Breaking Down 2025: AI’s Bubble, Inflection Points, and Survivors — 郑庆生 x 刘英俊 x 李彦男 x 李广平
Summary
2025’s inflection point was not the release of a single model, but AI moving from “picking up the next line” to “going straight for the outcome”: DeepSeek lifted market sentiment, Devin made tasks executable asynchronously, Claude 3.5 opened up coding and Agent workflows, and Nano Banana Pro gave ordinary consumers an immediately tangible “iPhone 4 moment.” 刘英俊 was shaken before DeepSeek: Devin was “the first product that let you leave the screen while it completed a task.” Models and products are still evolving rapidly on parallel tracks; betting on a static capability lead is not enough.
The potential business-model shift this cycle is not merely the emergence of another attention platform, but a move from “real-estate businesses” to “labor businesses.” The internet standardized and distributed goods, services, and content; 郑庆生 imagines a future in which white-collar labor is extremely abundant and platform value shifts toward connecting people’s full intent with labor. 刘英俊 believes the supply-side revolution has not truly arrived yet, but “it’s getting close.”
Total Record and the hardwaring of AI could create a long-term data flywheel: meetings, calls, operations, and life processes that were never previously recorded are becoming raw material for analysis. 刘英俊 calls the internet’s past digitization “shallow digitization” and this wave “deep digitization”; his principle is “record it first, figure out why later.” One of 郑庆生’s clearest expectations for 2026 is that “both data and intelligence become hardware,” solving collection and execution at the same time.
AIGC has not directly wiped out UGC as originally expected; the more realistic path is “humans driving AI,” with human territory continuing to shrink while people still own intent, professional judgment, and the finishing touch. Recommendation data tells models what outcomes are popular, but may not contain the scripts, storyboards, lighting, or performances that produced them; 郑庆生 compares AI to “a doctoral student who excels academically” or “the greatest common denominator.” 刘英俊 leaves open another possibility: know-how may eventually sink into the infrastructure layer, and humans may no longer need to understand the process at all.
The adoption inflection point for AI content is not whether it can convincingly pass as human-made, but whether users still care: Sora 2 videos carry watermarks and are obviously AI-generated, yet comment sections discuss only the plot and the memes. This resembles humanity’s transition from not understanding film editing to accepting flashbacks, close-ups, and complex montage; the next content paradigm may merge drama, anime, and games into one class of real-time, interactive, generated experiences—“as long as you enjoy it.”
The biggest danger in early-stage investing is not having no methodology, but mistaking the necessary conditions written by survivors for sufficient conditions that can predict the future. During periods of upheaval, be extremely optimistic about broad trends while remaining open-minded about specific products; 刘英俊 says “our descriptions are often toxic,” and that there is no such thing as “the Douyin, WeChat, or Xiaohongshu of the AI era.” Model capability is a rising sea level: startups must build “boats” that rise with it, not “pillars” that will eventually be submerged.
Incumbent follow-on is not necessarily bearish; it may validate that the market is large enough, but each cycle still offers only a handful of tickets into the next generation of giants—few enough to count on one hand. Midjourney still has roughly $700M in ARR, while 科瑟 and ChatGPT were discussed as non-consensus phenomena; this suggests that brand, habit, recognition, and founder traits may run deeper than short-term technical differences. The hardest edge to replicate may ultimately be “an individual’s passion and resonance with the times.”
The guests do not deny that many people will drown in the short term, but believe “is it a bubble?” is not the most useful question: canals, railways, telecom, e-commerce, and the Hundred团大战 all came with bubbles, and liquidity itself is a benefit of innovation. 郑庆生 believes that if demand for compute approaches infinity while the unit cost of compute keeps falling, the industry’s five- to ten-year scale could still be materially underestimated by linear extrapolation. 刘英俊 predicts that by late 2026 or early 2027, early-stage investors may stop discussing AI specifically; he also predicts Doubao could reach 500M DAU, rank third overseas, and that WeChat will develop substantive AI capabilities.
Deep dive
1. The 2025 Turn Began with Devin; DeepSeek Amplified Market Sentiment
The host’s annual timeline starts with DeepSeek: Chinese models and Agents then appeared in rapid succession, lifting market sentiment sharply; 刘英俊, however, places the real inflection point at the beginning of the year, when he used a new version of Devin.
What stunned him about Devin was not its coding ability, but that it was “the first product that let you leave the screen while it completed a task”: the user gives it a goal and leaves the interface; the system plans, executes, and delivers the result. The host later summarized this capability as “going straight for the outcome.”
Two years earlier, the team’s discussions of Agents still felt “a little like writing a novel.” Devin made 刘英俊 realize that “it’s almost here.” 李广平 added that the process of assigning tasks to an intern in the morning, afternoon, and evening, tracking progress, and collecting results had been digitized by Devin into a product relationship.
2. ChatGPT, Rewind, and Devin Defined Three Truly AI-Native Products
The first native paradigm 李广平 identified was ChatGPT: for the first time since the internet emerged, ordinary people could talk to a system and receive clear, immediate, unconstrained, multidimensional feedback.
The second was Rewind in 2023, which allowed software to remain resident on a computer and phone, observing and recording everything the user did. Before AI, “no one would have wanted to design something like this”; the technology was simply still at a very early stage.
The third was Devin: it did not confine AI to a chat box, but turned delegation, waiting, follow-up, and delivery in real work into an Agent workflow. The three products respectively rewrote conversation, recording, and task execution.
3. Claude 3.5 Opened Coding; Nano Banana Pro Reached Consumers’ iPhone 4 Moment
On the model side, 郑庆生 points to Claude 3.5 as the most surprising inflection point: chatbots had previously relied mainly on context to “pick up the next line,” while Claude 3.5 opened the coding and Agent paradigm; combined with OpenAI’s o-series models, the exploration space became more three-dimensional rather than flat.
郑庆生 sees ChatGPT, Claude 3.5, and Nano Banana Pro as three approximate “iPhone moments,” and chose Nano Banana Pro as the defining product of 2025. At the time of the conversation, it had been released for only about two weeks, and users were still discovering new ways to use it.
Nano Banana Pro matters because visual and entertainment applications give nontechnical users an immediate visceral experience: compared with the original iPhone, iPhone 4 was no longer merely something geeks appreciated—everyone could pick it up and say, “That’s really cool; it’s really smooth.” This was seen as a tipping point.
4. AI Is Moving Digitization from Paper Records to Unrecorded Human Processes
刘英俊 calls the internet era “shallow digitization”: articles, addresses, phone numbers, and information on paper were moved online, while countless conversations, operations, and judgment processes were never captured in computable form.
The “deep digitization” of the AI era is beginning to collect meeting recordings, calls, screen activity, and even the processes of daily life, suddenly giving models vast amounts of material that did not previously exist. Processing capability came first; only then did recording itself become a potential asset rather than mere archiving.
The shift is even rewriting the language people use with machines: when leaving a message for a large model, 刘英俊 only cares whether all the information points were covered, not about sequence or midstream additions, because the model can reconstruct the intent. “You can express yourself unclearly and still get a result.”
5. The Internet Standardized Supply; AI Matches Full Intent to Final Delivery
The core story of the past 20 years was turning information into articles and recommendation items, goods into images and copy, and services such as coffee and milk tea into photos plus descriptions, then using platforms to distribute them in standardized form.
AI’s difference lies on the demand side: a person can express more intent, context, and constraints, and the system may move from full intent to task delivery rather than merely retrieving an existing item. Investment opportunities may therefore emerge across the entire “intent-to-execution” chain.
The basic unit of internet innovation was connection: roads, canals, telecom, and platforms can all be understood as network expansion. This cycle adds intelligence to the network, with people and AI coexisting and collaborating continuously inside business flows; lessons from the previous cycle therefore apply only partially.
6. Today’s Tool Boom Looks More Like Early Mobile Internet Than a Consumer-Platform Endgame
The discussion compares today’s AI applications with the early iOS ecosystem of weather, calendar, perpetual-calendar, and productivity tools: tool-oriented, To B-oriented, and simple, but not evidence that such products will dominate forever.
O2O, social media, short video, and mobile games only appeared later in mobile internet. AI may likewise not yet have reached the stage of technological parity, mature foundations, and a flowering of products. Today’s flashes of inspiration are merely preliminary experiments in a new paradigm.
Old rules such as network effects may not have stopped working; mature To C interaction simply has not yet formed, nor has a large-scale model emerged in which AI produces content for other people to consume. In the future, a person may have dozens of work or entertainment Agents running around them at the same time.
7. Technology and Products Are Evolving in Parallel, Making Static Moats More Fragile
PC and mobile internet infrastructure were relatively mature when applications exploded. In the large-model era, foundational technology and products are changing rapidly at the same time; model rankings, capability boundaries, and product forms continue to be rewritten.
Some technological advances in the past remained behind the scenes, and by the time the market confirmed a change, the product landscape had already taken shape. In this cycle, technology and products are changing almost in sync, so a static technical lead may not provide lasting safety.
A paradigm can also mature very quickly. Like LBS eventually disappearing into the foundations of Meituan and Didi, users remember the product’s value rather than the technical theme supporting it.
8. Sora 2 Takes “Is It AI?” Out of the Center of Content Consumption
刘英俊 recalls seeing large numbers of Sora 2 short videos in October and November: the watermark was obvious, viewers knew immediately that the videos were AI-generated, yet comment sections no longer discussed the model and instead debated whether the plot and memes were entertaining.
He believes the key question is not whether generated content can disguise itself as human work, but whether “users no longer care whether it’s AI.” As time passes, users may not be able to tell; even if they can, their mindset will already have changed.
Film history provides the analogy: early audiences could not understand editing and close-ups, and might even think the person on screen had been “cut in half.” Today viewers naturally understand flashbacks and complex montage. AI content may also undergo a sudden cognitive jump at some point.
9. New Interactions Often Feel Intrusive at First, Then Become Muscle Memory
Full-screen single-column feeds, autoplay, and autoplay sound in short video initially violated the old habit of actively selecting content, pressing play, and dragging the progress bar; they could even feel intrusive in public. Later, they became the most natural way to consume content.
Pinterest’s waterfall feed and the pull-to-refresh gesture in third-party Twitter clients illustrate how product paradigms form: a design first solves a real mobile constraint, the entire industry copies it, and users eventually forget what interfaces looked like before.
AI video editing, storyboarding, and workflows remain at a similar stage. Whoever makes the interaction good enough first may define the paradigm products adopt by default, rather than merely winning a temporary feature window.
10. AIGC Has Not Simply Replaced UGC; Humans Are Temporarily Becoming the Drivers
The host admits that three years ago he briefly thought “all UGC was going to hell”—AIGC would take over supply faster and more powerfully. Three years later, model output has increased dramatically, but high-quality novels and complete videos still require human intervention.
He summarizes the revised path as “UGC will slowly evolve into U driving this AI,” or AI Copilot UGC. The human stronghold will keep shrinking, and the “eye” that adds the finishing touch may shrink as well, but the process could take a very long time.
The view remains open-ended: the host says he does not know whether AIGC will eventually cross the threshold of independently producing high-quality content.
11. Recommendation Results Cannot Directly Reconstruct the Creative Process
郑庆生 compares AI to “a doctoral student who excels academically” and to “abstraction and the greatest common denominator”: broadly capable, but not necessarily knowledgeable about investing, medicine, or directing, leaving domain expertise and final judgment to specialists.
The host asks whether Douyin’s vast trove of preference data can directly guide models to produce more popular content. 郑庆生 says this is outcome data: it shows what is good, but not “how that goodness was created.”
Behind a good film are the script, storyboard, director, lighting, production design, costumes, and makeup. Without entering the production process, the platform lacks data on the intermediate steps. 刘英俊, however, suggests that know-how may eventually sink into the infrastructure layer like HTTP, leaving users neither understanding nor caring about it.
12. Manus Shows Models Actively Searching for Paths to an Outcome
郑庆生 describes asking Manus to screen all Chinese-founded startup projects over a given period. The Agent not only wrote its own program, but actively considered how “Chinese founders” should be defined, established principles, and then produced the results.
The striking part was not merely its reasoning or logic, but that “no one told it to do this, and it did it.” AI moved from executing explicit steps to creating tools and filling in definitions on its own in order to reach an outcome.
Claude 3.5’s coding ability also broke through the early chatbot boundary: initially, models could get even two-digit addition wrong, making them not fundamentally different from smarter search; later, the space for programming and planning opened up.
13. ChatGPT Is Near 1B MAU, Yet Real-World Adoption May Still Be Early
The host notes that ChatGPT is approaching 1B MAU, and that domestic products such as Doubao bring the absolute user base to a massive scale. Yet many peers around him are still adopting AI much more slowly than the industry imagines.
On China’s deduplicated penetration rate, 李广平 estimates that it “may still be below 10%,” while also saying 10% would already be substantial: search is a fixed need, and users defaulting to Doubao or ChatGPT is already a clear behavioral shift.
The disagreement is whether this represents new demand. 郑庆生 believes Doubao may primarily be upgrading existing businesses such as search, without creating a new demand in the way Baidu did; he nevertheless sees conversation and multimodality as potentially important product innovations.
14. Conversation Is an Underrated Killer App; Natural Language Is Approaching Coding
郑庆生 sees ChatGPT and Doubao as the biggest game changers and killer apps today. Their core innovation is not merely answering search queries, but using conversation to express intent and advance a task.
“Conversation itself is close to coding”: a four-year-old only needs to explain what they mean to have a model find information, explain the world, or complete something. Multimodal conversation lowers the barrier to accessing knowledge and tools even further.
This also explains why conversation could be larger than any individual feature: users do not have to learn software; the software adapts to an expressive form humans already know how to use.
15. Memory Is Turning Tools into Proactive Assistants That Understand Work Habits
刘英俊 feeds his articles, PPTs, and presentation materials into a large model to build a personal digital twin. For tasks such as cross-stage speaking engagements, the model can already draft in “almost exactly the same tone as me,” saving significant time.
He uses GPT for search, thinking, and organization, with almost every output passing through it. Since the beginning of the year, he has also been recording computer activity, phone meetings, and WeChat calls: “I don’t even know what these things are being recorded for, but record them first and figure it out later.”
One follow-up from GPT-5.1 became his aha moment: based on memory, the model proactively asked, “Do you want me to summarize this into a paragraph like I did before?” He replied only “Yes” and received a project description he could copy and use directly.
刘英俊 sometimes asks the model to list every long-term memory it has about him, then checks and corrects them one by one. The host believes that once proactivity becomes habitual, losing AI assistance for writing memos or handling tasks could feel uncomfortable.
16. Aesthetics May Matter Less; Preference and Intent Become the New Control Surface
郑庆生 offers a “hot take”: editors, directors, and screenwriters once used one group’s aesthetic preferences to influence another group; recommendation engines and AI reduce the weight of a unified aesthetic, making users’ own preferences, feeds, and intent more important.
The host asks whether this reinforces filter bubbles. 郑庆生 responds that recommendation systems stop showing content users dislike when they swipe away, whereas AI dialogue may proactively present different views, allowing users to continue debating and potentially break the bubble.
This is not a claim that AI will inevitably be neutral, but a comparison of interaction mechanisms: one-way distribution optimizes immediate preference, while sustained dialogue allows disliked content to remain as an object of further discussion.
17. AI Has Entered Children’s Memories Without Children Needing to Understand AI First
With a single prompt and photos of his two sons, 郑庆生 generates one or two adventure videos every week: the older brother is cast as a superhero, the younger brother follows and grows, and their relationship and coming-of-age themes enter the children’s own narratives.
The host summarizes this as “using AI to create new memories for children.” The children do not know what AI is, yet already treat the generated characters as themselves and their brother, deeply coupling lived experience with machine-generated content.
郑庆生 sees this as positive for himself, while acknowledging that the outcome may look different in 10 or 20 years. 李彦男 remains broadly optimistic about television, the internet, phones, and AI being labeled as forces that “ruin a generation,” arguing that technology has generally expanded the cultural supply ordinary people can enjoy.
18. Startup History Is Written by Survivors; Necessary Conditions Cannot Produce the Next Winner
The host reflects uneasily after observing multiple cycles: “When people in the next cycle learn from the previous one, those texts were actually written by survivors.” Failed founders have no media platform, so many paths and lessons simply disappear.
郑庆生 warns that later generations see only necessary conditions in the success narratives; what people actually need inside a new cycle are sufficient conditions, because investors must derive the future from the present rather than explain results that have already happened.
刘英俊 believes that violent industry change, constantly moving technology, and individual agency can cause the same starting point to produce entirely different outcomes. Early on, therefore, “there isn’t that much methodology.”
19. Livestreaming, Short Drama, and Robot Visions First Appeared as Controversial or Absurd
李广平 recalls that when livestreaming first appeared, many people could not understand why anyone would watch a stranger talk for a long time and interact with them. Short drama was also controversial. Today, these forms are distributed across many nodes of content and commerce.
刘英俊 believes that an especially valuable early-stage signal can be a founder’s ability to make people dream big, even when the vision is “crazy” or science fiction. 王兴兴 once said he wanted to “use robots to build robots,” a statement that impressed investors more than the business path that could be verified at the time.
His vision includes building robots larger than mountains and larger than Ultraman, as well as robots down to the nanoscale. The story sounds absurd, but expresses a long-term direction of cross-form exploration and embodies “individual passion and resonance with the times.”
20. Investing Resembles Physics: Absorb Outliers Instead of Pretending Theory Is Always Right
郑庆生 compares startup judgment with physics: classical mechanics and relativity were internally coherent within their respective stages, until new phenomena broke through their boundaries and forced people to seek a more general theory.
The investment industry similarly reverse-engineers patterns after outliers such as Douyin or speech rates succeed, expanding old frameworks. When actually placing bets, investors dynamically balance founder passion, extrapolation from existing systems, and odds; neither pure rationality nor intuition can make the decision alone.
郑庆生 believes traditional, highly calculable industries can be more data-driven. Technical waves and extremely early-stage projects, however, deserve more optimistic support for people who look strange, because experience-based filters may eliminate the genuinely new first.
21. Non-Consensus Is Not Weirdness; It Is Remaining Unbelieved Until the Data Arrive
Claw-machine livestreams once seemed novel and entertaining while combining internet connectivity, livestreaming, and offline behavior, but giants quickly copied them. This shows that a product can be a good idea without becoming a durable business or moat for a startup.
ByteDance and Douyin’s recommendation logic remained non-consensus for a long time. 李广平 admits that he once disagreed with ByteDance and Douyin and was eventually proven wrong by the facts. The value of a blind spot is precisely that no crowded trade exists before consensus forms.
Midjourney still has roughly $700M in ARR and is viewed by 李彦男 as a non-consensus phenomenon that has already happened: even as stronger image models continue to appear, its revenue is still growing. 李彦男 is willing to remain open to such evidence.
“科瑟” and ChatGPT were also discussed by 李彦男 as non-consensus phenomena. Many people believe ChatGPT could be easier to use and its features could be surfaced more conveniently, yet it has continued to develop in a relatively simple product form. This challenges the intuition that more features necessarily make a stronger product.
22. The Deepest Competitive Difference Ultimately Lies in Irreplicable Founder Personality
The host divides moats into layers: the shallowest is technical leadership and how long it lasts, followed by network effects, user habits, and brand, then cognition and non-consensus, with founder passion, personality, and long-term trade-offs at the deepest level.
“Investing is the best job I can find” and “investing is my best life” sound similar but represent entirely different motivations. The host summarizes the distinction with the Analects: “Those who know it are not as good as those who love it; those who love it are not as good as those who take joy in it.”
Giants cannot replicate this difference simply by adding budget. Technology, distribution, and features can all be caught up with; whether a founder regards the business as “the joy of life” continues to change the organization’s speed and choices.
23. There Is No “Douyin of the AI Era”; Using Old Names to Describe the Future Is Toxic
刘英俊 believes one can predict with high confidence that intelligence will become stronger and cheaper, allowing everyone to “get intelligence at their fingertips,” but cannot describe in advance what specific products will look like.
His warning is: “Our descriptions are often toxic; you carry too many traces of the previous era.” Looking for the Douyin, WeChat, or Xiaohongshu of the AI era is essentially asking the new paradigm to resemble old winners. “That thing does not exist.”
李广平 uses the anime concept of “deus ex machina” to describe future To C products: the truly important product may suddenly introduce rules no one previously imagined and find an entirely new source of inspiration rather than linearly upgrading along an old capability tree.
24. Each Cycle Adds Only a Few Names to the Giant List; Old Players Are More Often Forgotten Than Destroyed
Every innovation cycle asks again whether giants will capture more value or startups will gain opportunities. The historical answer discussed is that very few names truly enter the next generation of giants; the new entrants in mobile internet could be counted on one hand.
Tencent and Alibaba can accumulate across cycles, and Google can make “an elephant dance.” Founders are still competing for a handful of enormous tickets. 郑庆生 believes incumbent follow-on does not mean opportunity disappears; startups are simply competing with one department inside a giant.
The more common fate of the previous generation is “not being eliminated, or not being destroyed, but being forgotten.” Portals may still have users, but they no longer occupy the most important place in the public mind or represent “the strongest voice of the era.”
李广平 believes an organization’s ability to survive a declining business depends on founder spirit, corporate culture, and its own vitality. When the business foundation changes, these factors can continue multiplying the outcome; when the organization ossifies, the decline of the original product drags down the company with it.
25. Short Video Is Near the Cognitive Limit of the Attention Economy; AI Must Open a New Value Dimension
李彦男 calls short video one of the top five media inventions in human history, perhaps even a “text-level” change: high consumption ROI per unit of time, compatibility with livestreaming, e-commerce, knowledge, entertainment, and games, plus natural advertising monetization.
郑庆生 believes phones may also be near the interaction limit of portable displays: they fit in a pocket and can be pulled out instantly, while glasses or projecting an interface onto the hand may not be more natural than taking out a phone. Short video is close to “pulling you over to take a look,” simulating the most direct form of human cognitive transmission.
The next giant product therefore may not continue competing for eyeballs and user time. 刘英俊 believes the new paradigm could emerge outside the human field of vision, handling information and tasks users do not have time to process themselves.
26. AI Platforms May Shift from Attention Landlords to Callable Labor Markets
Internet companies historically resembled “online landlords”: they aggregated users, sold attention, and monetized through advertising, goods, and services. 郑庆生 believes that if ChatGPT were also required to become a general entertainment platform, it might not stimulate efficiency as effectively as scrolling through short videos.
He proposes another mainline: when white-collar labor becomes extremely abundant and each person has one or even N assistants, platform value will come from connecting intent with labor—“turning a real-estate business into a labor business.”
刘英俊 therefore believes the focus should not remain solely on capturing time and distribution channels, but on giving ordinary people better labor. Programming provides an early example: tasks that once required dense intellectual intervention can be handed to a computer to complete on its own.
27. An Asynchronous Cloud World Removes “Time Is Already Fully Booked” as a Product Ceiling
李广平 imagines future assistants, digital twins, and Agents existing continuously in the cloud and processing tasks asynchronously. When users are offline, another world continues running.
He imagines a “multiverse” filled with Bots working or producing entertainment content, where a person’s 24 hours are no longer the hard upper bound on system output. This gives Agent platforms a growth logic different from attention platforms.
刘英俊 believes the capabilities of general Agents and Assistants have not yet been released. They look “very stupid” today, much like early short video, which could only show a few clips and seemed boring. Once they cross an inflection point, they may create a genuine supply-side revolution.
28. Multi-Person, Multi-Agent Collaboration Could Become a New Container for Creativity
The host compares the four participants’ free-ranging conversation with four Agents having different pre-training, post-training, skill packages, and RAG knowledge bases colliding with one another; their different backgrounds cause the same prompt to trigger different associations in each mind.
郑庆生 notes that OpenAI has added group functionality to ChatGPT. The feature may be somewhat boring today, but could foreshadow a future in which groups of people and groups of AI discuss, divide work, and produce outputs together.
Imagine four colleagues joined by four Agents, or multiple Agents for design, Web Coding, and general tasks collaborating around the same podcast page. The result could be more creative than a single model, rather than merely “cattle forced to work.”
The host says, “All poetry is actually prompts”: a line of poetry contains little information but can call upon the reader’s own model. A good prompt may be 98% crap, but if 2% of it creates a flash of insight for someone, it has value.
29. Total Record’s Commercial Value Is Not Explained by the “Voice Recorder” Shell
The core opportunity in AI hardware is not another device, but discovering data scenarios that remain undigitized. In the past, people recorded information but could not analyze it at scale, so there was little incentive to record. Now that analysis capability has arrived first, collection has become viable.
The host believes meeting-minutes and recording products make sense both in terms of commercial PMF and future potential. The market often understands them superficially as “voice recorders” and even asks why phones cannot record. The real question is what services and decisions become possible once every conversation can be analyzed over time.
The host has also considered directly launching a project to preserve a person’s recordings, photos, phone information, and other sensor data over the long term. His judgment is that these materials could theoretically “reconstruct a person’s entire life.”
30. Misreading Character.AI and Short-Video Supply Reveals an Underestimation of Emotion and Creativity
李广平 was once extremely excited about Character.AI, believing that role-playing and discussing problems with AI in the same space could reach several hundred million or even 1B DAU. He later realized that it “can dress itself up like anyone, but it simply is not that person.”
He does not fully reject the direction, but believes the technology may not have been ready at the time. Talking to an “Einstein” Bot could provide the appearance and language, but not the experience users actually wanted—speaking with that person. Multimodality and long-term memory may gradually close the gap.
Looking back at short video, 刘英俊 admits that he once thought ordinary users could not produce high-quality content, only to discover dialect jokes, life skills, and all kinds of group creation forming previously unimaginable formats. “We completely underestimated human creativity.”
The host recalls that people early on also worried Chinese users were too introverted to hold up a phone and dance in the street. If the product structure is open enough, users move from consumers to creators, and the supply problem may be solved suddenly after education reaches scale.
31. “How Big Is It?” and “What If a Giant Does It?” Must Be Reasked of Early-Stage Projects
李彦男 recalls that when he first encountered projects such as Xiaohongshu, Bilibili, and Douyin, static data did little to reveal the end state. Around its $1B valuation, Xiaohongshu was still exploring between community and cross-border e-commerce; Bilibili’s early business model relied heavily on games; and Douyin once had only several million DAU.
A US bill-negotiation Agent is another example of cognitive bias: the system calls hospitals, insurers, banks, or parking operators on behalf of users to negotiate bills or “argue with them.” It initially did not sound like a mass-market need, but two or three companies later developed well.
For early AI companies, instead of asking only whether the business model already works, whether ARR is growing fast enough, or whether the sector is too crowded, investors should examine the team’s rate of progress, whether the product structure is open, whether the market is broad enough, and whether there is clear evidence that could kill the company halfway through.
郑庆生 believes giant investment can actually be exciting: it validates that the direction is large enough, while the startup is only competing with one department inside the giant. Chatbots, coding, general Agents, and video editing are all in this state of resonance and direct competition.
32. When the Models’ Tide Rises, Startups Must Become Boats, Not Pillars
刘英俊 relays an analogy from a founder: every model iteration is like a rising sea level. If a product is merely a fixed-height “pillar,” its capability ceiling will eventually be submerged by the foundation model.
If the company builds a “boat,” model improvements, falling costs, and even capital bubbles may lift it higher. Investment judgment therefore should not focus only on current functionality, but on whether the product can continuously convert underlying upgrades into greater value.
李广平 also warns against merely “using a new tool to build a carriage.” Some sectors have failed to gain traction through one or two previous cycles; the arrival of AI alone does not justify linearly extrapolating them into large opportunities. One must investigate whether the earlier failures came from technical limitations or from fundamentally different markets and user behavior.
33. Bubbles Fuel Innovation, but Long-Term Trends Will Not Guarantee Every Boat’s Safety
Asked whether the current market is a bubble, 郑庆生 says “it is not bubbly enough,” including in the US. Over a five- to ten-year horizon, demand for compute from asynchronous cloud systems, multiple Agents, and continuous interaction may still be severely underestimated.
李彦男 calls this “a very typical pseudo-question”: the Hundred团大战, e-commerce, portals, telecom, cable television, railways, and canals all went through bubble cycles. If bubbles are unavoidable during infrastructure expansion, identifying their temporary existence does not directly guide action.
李彦男 believes bubbles may even be a “benefit or reward” for founders and investors, because ample liquidity allows large numbers of innovations to emerge simultaneously. If every project always had to raise money at seemingly rationally low prices, there would be too little exploration.
Their optimism has explicit conditions: compute demand approaches infinity, unit compute costs continue falling, and industry growth may be exponential rather than linear. The host simultaneously stresses that a valid long-term trend does not guarantee individual safety; many people still died at sea during the Age of Exploration.
34. Superindividuals Are Rewriting Organizations from Collections of Jobs into Collections of Intent
On whether to start companies personally, 李彦男 wants to rebuild ACGN by combining video, interaction, and games; 郑庆生 does not want to be a CEO and prefers operating with fewer than 10 people or even alone; the host wants to build hardware that records an entire life; 李广平 focuses on real-time, interactive multimodal applications.
郑庆生 mentions a company with only 3 people: Agents monitor data and write code around the clock, while humans review the work in the morning. After the proliferation of AI tools, every startup idea deserves the first question: “Can one person do this?” A superindividual may even form a supercompany.
The host also describes a company that no longer strictly distinguishes among product, planning, coding, and engineering. Anyone with an idea can launch it; within 3 or 4 days, it goes live and receives traffic and feedback. An intern used Gemini 3 to build a website, connected it to the Nano Banana service, and turned it into one of the company’s most profitable businesses.
The early business generated roughly $10K-$20K per day, leaving the intern “feeling like it was a dream.” The case shows that job skill trees and seniority are losing explanatory power; people who can express intent, mobilize tools, and close the loop quickly are gaining greater leverage.
35. AI Will Revalue Talent, Careers, and Venture- Studio Organization
李广平 observes that top researchers may receive packages worth hundreds of millions of dollars, perhaps the first time in human history that an individual employee’s returns have far exceeded those of many CEOs. Tool leverage has amplified the output of a few people, making the very best “superhumans.”
郑庆生 believes traditional hierarchies were already behind information technology, and the gap is widening further. Future companies may consist of 2 or 3 key people plus large numbers of AI systems, with boundaries among product, growth, operations, and engineering continuing to weaken.
刘英俊 believes Venture Studio may be entering a better era: one person can try multiple directions, validate, monetize, or exit within half a year, then switch to the next opportunity. A career is no longer a fixed lifelong identity, but a temporary organization formed by an entrepreneur with a group of people.
The host summarizes the shift as follows: “A huge increase in subjective will multiplied by tool capability may become an explosion of creativity.” On whether ordinary experience remains worth having models learn, 郑庆生 emphasizes expert knowledge, while 李彦男 believes memes, music, and humor still require ordinary people to discover and label them.
36. Raw Data, Databases, and PPTs May All Be Redefined by AI
The host believes enterprise software historically required data to be recorded at preset levels of granularity, fields to be defined, ERPs to be connected, and system interfaces to be maintained. If raw information is preserved comprehensively enough, AI could theoretically reconstruct workflows and views afterward along different dimensions.
郑庆生 goes further, suggesting that the future may no longer require data formats to be defined as they were before, and that many aspects of databases could change. He also acknowledges that this is merely a directional “hot take”; the underlying infrastructure shift has yet to unfold.
郑庆生 relays a CEO’s response to the question of how to make AI produce better PPTs: “Why do we still need PPT?” A deck is an old-era container for compressing and presenting information sequentially. In the future, information that needs to be delivered could sit inside a single container or AI tool, appear as a video stream, and allow the recipient to ask questions in real time and see different versions.
The host believes information authorization could also shift from “show it to you or do not show it to you” toward progressive disclosure. LinkedIn must display large amounts of personal information the other party does not care about; AI could answer only the current question, allowing the user to open up necessary information step by step through interaction.
37. 2026 Bets Center on Hardware, Multimodality, Doubao, and WeChat
郑庆生 expects “both data and intelligence to become hardware”: one class of devices continuously records life, while another receives instructions and actually gets things done. By the next recording, the system may already be able to generate his personal script based on his entire history.
李广平 is betting that real-time, interactive multimodality will cross the product threshold. Traditional categories such as drama, games, and anime may merge into one type of dynamic content, generated instantly in response to user interaction—“as long as you enjoy it.”
李彦男 believes that if video models reach PMF similar to Nano Banana Pro in speed, real-time performance, and cost, the API release will trigger a “product-manager carnival.” Large numbers of interactive formats previously constrained by production costs will appear simultaneously.
刘英俊 makes the more aggressive prediction that by late 2026 or early 2027, early-stage investors may stop discussing AI specifically. AI will become a ubiquitous default that no one pays particular attention to.
38. Doubao Is Pegged at 500M DAU; WeChat Is Seen as a Critical AI Container
刘英俊 makes a specific prediction: Doubao will reach 500M DAU and rank third overseas, behind GPT and Gemini. This is not a statement of its current position, but a bold forecast to revisit a year later.
WeChat is selected by multiple participants. 李彦男 believes many of today’s ideas about AI phones, smart assistants, virtual humans, and digital humans could be implemented within WeChat without much additional effort.
The host believes Tencent’s slowness may be an advantage here: it “almost never makes mistakes” and often waits until PMF is clear before entering. Video Channels launched nearly 6 years after Douyin yet still approaches 400M DAU, which was cited as evidence that Tencent’s moat remains powerful.
The discussion’s conclusion is that WeChat may become a critical container for AI penetrating existing daily life. At the same time, the products truly belonging to the new cycle—and still nameless today—will emerge from startups competing for those few enormous tickets.