching yesterday, “a lot of people are looking for me and inviting me to sit at the table.” When Manus exploded, he “didn’t really feel anything” and focused on doing his own work. His material ambitions are specific and strange: build a house in the mountains and commute by helicopter. “There’s no one in the mountains. It’s comfortable. I would feel safe.” His motto remains unchanged: “Go boldly. Don’t be afraid. No one cares—and even if someone does, what does a person amount to?”