“Pessimists Are Right, Optimists Succeed”: A Conversation About the AI Industry with My Friend 亚婷

Summary

• 庄明浩把早期 VC 归结为一场概率游戏：一轮泡沫即使“全军覆没”，也不意味着当时拒绝下注就一定更高明。 If the metaverse had worked in a parallel universe, a handful of winners would have been enough to cover the losses; the real question is whether investors can accept a process of betting first and validating later. “Pessimists are often right, but optimists succeed.”
• AI has turned startup trial and error from one long-term bet into a series of shots: the same amount of money that once funded one stage might now let a team take 6 shots. The fundamentals—user demand, team fit and the like—have not changed, but a product can now be approved or killed in 1 or 2 months. Investors’ stage assessments, team requirements and the experience baggage that comes with age all need to be recalibrated.
• When an application company reaches $100M in ARR quickly, that may be exactly when foundation-model companies begin sealing off its ceiling. After Cursor and Lovable, capabilities such as OpenAI Codex quickly moved downstream; after Harvey and Legal Run made the fastest-to-$100M-ARR lists, Claude entered law firms as well. Vertical data, workflows, skills and APIs are not durable moats—at best, they are temporary ones, and nobody knows how long they will last.
• Model competition is moving from 3 main tables—language, multimodal and coding—to a single table, with resources and strategic trade-offs deciding who wins each phase. 庄明浩 believes Anthropic forcibly turned coding into a standalone track, while Google gained a temporary edge through multimodality. On the video side, Seedance 2.0, Kuaishou, Alibaba, Google and more than 10 other companies are all placing bets; the leading players share one trait: “There will be no hesitation.”
• The greatest danger for companies is not failing to calculate AI’s ROI for now, but waiting indefinitely for a benchmark company to provide the right answer. AI is like dribbling a basketball: “You cannot learn to dribble by watching videos.” Companies have to start using it and develop the feel. The debate has already shifted from “can we, should we” to “what have we built, and how much did we sell it for?” That shift itself shows the industry entering a phase that puts more weight on delivery.
• The US-China AI competition is not just about model scores, but the full stack of data, algorithms, compute, energy and an autonomous ecosystem. 庄明浩 says the 5 largest US companies may control more than 75% of global compute, while China’s total compute combined would rank only 4th or 5th. DeepSeek V4 was delayed by roughly 4 or 5 months for work with Huawei; the ecosystem value of that effort may matter more than whether a single model reaches the top of the leaderboard.
• 庄明浩’s thesis is that once AI crosses the capability line, human scarcity may shift from technical skill to intuition, expression and relationships—but the transition will be full of friction. The 李世石-AlphaGo story shows that “fairness” can be maintained only through human-imposed time limits, handicaps and input restrictions. If AI can also penetrate the in-game intuition Faker built through thousands of rounds of training, humanity’s last defensive line may advance as well. 庄明浩 is pessimistic about the end state for employees, but stresses that inertia and organizations will keep the middle period going: “What comes next is what matters”（“然后的事情才重要”）.

Deep dive

1. “The Art of Slaying Dragons” Is a Deliberate Expression After Old Experience Stops Working

• 易亚婷 recalls that when the two met in 2018, 庄明浩 was the big shot salespeople would add on WeChat before his talks even began. Whenever she tried to include his professional title, he would leave only “The Art of Slaying Dragons主播.” 庄明浩 says titles matter less and personal expression may matter more; in informal settings, he usually keeps the more accommodating identity of “podcast host.”
• The show’s name came from a friend’s reaction to an episode about ByteDance’s growth history: user growth, operations and mobile-internet methodology were once dragon-slaying skills, but “the world no longer has dragons, so everything you know is obsolete.” 庄明浩 decided to preserve a space for expressing experience that had lost its mainstream market.
• 《屠龙之术》 launched around September 2024. At the time, perhaps only half of the tech industry’s conversation belonged to AI; today, the share may be as high as 90%. The show has gradually shifted from carrying the old history of mobile internet to becoming a window onto an industry that is now mostly about AI.
• 易亚婷’s opening contrast explains the show’s value. When ChatGPT launched, 庄明浩 spent 6 hours analyzing 200-300 YC-backed companies one by one. Several years later, companies he had favored had risen from $50M to $10B in valuation, yet he still describes himself as the “professional caulker” who fills in when media outlets cannot find an expert.

2. Faster Theme Rotation Turns a Veteran Investor’s Experience into Baggage

• 庄明浩 joined 经纬 in 2011 and left in the second half of 2015. His first stint covered communities, e-commerce, games, advertising and platforms, followed by the generation of content companies built around anime, esports, ACG and Bilibili. He then spent more than 3 years in game livestreaming, failed at a startup and returned to 经纬 in June 2019.
• His second return coincided with the metaverse’s shift from an ignored niche to an industry-wide land grab. Early-stage VCs might once have seen 1 major theme every year or 2; later, the theme changed almost every quarter, with each new one pushing the old one completely out of view. If an investor missed the first step, they often could not participate in any of the steps that followed.
• The flood of dollar-fund capital and fierce competition in 2020 and 2021 pushed valuations beyond what 庄明浩 could accept. A company might have little more than a team, yet raise tens of millions of dollars at a valuation in the hundreds of millions. He admits he could not persuade himself to invest at those prices and on that timetable, so he left.
• As he was leaving, his advice to the bosses was not to replicate his caution, but to “find a few young people, not old guys like us”: carry less baggage, scan the entire market and make a few blind bets first. He could recommend that strategy from the sidelines, but he could not execute it himself.

3. VC Bets on Probability, Not on Proving Itself Right Every Time

• When 易亚婷 asked whether that wave of crazy projects had ended well, 庄明浩 answered directly: “That whole wave was wiped out.” But his conclusion was not that caution had won, because VC “doesn’t need to guarantee that every investment works”; it only needs 1 outcome to succeed.
• His counterfactual is simple: if the metaverse had actually taken off in a parallel universe, all those investments would have seemed worthwhile. Investment decisions are therefore not just about factual judgment; they also reflect personality, risk appetite and whether someone can accept being wrong most of the time.
• Around age 35, 庄明浩 gradually concluded that his preferences no longer fit the early-stage betting environment. When 易亚婷 asked whether investment principles should remain consistent, he acknowledged that user needs, how those needs are met and team fit remain the underlying answers. But new people, new forms of education and new ways of understanding require investors to revise their judgment parameters.
• AI has further shortened the validation cycle. A decision that once required months of deep work before anyone knew whether to continue can now be stopped and restarted within several months, or even 1 or 2 months. The same money has shifted from funding 1 stage to supporting 6 stages—6 shots. What changed is not strategic common sense, but the yardsticks for stage assessment and team execution.

4. AI Has Already Entered the First 3 Stages of Knowledge Work

• Asked which parts of “collecting information, finding patterns and explaining them in plain language” AI cannot replace, 庄明浩 answered: “All of them. It can do all of them now, and it’s getting better.” He used a chef’s workflow to break down his own content production and show that replacement is not an abstract proposition.
• The first step is buying the raw materials: rapidly browsing and collecting information every day, then putting it into a knowledge base. The second is washing and chopping. AI can already perform OCR on complex English charts, understand their meaning, explain trends and interpret their implications. The organizing work that once demanded experience and patience, in his view, is “no problem at all” anymore.
• The third step is the head chef’s coordination: deciding how many dishes to make, what ingredients each requires and the order in which several pans should be used. 庄明浩 says he and AI are roughly “one-to-one” at this stage. The fourth is cooking in the narrower sense—the final output, such as a PPT or a word-for-word speech script. He initially said he barely lets AI touch this stage, but after further questioning estimated that the final output is already roughly one-to-one, with some manual work retained.
• Since NotebookLM and Nano Banana arrived, turning complex text, images and tables into visual content has become sufficient for non-serious settings. He calls the relevant presentations “the last handcrafting”; the insistence on doing more by hand is now mainly a self-imposed standard for an old-school craftsman, not a firm technical necessity.

5. The Main Geopolitical Table Is Still Finance, Manufacturing and Other Traditional Industries

• Asked why Trump had come to China amid recent US-China interactions, 庄明浩 refused to pretend he knew: “How would I know what great-power competition is about?” But he observed that the delegation may have included only 黄仁勋 or 马斯克 as people directly tied to AI—and 马斯克 does not do much AI anymore. Most of the companies still came from traditional finance and manufacturing, including Visa, banks and Boeing.
• The tech industry has focused intensely on whether chip controls might ease, but he believes the issue “wasn’t even discussed,” or at least was nowhere near as high a priority as industry participants imagined. The 2 heads of state were handling a much broader relationship; “we may be over-dramatizing our own importance in our expectations.”
• AI has instead created a striking change in the media layer. Synthetic images of 黄仁勋 carrying a backpack, moving Moutai or hauling chips are both creative and plausible, making them easy for ordinary viewers to believe at first glance. 庄明浩 believes image models crossed the point in 2025 where people could no longer find flaws with the naked eye; Nano Banana was an important inflection point in combining generation and understanding.

6. Once Technology Crosses the Line, Neither Copyright nor Employment Can Be Solved by Technology Alone

• 易亚婷 points out that with traditional Photoshop, the cost of creating a fake and the cost of reviewing it were roughly 1:1. AI now allows generation at a scale far beyond platforms’ review capacity. For now, she believes the only option is for governments to require labels or watermarks on AI content. She also argues that the existing copyright system may stop working for a period of time, while “nobody knows” what the replacement will be.
• 庄明浩 adds that when video generation involves protected characters such as Pikachu, copyright ownership is itself a problem, and overseas companies will keep filing lawsuits. The hard part is no longer whether models can generate the content, but how law, society and culture redraw the boundaries. “The technology itself has long since stopped being a problem”; institutions simply have no answer yet.
• The disruption to Hengdian, live-action dramas and traditional production workflows will not stop because any particular company exercises restraint. Even if Google or ByteDance does not pursue it, another company will push the technology across the line. “It’s like Pandora’s box: once it’s opened, there may be no going back.”
• 庄明浩 rejects the breezy claim that “new jobs will always appear.” Nobody knows when the next equilibrium will form, what it will look like or how many people will share in the new opportunities. The talent pipeline is especially exposed: once entry-level work is eaten, middle-aged workers lose the juniors below them who need space to practice, while university graduates lose the first rung into the industry.

7. $100M in ARR Is Both a Success Milestone and a Target Line for Model Companies

• On the limits of foundation-model companies, 庄明浩’s answer is nearly brutal: “Right now, it looks like there are no limits.” Once an AI application proves that it genuinely works, model companies follow the revenue signal into the category instead of treating it as an uninteresting vertical niche.
• Coding companies such as Cursor, Replit and Lovable quickly reached $100M, and even $200M, in ARR. OpenAI products such as Codex followed soon after. The original companies may continue to grow, but “their ceiling has been sealed.”
• Legal AI repeated the same story. After Harvey and Legal Run appeared on lists of the fastest companies to reach $100M in ARR, Claude launched a legal application for law firms. Data security, privacy and compliance requirements in the legal industry may provide some breathing room, but they may not form a stable, permanent wall.
• 易亚婷 offered the counterargument on behalf of Agent founders: vertical data, workflows, large numbers of skills and APIs are all accumulated assets. 庄明浩’s answer preserves the time dimension: these are not true moats, at most temporary ones. How long the phase lasts and how deep the moat can become are both open questions. The most realistic joke is: “Better not get there first—once you do, people will start watching you.”

8. The 3 Main Tables Will Eventually Become 1; Each Phase Is About Which Leg Cannot Be Allowed to Fail

• 庄明浩 divides the AI battlefield into 3 main tables: language, which has evolved from language models into reasoning; multimodal; and coding, represented respectively by OpenAI, Google and Anthropic. Coding was originally a subset of language, but because it fits Transformers plus reinforcement learning so well and its revenue was growing so quickly, Anthropic “forcibly” turned it into a standalone track.
• The 3 routes are like climbers approaching the same summit from different faces of a mountain, and may ultimately merge into 1 table. For now, no company can do all 3 best. OpenAI is currently more focused on competition with Anthropic in coding and Agent, temporarily putting multimodality behind it; first and foremost, this reflects resources that cannot cover everything at once.
• Google was seen by the market as the winner from the second half of last year through around this year’s Spring Festival, thanks to Nano Banana and its multimodal strength. In the several months after Spring Festival, Anthropic was winning because “coding was winning.” Google can make trade-offs, but it cannot allow its coding leg to lag too badly.
• Pure model research is no longer about holding back a single “big move” in advance. Teams now layer on versions and add features every month, then watch for positive feedback from the market, users and competitors. Coding’s rise itself was unexpected; whether a 4th route will emerge in the future is likewise impossible to determine in advance.

9. Video Models’ Capital Burn Will Eliminate Players Unable to Promise “No Hesitation”

• The video field already has ByteDance’s Seedance 2.0, Kuaishou, Alibaba, Google, as well as more than 10 named model companies including PixVerse, Vidu and HiDream. A huge market does not mean it can support more than 10 foundation models; competitive pressure over the next 2 or 3 years will be higher precisely because of that.
• 庄明浩 repeatedly uses the same qualifier for ByteDance, Kuaishou, Alibaba and Google: “There will be no hesitation.” Video connects to their core businesses, platforms or multimodal strengths, so competitors cannot expect these giants to exit voluntarily because of short-term losses.
• In his view, Imagen 2 did not necessarily produce a comparable leap in pure generation capability. What mattered more was its ability to automatically fill in visual details using world knowledge. Nano Banana combined image generation and understanding; the old review process of searching for flaws is now difficult to sustain. Once this class of capability crosses the line, it will keep moving forward.

10. Companies Can Learn AI Only by “Dribbling the Ball”; Waiting for a Standard Answer Means Leaving the Game

• 庄明浩’s advice to business owners who do not care about AI is: “Don’t get anxious. Start using it and figure out the rest later.” AI is like playing basketball or dribbling a ball; you cannot develop the feel by watching instructional videos. Nor should companies set aside a large block of time to “learn AI.” They should put it into their daily actions and keep using it.
• 易亚婷 offered the common waiting strategy on behalf of business owners: let leading companies chart the path first. 庄明浩’s response is that the speed of change means there is no stable benchmark. Wait long enough, and you end up like a spectator watching Kobe and Jordan play, having completely lost the desire to step onto the court.
• She continued: many organizations of 100 or 1,000 people are still stuck at the chatbot stage and are living comfortably on inertia; even adding AI seems to change little in the short term. 庄明浩 acknowledges that organizational change will not happen overnight. For now, there are also no good answers to “what exactly has coding built, how much did it sell for, and what is the ROI?”
• But the object of debate has changed. The industry is no longer arguing about whether AI “can” or “should” do something; it is asking about output and ROI. Nadella’s example of AI adding perhaps 10% to GDP remains distant, but GDP is a lagging indicator, and organizational transformation takes time. The absence of a visible macro number cannot disprove changes already underway.

11. AI Will Push Organizations Toward Super Platforms and Super Individuals

• 庄明浩’s pessimistic scenario is that the future may contain only foundation models and super-individuals, alongside a small number of enormous organizations. The middle layer that survives on information friction faces the greatest challenge. Intermediaries exist because information is incomplete; if AI pushes friction close to zero, middle structures lose their economic rationale.
• The games industry offers the clearest example. One end of the future may consist of massive online games, while the other consists of tiny independent games created purely for expression. The commercially viable middle-sized game category shrinks: a product either becomes a plug-in inside a large platform or retreats into a small game driven purely by traffic.
• A Roblox-style platform provides the physics engine, AI engine, development tools, payments, operations and cloud servers. Developers need only build a mod or module; they no longer need to assemble a full company. The more advanced the AI capability, the more capital it requires to develop, while users may have to enter the ecosystems of giants. The divergence therefore reinforces itself.
• This is not a definite forecast about headcount, but a possible organizational shape: a tiny number of OPCs coexisting with giant companies. 庄明浩 repeatedly retains “I don’t know” and “it’s possible,” because the middle state will be challenged but may still persist for a long time through inertia, organizations and other complex forces.

12. Compute Is the Hard Constraint in US-China Competition; DeepSeek Is Betting on an Autonomous Ecosystem

• Asked about “de-dependence,” 庄明浩 says DeepSeek is advancing a partnership with Huawei. DeepSeek V4 was delayed by roughly 4 or 5 months, with part of that time devoted to this work. In an era when monthly updates have become standard, the effort itself represents an investment “not driven by immediate commercial interest.”
• He believes the significance of this work may exceed whether V4 becomes the strongest model in the market at that moment. Its ecosystem benefits may begin to emerge in the second half of the year; potential investors in DeepSeek and its fundraising narrative are also tied to this path toward autonomy.
• The compute gap remains enormous. Based on statistics he cites, the 5 largest US companies may account for more than 75% of global compute. If China’s entire compute base were combined into a single entity, it would rank only 4th or 5th globally. The US advantage comes from GPUs; China’s usable relative advantages are electricity, energy and potentially storage and other components where breakthroughs may still come.
• Data was already a central point of contention before the current AI cycle; the TikTok episode took place in 2021. Algorithms are harder to keep absolutely secret because “half the engineers in Silicon Valley may be ethnic Chinese or from China,” and R&D has already shifted from individual heroics to team collaboration. 庄明浩 uses 姚舜宇’s summary to explain why China is catching up so quickly: “Algorithms have no secrets”（“算法没有秘密”）.

13. When Capability Is No Longer Scarce, Human Value Shifts to Intuition, Relationships and Expression

• Alignment first gets stuck on the question of “who are humans?” Give the same prompt to Kimi, 豆包, DeepSeek, 元宝, ChatGPT and Claude, and the answers differ not only in wording but also in gendered feel, emotional tone, optimism and rationality. There is no single line that can be drawn around the “correct” value system, so model differences cannot disappear entirely.
• 庄明浩 leaves open both sides of the question of whether model gaps will widen or narrow. Once a benchmark passes 80 points, it may be “basically good enough.” But when DeepSeek R1 appeared, people said the same thing; 1 year and 4 or 5 months later, progress was still accelerating. The arrival of GPT-5 produced the same debate without stopping the race. Old benchmark results do not say much; the discussion ultimately returns to the human experience.
• 李世石 offers a case study several steps ahead of where most people are today. He went from dismissing AlphaGo, to doubting himself in the second game, breaking down in the third, rebuilding an advantage in the fourth and admitting that it was the only—and last—time in his career he had won with something like an underhanded move. He later asked AI to spend 20 seconds per move while he had unlimited time, then gave it 2 stones, calling that fair. After taking several weeks to win 1 game, he removed the handicap, and reality shattered him again.
• Had it not been for 马斯克’s recent changes at xAI, an xAI model might have played against T1, Faker’s team, this year. To make the contest fair, the model’s latency in operating a computer would have to be limited. DeepMind is testing a related model that can only view the screen and operate the mouse and keyboard; it cannot read backend data.
• During the show, the host asked an AI whose identity was not specified: if you wanted to guarantee a win against Faker, what question should you ask Faker? The answer focused on how to tell whether an opponent is nervous and how to notice that you are being targeted—abilities built through thousands of repetitions, physical condition and in-game feedback, rather than through operating rules. Human beings’ last defensive line may be this kind of intuition.
• Music pushes the end state even further into the cold. Pleasant melodies are merely finite permutations and combinations; existing songs are already enough to occupy several generations of listeners, and music models can keep arranging combinations within a range that does not offend. The reason to keep creating is therefore no longer to produce 1 more song, but “what do you want to express, who do you want to hear it, and what feedback do you want?” What lies below the waterline gets submerged; sincerity just above the surface earns disproportionate distribution.
• In the rapid-fire questions, 庄明浩 admits that if there is only 1 conclusion about the future of office workers, it may be “a dead end.” But this generation may not live to see the extreme endpoint; organizational inertia may be enough to carry many people to retirement. Human collaboration will not disappear either. Pure information exchange is easiest to replace, while eye contact, smell, bodily feedback and the atmosphere of being there will become increasingly valuable.
• Taken to the extreme, people may spend their time in nutrient pods and virtual worlds, with some form of digital immortality becoming standard. The picture is cyberpunk, but it is not an answer for action today. 庄明浩 leaves not an optimistic guarantee but a question after passing through the fear of the endpoint: “Then what? Do we just stop living? … What comes next is what matters”（“然后的事情才重要”）.