Summary

Intent’s first step is not to copy WeChat, but to solve the urgent need of cross-generational, cross-language families who have plenty to say but cannot communicate it, using context-aware translation. Brandon’s core use case is second-generation immigrants in the US exchanging voice messages with elders who speak only Cantonese, Spanish or Vietnamese; an IM product can use conversational context, relationships and tone to distinguish Japanese plain form from honorifics, and “LLM” meaning a Master of Laws from “large language model.” “I only need to focus on whether these users’ problems are actually being solved.”
The underlying asset is IM’s unmatched interaction frequency and trove of life data, not merely a translation feature. Brandon’s comparison: WhatsApp MAU send roughly 60 messages a day, ChatGPT MAU ask roughly 10 questions, and Google Search gets roughly 3 searches; the latest figure for daily Google searches per MAU may still be changing. Instagram users post fewer than 0.1 times a day, with roughly 0.3–0.4 Stories. Social conversations continuously create triggers while naturally accumulating gifts, relationships, status and language habits, which is why he calls IM a platform that “comes close to a complete record of personal life” with almost no effort.
The structural opening in overseas markets comes from open relationship graphs and a fragmented IM landscape, not Silicon Valley halo. Brandon’s view is that Chinese users’ relationships are concentrated in WeChat, making startups “impossible to beat”; overseas relationship graphs are more heavily embedded in phone address books, and every app may be able to read, add and access them. WhatsApp has just over 100M users in the US, Messenger may account for roughly 50%, and iMessage is used mainly in relatively affluent regions. “These IM products are nowhere near as concentrated as they are in China,” leaving room for a new product.
Intent has not yet proven PMF; growth and getting both sides to install remain the key unvalidated weaknesses. The host’s central challenge was: Sogou Input Method, Typeless, Aqua Voice and Wispr Flow can already turn Chinese directly into English, so why persuade the other person to install a new IM app? Brandon’s answer is that relationships and context can produce more accurate translation, while the company first co-builds with seed users and power users, then spreads through their real communication channels. The company has 20 people, no profits, and only several hundred paying users across previous products, but enough funding to keep exploring.
The long-term path is to turn intent in a conversation directly into action, but the team is starting with “no credit is better than taking credit.” The vision includes ordering delivery after a conversation, comparing Uber and Lyft, generating shopping lists, explaining messages users cannot understand, and even recommending people to talk to at that moment. The host noted that a wrong recommendation feed can be swiped away, while a wrongly triggered card in a chat keeps interrupting the user. Brandon acknowledged there is no ready-made solution; both models and interactions must iterate, and any intelligence must be built on the foundation that “messages can be delivered reliably.”
The founder signal is a refusal to pivot to products with easier-to-realize revenue after 7 consecutive attempts at chat products; the cost is materially delayed commercial validation. The team once built an AI document tool that generated overseas demand, and made money from a menu-translation app whose videos reached 1M views on TikTok. But investors told him: “We don’t care whether you have $1M or several million dollars in ARR.” Their original bet was on a product better than WeChat. The company previously raised roughly $2M from US institutional investors and recently received investment from 2 undisclosed internet heavyweights; one offered only this assessment: “It’s good that young people dare to take on something this big.”
Brandon has turned his greatest founder-market friction into founder-market fit while openly acknowledging the odds of failure. He ranked at the lowest level in Tsinghua’s English placement system; after arriving in the US, he could not even name Subway’s bread options and once ran every text message through ChatGPT, making language barriers the source of his product intuition. His exact words to investors were: “I don’t have a PM; the company keeps losing money; I can’t build an IM… We’ve failed 7 times, and the 8th could fail too.” If he succeeds, it can only be because he truly understands and solves the user scenario; if he fails, the primary causes will be insufficient user understanding and organizational capability.

Deep dive

1. Intent’s First Beachhead: Cross-Language Families

Brandon’s one-line definition of Intent is not “AI social,” but “a tool that helps people communicate better with one another.” The current product is an instant-messaging app with an “extremely seamless translation feature”; translation is only the first step toward the long-term vision.
The best use case he gives users is: “With this, you can communicate with your grandmother.” A second-generation immigrant in the US may be comfortable in English, while their grandmother speaks only Cantonese or Spanish, or cannot type at all. They can both send voice messages and read what the other person said in their own language.
This scenario came from early user research, not ad copy, and seed users continue to provide feedback around it. The team has 20 people and no profits; a previous product had several hundred paying users, but Brandon considers that scale “negligible.” Existing funding is enough to ensure the product will not suddenly disappear.

2. His First Startup Began with a Pitch Deck—and Failed Because None of the 3 Founders Had Conviction

With too few investable jobs for a biology major, Brandon was pushed toward a VC internship. There he learned the importance of sourcing, then joined the “President’s Cup” with 2 Tsinghua classmates he had met through gaming. The team wrote an overnight BP for a game-creation platform, claiming it would build a tool better than Unity and Unreal despite having neither a product nor any understanding of game development.
When investors asked how much money they needed, the young CEO named RMB1.2M, which he thought sounded like a large number. The investor simply replied: “I’m in.” The unexpected funding got the company started but did not create real conviction; after setbacks, none of the 3 founders had the energy to keep going.
Brandon and the CEO were both around 20 at the time, and the product was never built. By early 2022, the CEO was barely involved, so Brandon gave up all his equity and salary and left. He attributes the failure to circumstance-driven entrepreneurship and a lack of maturity.

3. Dropping Out Was Not a Hero Story; It Was His Confirmation That the Existing Path Had No Opportunity Cost

After leaving the company, Brandon considered continuing school, investing, taking a job and entering the civil-service selection program; starting another company ranked last. The turning point came during a retention interview at Source Code Capital. Asked about the biggest failure of his university career, he answered: “That I’m still studying here,” because he was in his third year and still had no idea what he really wanted to do.
The interviewer reminded him that Steve Jobs and Bill Gates “dropped out because they had something to do,” not for the sake of dropping out. Later, during a regular meeting on a boat at Guizhou’s Xiaoqikong scenic area, he suddenly said, “I’m going to drop out.” He had no investors and only a vague direction related to “group chats and chatting,” so he first took a one-year leave of absence.
The takeaway he retained from biographies was that “doing something big and doing something small are equally difficult.” Being a founder and delivering KPIs at a company both require burning yourself out; the former at least lets you “burn happily for it.” Classmates also kept asking, “Have you found anything to do lately?” which made him feel he could bring together a group of talented but bored people and try something.
The real safety net was not savings. After being recommended to Tsinghua in his senior year of high school, he was kicked out of the classroom by his homeroom teacher, contacted professors on his own, moved to Beijing with a few thousand yuan to rent a room, worked and rebuilt his social network. That experience convinced him that “I’ll never starve, no matter what.” The high hourly income from competition training further eliminated his dependence on a degree or a salary worth several hundred thousand yuan.

4. The Original IM Thesis: High-Traffic Assets Had Not Unlocked Chat’s Commercial Value

The second startup raised “several hundred thousand dollars” from MiraclePlus and a Beijing angel firm. Before the dollar-denominated structure was completed, Brandon and his co-founder fronted several hundred thousand yuan to get the company going. Although global travel was restricted at the end of 2022, he wanted to build globally from Day One, first studying Singapore and South Korea before returning to Beijing.
He saw that chat apps had extremely long usage times, very high open frequency and enormous traffic, yet Tencent and Meta had long treated them as defensive products. WeChat’s mini-programs, mini-games and Moments ads, along with Meta’s business accounts, had worked, but chat itself still did not resemble the native content advertising of Douyin or Instagram, where “just keep scrolling and it’s a money-printing machine.”
Brandon was not proposing to stuff ads into private chats. He was asking whether the highest-frequency user scenario—chatting—could produce a form in which product value and commercial value genuinely aligned. He believes WeChat has done a great deal of excellent work, but much of its value is still realized by taking users away from the chat interface. “Chat itself” remains underdeveloped.

5. Eric Yuan’s “It’s Still a Bot” Exposed the Lack of a User Reason in the Early Plan

After Qi Lu mentioned the project to Zoom founder Eric Yuan, Brandon planned a trip to the US. He had expected the meeting to feel like visiting a friend; only after arriving did he realize that a major founder’s schedule had to be booked in advance. He eventually waited 2 months for a 1-hour conversation.
The team imagined a system behind every group chat that understood context and member identities and automatically completed tasks in the background—roughly what would later be called a multi-agent system. But Brandon could not explain why users would use it or how it differed from an ordinary bot. Eric Yuan’s conclusion was: “At its core, this is still just a bot in Discord.”
A real-world experiment supported the criticism. Baidu opened the Tieba API, allowing the bot to launch in the Sun Xiaochuan forum, which had more than 1M users. Many people played with it initially, but the project ended after a single phase because it had no practical value. A Tsinghua computer-science professor later used the team’s customizable web chatroom in class, producing the team’s first product launch, revenue and real user feedback.

6. From a Stanford Floor to a Berkeley Hillside, Silicon Valley’s Role Was Demystification and Survival

After launching the Tsinghua project, the team tried to enter Stanford. Brandon slept on a student dormitory floor for a semester, later fell out with a friend and was briefly homeless in Palo Alto. Meanwhile, a US institutional investor was introduced through a friend and decided to invest roughly $2M after 2 meetings, giving the team the means to establish a company in the US and cover its expenses.
His reason for staying in Silicon Valley was not that “people there are stronger.” It was the opposite realization: many founders were “just like this.” They were not necessarily mature, and their understanding of internet monetization was sometimes worse than that of some random person in China. The startup stories of WhatsApp and Airbnb completed the demystification: “If they can do it, so can I.”
An investor lent him a white house overlooking the ocean in the Berkeley hills for free. Brandon might stay inside for 1 or 2 weeks, listening to Zhang Xiaolong’s public WeChat lectures, covering the walls with more than a dozen sheets of white paper and greeting the deer in the yard. Most of his systematic thinking about IM took shape during that quiet period.

7. Social Brought Early Capital; Solitude Made the Product Thesis Converge

Responding to the criticism that he was “socializing every day instead of properly building the product,” Brandon admitted that meeting people after arriving in the US in 2023 did help him obtain information, support and his first US funding. A friend heard at an event that a fund wanted to invest in AI and referred him; he became the first deal in that fund’s batch, so the fund moved relatively quickly.
Once the survival crisis was resolved and his work visa secured, he began going out less. His landlord was willing to host startup events at the Berkeley house, so as the venue manager he could still meet guests in the living room without actively traveling around. His conditional advice for zero-to-one founders is: when the direction is unclear or resources are scarce, create more collisions; once you have figured it out, put your head down, build it and then accept the world’s feedback.
Two undisclosed internet heavyweights later invested in Intent. One seemed barely interested in the feature pitch and only said, “It’s good that young people dare to take on something this big.” The other asked nothing about Tencent, Meta, market size or “why you,” and focused exclusively on how to observe users and conduct user research, making Brandon feel that the investor was first and foremost “a very good product manager.”

8. There Are Many Global IM Products, but “Helping People Communicate” Was the Anchor He Found Only This Year

The team systematically studied KakaoTalk, LINE, ShareChat, Zebra, Yalla, Zangi, Telegram, Signal and an IM product developed officially in Iran. Brandon’s early discovery was that the world does not consist only of WeChat and WhatsApp. Many products unfamiliar to Chinese users also have more than 100M users, proving that IM can carry enormous, highly regionalized traffic.
But the existence of many large IM products did not mean they had found a product. Last year the team was still building prototypes for chat summaries, automatic contact notes, AI search and personal assistants. Only this year did Brandon move the anchor from “what AI feature should we add to IM?” to “help people communicate better with one another.”
He still refuses to claim that he knows the complete form of the next-generation IM: “I don’t think I do.” The change is that the team now has a standard for judging features: do they improve receiving messages, sending messages and both sides’ actual understanding of one another, rather than merely looking sufficiently AI-driven or pointing toward the future?

9. Translation Must Use Relationships, Context and Tone to Offset the Cost of Bilateral Installation

After moving to the US, Brandon had ChatGPT process every English text message before he sent it. The difficulty was not only word meaning, but also tone and American young people’s abbreviations; when he first saw “low-key,” he did not even know what it meant. He emphasizes that even when 2 people speak the same language, they may still fail to understand each other.
The host’s challenge went straight to cold start: Sogou Input Method can convert Chinese to English, while Typeless, Aqua Voice and Wispr Flow on Mac can turn Chinese speech directly into English. None requires the other person to install a new IM app. If Intent adds another layer of installation friction, its translation experience must be materially better.
Brandon’s answer is that input methods cannot access the full context of a relationship. Japanese and Korean require a judgment between plain and honorific forms; “LLM” can mean a Master of Laws or a large language model; and what “that” refers to may depend on the conversation history. IM can also learn both sides’ language habits and interaction styles. He therefore believes only by controlling the communications tool itself can translation be made truly comprehensive.
The user who best validates this is a Vietnamese mother: her child was born in the US and speaks only English, while she speaks only Vietnamese. They live together but can communicate only about “Are you back yet? Have you eaten? Have you slept?” Brandon considers the feedback that they “clearly have so much to say” a more valuable positive signal than praise from the industry.

10. Large-Model Translation Is Often Already Very Good, Making “Better” Hard to Measure

The team’s most frequent recent discussions concern translation quality, but it has not yet found a satisfactory benchmark. Traditional NLP offers metrics based on word correspondence, word vectors and sentence vectors. Large models can already translate very well much of the time; the more they try to evaluate subtle correctness in tone, relationships and context, the harder it becomes to establish a standard.
The paradox Brandon sees is that they may have no choice but to use a more advanced model to evaluate the current model, creating a self-referential loop. Errors may also come not from model capability but from missing context. The same sentence can require completely different tones when used with family and with colleagues.
He does not believe studying translation has pulled the team away from the IM vision. Translation is fundamentally about understanding a passage and accurately transferring the information to another person, which is “similar in kind” to understanding intent. Apple and WhatsApp have added translation enhancements, but based on the team’s user research, users still do not seem able to use them well. Overseas users may also be unable to treat WeChat as an out-of-the-box tool they can use smoothly. That is precisely the space Intent wants to test.

11. The Cold-Start Strategy Is Not Mass User Acquisition, but Community-Led Replication of Seed-User Demand

Before launch, Intent spent substantial time on user research and continues to involve early users in product development. Some power users are willing to promote it proactively. The team plans to iterate until it is satisfied, then ask these users where they normally get information, because “their communities must be the same channels they use to get information.”
External criticism clusters around 4 points: building an IM from zero is impossible, translation giants can replicate the feature casually, the product is “not sexy,” and a two-sided network cannot get started. More aggressive critics dismiss the young founder’s grand narrative as a scam. Brandon offers no macro rebuttal; he reduces the question to whether specific users are willing to use the product with their families.
That means the current growth path remains narrow-scenario-driven rather than powered by an existing network effect. The team has tested at a scale of more than 1,000 users, but admits it is still short of “frictionless daily use with nothing going wrong.” It is therefore delivering relatively simple translation first rather than launching every complex demo at once.

12. The 7 Failures Were Actually 7 Attempts at Chat Products—and 7 Steps Toward User Demand

Brandon defines “one failure” as building at least one client and frontend capable of sending and receiving messages. The first was a crude web chatroom supported by a Python backend, later upgraded with a UI. Tsinghua required public-cloud stability, while Stanford brought different user habits and needs, forcing many components to be rewritten.
The team later learned app development and built its own backend and mobile clients. Toy-level products helped validate multiple prototypes. It then built LingoGram, a third-party client based on Telegram, adding chat summaries, translation and tone polishing before moving into Intent.
His summary is “two-way convergence”: on one side, learning how an IM system works; on the other, moving closer to real demand through repeated attempts and user rejection. “Start from the user’s real needs” sounds simple, but it took multiple rebuilds to turn that principle into actual product discipline.
Progress was not a linear series of failures. The early web chatroom had several thousand users, and LingoGram had tens of thousands. The team now has a large body of first-hand user research. Brandon says the internal feeling recently became that they were “getting on the right track,” while also acknowledging that many colleagues remain anxious.

13. 2 Easier-to-Monetize Products Proved His Ability but Did Not Change the Company’s Mission

When the Stanford project failed and the team fell apart, Brandon built an AI document-organization tool to solve his own need while applying for a work visa. After launch, overseas users proactively wrote in asking whether they could pay for a business-plan version. It was the team’s first clear validation from an overseas market.
The team’s first mobile app was a menu-translation tool written by the co-founder over a single weekend. Understanding the literal meaning of “couple’s lung slices” or “four cheese” still does not tell you what kind of dish it is. The product later made a lot of money, and an explainer video received more than 1M views on TikTok.
Brandon once told investors that building a US B2B SaaS company could at least earn back the investment. They replied: “We don’t care whether you have $1M or several million dollars in ARR.” Their original investment was because he wanted to build a product better than WeChat, not a safe business with a limited ceiling.

14. IM’s Frequency and Effortless Record-Keeping Make It an Important Data Entry Point

Brandon’s frequency comparison: WhatsApp MAU send roughly 60 messages a day, ChatGPT MAU ask roughly 10 questions a day, perhaps more recently; Google Search gets roughly 3 searches; Instagram sees fewer than 0.1 posts per user per day, with roughly 0.3–0.4 Stories. IM gives users a much higher density of voluntary expression than most To C products.
The reason is not only urgent need, but the continuous external triggers created by a social setting. Users still have to formulate a question for ChatGPT, whereas with a friend they will “unconsciously bring out many things” while responding. Brandon believes that if he dictated to ChatGPT, he would never generate as much information as he does in a conversation with the host.
These conversations also create a life archive with almost no additional labor. People find it difficult to write a diary every day or tag what they did each hour, but they keep chatting. Language habits, relationship status and gifts sent last year may all be buried in IM.
Existing search has not unlocked the value of these records. Someone may write “Wednesday,” while a user searching for “Wed” may never find it. Brandon has no complete answer yet, but believes IM is currently the platform that records the most of a person’s digital life with the least effort. Better understanding and retrieval therefore represent a long-term opportunity.

15. The Long-Term Product Is Not 10 More Toggles, but Closing the Loop on Intent Inside the Conversation

Intent’s vision is to “understand you extremely well”: help users send and receive messages correctly and resolve the intent carried by those messages. If 2 people say, “Let’s order delivery,” the system completes the order with authorization. If a user receives something they cannot understand, they tap once for an explanation and continue the discussion in the original thread.
Other ideas include recommending people who are appropriate to talk to at that moment, automatically organizing shopping lists, comparing Uber and Lyft, reducing manual grouping and contact notes, managing unknown numbers and storage, and preventing users from trying 7 or 8 keywords repeatedly to find one old message.
Brandon believes WeChat is capable of building many of these features, but its 1B-user installed base is also a constraint: “Any tiny change affects the daily lives of 1B people.” Intent’s disadvantage is having no users; its corresponding advantage is a blank sheet of paper, allowing it to redesign interactions without maintaining existing habits.
His reservation about mini-programs is that dynamic cards may already be close to an excellent answer, but clicking one moves users to another page and “your brain suddenly starts buzzing,” so they may even forget what they were trying to do. Intent wants to keep actions inside the conversation flow, while acknowledging that its own approach could also be wrong.

16. The Biggest Product Risk of Proactive AI Is Not Insufficient Capability, but Disruption After Misreading Intent

Brandon compares intent recognition with recommendation algorithms. Early Toutiao was not as good as today’s Douyin, but it had to establish evaluation metrics through real use, exploring step by step with early adopters. The team cannot simply imagine in a lab what the correct behavior should be.
The host’s rebuttal is worth preserving: when a recommendation is wrong, the user can swipe it away; if ride-hailing or food-ordering cards keep popping up in a chat, the interruption cost is far higher than seeing one unwanted video. Whether proactive AI helps or interrupts cannot be waved away with “the model will get smarter.”
Brandon’s candid answer is that there is no particularly good solution yet. The team can only start from user experience and make the feature more restrained. “No credit is better than taking credit”: Intent must first be a stable, reliable communications tool. Before that, no additional value is worth compromising message delivery.
He agrees with WeChat-style decentralized discovery: if a friend has not shared something, users may not even find the entry point to a feature. Ultimately, both higher model accuracy and continued interaction-design exploration are needed to reduce disruption, but the only way to determine which prompt works is to test it in real scenarios.

17. Overseas Fragmentation, Language Disadvantage and Small-Team Learning Speed Define the Bet

Brandon believes a new IM is more likely to emerge overseas than in China. China has WeChat, an exceptionally powerful product, making the opportunity for startups “quite slim.” Overseas relationship graphs are more heavily embedded in phone address books. WhatsApp has just over 100M users in the US, Messenger may account for roughly 50%, and iMessage is used mainly in relatively affluent regions; no single product has absolute dominance.
He even offers a conditional competitive judgment: if Meta’s Family of Apps is led by Alexandr Wang in the future, that would be “a major positive for Intent.” This reflects Brandon’s negative assessment of the innovation capabilities of Meta’s current products, though the program does not develop the argument further.
One advantage of his Chinese background is that friends can continuously supply information about China’s internet industry. Brandon believes China’s internet is “significantly better than the US at operations and monetization.” Silicon Valley’s difference is that founders face less interference from people telling them how to do things and can make decisions and bear the consequences themselves. But serving overseas users is the primary reason the team remains there.
A more direct source of founder-market fit is language disadvantage. He ranked at the lowest level in Tsinghua’s English placement system, did not know how to use a credit card after arriving in the US, and could not name Subway’s bread options, remembering only to say “no pickles.” This turned “defamiliarization” into an observation tool, and converted his greatest disadvantage into direct intuition for the pain of cross-language communication.
The company maintains a “small and sharp” team of 20. The co-founder handles technical management and forward-looking exploration while also participating in product strategy. Brandon points to Telegram’s 1B MAU with roughly 30 developers, and WhatsApp’s 450M MAU and roughly 35 employees when it was acquired, to show that he does not plan to build IM by piling on headcount.
The hiring criteria, in order, are belief in the mission of communication, strong learning ability, and sincerity and kindness. Designers learn to write code; a product colleague with a psychology background uses GitHub to manage PRDs; Brandon, whose background is in biology, learns “almost everything.” The team wants to help members reduce the burden of living costs and then reward them through long-term shared outcomes.
On the outcome, he deliberately keeps expectations low: “I don’t have a PM; the company keeps losing money; I can’t build an IM… We’ve failed 7 times, and the 8th could fail too.” If it succeeds, the only core reason will be understanding user scenarios and solving them with a good product. If it fails, the likeliest cause will be insufficient user understanding, followed by Brandon’s own and the organization’s overall execution capability.
His conclusion for university students follows the same logic: GPA, graduate-school recommendations, final exams and “I can only start after I learn it” are not the real constraints. With ChatGPT, YouTube and rapid-development tools, you can build something in a weekend and have users “slap you in the face” the following week. The shortest path to learning remains: do exactly what you want to do.

Robot Special: Hardware and Algorithms Behind Dexterous Hands The Unicorn Founder Who Delegated Everything.