No Middle Ground: AI Rivalry and White-Collar Workers' Engels Pause
Summary
GPT Image 2 has moved text-to-image from “can draw” to “understand the world, then draw it,” leaving design and production-art jobs with the waterline nearly at their necks. From a single prompt, it can recreate a Douyin livestream, a game-collaboration poster, a company logo and copyright line, and even get miHoYo’s capitalization right; when asked to give EVA’s Shinji Ikari a phone, it independently found the Da Nei Mi Tan cover and the words “Da Nei Mi Tan, sincerely for you.” The model was free at the time and led the rankings by a mile. 庄明浩’s conclusion was blunt: design, graphics production and e-commerce creative work are already at neck-high water—“humanity has been utterly routed.”
The controversy over Meta’s acquisition of Manus is not an ordinary exit, but a sign that the China-U.S. AI contest is beginning to eliminate the “middle ground.” By the show’s account, Manus reached $100M in annualized revenue on a monthly price of roughly $20, then received a rumored $2B-$3B all-cash acquisition offer just six months after raising at a roughly $500M valuation. The deal was reportedly closed, proceeds distributed and employees transferred to Meta, but there has been no official confirmation. 庄明浩 believes taking the offer could be rational for the company and its shareholders, while regulators are focused on the precedent set by a team moving from Wuhan to Singapore, taking Benchmark’s money and then selling at a premium: “There is no middle state; everyone is being forced to pick a side.”
The show’s macro map is a bipolar world with weaknesses on both sides: the U.S. holds the algorithmic edge, China holds the Token and cost edge, and there is no third pole or easy winner. Europe, Japan and South Korea will keep trying, but 庄明浩 expects two rulebooks and two ecosystems; the U.S. has already acknowledged that it cannot easily beat China on cost and will therefore work to preserve its technology lead. Meanwhile, Stargate originally targeted 5 GW of capacity coming online in 2025, but the show says actual capacity was only 0.3 GW, while 11 U.S. states were discussing data-center restrictions. Beyond algorithms, power, land and social license are equally binding constraints.
AI video has turned short dramas from a content business into an industrial pipeline priced by the second and culled by ROI. Before Seedance 2.0 launched, Douyin-linked short-drama platforms were spending about RMB30M a day on traffic; in less than two months, the show says that had risen to more than RMB100M, approaching the roughly RMB150M daily average implied by China’s 2025 box office. Alibaba’s “Happy Pony” costs about RMB0.4 per second, while Jimeng’s highest-end version may cost roughly RMB1 per second; even four attempts per successful generation amount to only several hundred yuan per minute. As long as the pipeline delivers a 1.1 ROI, spending keeps rising: “It can’t be stopped.”
“Record revenue, further headcount cuts” is becoming the new operating function for technology companies, and AI replacement has moved from forecast to actual labor action. The show says global tech giants have already cut nearly 100,000 jobs, with some companies reducing headcount by 10%-20%, even as Apple, Meta and Microsoft post record revenue. Google is generating 75% of new code internally with AI, while 庄明浩 estimates Anthropic could be at 90%-95% or higher. Meta is also recording employee clicks and workflows to “distill” its workforce; the show cited 8,000 layoffs followed by another 8,000 in the first wave: “If nothing major goes wrong after the cuts, keep cutting.”
This revolution is hitting the white-collar knowledge system first, and the worst analogy is not a brief period of pain but an “Engels Pause” that could span a generation. During the steam-engine and textile revolutions, GDP grew while worker wages may have stagnated and employment deteriorated for 40-50 years; this wave is directly compressing the execution cost of coding, design, translation, content and research. 庄明浩 rejects the glib idea that “a page of history is a person’s entire life”: “Who wants to be treated as a page?” He also acknowledges that anxiety cannot change the environment. Individuals can only adopt AI faster, find their boundaries and build their own positive feedback loops.
What makes even model researchers vulnerable is not the Chatbot, but L3 Agents evolving toward L4 “researchers.” Harness is not a thin wrapper; it puts a steering wheel, brakes, memory, permissions, task decomposition and acceptance criteria on the model’s F1 engine, with the goal of having the model execute and correct its own work. If Coding is the first field to develop an AlphaGo Zero-style self-training mechanism, people who research AI may no longer be needed in 2 years. 庄明浩 kept returning to one conclusion: “This wave still can’t be stopped.”
The investment theme is broadening from GPUs into optical communications, CPUs, power and domestic adaptation, while the only controllable lever for individuals is to build a small personal feedback loop as quickly as possible. The CPU-to-GPU ratio was roughly 1:8 in training, could reach 1:4 in post-training and approach 1:1 in the Agent era; the show said Intel rose 20% in a single session after its earnings report, reaching a P/E of roughly 100x, while the first-quarter A-share slogan was “believe in light.” DeepSeek V4’s main weapons are domestic-chip compatibility and cost performance, followed by a direct cut to 25% of the original price. For individuals, the strategy is not to predict the end state, but to use OPC, one’s own business and genuine preferences to answer: “Why should anyone choose you?”
Deep dive
1. GPT Image 2 Is Calling on World Knowledge, Not Just Generating Pixels
Four or five months ago, the two hosts were still discussing Nano Banana imitating 井上雄彦’s Vagabond. This time, GPT Image 2 changed the experience from amusing to “hair-raising.” 相征, who studied drawing, described the waterline as up to the chin; 庄明浩 simply said, “It’s terrifying.”
The most representative test used a single prompt: “Draw a middle-aged woman selling one-day trips to Beijing in a Douyin livestream.” The model filled in the viewer count, hearts, gifts, comments, shopping cart and product panel. The result looked like a real livestream rather than an abstract “livestream scene.”
Another poster combining multiple games, including Black Myth: Wukong and Genshin Impact, got not only the characters and composition right, but also each company’s logo and the English copyright line. 相征 was most surprised that miHoYo appeared with lowercase “mi” and uppercase “HOYO”: “I didn’t say anything.”
2. Free Models Are Already Good Enough to Absorb Large Amounts of Commercial Art and Stylized Production
相征 asked only for “EVA’s Shinji Ikari sitting on a train after battle, listening to Da Nei Mi Tan.” The model put the show’s cover on the phone and wrote “Da Nei Mi Tan, sincerely for you” in the background. His only revision was to ask the model to turn the phone slightly away from the camera so the pose looked more natural.
庄明浩 gave the model 3 photos of himself in Dubai and asked it to restyle them in the manner of 井上雄彦, 荒木飞吕彦 and 原哲夫. The 3 images captured realism, clean proportion and hard-edged masculine linework respectively. The model was no longer applying filters; it first had to understand 3 manga artists and then transfer their styles onto the same person.
GPT Image 2 was free at the time, Chinese-text errors had fallen sharply, and it was leading the rankings by a wide margin. 庄明浩 allowed that competitors would catch up quickly, but his conclusion on the current state was direct: Shinji images, advertising pages and full e-commerce-store creative packages could all be handed straight to the model. “Humanity is finished—we’ve been utterly routed” (“人类完了，一败涂地”).
3. Manus Was the First to Package an Agent in a Product Consumers Would Pay For
庄明浩 described Manus as the first application anywhere to make the AI Agent people had imagined look broadly real. It does not train a base model; it calls models such as Claude, orchestrates the workflow and consumes Token, while the user provides the task and collects the result.
The same team had previously built the browser extension Monica, which made foundation-model capabilities easier to use. But 庄明浩 still viewed Monica as close to a “wrapper.” Manus’s step-change was hiding multiple models, multiple steps and final delivery inside a single task.
The product launched around March last year. Its base subscription may have been roughly $20 a month, with additional credit packages. By the show’s account, it was at one point among the fastest AI application companies to reach $100M in annualized revenue, and revenue may later have “doubled.” But high revenue does not equal profitability, since costs such as Token usage were not disclosed.
4. The Wuhan-to-Singapore Capital Path Created the Transaction Risk
Founder 肖弘 had already started and sold 2 companies. 真格基金 had followed him for more than a decade, and Manus was his third startup; Tencent later joined the shareholder base. 庄明浩 stressed that this was not a speculative project assembled overnight, but a team that had spent years iterating on software products.
Manus never operated for users in mainland China from the outset, because consumer AI services in China involve filing and content-review requirements. The team once wrote “stay tuned” on a Chinese-language webpage and said it was working with Qwen, but the international site consistently used overseas models and charged overseas users. The domestic partnership never truly went live.
真格 and Tencent are Chinese institutions, but the investment was in dollars. The company also used a Cayman-Hong Kong-onshore VIE structure. Around May to July last year, U.S.-only fund Benchmark led a new round at a valuation of roughly $500M; the investment was later questioned by the U.S. Congress.
To complete that step, the company moved its core operations from Wuhan to Singapore. The show said roughly 100-plus people relocated, while those who did not move left. 庄明浩 compared it to a “washed crab”: ordinarily a routine corporate arrangement, but after the DeepSeek boom, media coverage elevated Manus into a “national-fortune-level” story. Attention from CCTV and Xinhua meant it was no longer just a company.
5. Meta’s $2B-$3B Offer Put Corporate Rationality on a Collision Course with National Logic
According to unconfirmed market rumors, Meta offered $2B-$3B at the end of last year to acquire Manus outright, in a deal resembling Facebook’s acquisition of Instagram. For a company that had raised at a roughly $500M valuation just 6 months earlier, it was an exceptionally difficult offer to refuse.
庄明浩 repeatedly limited his judgment: viewed only from inside the company, founder and existing shareholder group, accepting could be rational. Regulators, however, are not looking at a several-fold exit return in 6 months. They are looking at the precedent of a Chinese team becoming a Singapore company and then being bought by a U.S. giant.
Based on 庄明浩’s account and the Caixin report he cited, the transaction had probably closed and the money may even have been distributed. Employees had already been working at Meta for months, and the person in charge had reportedly become a business-unit VP. But there are no complete public documents on the amount, completion status or official position, so the show consistently described the deal as rumor and media reporting.
6. Regulators Are Trying to Stop the Precedent, Not a Single Data Leak
Supporters of the deal argue that commercial companies exist to maximize returns and that China should not block a transaction if domestic buyers were unwilling to match the price. Opponents use a defense-industry analogy: if the company made aircraft or missiles, “sell to whoever pays” would not be an acceptable explanation. Neither host tried to erase the disagreement.
庄明浩’s core view is that Manus sits on the front line of the China-U.S. AI contest even though it does not build a base model and has almost no Chinese user data. Once a strategic question becomes a “two-pole contest,” implementation details such as application-layer compliance and data security are no longer decisive. “It wasn’t caused by those issues; it was caused by the core issue.”
If the acquisition goes through, other AI founders may see a viable path: move offshore, become an overseas entity, take U.S. capital and sell to a U.S. giant. From a regulator’s perspective, that demonstration effect could justify action, even a “kill one to warn the others” response.
相征 summarized the setup this way: the company’s original compliance handling had flaws, while the media’s “national-fortune-level” narrative amplified its political visibility. 庄明浩 agreed that this was possible, but placed it within the larger premise: “AI is, to some extent, defense industry.” Reality is forcing everyone to choose a camp before debating the principles.
7. TikTok Showed How Quickly the Key Battlefield Can Abandon Decorum
相征 viewed the TikTok episode as a fundamental blow to his faith in free trade, fairness and justice. The U.S. was not negotiating over price; it was demanding the breakup of a platform used by huge numbers of young people and potentially important as an election battleground. “Today I’m taking yours away.”
周受资 repeatedly explained to Congress that he was Singaporean, yet was still asked whether TikTok connecting to a family Wi-Fi network would leak data and whether it could function without a network connection. 庄明浩 believed there was “no way to prove innocence,” because the conclusion had already been determined by national positions.
The same logic extended to the U.S. gay social platform acquired by Kunlun, Tencent’s wholly owned Riot Games, its roughly 40% stake in Epic Games, and Tencent-backed Roblox. Ordinary players might ask, “Is a game really that important?” The show’s cold answer was: “It is.”
Chip restrictions follow the same playbook. 庄明浩 noted that Singapore appeared to account for roughly 30% of Nvidia’s demand, an unusual figure whose reason was difficult to guess. Once the U.S. threw the first punches on chips and TikTok, China could not simply stand there and absorb them: “If you don’t want decorum, then neither do we.”
8. Software Companies That Have Already Closed a Deal Can Hardly Return to Their Pre-Transaction State
The only official language, as relayed by the show, was reportedly a demand that the parties rescind the transaction and restore the prior state. It did not specify whether “the parties” meant Meta, Manus, the founder or the shareholders, nor how the money or employees would be returned.
庄明浩 criticized self-media accounts that casually wrote “the $2B will be returned through the original channels”: “Where is that channel? Do you know?” The software had been continuously rewritten, AI Coding was accelerating code iteration, employees had been in their new jobs for months, and funds may have passed through multiple layers back to fund investors. Restoring the deal was not like moving a factory back.
Meta could theoretically argue that the transaction complied with U.S. law and that the contract had been performed, but it would still face consequences in China and among overseas advertising clients. Foreign media also reported that Meta might be willing to return the assets. OpenClaw, the open-source version of Manus, prompted jokes that Zuckerberg had “used it for months and could still return it,” but 庄明浩 rejected the idea that Manus’s value had fallen to zero.
9. VIE’s Old Foundation Is Being Torn Up by Two Capital Rulebooks
Since 庄明浩 entered the internet industry in 2009, the VIE model pioneered by Sina had been the default path: Chinese companies raised dollars, burned money onshore, then listed in the U.S. or Hong Kong, allowing global investors to participate and exit. “Today we’ve discovered that the foundation itself is being pulled out.”
Didi went public in the U.S. because of cash-flow pressure and the arrival of its listing window, but failed to coordinate with the relevant authorities. Its business also directly involved transportation, public welfare and massive amounts of data, leading to regulatory intervention and app removals and severely affecting Chinese companies’ access to U.S. listings. The Manus episode involves the foreign-investment security-review rules that require the transaction to be rescinded and the prior state restored.
The show noted that Zhipu and MiniMax have listed in Hong Kong, while Kimi and StepFun were rumored to be preparing listings. Foreign media reported that regulators wanted the latter companies to dismantle their existing dollar structures before listing and remove the shares held by purported U.S. investors. 庄明浩 explicitly warned that these reports were “not guaranteed to be true”; the request may amount only to verbal guidance or an implied signal.
His conclusion is that this is not a discrete “threat” to resolve, but an “environment” to adapt to. A startup must either be fully and cleanly on one side from day one or fully and cleanly on the other. But even the most senior lawyers cannot clearly define what “clean” means.
10. Founders Are Being Forced to Answer Not a Product Question, but Which Side They Stand On
肖弘 is simply a young programmer who graduated from Huazhong University of Science and Technology, built software companies in succession and sold them. Yet he has had to confront the question “Which side are you on in the China-U.S. contest?”—a question far outside ordinary startup logic. 相征’s dark joke: “If he can pick the right character in League of Legends, that’s already the limit of his abilities.”
The more practical contradiction is that the show says 50% of Meta’s engineers are Chinese, while the China-U.S. AI contest inherently involves both Chinese people in the U.S. and talent in China. They cannot naturally be divided into 2 groups, but national competition demands that everyone choose. Even splitting onshore and offshore operations into 2 brands may have “no practical meaning” once the structure is scrutinized.
11. China and the U.S. Will Probably Occupy Different Quadrants, but Neither Can Win Outright
庄明浩’s coordinate system has algorithms and technical R&D on one axis, and Token and compute cost on the other. China may own the low-cost, large-scale supply side, while the U.S. maintains the lead in models and research. He sees a complete victory for either China or the U.S. as “unlikely.”
Europe is still trying, as are Japan and South Korea, but the show sees no genuine third pole in the near term. 相征 relayed the idea that China and the U.S. could fight while jointly harvesting the rest of the world; 庄明浩 did not treat that as a conclusion, but agreed with the bipolar structure itself.
The U.S. strategy is therefore to acknowledge that it will struggle to beat China on cost while preserving its algorithmic lead, using looser approvals for land and data centers to narrow the cost gap. China will turn compute, Token scale and power into advantages. Neither side can publicly concede.
12. Data-Center Expansion Is Running Into Electricity Prices and Politics Before Algorithms
The show said OpenAI’s “Stargate” project, advanced with the Trump administration, SoftBank and Oracle, originally targeted 5 GW of data-center capacity coming online in 2025. Actual capacity was only about 0.3 GW, less than 10% of the target. Physical construction, power supply, permits and land cannot move monthly like model updates.
11 U.S. states have discussed pausing or banning data-center construction. 庄明浩 used Minnesota as an example: the state legislature passed a bill pausing new construction until November 2027, but the governor had not signed it and might oppose it. Over the past 3 years, Minnesota had the fastest-rising electricity prices among the 50 states.
For residents, the issue is not abstract AGI but “I can’t afford to run my air conditioner.” For governors, rejecting projects means investment and industry may move to another state. Environmental protection, animal protection and land permits are also constraints, while storage and memory capacity expansion require another 2-3 years.
Social conflict has already moved beyond policy debate. The show said Sam Altman’s home had been attacked twice, once with an incendiary bottle and once in a shooting; many anti-AI groups have also emerged across the U.S. The hosts used this to explain that fairness, justice and freedom are expensive to maintain. When jobs, electricity bills and great-power competition all apply pressure, decorum is the first thing to go.
13. While iQIYI Faced a Public Trial, RedNote Was Already Executing a Colder Version
After iQIYI CEO 龚宇 discussed AI copyright for actors, a trending topic read “iQIYI has lost its mind,” and the stock was described as falling off a cliff. 相征 even speculated that it might already be below RMB1. Long-form video remains tied to actors, professional production and legacy copyright structures, so public anger is directed at it.
By contrast, the ByteDance ecosystem—Douyin, Douyin Lite, Tomato Audiobooks, Soda Music and RedNote short dramas—could reach an estimated 900M-1.1B users, according to the show. RedNote’s share of the short-drama market was estimated at roughly 80%, effectively a monopoly.
At a government-hosted online audio-visual conference, RedNote announced RMB500M in support for live-action short dramas. Almost simultaneously, the app rankings removed the distinction between AI dramas and live-action dramas. 相征’s read was that the RMB500M represented “the attitude from above,” while the product move was candid: the AI wave has arrived.
14. Seedance 2.0 Pushed Short-Drama Traffic Spending to Box-Office Scale
Before Seedance 2.0 launched, Douyin-linked short dramas were spending roughly RMB30M a day on traffic. In less than 2 months, the show said that had risen to more than RMB100M, nearly 4x. Once technology and content quality cross a threshold, capital does not migrate gradually; it floods in immediately.
For comparison, China’s total box office in 2025 was more than RMB50B, or roughly RMB150M a day. Short dramas had already surpassed movie box office in 2024, and this AI-short-drama wave exceeded that daily scale within months and kept growing.
庄明浩 described leading companies with roughly 5,000 employees, 800 teams and about 5 people per team—roughly 600 lines running simultaneously: script to image, image to storyboard, storyboard to video cut, and video cut to final production.
Finished content enters a “time furnace.” Traffic buyers purchase distribution to test it, and as long as ROI reaches 1.1—RMB100 spent to bring back RMB110—they keep increasing the budget. Aesthetic quality is not the first gate; the ability to generate returns is.
15. Once Content Is Priced by the Second, Live Action Starts to Look Like a Luxury Good
Around 2023, a 30-plus-episode short drama running roughly 100 minutes cost about RMB50,000-RMB100,000. As the industry heated up, that rose to RMB300,000-RMB500,000; by the end of last year, a top live-action short drama could cost RMB3M, or roughly RMB30,000 per minute. 庄明浩 called that “the final glory.”
AI video takes the price directly down to the second: Alibaba’s “Happy Pony” costs about RMB0.4 per second, while Jimeng’s most expensive version may cost roughly RMB1 per second. Even assuming 4 attempts for 1 successful generation, the cost is still only several hundred yuan per minute. “Can it get any cheaper than a few mao for my one second?”
The cost structure may shift from actors, locations, cinematography and directors to roughly one-third script, one-third generation attempts and one-third production. Fantasy, period drama, mecha and outer-space themes—the most expensive genres in the past—may actually be the best suited to generation.
相征 mentioned a Hengdian “king of drama,” an actor who landed endless work during the live-action short-drama boom because he was suited to playing fathers and grandfathers, only to see his work suddenly fall to zero. When the role is unimportant, the appearance can be described and even the lowest price cannot beat generation, live action becomes a luxury good.
16. AI Is Eliminating Old Jobs While Handing Creative Licenses to Smaller Teams
相征 once got hooked on a Douyin drama about a forest girl meeting elves and monsters. The effects were “too delicate” for him to understand how free content could afford them, until he saw the label “This video was generated by AI.” His surprise shows that viewers may now be drawn in by the content before realizing how it was produced.
The mood at the Game Developers Conference has also changed. Last year the discussion was whether to use AI; this year the default answer is already yes, with the debate focused only on how to use it more efficiently. Small teams can therefore make the mid-sized games once reserved for large teams, and mid-sized teams can make large projects. Conversely, large teams themselves may no longer be necessary.
The show reused an earlier rough estimate: a project like Detroit: Become Human might once have required tens of millions of dollars, while an AI version could be brought down to hundreds of thousands or a few million yuan. The figure is not precise; the point is that young creators can finally ask “What do I want to make?” before cameras, actors and post-production budgets veto the idea.
17. Software Scarcity Is Moving from Execution to Intent Design
The narrative behind the decline in U.S. SaaS stocks over the past 1-2 quarters is straightforward: if users only need to tell AI what they want and it can build a tool on the fly, why keep paying for expensive software subscriptions? 相征 pays several thousand yuan a year for Adobe and asks: “Why?”
A triangle described by a senior Anthropic executive is inverting. In the past, ideas accounted for 10%, while implementation, maintenance and other execution accounted for 90%. In the future, requirements, purpose, boundaries and audience fit will account for 90%, while execution falls to 10% and is handed to AI. What is scarce is no longer the ability to make something, but knowing why to make it, whom it is for and where to stop.
18. Tech Giants Have Proved That Layoffs Can Coexist with Record Revenue
庄明浩 said global tech giants have already cut nearly 100,000 jobs, and 10%-20% reductions at a single company are no longer minor adjustments. Model capabilities have crossed the thresholds for language, Coding and multimodality at the same time, forcing management teams to recalculate labor needs.
Several tech giants were releasing quarterly results on the night of the recording. Apple, Meta and Microsoft all posted record revenue. 相征 completed the causal chain from the boss’s perspective: “Revenue hits a record, then I cut people and find there’s no impact. Why wouldn’t I keep cutting?”
Meta’s approach is more direct: record every employee click, action and workflow, then use AI to train a replicable behavioral double. The show mentioned 8,000 Meta layoffs, followed by another 8,000 in the first wave. If the distillation works for a period and the business has no major issues, the cuts will continue.
China already has unnamed examples of employee doubles. After a colleague left, the account continued replying in a similar tone: “I’m the digital double of [colleague].” The show had also previously mentioned a large company whose game revenue was growing even as much of its several-thousand-person outsourcing team was cut. The first phase of layoffs was largely completed before May 1: “The pace of layoffs has to be fast.”
19. White-Collar Workers May Be Entering Their Own “Engels Pause”
Earlier industrial revolutions mostly replaced muscle and basic productive capacity. This one is going straight at the white-collar knowledge system: language understanding, code, images and video are all improving together, and tasks once performed by people can already be executed faster and more cheaply by Agents.
“Engels Pause” describes the 40-50 years in which steam engines and textile machinery drove rapid GDP growth while workers’ wages stagnated or fell. 庄明浩 believes an entire generation may not live to see the new equilibrium, and condemned the phrase “a page of history is a person’s entire life”: “Who wants to be treated as a page?”
His response is not grandiose: use AI more, feel out its boundaries and improve personal efficiency. First determine what you want and whom you envy, then break down whether part of that path can be replicated. Anxiety is rational, but “anxiety is useless here”; the only real lever is the individual’s micro-level choice.
20. An L4 “Researcher” Is Making Even Frontier AI Researchers Worry About Their Jobs
相征 had assumed that the people training the top models would be safest. 庄明浩 said researchers at leading model companies were already worried they might not be needed in 2 years. They sit closest to the frontier and therefore see the limits of model capabilities first.
He summarized OpenAI’s levels as follows: L1 is the Chatbot, L2 is reasoning, L3 is the Agent and L4 is the “researcher.” The industry is now at L3, while L4 would perform the work researchers do today—figuring out how AI can improve AI faster.
R&D is therefore shifting toward “autonomous evolution”: define the task, set the evaluation criteria, let the model execute and feed the result into the next round. The logic is similar to AlphaGo Zero, which stopped learning from human games and played against itself. Coding is likely to be the first field to cross the threshold.
21. Harness Turns the Model Engine into a Self-Driving Vehicle
庄明浩 explained Harness with a car analogy: a reasoning model is an F1 engine, but real work still requires wheels, a driveshaft, steering, brakes and a chassis. The system must specify how tasks are decomposed, what tools can be called, how far permissions extend and how results are accepted.
OpenClaw, Codex and Claude Code all belong to this outer layer, but they are not simple “skins.” What memory is stored, how long it is kept, how much weight it receives, how context persists and when the system communicates with a human will all determine how far the same model can run.
The show said 75% of new code inside Google is already generated by AI. 庄明浩 estimated that Anthropic may be at 90%-95% or higher, which could explain how Anthropic released more than 100 features and products in the past 3 months.
The example of a small software tool is more revealing. An event has 120 seats arranged in a 10×12 grid; tell the model to randomly draw seat numbers and it can build a raffle tool in minutes. Tasks that once required buying a mini-program or hiring a developer no longer even require a complex model.
22. The Agent-Era Trade Is Broadening from GPUs to Optics and CPUs
After memory and hard-drive production could not expand fast enough, the new bottleneck in the first quarter became data transfer between cards. With 100,000-plus cards exchanging data simultaneously, traditional cables are too slow. Optical communications, optical chips, optical transmission and related suppliers immediately became an A-share theme: “Everyone needs to believe in light.”
Asked whether the hottest A-share company recently was 中际旭创, 庄明浩 said yes. His point was not a single stock, but the fact that once consensus forms, the market “has to overcorrect”—pricing in the future trend first rather than waiting for current production, revenue and profit.
Intel rose roughly 20% in a single evening after its earnings report, finally exceeding its price during the 2000 internet bubble, with a P/E of roughly 100x. AMD rose as well. CPUs, previously considered laggards, regained a narrative because Agents depend more heavily than training on task orchestration.
The CPU-to-GPU ratio was about 1:8 in training, may reach 1:4 in reinforcement learning and post-training, and could rise to 1:1 in the Agent era. Complex tasks must be split, assigned and verified, requiring a CPU “head chef” to coordinate while GPUs act more like the line cooks chopping and sautéing.
23. DeepSeek V4’s Weapons Are Domestic Compatibility and Price, Not a One-Off Shock
庄明浩 believes people should no longer expect DeepSeek V4 to recreate the “big move” of its V2.1 launch. Model competition has entered a monthly-update cycle; holding back for a single long-cycle release cannot keep pace.
V4 focuses on language, Agents and Coding, with stronger compatibility for ecosystems such as OpenClaw. As of the discussion, it still had no multimodal capability. 庄明浩 calls language, Coding and multimodality an “impossible triangle” that makes it difficult for any provider to lead in all 3 simultaneously.
The more direct advantage is domestic-chip compatibility and cost performance. The official launch price was competitive but not dramatically different; several days later, the company cut it directly to 25% of the original, reducing cost to one-quarter. That is DeepSeek’s strategic significance within China’s Token supply system.
Jensen Huang’s argument against restricting Nvidia sales to China fits the same logic: export controls force Chinese model and chip companies to refine their products together under constrained conditions. Once they succeed, they may create an even stronger domestic substitute. 相征 put the argument more bluntly: “Wasn’t all of this caused by the U.S. government?”
24. OPC’s Real Foundation Is Not Infinite Productivity, but “Who You Are”
As productivity in some areas approaches unlimited supply, OPC—doing something on one’s own—is being revisited because AI can fill in execution that once required a team. But 庄明浩 dislikes presenting it as easy entrepreneurship. What matters is identity, preference and clear choices.
The new equilibrium may take 40 years to arrive. The more realistic task for this generation is to build a small loop that can keep running. Even a rough first step can gradually compound income, influence and agency if it generates positive feedback beyond the step itself.
People used to define themselves through companies, jobs, income and assets. The next question may be the reverse: “Who were you before all that?” The philosophical reminder cited by the show is that people are not wrenches or pliers; they arrive in the world by accident and ultimately answer only to death. Every deliberate choice therefore deserves to be taken seriously.
25. Once Content Is Infinite, “Why Listen to You?” Becomes Scarcer Than Production Capacity
AI music is now strong enough that tens of thousands of tracks may be uploaded each day, yet almost nobody seems to listen. 庄明浩 therefore asks: when supply approaches infinity, why would a user listen to any one track? The creator’s purpose, identity and relationship with the audience may matter “several orders of magnitude” more than production quality.
A company 相征 knew once used several hundred people to manufacture Douyin hits. Team leads set the theme each day, each worker submitted 1-2 songs, and the songs were recorded, uploaded and promoted the same day. Once the data came back, the budget was increased; a small number of tracks ultimately became the public soundtrack of an entire year.
That factory is now using AI as well. Hundreds of “music workers” are already being replaced, not merely at risk of replacement in the future. The industry can no longer be observed quarter by quarter; model, algorithm or operations changes can send it from one extreme to the other in 1-2 months.
26. The Effective Educational Response Is Not to Ban AI, but to Let Children Spar with It First
庄明浩 has a child about to enter seventh grade practice English debate directly with AI. Since third grade, the child had relied on teachers for prompts, simulations and critiques; now AI can serve as both opponent and source of new arguments. “If you can beat it in debate, humanity is basically fine.”
He used 樊麾’s 10-game internal test against early AlphaGo to illustrate this kind of training. On the first day, after losing an official game to a blunder, 樊麾 won a fast-paced test game and still thought the machine was “not that impressive.” From the second day onward, however, consecutive losses rapidly turned confidence into doubt.
By the third and fourth days, his confidence had completely collapsed. Even amateur players at DeepMind could not understand why he was making those moves. Against a machine without emotion, every tension, suspicion and piece of feedback was reflected straight back at him. By the end of the fifth day, a 3-time European champion had become someone wondering whether he even knew how to play Go.
The industry later suspected that 樊麾 had been paid to lose. He instead joined DeepMind and waited for 李世石 to validate the result. The value of the episode was not merely that the machine was stronger, but that it showed how humans confronting superiority move through contempt, rationalization, psychological collapse and a renewed understanding of themselves.
27. 李世石’s Fourth-Game Win Was the Show’s Nonstandard Answer for Humanity
In the 5-game match against 李世石, the first game could still be explained as a small mistake. In the second, when AlphaGo played move 37, professional players initially thought the move was random. 李世石 may even have smiled faintly after returning from a cigarette break, but grew increasingly grave and spent roughly 13 minutes considering his response.
After losing the second and third games, he sounded close to tears at the press conference; his existing understanding had been shattered. He remained behind for much of the fourth game, then played a move that surprised every commentator, possibly exposing a weakness in that version of the system. After suddenly taking the lead, he held on to win.
樊麾 ran down from the referee’s seat and gave him a thumbs-up. 李世石 said the game was the most precious asset of his life and that the move could not have been exchanged for anything. 庄明浩 cared less about whether humans could ultimately win than about the fact that, after his confidence had been destroyed, one person could still make an irreplaceable choice through his own agency.
Ten years later, a Korean program brought 李世石 together with League of Legends GOAT Faker. Elon Musk’s Grok is trying to challenge T1 through vision and keyboard-and-mouse control, while Google is teaching models frame-by-frame reactions in King of Fighters. If one day it becomes impossible to tell whether the ID on the other side belongs to a person or an AI, 相征 asked what to do. The honest answer from both hosts remains: “We don’t know.”