3-Hour Interview: YouWare Founder 明超平 on Fire-Stick Gorilla Agents
Summary
明超平 is betting that Coding will evolve from a niche programming skill into a mass-market creative medium, and that YouWare should become “the YouTube and Instagram of the Coding era.” His analogy is that the phone camera turned photography into casual picture-taking: model companies such as Anthropic may play Sony, continuously supplying more powerful “modules,” while the real application opportunities emerge in the Instagram, Snapchat and TikTok layer. For now, YouWare combines tools and community so that when users see quizzes, mathematical visualizations, games, résumés and websites, they immediately understand what they themselves can create.
His test for an AI-native product is demanding: across jumps from GPT-3.5 to 3.7, 4.0 and 5.0, if the product does not deliver a 5x or 10x experience improvement, it is probably not meaningfully tied to the era’s biggest variable. Early in his startup journey, however, he repeated the bitter lesson: to compensate for model instability, he stuffed Figma and Notion workflows with templates, CSS, JS and crawlers, turning “80% AI” into “20% AI.” After the Lunar New Year, he proactively killed a product that had not yet launched, returned to a more open Coding Agent, and rebuilt the old capabilities more elegantly through MCP servers and tools.
His measure of AI application value has shifted from “the more tokens consumed, the better” to per-token valuation: how much useful value each token ultimately creates. A chatbot must be triggered by the user one turn at a time; Deep Research can expand one prompt into roughly 100 tasks; and an Agent network may let multiple Agents consume intelligence together. But “writing furiously” does not equal value. The sustainable direction is improving the conversion of tokens into economic outcomes. Cancer cures and the Riemann hypothesis may justify unlimited energy consumption; ordinary consumer tasks must gradually become more efficient.
To him, a “wrapper” is not a thin layer around a model but the container and environment that shape user behavior; giving users nothing beyond an empty prompt box is even “irresponsible.” Midjourney’s breakthrough was not just the model, but Discord, where users could view one another’s work, copy prompts and stack ideas, eventually evolving from “a dog” into descriptions hundreds of words long. The environment itself provided education, distribution and network effects. YouWare therefore treats community as the environment in which creator content, experience and demand accumulate, before distilling them into tools and Agent capabilities.
He objectively leans toward a future ecosystem in which multiple vertical Agents are orchestrated, though personally he still wants to become the super-portal. If a super-portal turns PageRank into Agent Rank, many consumer companies could be compressed into callable B2B APIs and provoke resistance from the ecosystem. Vertical Agents would first need to enter the invocation network, then earn ranking through being “the best” or delivering 80-point performance at half the cost. YouWare accepts that it must first become an orchestratable node, while retaining the ambition to become an OS Agent.
He believes today’s Agents are nowhere near their breakout, more like “a gorilla that has just picked up a fire stick”; the real inflection point comes when large numbers of Agents gain identity, permissions and controllable context and begin collaborating. Agent IDs, password locks, trust, bidding and tiered authorization will emerge; humans will look more like directors and judges, using taste and aesthetics to drive repeated iteration among multiple Agents. Foundation-model companies are “building smarter people,” while application companies give those people environments, experience and professional skills. In the near term, startups have no reason to imagine they can redesign base intelligence.
Dynamically generated UI is his long-term path toward the next OS: fixed apps and fixed workflows may give way to interfaces generated in real time for each task, device and context. In his experience, phones and computers cover only about 20% of the context in daily life, while a phone’s sensors spend most of their time in a pocket. The value of new hardware is therefore not to “replace the phone,” but to capture life context that the giants cannot access. He dismisses claims that “glasses will replace phones” as “complete nonsense.” The path may take 10 or 20 years; for now, the practical task is to build a product, establish a company and survive.
The company was founded on October 10, 2024, launched YouWare in March 2025, and had 20 employees and losses at the time of the interview; 明超平 said half the team was in R&D, described the financing as an angel-plus round, and answered the valuation question with “$8,000.” His base case for the team is not victory but “a 99.9% chance of dying,” followed by a daily search for ways not to die. Major decisions must surface dissent, and investors serve as a “mirror” reflecting hesitation, wavering and ego. His core bet ultimately reduces to two lines: Coding will become the creative medium of a new era, and dynamic OS may create a window to break through Apple and Google.
Deep dive
1. From Product Manager to CEO of a Creative Agent
明超平 was born in 1995 and was 29 at the time of the interview. He worked at OnePlus from 2018 to 2020, at ByteDance’s Jianying/CapCut from 2020 to 2023, and joined Moonshot in September 2023. After roughly a year working on Kimi, he left in September 2024 to start a company.
The company is called Xingye in China and YouWare overseas. It was founded on October 10, 2024, and launched YouWare in March 2025. It was in an angel-plus financing round and still losing money. Asked about valuation, he answered “$8,000”; the interview did not elaborate.
This is his first CEO role. Later in the interview he said the company had 20 people, roughly half in R&D, plus 2 people in product and operations and an HR lead. The product is not a general-purpose Agent, but a vertical Coding Agent for creative expression.
2. A Personality Shaped by Freedom and Hard Boundaries
明超平 was raised by his grandmother and was rarely controlled in his grades or day-to-day choices, leaving him “freewheeling” for much of his early life. His grandmother was strict about manners: everyone had to be present before eating, he could take only food in front of him, and certain lines in speech and conduct could not be crossed. “You cannot cross the lines; beyond that, figure things out yourself.”
In middle school, a good teacher led him to study seriously “to repay the teacher.” His grades fell after he began dating in high school, then improved again because his girlfriend was a strong student. He never presented himself as a top student, only as “middle of the pack.” He chose Wuhan University partly because it was “free and romantic.”
3. Debate Made Him Assume Every View Has an Opposite
As an automation student at Wuhan University, he spent most of his freshman and sophomore years debating, at peak intensity entering 14 debates in a week. The 2-year training left him not with rhetorical tricks but with the assumption that no view is permanently correct: everyone uses a different model to understand the world, and while objective truth may exist, people can only approach it indefinitely from multiple subjective angles.
He brought that habit into the company. Product data and market facts matter, but the team must also articulate a view, because information without a view has “no consumption value and no discussion value.” At the same time, “always know that your view is wrong” and actively search for its opposite form.
张小珺 asked how a company can make decisions if every view has both a positive and negative side. 明超平 did not pretend to have a scientific formula: the final answer is still preference and intuition, but that intuition must return to the user scenario and be corrected by data and opposing views.
4. “Users Above Ego” Is the Final Constraint on Intuition
The team motto is “users above ego.” It is not enough for someone to explain the data; the real question is whether they have spoken with a user for an hour, spent every day in the Discord server, and watched how users create, share and respond to comments.
Everyone puts their view on the table first, then returns to real-world scenarios to find intuitive consensus. Most of the time, 明超平 still makes the final call because the team trusts his product intuition, but that authority must come from sustained contact with users, not title or rhetorical advantage.
5. Fourth Speaker Training Is About the Third-Party View
明超平 most often debated as the fourth speaker. The fourth speaker goes last and can reverse the outcome. He cited star debaters whose first 40 minutes might be poorly argued, only for the final 3 minutes to use an emotional story that made judges forget the earlier arguments. He was not claiming that this described his own performance.
The fourth speaker’s real audience is not a thoroughly prepared opponent who may not be persuadable, but the judges and audience hearing the topic for the first time—“in essence, your users.”
This became a product principle: be able to “turn yourself into an idiot in one second.” Users do not know the company’s story or have the manual. If opening the product still requires a toast or pop-up explaining how to use it, the product has not achieved natural, out-of-the-box usability.
He sees Apple as the archetype of this product philosophy: it does not rely on manuals or treat pixels, frame rates and clock speed as the experience itself. Simplicity is not a visual style; it means a third party can understand the value without preparation.
6. Autonomous-Car Racing Gave Him Engineering Intuition and a Growth Mindset
In his junior and senior years, he joined a 4-person autonomous-car racing team and served as captain, responsible for software. The 2 cars had to use cameras to identify the track while communicating, avoiding obstacles and coordinating overtaking. He often slept in the lab for 2 years, working hands-on with C, C++, sensors, chips and hardware-software coordination.
Improving the car from 20 seconds to 19 seconds provided immediate feedback on every line of code and every bug, creating an exceptionally clear reward model. “It wasn’t like research; it was like playing a game.” Coding, soldering boards and coordination all served a visible improvement in speed.
The longer-lasting lesson was that “there is nothing you cannot learn.” Bluetooth, cameras and control algorithms could all be filled in through research and the right people. From then on, his fear of difficult problems fell close to zero.
7. The Best Debater Can Also Have the Most Losses
明超平 was once both the team’s most frequent best-debater winner and its biggest loser. He would rush to stand up during questioning and open debate, appearing to shield teammates under pressure, but the behavior also contained arrogance: he took away their chance to speak and escape traps, making himself visible while weakening team coordination.
In a senior’s retirement match, he talked the team out of a game they had been leading. They had accumulated advantages at every stage, but in the final 3 minutes he tried to repeat every winning point, accelerated his speech and lost the main line. “The more you say, the less information the audience absorbs.”
He maps this directly onto products. Judges usually have only 1 or 2 questions, and users often need only 3 functions. Roughly 90% of features in most software are never opened; adding 100 features only obscures the real point of differentiation.
8. No Longer Needing to Persuade Others Made Him More Cautious About UGC
Two years of debate pushed him from a strongly extroverted personality toward introversion. After talking too much, he realized that persuading others was neither necessary nor easy because everyone has a different world model. Once changing other people stopped being the goal, he naturally spoke less.
The same restraint extends to community governance. Strange, suggestive, funny or aesthetically unappealing content may be popular with users. He does not want to suppress it simply because he dislikes it: “I don’t judge the people around me, and I won’t judge users’ tastes and preferences.”
9. OnePlus Trained His Product Sense in a Data-Poor Environment
After reading Steve Jobs’ biography, he switched from automation to product management, confident that “understanding technology should make me stronger than an average product manager.” At OnePlus he discovered that the organization emphasized product sense without being able to define or teach it.
The phone company’s data infrastructure was weak. Early on, it did not even have complete analytics, with data kept in Excel. When experiments were difficult to run, product managers had to rely on subjective judgment, user feedback and direct observation, which forced him to train his intuition.
A mentor he called a “genius product manager” took new hires through an entire subway ride and shopping-mall visit, observing battery levels, earbud usage, app switching and rush-hour behavior. Heavy smartphone users in the office were not the whole market: “Your users are on the subway and in the mall.”
10. Battery Percentage Is a Psychological Experience Curve, Not a Linear Number
The team initially assumed people leaving home in the morning should be near 100% battery. In the field, they found morning battery levels could be as low as 10%, with the phone already showing red. Without complete data, they at least learned that users’ real state could be the opposite of the conference-room assumption.
More importantly, battery specifications and battery experience are not equivalent. Users are most sensitive at 95%-100% and 0%-5%. A drop from 100% to 99% immediately after unplugging can make them suspect the battery, while an unusually durable final 5% can ease the most anxious moment.
The ideal display may therefore not decline linearly but follow an S-shaped curve; the final displayed 5% could correspond to a real 10%. His point was not to encourage fakery, but that “experience is both subjective and objective.” Products serve human perception, not just a paper specification of 4,000 or 5,000 mAh.
11. ByteDance Turned Product Judgment into Muscle Memory Through Fast Feedback
His immediate reason for leaving OnePlus was one phone per year, with responsibility for only one module and feedback arriving too slowly. Jianying shipped a version every 2 weeks, with up to 10 features and 5 experiments per release. Over 3 years, he accumulated dozens or even hundreds of cycles of judgment and correction.
When he joined ByteDance, he did not understand DAU, VV, MAU, PV, UV, or 7-day and 30-day retention. He felt “terrible.” Only after learning the metrics and their relationships did he enter ByteDance’s real product operating system.
High-frequency experiments eventually formed data intuition. When a proposal appeared, he could roughly judge whether metrics would rise or fall and within what range even without an A/B test. Those hundreds of prior-posterior loops were, in his view, ByteDance’s strongest training asset.
12. Data-Driven Processes Raise the Average but Cannot Validate Disruptive Innovation
He compresses the ByteDance method into a chain: write down a clear judgment and expected result; execute with the roughest, fastest MVP—“if you can avoid doing it, don’t do it; if you can avoid doing it today, don’t do it today”—then compare data and user feedback against the prior and review the gap.
This infrastructure can help a 60-point newcomer deliver an 80-point result within 1 or 2 weeks. If an experiment is down 20%, even an inexperienced person will not casually roll it out. Google and Facebook used similar systems to help people from different cultures and backgrounds deliver consistently.
Asked about the weakness, 明超平 admitted that the method can grind down the best ideas. Disruptive products such as the iPhone and Tesla cannot first be validated by A/B testing. Data is only a “rearview mirror”: it clarifies the turns already taken but cannot tell the driver where to go next.
13. When Intuition and Data Conflict, He Ultimately Chooses Intuition
The team now uses views and intuition to determine direction, then validates through rapid experiments and data. But short-term growth does not automatically equal long-term value, and short-term decline does not automatically mean a decision was wrong. Some decisions must be treated as “faith judgments,” with the CEO carrying the risk.
When the 2 methods genuinely collide, he openly says he “trusts intuition.” Most important decisions work this way. OnePlus gave him simplicity and subjective experience; ByteDance gave him scalable experimentation and a higher floor. He is trying to put both into the same decision loop.
He especially admires 张一鸣 for continuing to discuss tiny startup projects firsthand. Some extremely hungry teams discover that 张一鸣 has also spoken with most of the projects they have seen. “Someone at that level is still talking to projects like these” reflects rare humility and hands-on ability, in his view.
14. Inside a Powerful Platform, Choosing the Right Direction May Account for 90% of the Outcome
He largely agrees with the criticism that ByteDance has produced many entrepreneurs but no universally acknowledged breakout success. ByteDance’s infrastructure is so strong that it can obscure an individual’s ability to operate alone and make platform achievements look like personal ability.
His retrospective conclusion is that “the decision accounts for 90% of the value.” Whether to join a new team in 2019 or 2020, whether to join Kimi in September 2023, and whether to stick with video editing despite doubts about its ceiling mattered more than execution after entering the right environment.
As long as someone is reliable and hardworking, a strong system makes very poor delivery unlikely. What differentiates entrepreneurs is whether they can spot weak trends and place a bet before entering the system—not whether they later attribute scale to their own ego.
15. He Left ByteDance to Give His Scarcest Time to AI
He resigned in July 2023 before studying entrepreneurship and AI. The reason was not that he could explore opportunities while working, but that he could not accept “drawing a salary here while studying my own thing.” If he stayed at Jianying, 90% of his energy would still go to meetings, reports and biweekly materials.
After cycling in Japan in August, he spoke with 杨植麟 from noon until evening in September. They mainly discussed life, music, art, skiing and personal experiences; only over pizza did they formally discuss technology. Among the teams he had encountered, 杨植麟’s was the only one that could explain token prediction and “compression is intelligence” clearly enough for him to follow while eating.
The key analogy he retained was E=mc²: enormous information compressed into a simple formula is intelligence. The Feynman technique likewise compresses a complex system into one sentence that another person can understand. 杨植麟’s fixation on AGI and the “light in his eyes” when discussing Geoffrey Hinton eventually led him to join Moonshot.
16. Moonshot Gave Him Near-Complete Product Freedom
On his first day at Moonshot, its 3 co-founders told him, “Please, please, help us do To C.” The team let him place the overseas product group in Shenzhen; many colleagues did not even know the Shenzhen base existed. Even as a CEO today, he doubts he could reproduce that level of trust.
He learned in both directions with the team. Every 2 weeks in one-on-ones, he shared product methods while the other person explained papers and technology in the simplest possible language. Team members even believed that if they could not explain something clearly, the responsibility was theirs, not that the product manager was “too stupid.”
Late in his time at Kimi, he also worked on plugins and early versions of Computer Use and Agent exploration. The plugin unexpectedly spread, but the more important outcome was learning which things an AI product should restrain itself from doing.
17. Noisee Emerged from a Sharing Gap Among Suno Users
At the end of 2023, before Suno broke out, he spent long periods in Discord and saw that music creators had a much stronger desire to communicate and share than users of ordinary AI tools. People who could not compose were generating multiple songs a day; those works could not stay forever on a phone or inside a channel.
He observed that traditional musicians often uploaded music to YouTube with a single static cover image for 2 minutes. At the same time, video models had poor character and style consistency, making continuous stories unworkable but making trailers and MVs—where consistency was unnecessary—a good fit.
That led to Noisee: Suno let ordinary people generate music, while Noisee turned the link into a complete, shareable MV. Instead of trying to eliminate the model’s weakness, the team repriced it through a content category where “consistency is unimportant, and jump cuts add imagination.”
18. Manual Editing Was Noisee’s Most Effective Demand Test
Before the product existed, the team posed as independent developers, found users’ work in Suno Discord, manually edited it into MVs with Jianying, and asked, “Do you like it?” The next day, 10 users requested links, so 2 interns edited 10 clips and returned them one by one.
Discord generally prohibited posting one’s own links in another person’s channel, so he never dared to drive traffic proactively. But administrators saw that users genuinely liked the work and later posted Noisee’s product link themselves, bringing the first seed users.
The case reinforced his third-party perspective: instead of building a complex system first and then proving demand, use existing tools to deliver the finished work users actually want to share, and let community administrators and creators provide the signal themselves.
19. Suno v3’s Breakout Pulled a Complementary Product Along With It
In March 2024, Suno released v3 and, according to 明超平, quickly reached several million DAU. “Suno makes the music—Noisee makes the MV—then upload to YouTube” formed a natural combination, allowing Noisee to break out on Suno’s growth.
Overseas VCs, Warner Music, Spotify and Suno all made contact. The VC he named was HSG. Suno spoke with the team twice, and he guessed it might have acquisition intentions. The Information later disclosed that Noisee was backed by Chinese foundation-model company Moonshot, after which discussions did not continue.
He emphasized that users called it “Sora-like” not because the model was stronger, but because the delivery was more complete. Others offered 4- or 5-second clips; Noisee delivered a 1-minute MV with transitions, music and visual rhythm. “What we need to deliver to users is a result.” Adding sound to video improved the experience by far more than 2x or 3x.
20. Shutting Down Noisee Concentrated Resources After Its Non-Consensus Edge Disappeared
Kimi broke out during the same period, and the company decided to concentrate resources. 杨植麟 and others were unusually indirect on the phone because 明超平 had led the product. He asked directly, “Are we going to shut it down?” and then said he “firmly supported” the decision, not viewing it as a forced sacrifice.
张小珺 called it a shame. 明超平 instead thought the product not only should have been shut down, but perhaps should never have been built. Multimodal understanding is related to intelligence, but he believes “multimodal generation does not produce intelligence,” putting it further from the company’s core dream. That was his judgment, not a correction of the facts.
More importantly, Sora had already shown the leading path, and Kling later produced a good model as well, erasing the original non-consensus. Unless a founder can pursue a different route, such as autoregressive generation, to deliver lower cost, greater speed or better quality, repeating the same path only burns resources. What truly saddened him was discovering one day after leaving that the domain had finally stopped opening altogether.
21. The Hardest AI Product Skill Is Knowing When Not to Polish
The principle he learned at Moonshot was that traditional internet work on UI, UX, templates and scaffolding can create short-term improvements worth pursuing. In the AI era, however, many things should be left alone because the next model upgrade may erase the engineering entirely.
After starting his company, he repeated the mistake. Models generated websites and games that were not beautiful and occasionally errored, so he had designers prepare roughly 50 CSS and JS templates to stuff into context, hoping to force the output to be attractive and stable.
The product became more like an engineering system than an AI system. The bitter lesson was that a founder’s most familiar traditional product skills can become path dependence that prevents the model from releasing its capabilities.
22. His First Startup Intuition Was Right, but Six Months of Chasing Stability Led Him Astray
After leaving Moonshot in September 2024, he formally began approaching institutional financing, partly to avoid having his departure and fundraising affect Kimi while it was still at the center of attention. The company was founded on October 10, 2024, and team members joined after understanding the risks, allocation and direction.
The company already had a working demo during fundraising, with a form close to the later YouWare: users created websites, games and other works in natural language, then interacted, shared and remixed them in a community. The original BP already described this vision.
But the models were uncontrollable and unstable. He suspected the timing was wrong and switched to vertical products such as Figma-to-prototype and Notion-to-landing-page. Crawling file structures and handling large numbers of corner cases turned the product from AI into engineering, creating a 6-month detour.
23. MCP Made Complex Vertical Capabilities Natural Again
Implementing Figma and Notion conversion had required separate crawlers, structure parsing and edge-case handling. After returning to YouWare, the Agent could call MCP servers and tools to retrieve data, then reconstruct it through Coding, with one framework covering multiple sources.
The same functionality reappeared, but the mechanism shifted from hard-coded scenarios to Agent orchestration. “Very elegant” was an important signal to him that the direction might be right: complexity had moved into general capabilities rather than continuing to pile up in product engineering.
24. Can a Model Upgrade Amplify the Product? That Is the AI-Native Test
Using jumps from GPT-3.5 to 3.7, 3.7 to 4.0 and 4.0 to 5.0, he asks teams how much the product experience actually improves every time the base model gets stronger. If the change is small rather than 5x or 10x, the product is probably not AI-native.
The point is not to chase a label, but to align the company with the era’s biggest variable. ByteDance was built on the combination of mobile devices, bandwidth and recommendation algorithms. One of today’s biggest variables is model intelligence; if it has little to do with your product, you are already far from becoming a large or great company.
After the 2025 Lunar New Year, he repeatedly briefed shareholders on progress. In the process of saying the same things over and over, he suddenly realized the product was wrong. Even as the team was enthusiastically discussing the next step, he stopped the unreleased plan and admitted that the traditional product manager’s instinct to “polish” had overwhelmed his original intuition.
25. A Product Built Overnight Solved One Break in the Chain: Works Could Not Be Shared
Around 10 p.m. one night in March 2025, he was scrolling through Twitter and saw many people using Grok 3 to write games but only sharing screen recordings. Interactive works of the new era were incompatible with old social media and needed a new container that could run, be experienced and be shared directly.
The first version, completed around 1 or 2 a.m., was only “HTML to website.” Users could paste code generated by DeepSeek or ChatGPT—code that previously ran only on localhost—and immediately get a website friends could access. He thought the product was too rough and, when recommending it publicly, did not even admit he had made it.
The minimalist launch reconnected with the original premise: solve the creator’s real sharing bottleneck first, then explore Agents, community and creative capabilities instead of assuming a complete workflow upfront.
26. Token Consumption Was the First-Stage Metric; the Next Stage Is Unit Value
Another question he asked himself after the Lunar New Year was whether AI products should track DAU, retention or a new metric. His initial answer was the speed of token consumption because it reflected the scale at which a product called intelligence and completed tasks. Later, after watching interviews and talks by the Manus team, he felt the conclusion was “uncannily aligned.”
After using multiple Agents, he revised the metric to per-token valuation. Writing a website with 1M tokens is not better if most of them go to ineffective trial and error. Like oil or electricity being converted into mechanical work, mature systems must pursue higher conversion efficiency.
He does not treat efficiency as absolute. If an Agent can cure cancer or solve the Riemann hypothesis, the number of tokens—or even “half the battery”—does not matter. But most ordinary user tasks do not have such terminal value, so cost and useful output ultimately must be optimized.
This also explains the stages of AI products: first pursue speed and let Agents such as Devin “write furiously”; once scale forms, determine whether 30% or 70% of those tokens actually created value.
27. The Through Line from Chatbot to Agent Is Expanding One Trigger into Sustained Work
Early base-model products mainly completed copy and marketing emails, with token consumption constrained by low-frequency use cases. Instruction models and ChatGPT turned the model into a general searchable, question-answering dialogue system, broadening the range of use.
A chatbot triggers the model once per user prompt. Deep Research can expand one prompt into roughly 100 subtasks and work at nearly 100x the scale. Manus, Devin and today’s many Agents, including Lovable, continue along the same line.
NotebookLM showed him that applications do not need to consume only text tokens; text, images, video and audio can enter the same workflow. A2A and Agent networks take the next step, allowing multiple executors to collaborate on harder problems.
28. The Value of the “Wrapper” Is Shaping Behavior, Not Hiding the Model
明超平 rejects dismissing the application layer as a “wrapper,” preferring to call it an environment and container. The same person changes their volume, movements and speech in a street-food stall versus a French restaurant; the environment itself changes behavior.
Giving users only a chatbot input box is therefore “irresponsible.” Creative users often do not know why they should use the product or what to enter, and the product cannot push the entire burden of exploring the model’s boundaries onto them.
YouWare’s input box, work feed and remix paths jointly provide guidance. The goal is not decorative complexity, but to give users creative motivation upon entry, provide inspiration and teach more advanced expression through other people’s work.
29. Midjourney Proved That Community Itself Can Be the User Manual
When he first encountered Midjourney at the end of 2022, most users could enter only “a dog,” “a cat” or “a girl.” If everyone received only an isolated tool box, overall usage might still be at that level 3 years later.
Discord let one person’s corgi, sunglasses and cape inspire another to switch to a border collie. Terms such as “surrealism,” lighting and image quality were copied into the next prompt. Short phrases evolved into descriptions hundreds of words long, with creativity and skill accumulating through public imitation.
That is the network 明超平 wants to replicate. Community does not merely display results; it helps users understand the capability boundary. High-quality work attracts better designers and creators, whose experience can then be distilled into the tools and behaviors required by Agents.
30. Coding Should Be Like Taking Pictures, Not Remain Photography
He deliberately defines YouWare around “creation,” not the more technical notion of Coding. Photography evokes DSLRs, ISO, shutter speed and professional equipment; taking pictures is something done casually. Traditional Coding is photography. AI Coding should become the picture-taking activity anyone can do.
The phone camera turned people who did not understand shutter speed into “mobile photographers,” while natural-language programming is creating web coders who did not previously exist. The important signal is not only higher productivity among existing programmers, but the rapid growth of “people who could not do this before and can do it today.”
After cameras emerged, daily photo volume rose from millions and tens of millions to hundreds of millions. He expects code works to expand exponentially as well. Coding is therefore not simply a tool upgrade for programmers, but a transformation in creative supply, creator composition and distribution.
31. Model Companies May Be Sony; Platform Opportunities Sit Downstream in the New Medium
During the mobile-photography era, Sony continued supplying camera modules. As pixels, noise and white balance improved, Instagram first optimized images with filters, then hosted massive volumes through the feed, followed by Snapchat, TikTok and TikTok Live.
In today’s analogy, Anthropic may play Sony, continuously strengthening Coding capability. If startups all compete around the camera itself, they will miss the new platforms and social relationships that emerge as supply explodes.
YouWare’s bet is to find the YouTube or Instagram of the Coding era: reorganize text, images, video and interactive software, and solve where creation, consumption, remixing and distribution happen, rather than build another foundation model.
32. Tools and Community Are Only the First Stage; the Real Content Category Remains Unknown
The practical path is to place tools and community together, giving web coders a simple, effective creative platform that continuously provides inspiration. At the same time, the team is watching where the creative process gets stuck and whether users ultimately make websites, games, apps or something else.
He wants growth to come from good content, entertaining content and word of mouth, so the team must find a viral category like early Douyin’s lip-sync videos. Short video initially meant dancing, singing and hitting the beat; only later did it absorb quantum physics. New media boundaries also expand outward from the high-sharing use cases of the early period.
There is no answer yet. He openly says there may be no “lip-sync moment” at all, and the team can only keep observing. But the diversity of overseas user creations is already sufficient to make continued search worthwhile.
33. YouWare’s Aha Moment Is Showing Users, “So This Is What You Can Create”
After entering, users see an English teacher making quizzes, a math teacher building trigonometric-function visualizations, and others making landing pages, games, personal websites, portfolios and résumé sites. Many users’ first reaction is not how powerful the model is, but “I didn’t know so many people were creating things like this here.”
张小珺 questioned whether YouWare was too broad because YouTube and Instagram each impose constraints on their medium. 明超平’s answer was to observe real supply rather than commit to a category prematurely: Agents solve creative pain points, while content reveals the forms with the strongest distribution.
He is also not committed to building an “AI Douyin.” Next-generation content should improve information density and ease of consumption, but may no longer be called short video. If the Coding Agent becomes smart enough, it might even propose ideas in conversation while the user simply answers yes or no like a director.
34. A Creative Agent Must Balance Efficiency with the Joy of Building with LEGO
He defines YouWare as a vertical Agent for creation because he cannot build a general-purpose Agent, his capabilities are limited, and that path is outside his long-term interests. His experience at Jianying, along with his preference for documentaries and art films, makes him more focused on expressing ideas without friction.
He is not pursuing a pure efficiency tool. If a fully assembled LEGO set is handed over, 80% of the value disappears; the real value is the 2-hour flow state. The ideal product lowers the professional barrier while preserving the pleasure of assembling something yourself.
Image, video, audio and text-generation tools may all be called by a Coding Agent in the future and combined into interactive websites or content. Whether to add a capability depends on user demand and valuable use cases, not on the need to claim the product is general-purpose.
His starting point is simple: Coding is an advanced capability held by only a few people. Helping an unfamiliar shop owner use Jianying or giving an ordinary person a personal website creates the satisfaction that “the product really helped someone I don’t know.”
35. Super-Agents and Vertical Agents Correspond to Two Political Structures
He uses human society to understand Agents: one highly intelligent center that dispatches all resources resembles a system optimizing for efficiency and upside; multiple actors that check and replace one another sacrifice efficiency but protect the downside.
In the first future, OpenAI or Gemini could function like an OS, understand intents such as “build a website,” “make a design” or “book a hotel,” and then find the right Agent. In the second, each field would have its own entrance and users would interact directly with multiple vertical Agents.
Personally, he naturally chooses the first and wants to be the dispatcher. Objectively, he leans toward the second because different categories will form their own orchestration layers. It is unlikely that only one top-level entrance remains, though the hardware layer may still have an Apple-like overall portal.
张小珺’s challenge hits the commercial structure: if PageRank becomes Agent Rank, every consumer company could be compressed into a B2B API. 明超平 believes that would trigger resistance—founders will not willingly lose knowledge of their users, take orders only by token and hand the relationship entirely to an upstream platform.
36. Agent Rank Will Reallocate Traffic, Profit and Corporate Identity
In the super-portal path, choosing Lovable over another Agent for design, or YouWare A over YouWare B for website creation, creates Agent Rank. Ranking could be based on performance, cost, click-through rate, conversion, consumption and advertising.
A vertical Agent must first enter the invocation network; otherwise, “your value is not enough.” Once inside, it can occupy 1 of 2 main positions: best-in-category performance, or 80-point performance at half the cost.
He sees professional experience as a moat. Between a Tsinghua Yao Class graduate and himself, if the task is producing a design draft, users might choose the person with 10 years of design experience because he can deliver faster and cheaper. Agents can likewise use data, behavior, user feedback and know-how to offset differences in general intelligence with experience.
37. Today’s Agents Have Just Learned to Use Stones and Sticks
明超平 says Agents are “like a gorilla just starting to pick up a fire stick and smash things,” roughly at the stage of Homo erectus learning to use stones, wood and fire. Discussion of a mature Agent society is far ahead of actual progress.
His product principle is to believe 2 things at once: “the model will improve,” and “the model has nothing to do with you.” Foundation-model companies advance base intelligence; application companies accumulate environments, experience, tools and users, giving Agents professional skills.
In the near term, startups do not need or have any reason to optimize around the base model. That is not because the model is unimportant, but because startups have limited resources. He compares it with Apple: the MacBook used Intel chips for more than 10 years before Apple had enough resources and accumulation to move to the M series.
38. Foundation-Model Companies Build “People”; Application Companies Train Professional Experience
He describes model companies as “building smarter people,” raising IQ from elementary school through middle school, university and a PhD. Application companies place those people in specific environments and use experience and tools to adapt them to production needs.
There is therefore no absolute boundary between models and applications. A model itself can be a product. The test remains whether it creates user value: raising social efficiency, improving people’s lives or taking on dirty and difficult work.
A model company may build one product or pursue several. He sees Kimi and MiniMax as different strategies, not a question of right versus wrong. Perplexity, ChatGPT and Gemini can also follow different paths to solve search; not everyone needs to climb the mountain by the same route.
39. The Opportunity in New Hardware Is Filling the Context Gap, Not Replacing the Phone
If an Agent is to become an OS, it needs sufficient context, tools, intelligence and real-time interaction, which may eventually require a new hardware carrier. But he strongly rejects the stories that “glasses will replace phones” or “next-generation hardware will eliminate phones,” calling them “complete nonsense” and “a money-making scam.”
The phone has not replaced the PC, and different devices are more likely to complement one another. Phones have powerful computing and sensors; the problem is that they spend most of their time in a pocket, making their sensors nearly invisible to the surrounding environment.
By his subjective estimate, computers and phones capture only about 20% of the context in daily life. Meals, conversations and trips generate large amounts of uncaptured information. Whoever can obtain the context Apple, Google and Microsoft cannot access may establish a non-consensus entry point.
If a startup builds only an app within an existing phone system, it is often just a small subset of the platform’s complete data universe and ultimately helps the giant validate PMF. A single Android update could absorb the startup’s work.
40. The Next OS Interface Will Be Generated in Real Time for Each Task
His imagined OS is not built around fixed apps and prewritten workflows, but is “alive.” Driving versus parking, landscape versus portrait mode, and meetings versus solitude should reorganize the interface around the context even on the same screen.
In a meeting, an Agent could instantly generate a page showing subtitles, notes, progress and supporting material at once. When a website opens in 2 seconds, the system may eventually do more than transmit service data; it may generate and render the most suitable UI for that moment.
This path depends on a Coding engine because every interface must be driven by code and rewritten in real time. It may take 10 or 20 years, or may not happen even in 20 years. The practical prerequisite remains having a product and a company, moving toward the goal without dying along the way.
41. Agent Networks Need Collaboration, Identity and Trust to Break Out
He expects small tribal-style networks to gradually appear over the next 1 or 2 years, with Agents calling, dispatching and coordinating with one another. Once they scale, questions of identity, permissions, trust, bidding and management will emerge. Agents will need an “ID card” and a “password lock.”
Context cannot be shared indiscriminately. A person does not hand a supplier every document from the past 20 or 30 years; an Agent should likewise open only a portion based on the recipient, task and risk. Authorization scopes, collaboration relationships and information isolation across Agents will become infrastructure.
Google has proposed the A2A protocol, but in reality the system is still mainly API calls and has not formed autonomous cooperation among Agents. In his grassland analogy, the “savages” of East Africa have not even encountered their counterparts in West or South Africa; most Agents today remain “solitary savages.”
That is why he believes Agents are “far from exploding.” Different fields must first create enough value, then connect and collaborate. Only when combined results clearly exceed what a single Agent can do will the network effect truly emerge.
42. Humans Will Ultimately Look More Like Directors, Judges and Aesthetic Arbiters
明超平 likes the director analogy. Humans do not need to complete every step themselves; they judge whether the result is good or beautiful. After Agent A and Agent B hear that “the boss thinks it doesn’t look good,” they collaborate, improve the result and send a new version back for human review.
He believes visual presentation remains the most efficient. Around 80% of human information, he says, comes through vision. Even if inputs can be voice, gestures or other modalities, complex results still need to be presented through an interface.
A fixed interface cannot contain the enormous heterogeneous information an Agent may obtain. Today it may be video, tomorrow a website, and the day after an image, text or paper. The presentation layer must ultimately change dynamically and may need millisecond-level adaptation.
He sees 2 most important layers in the value chain: the front end, where the system decides “who gets the task,” and the back end, where the result is presented. The former allocates economic value; the latter directly owns the user relationship.
43. From 2022 to 2025, the Industry Shifted from Models to Agents
In 2022, large companies were still debating whether to do AI and how to combine it with existing businesses. Organizations would not immediately abandon their core operations, and decisions and implementation often took 6 months to a year.
In 2023, ChatGPT “knocked the industry unconscious.” The focus shifted to how to catch up and whether to go open source or closed source, followed by the debate over models versus applications. By 2025, the center of discussion had clearly moved to Agents.
His assessment of DeepSeek and Manus was restrained and direct: “Both are excellent, both are very capable companies—learn from them and respect them.” Although investors often compare him with Manus founder 肖弘 and an intermediary had introduced them, they still had not formally met at the time of the interview.
44. Anthropic’s Coding Conviction Contrasts with OpenAI’s Multi-Front Expansion
Discussing Sonnet 3.7 and 4.0, he said Anthropic had “engraved Coding into the model,” executing with determination and precision. He might use Gemini more for everyday questions, but for Coding he chooses Sonnet without hesitation.
The host asked about “OpenAI’s $3B acquisition of Windsurf.” 明超平 did not verify the transaction, instead arguing that OpenAI is trying to do too much: AGI, models, applications, hardware and the user entrance. An acquisition can directly shorten the time and cost required.
In his view, OpenAI’s biggest competitor is Google. Google has richer systems, users and context, making it difficult to build a durable independent moat by seeking only local opportunities within its platform.
He still supports the coexistence of model-level and application-level products. Technology, community, tools and content are all different means of creating value. The market should not force every team onto the same mountain path simply because one route is temporarily ahead.
45. Global Products Can Be Unified; Content Must Be Regionalized
He believes “good products are naturally global.” Core value and methodology can cross markets, but content, language and interaction habits must be localized. Japanese users cannot be expected to consume Chinese content indefinitely, and a Chinese community cannot simply be transplanted to the US.
Cultural, aesthetic and usage differences between overseas and domestic markets are large. The only constant is staying close to users. Noisee gained traction because he spent so much time in Suno Discord, not because a China-based office theorized about overseas demand.
He describes the new generation of Chinese founders as more open, confident, relaxed and rebellious. They no longer assume they can only build copycats, nor do they need products to deliberately look like they came from an American team. YouWare’s Chinese-knot logo expresses its value of connecting people, ideas and works.
46. The Business Model Serves Creator Alignment, Not the Agent Label
He agrees that charging by service outcome is one model for Agents, but rejects defining YouWare solely as an Agent product. Subscriptions, service fees, advertising and other methods can coexist; “we will most likely do advertising.”
The standard is whether the company and creators can win together: the platform earns enough profit to keep investing, while creators receive enough incentive. A business model is not an identity statement; it is a mechanism for keeping the value loop running over time.
He quoted Jobs’ internal formulation for Apple: first, build the world’s best personal computer; second, make enough money to continue doing the first thing. YouWare has no ideological purity about revenue, as long as it does not compromise the ability to keep creating.
47. As CEO, Product Gets Roughly Half His Time, but the Problem Space Is Wider
When he left Kimi, 雨桐 warned him that as CEO he might not be able to spend even 50% of his energy on product. That proved broadly true, but the other 50% is not pure burden; it includes hiring talent, launching, fundraising, acquiring top-level insight and calibrating market understanding.
He admits he is “not aggressive enough.” When he discovers an experience problem at night, one side of him wants the team to work until 2 a.m., while another imagines the employee asking, “Will the earth stop turning if we don’t do it today?” Most issues wait until the next day; he asks people to stay only when the fire is genuinely urgent.
The team is gradually learning to live with uncertainty. Business direction and demand priorities can change at any time. “Knowing something is wrong and still refusing to change is the biggest mistake.” During the May 1 holiday, employees understood his concern, volunteered to work overtime and said overtime pay should be provided.
48. Dissent, Low Ego and Investor Mirrors Form the Governance System
For major decisions, he requires at least 1 opposing view, from inside or outside the company. If everyone unusually agrees that something is good, he becomes more afraid. The ratio does not matter; what matters is fully hearing the concerns and the different models behind them.
He does not treat employee disagreement as a challenge to authority. Many arguments are not fact conflicts but ego conflicts—“Who am I, and why do you get to say that?” Returning to facts often removes the emotion. A CEO who actively lowers his ego also shapes how the team communicates.
When fundraising went smoothly, he once felt invincible at night, then reminded himself to “slap himself awake immediately.” Ego will not disappear; he can only keep fighting it.
Investors are a “mirror” for him. In his head he may swim like 孙杨 or play basketball like Kobe Bryant, while on video he may look awkward and uncoordinated. External investors can reflect the hesitation, wavering and lack of resolve hidden by the team’s internal halo.
49. He Understands Entrepreneurship as Game Records, Surfing and Free Solo
In Go, reviewing a game record is not copying the final board, but understanding why the first, second and 50th moves were made. Copying Douyin, WeChat or YouTube’s interface today is meaningless; if the 50th move is wrong, the 100th move will never exist.
Product prioritization is the order of moves: what must be done today, what should wait 3 years and what should never be done. When 张小珺 asked who the opponent on the board was, he paused and said, “It seems to be myself”—a hesitant self continually playing against an overly aggressive self.
He downplays team competition and prefers alliances that expand the overall market. The real competitors “have not discovered us yet,” while companies that should not be competitors already treat them as one. Like Free Solo, looking left or right while climbing wastes attention; you can only focus on the next handhold and foothold.
The company’s default scenario is “a 99.9% chance of dying,” not pessimism that leads to stopping, but a daily search for ways not to die. The most painful moments include rejecting a product the team is enthusiastically building and being overwhelmed by a string of important questions from investors that he had never considered.
50. The Final Bet Is Not Effort Itself, but Capturing Weak Trends
明超平’s worldview is that an individual and a 20-person team can do very little. To serve more people, they cannot rely on working 24 hours a day; they need to be empowered by important trends in the world.
He compares entrepreneurship with surfing: spend 80% of the time sitting and watching for where the wave is coming from; accelerate when it arrives to stand up, because effort is useless when the wave is absent and fighting against it gets you knocked down. Everyone wants to “follow the trend,” but the difficult part is that early trends are extremely weak.
His key bet is that “Coding will become a creative medium of a new era, not simply programming.” His longer-term bet is that an OS with dynamically generated interfaces may have a chance to break through fixed systems built by Apple and Google.
Both judgments carry clear uncertainty. The OS may take 10 or 20 years, or may never happen, and the company will probably fail. But if the trend holds, YouWare hopes to be both a specialist node called by the Agent network and, eventually, a layer capable of owning the task entrance and user interface.