Summary

Cong Guangle sees mobile internet as a finite, Age-of-Discovery-style set of opportunities, while AI is a longer-term productivity revolution. The former was driven mainly by connected users, personal devices and sensors, with quantitative change producing qualitative breakthroughs; China’s core map was largely completed in roughly five or six years, from 2012 to 2018. AI works in the opposite direction: qualitative breakthroughs produce quantitative expansion. Without capabilities such as Reasoning and Coding, many Agent use cases are not a matter of scoring 60 versus 90, but of not existing at all. The opportunity therefore lies not only in migrating existing applications, but also in new supply, interactions and platforms unlocked when model capabilities cross key thresholds.
Cong Guangle is putting 70%—80% of his resources into an AIGC system that replaces content creators, rather than a standalone video-generation model. The system starts with an account’s persona and stable elements, then moves from “the point of inspiration” to “the line of a topic,” “the surface of a script” and “the volume of production,” covering creation, production, tool orchestration, publishing, feedback and commercialization. A video model is merely one production step, comparable to photography or CG. “No matter how many steps there are, they are still finite”; the key is to connect goals, logic, knowledge and means into a closed loop.
The other 2 “infinite games,” still at the research stage, are an AI App factory and a community platform that generates interactions. The former starts from demand across the web, automatically initiates projects, generates PRDs, writes code and launches products. The latter lets community organizers generate components such as identity verification, LBS maps and check-in systems, extending communities from information exchange into real-world behavior and demand fulfillment. He summarizes the 3 lines as AIGC, AIGI and “AIGI-assisted UGI,” all aimed at studying how AI can replace or augment human supply.
When assessing the value of an AI startup, Cong Guangle is more likely to ask “What is the moat?” than “Is it AI-native?” He does not consider speed a moat because “speed is too linear.” More durable answers include proprietary data, dependence and stickiness created by concentrated users, domain integration complex enough to deter followers, or a flywheel built on real feedback. Products that merely “patch” a large model, and lightweight tools based only on personal intuition, may follow the path of mobile-era ROMs, jailbreaking and simple course schedules: useful in the short term, but limited in long-term value.
The core method ByteDance left Cong Guangle with was to define products through “felt experience, logic and imagination,” then use data to solve supply-demand matching. Jiujiufang and later Toutiao first scraped and migrated PC content, solving the mobile “chicken-and-egg” problem, then moved from product-name-based categories to personalized recommendations and a unified platform. The host recalls meeting Zhang Yiming in Guomao in 2015 or 2016, when Zhang said Toutiao’s VV had surpassed news PV and that the company would go all in on short video over the following 2 or 3 years. “Strategically, it was simple, grand and fundamental; tactically, it was executable and aligned with the team’s DNA.”
The high margins of Piaoquan Video led the host to call it a “ticket warehouse” and “grain warehouse,” giving the team room to absorb the uncertainty of long-term AI R&D. Launched in 2018 during the mini-program window, the product evolved from a video-production and storage tool inside WeChat into a platform with distribution algorithms and monetization. Revenue and profit were not disclosed; the host only said annual profit was a figure “everyone would envy, even feel jealous of.” The team has 50—100 people: roughly 50—60 in Beijing and 30—40 in Changsha.
Cong Guangle is looking not for the people who can monetize consensus fastest, but for collaborators willing to participate in foundational invention and pursue “non-deterministic goals.” The company no longer measures progress solely by daily output, but also by learning speed, iteration speed and accumulated team understanding, because innovation may show no linear progress for a long time before “suddenly multiplying by 10 or 100.” The billion-dollar opportunities he favors include models and infra, personal and professional tools, supply-side replacement across industries, and platforms born from new intelligent interactions.

Deep dive

1. Jiujiufang Was Cong Guangle’s First Look at “Enumerative” Product Innovation

Cong Guangle is 35. He says he has been an entrepreneur for 12 years and has spent 7 years building Piaoquan. Earlier, in 2011, during his junior year of college, he joined Jiujiufang, founded by Zhang Yiming. What attracted him was both the novelty of mobile internet and a campus project he had built himself: using a phone camera to identify buildings, then providing navigation and information.
Jiujiufang grew out of demand for property information, but it did not limit itself to a single transaction gateway. The team explored second-hand homes, new homes and rentals in parallel, along with property news, home-viewing diaries, agent searches and other content, community and utility formats. His summary: “It was somewhat of an enumerative logic”—build out every visible opportunity first, then use experiment results to decide which path to deepen.
The experience taught him to distinguish domain specialists from technology-oriented talent. The former are better suited to solving industry problems such as agent services and offline transactions; the latter are better at “capturing big opportunities and universal problems” and solving them through technological innovation. Had Zhang Yiming continued with Jiujiufang, he might have improved only the property-information layer. Moving into content created room to reshape supply, distribution, format and monetization at the same time.

2. “Felt Experience, Logic and Imagination” Form a ByteDance-Style Product Philosophy

Cong Guangle sees product management as the job of “defining problems,” while engineering is responsible for “practicing against the problem.” Definition starts with individual felt experience: what problems in daily life are worth solving and remain poorly solved. It is not a complete answer, but it is the entry point for discovering real demand.
Logic is a ladder. By weighing competitors, new experience minus old experience, migration costs and other factors, it determines whether a direction is worth pursuing and how it should be executed. Imagination is the ceiling: “When you can imagine many, many innovations, even the extreme possibilities,” you can construct disruptive solutions.
He believes early mobile products were generally obsessed with feature and interaction innovation. What made Jiujiufang and ByteDance different was that they “fully understood and leveraged data.” For content products, they did not only ask how to get users to publish; they also pulled in the broad supply of the PC internet and optimized it for mobile reading, using the migration momentum from PC to mobile to solve cold start.
Once supply became large enough, search, subscriptions and categories were no longer the only answers. The problem became efficient matching between production and consumption, making personalized recommendation the more fundamental solution. Data was not an add-on optimization; it was part of product definition.

3. ByteDance Succeeded Through Incremental Execution—and by Keeping Its Eyes on the End State

Cong Guangle considers ByteDance’s success “fairly inevitable,” but only under the timing of PC-to-mobile migration: the connected population surged, devices became personal, and more sensor data was continuously collected. ByteDance made exceptional use of these variables in product definition, interaction design and supply-demand matching.
It did not bet on a unified platform from day 1. Instead, it used product names such as Jiongtu, Neihan Duanzi and Jingpi Yulu to split demand, letting each category complete its own supply-demand matching before gradually converging into Toutiao. The path addressed immediate constraints while retaining a clear view of the “end game.”
Short video was that end-game variable. Before Douyin, ByteDance had Huoshan, Xigua, Toutiao Video and the earlier “Tonight’s Must-Watch Videos.” The host recalls meeting Zhang Yiming in Guomao in 2015 or 2016, when Zhang said Toutiao’s VV had surpassed news PV and that the company would go all in on short video over the next 2 or 3 years. Mobile-native UGC video was seen as the next supply-side transformation.

4. Not Joining Toutiao Was the Result of Real-World Constraints and Deliberate Catch-Up

When Toutiao was being incubated from zero, Cong Guangle remembers Zhang Yiming inviting him to join, and he participated in several months of early product exploration. He ultimately did not stay for 2 reasons. One was that he had to complete his graduation project, while the team was taking only half a day off each week; the school schedule conflicted with the startup’s pace.
The other was that he felt the thinking of the technology team around him was too close to his own, and instead wanted to learn commercial operations and vertical-industry methods from the Anjuke executive who had taken over Jiujiufang. Looking back, he explicitly acknowledges that his memory may not be perfectly accurate, rather than constructing a definitive story to justify the decision at the time.
The Zhang Yiming he remembers did not project grandiosity. He was “calm, but possessed enormous courage and deep accumulation.” Cong speculates that Zhang may have defined content modalities, production and matching across both horizontal and vertical dimensions. Strategically, the approach was simple, grand and fundamental; tactically, it started with scraping and the team’s own DNA. Zhang was also “extremely hardworking.” During a team retreat in Zhangbei, back pain kept him from sleeping, which made Cong feel how much he had invested.

5. Hongdian and Zuiyou Validated the Value of New Modalities and New Interactions

Cong Guangle started building Hongdian Live in August 2013. At the time, he viewed livestreaming as a “modality” for transmitting information in real time, rather than a single vertical. It could become an education, meeting or sharing SaaS tool, and could also evolve into a live-show business, a gaming-streamer platform or an e-commerce platform.
Mobile devices made it easier to capture audio and video in a more personal way, turning media capture into a more universal capability. Hongdian could serve as a tool, but he had not yet figured out which vertical SaaS category it should serve. He was also more interested in exploring UGC, so he left and began building Zuiyou.
He saw 2 ways to promote UGC: provide production tools, or bring together people who share a culture and identity, then give them new interaction components. Short-video tools had not yet meaningfully lowered the creation barrier, so Zuiyou chose the second path.
Zuiyou combined bullet comments, voice, image remixing, polls, chain responses, upvotes, downvotes and top comments within a community, while adding recommendations, groups and better distribution and connection. It changed the PC-era path of relying on search engines to find forums. Cong believes its early “unique group effect” and content culture were real, though the product later developed strategic differences.

6. The Community-versus-Content-Platform Split Was Fundamentally a Trade-Off Between Efficiency and Ecological Space

His partners wanted to “thin out” Zuiyou into a content platform, improving consumption and distribution efficiency. Cong Guangle wanted to preserve complex group organization, communication and collective production. As with Kuaishou’s shift from a 2-column feed to a 1-column feed, the change would reduce discovery and improve consumption efficiency, but also alter the community’s character.
His observation is that Chinese mobile communities either remain vertical over the long term or become content platforms once they expand. Reddit and Discord can preserve decentralized organization. On one hand, that fits the way information is separated among different ethnicities, religions and interest groups; on the other, they are willing to “let it grow freely” instead of requiring the platform company to directly operate every group.
WeChat groups demonstrate the same model. The richness of industry groups and associations does not come from “the will of WeChat.” But WeChat already satisfies roughly 70% of the interactions and needs of many domestic communities. Independent community startups carry retention, scale and monetization targets, making it difficult to practice long-term benign neglect—and even harder to design a distribution mechanism that does not conflict with group interests.

7. Piaoquan Video Filled a Missing Layer in WeChat’s Network

When deciding to build Piaoquan, Cong Guangle did not start from the sight of a major mobile-internet opportunity. He believed that at the level of universal human consumption and demand, more than 95% of mobile internet’s problems had already been solved: people had WeChat for connecting with one another; Alibaba, JD.com, Meituan and Didi connected people with goods and services; and ByteDance had pushed people-to-content interaction to Douyin.
During his Hongdian years, he noticed users placing the livestreaming tool inside WeChat to listen. That led him to realize that WeChat was not an ordinary app, but a sufficiently fundamental and diverse social network. Red packets and payment capabilities could generate countless transactions; similarly, a startup could add a new capability to WeChat and create a new set of use cases.
Before Channels appeared, video on WeChat was more like a file: it lacked centralized storage, comments, data infrastructure and a creator monetization system. Piaoquan first provided video production and storage tools, then expanded into a platform and distribution layer. It entered a gap in the network’s infrastructure, rather than building another isolated content app.
Mini Programs launched in 2017, and Piaoquan entered quickly in 2018, then moved smoothly along the path of “tool—tool radiating into use cases—platform—distribution and recommendation—monetization.” Cong believes this involved fewer variables than inventing a community or short-video platform from zero, giving each step a higher probability of being right and producing better results.

8. Early Attempts at Generated Video Had the Right Direction, but Not Enough Productivity

Around 2020, after Channels appeared, the team concluded that Piaoquan could not carry the ultimate mission of a WeChat video platform and began looking for a larger supply-side variable. Cong remembers that GPT-2 or GPT-3 and the concept of DALL·E had just appeared, leading the team to conclude that video production might shift from shooting and editing to generation.
The early product let users enter text, then used NLP to retrieve and match images and video assets, combined with TTS to generate explanatory, expressive or showcase videos. It already had tens of thousands of creators, and their workflows were genuinely different from traditional shooting and editing.
But his retrospective does not dress up the early exploration as a victory: “The concept was right, and the trend was right,” but the underlying productivity had not advanced in tandem with products such as GPT and Midjourney. The effort therefore did not produce the variable-level impact he had expected.

9. A Real AIGC Product Must Understand Creators, Not Merely Generate Shots

The team’s main direction today is to fully deconstruct creators, creative methods, content elements and tool orchestration, then build an “AIGC system that replaces creators.” The system learns workflows, patterns, combinations of elements and user preferences, assigning weights to different methods rather than merely generating an individual video.
An account must first define its persona, invariant elements, creative domain and core methods, then generate specific content through “the point of inspiration, the line of a topic, the surface of a script and the volume of production.” Only then can the output maintain account continuity instead of restarting randomly each time.
Cong uses film and television production to explain the boundary. Screenwriters and directors are closer to creation; lighting, cinematography, costume and makeup, assistant directing and acting are closer to production. A video-generation model is a production capability, comparable to CG and similar technologies. It does not equal understanding creative intent, identifying the features humans like or completing the combination.
The system must also predefine goals and carriers for publishing, promotion and monetization, then improve through reinforcement and feedback while running. He sees the problem as grand but still finite: “No matter how difficult the step,” goals, logic, knowledge and means can connect it into a chain.

10. The AI App Factory Must Automate Entrepreneurial Judgment as Well

The second direction extends the “App factory” he knew early in his career, but its scope is no longer limited to content categories. The system will use demand across different fields on the web, combined with current product capabilities and unmet problems, to automatically form directions and initiate products.
The subsequent chain includes automatically generating PRDs, implementing code and launching products. The goal is a “real App factory, or tool factory.” It puts demand capture, product definition and engineering implementation into the same system, rather than simply handing predetermined requirements to a Coding Agent.
Cong Guangle believes ByteDance-style App factories were essentially category expansion within content. If AI’s productivity is combined with entrepreneurial, product and demand judgment, the same process could be repeated across industries. This direction remains primarily at the research and thinking stage.

11. AI-Generated Interactions Could Turn Communities into a “Non-3D Metaverse”

The third direction draws inspiration from Craigslist’s evolution. PC websites could provide only forms, filters and comments; in the mobile era, dedicated interactions such as LBS, dispatch and matching gave rise to vertical products such as Uber and Airbnb. “Interaction determines how much better the experience in a field can become.”
If AI can give every group organizer customized functions, communities would not have to stop at posts and comments. An anime community could design a question-and-answer adjudication system; an alumni group could connect student IDs, facial recognition and China’s academic-information database to verify member identities and relationships.
He gives a birding example: every birdwatcher’s uploaded video or image could carry LBS data, while a dedicated map would contain only bird records. In the past, the group would have had to wait for a centralized company to decide that the market was large enough to justify developing a dedicated product. With AI, niche groups could generate this information model and interaction themselves.
Cong calls this “something like a non-3D metaverse effect.” It does not depend on a game physics engine, but on functional components, filtering and verification systems, allowing groups not only to exchange information but also to generate behavior and satisfy specific needs.

12. An “Infinite Game” Requires Companies to Accept Nonlinear Progress

Of the 3 directions, the creator-replacement system receives 70%—80% of resources, while the AI App factory and generative community remain primarily research projects. The latter 2 better fit his preferred “infinite game”: they may not produce short-term results, but their end states must be large enough, and they connect with his past work in content, tools and communities.
He summarizes 3 entrepreneurial choices: start from individual felt experience or problems in one’s industry; enter an explicit consensus market such as the sharing economy or the “100-teams battle,” then compete through faster and stronger execution; or, like WeChat, start from foundational problems such as communication and social connection, gradually constructing general capabilities. He clearly favors the third, while acknowledging that it is less consensus-driven in China.
On hiring, he believes a 1-person company is better suited to personal exploration, realizing individual will and transforming specific fields. A team is better for people interested in invention and fundamental change across broad domains, and willing to undertake open-ended exploration together.
The company is also rewriting its definition of progress. Deterministic businesses produce output every day; invention does not move forward at the pace of “1 step per day.” Cong allows work to be measured by learning speed, trial-and-error speed and accumulated understanding, because after a long period of exploration, the team may “suddenly multiply by 10 or 100.”

13. AI Opportunities Are Plentiful; What Is Scarce Is Matching the Era’s Variables with Real Moats

Cong Guangle compares mobile internet to the Age of Discovery: population inflows and more convenient devices amplified existing capabilities. Douyin could theoretically have been built in 2012; the difference was that network and other conditions moved from 60 to 90. Once the territory was explored, the market entered an iteration phase. AI is more like the Industrial Revolution and the electrical revolution: foundational capabilities are still developing, applications are broad and the innovation cycle may be very long.
AI’s key leap is the difference between “can” and “cannot.” Without Reasoning or Coding, many Agents cannot be built at all, rather than merely offering a worse experience. Large models therefore look more like handset makers and operating systems. New platforms should also emerge from the interactions and behaviors enabled by productivity, rather than from another sociological segmentation of groups such as mothers and babies or gamers.
The large-company directions he favors, in order, include models and related infra; tools for individuals and for different professions or businesses; supply-side replacement for drivers, creators, developers and other industries; and platforms built on new intelligent interactions. But an industry label is not an answer. End-state value still depends on the depth of supply transformation.
His negative list includes products that merely patch large models or optimize for a specific use case, as well as lightweight tools with only shallow felt experience. They may repeat the trajectory of ROMs, jailbreaking, app stores or simple course schedules. A real moat should come from data, dependence created by a concentrated user base, deep domain integration or a feedback flywheel. For him, WeChat’s File Transfer Assistant already solves most note-taking needs; only when AI can summarize, provide feedback and improve information efficiency can products in this category become genuinely meaningful.

140: Deep Potential’s Zhang Linfeng and Sun Weijie: AI for Science, from the Beginning to Now Chadd Garcia breaks down WaterBridge's post-IPO value creation story $WBI