Vol. 89 AI Industry 2025 Annual Review Supplement (The “We’re Not Waiting for V4” Edition) — 70-Page PPT, Solo

Summary

• 庄明浩’s topline call is that AI crossed the application threshold again around the 2026 Lunar New Year: “It can’t be stopped now” (挡不住了). OpenClaw gave Agents access to local files and instant-messaging channels, Seedance 2.0 pushed video generation toward industrial restructuring, and Chinese models entered a dense release cycle; yet genuinely high-frequency AI users globally remain a tiny minority, so “everything, every last thing, is really only just beginning.”
• Microsoft, Google, Meta and Amazon are expected to push 2026 capex to $630B, up 68% year on year. The four spent $222.8B in 2024 and $376B in 2025, accelerating for 3 straight years; Meta’s capex now exceeds 54% of revenue, while the tech giants’ bond issuance is growing faster than investment, with Google even issuing century bonds—“everyone is worried, but their spending says otherwise.”
• NVIDIA’s problem is no longer earnings, but what can create incremental expectations after a run of consecutive perfection. Quarterly revenue reached $68.1B, up 73%, gross margin rose to 75% and profit topped $43B, yet the stock still fell 5.3% the next day; after 14 straight quarters beating expectations, “beating expectations has become ordinary,” with shares range-bound near $180 for 6 months and the real indicators shifting to GPU rental rates, depreciation useful lives and ecosystem financing.
• What OpenClaw revealed is not demand that can be satisfied by 1M users, but a potential 1,000x compute shortfall once everyone has an Agent. 庄明浩 estimates deployed users in China may number only several hundred thousand and are unlikely to reach 1M, yet have already congested model, cloud and API platforms; covering more than 1B people would require the user base to expand roughly 1,000x, while task duration and complexity would continue rising—“scale today’s compute investment another thousandfold and the number becomes utterly unimaginable.”
• Software stocks have already priced a long-term scenario in which traditional SaaS protects its installed base while Agents take the incremental growth. Software and SaaS ETFs have fallen 20–30% over several months, and SAP, Salesforce and Workday have received no valuation reward despite continued growth; the “2028 doomsday thesis” links layoffs, faster AI adoption, white-collar consumption, intermediary profits and debt crises into a closed loop, but 庄明浩 draws a clear line: capex, debt, penetration and rising unemployment are already happening, while ROI collapse, consumer-spending collapse and large-scale debt blowups are not.
• OpenAI’s announced $110B financing has only $35B in true upfront funding; much of the balance depends on AGI, an IPO or compute-buildout milestones. Amazon pledged $50B, NVIDIA and SoftBank $30B each, but the initial tranches are only $15B, $10B and $10B respectively; meanwhile OpenAI’s weekly active users reached 900M, 2030 revenue is expected at $284B, cumulative burn could reach $218B, and capital markets are already trading an IPO in 2026–2027.
• Chinese open-source models’ usage and leaderboard positions have shifted from a catch-up narrative to observable share gains. In February, MiniMax M2.5 processed 5.44T tokens on OpenRouter and Kimi K2.5 processed 4.2T; 8 of the top 10 open-source text models were Chinese, and all 6 leading Code models were Chinese; more disruptive still was Seedance 2.0, which 冯骥 called “Kill the game” and said would push the cost of general video production toward marginal compute cost.
• 智谱 and MiniMax rose from roughly HK$40B–50B to peaks near HK$300B, and valuation spillovers have reached private markets. Using 1% of OpenAI’s roughly $840B valuation as a “unit of account,” the two briefly rose from 1 unit to nearly 5; Kimi then raised $700M at a $10B–$12B valuation, StepFun completed RMB5B in financing and was reported to be preparing a Hong Kong listing, but contemporaneous declines in SenseTime, Unisound and others remind investors that “we see the new names’ smiles, not the old names’ tears” is not the full picture.

Deep dive

1. “It Can’t Be Stopped Now” Is a Threshold Call, Not an Emotional Slogan

• The PPT had originally set aside part of the deck for DeepSeek V4, which still had not launched by March 9. 庄明浩 decided to publish the “We’re Not Waiting for V4” edition first, because the industry is moving so fast that if the material did not go out now, it could already be obsolete again.

• Looking back through his Lunar New Year chat logs, he found himself repeating the same 4 characters: “It can’t be stopped now.” Many developments are no longer experiments or waiting for the final barrier to give way; model capability, application experience, capital investment and organizational change are accelerating at once.

• The piece maintains its U.S.-China comparison throughout: on the U.S. side, capex, NVIDIA, OpenClaw, software stocks and OpenAI financing; on the China side, dense model launches, Seedance 2.0, DeepSeek V4, red-envelope campaigns and the capital-market pricing of model companies.

2. The Four Giants Have Pushed 2026 Capex to $630B

• Capex at Microsoft, Google, Meta and Amazon rose from $147B in 2023 to $222.8B in 2024, up 55%; it increased another 65% to $376B in 2025, and 2026 guidance is up 68% again to $630B, or more than $150B per company on average.

• 庄明浩 emphasizes that this is not acceleration from a low base, but another step-up after “2025 was already an absolutely crazy year.” Over the past 3 years, annual investment has grown roughly 1.72–1.78x, with an increasingly material contribution to U.S. GDP.

• The capex-to-revenue ratios are equally aggressive: more than 54% at Meta, roughly 40% at Google, 33% at Microsoft, 25% at Amazon and just 3% at Apple. In the latest quarter, Amazon, Microsoft, Google and Meta raised capex by 44%, 89%, 95% and 48% respectively, while Apple cut it by 19%.

3. The Capex Trade Has Spread to Bonds, TOTO and Apple’s Strategic Choice

• In 2025, the tech giants’ bond issuance grew faster than capex, and Google even issued a century bond. The maturity sounds absurd, but investors bought it out rapidly because the yield was attractive. 庄明浩’s warning remains: “You never know what debt will set off.”

• The stock-market narrative moved from NVIDIA to data centers, power and storage, then spread to TOTO. The market story was that its ceramic materials could be used in trays for GPUs inside data-center racks, sending the stock up nearly 40% at one point this year; listeners were often left with an “ah” or a wry smile, because the AI chain had been pushed all the way to a toilet maker.

• Apple remains either the most elegant contrarian or the clearest cycle miss—there is still no answer. It has confirmed a partnership with Google on AI models and is rumored to be considering hosting the related data on Google Cloud; if Apple wants to reverse its weakest growth trajectory, whether Cook’s successor makes AI one of the new CEO’s first 3 big bets could be a key 2026 variable.

4. NVIDIA Keeps Delivering, but the Stock No Longer Rewards Perfection

• NVIDIA’s quarterly revenue was $68.1B, up 73% year on year; gross margin rose from roughly 72% in the prior quarter to 75%, and quarterly profit topped $43B. Data center is now overwhelmingly dominant, while gaming GPU revenue has fallen to a single-digit share.

• The next day, NVIDIA fell 5.3%, Broadcom fell 7%, and other mega-cap tech sold off broadly. 庄明浩 says 1 day cannot explain everything, but over the past 2 quarters, “no matter how explosive the earnings were, the market was not buying it as much.”

• From roughly $140–$150 early in 2025, through a 17% one-day drop after DeepSeek R1 launched, a tariff-driven fall below $80 and a rebound to around $180, NVIDIA has since traded sideways for nearly 6 months. 庄明浩’s earlier view was that below $100 was a buy with your eyes closed; the question now is whether the range-bound tape marks a peak or the stock is “brewing something big.”

• The company has beaten expectations for 14 consecutive quarters, spanning the 3.5 years since ChatGPT launched. “If every quarter beats expectations, then it seems not to be a beat anymore.” Revenue, profit and forward revenue expectations are unlikely to provide much new information, so the market needs to find the next set of leading indicators.

5. The Scale of “Overinvestment” Is Undefined, and Common Metrics Are Being Warped by Ecosystem Strategy

• Every technology revolution goes through infrastructure overbuilding, but no one knows whether AI’s “overinvestment” means 10x, 100x or 1,000x the current scale. Demand urgency appears nowhere near gone, yet current investment is already hitting human constraints in power, construction and financing: “Can you imagine another 1,000x?”

• Lease prices for older cards such as the A100 and H100 have remained stable for more than a year, showing that legacy compute is still undersupplied. But NVIDIA is also supporting neoclouds and investing externally, including through CoreWeave, Intel and Nokia, to keep the ecosystem stable; that means GPU rental prices may be “controlled to some extent by NVIDIA.”

• Competitors are using similar playbooks. After its deal with OpenAI, AMD formed a roughly $100B partnership with Meta, tied to options on as much as roughly 10% of the company; Google is using TPU investment, lending and support for new neoclouds. The routes differ, but the goal is the same: turn chips into a complete ecosystem rather than simply sell hardware.

• Gamers have become the casualty of the supply allocation. NVIDIA barely discussed gaming at CES; market rumors say it will not launch a new 60-series this year and may even restart production of the 3060. 庄明浩’s conclusion is blunt: “If you were 黄仁勋, you would do the same. The gamers just have to take the pain.”

6. OpenClaw’s Breakthrough Is Permissions and Access, Not the Agent Concept

• OpenClaw launched at the end of last year but began runaway growth from late January onward. When 庄明浩 made his chart, it had more than 200K GitHub stars and had surpassed Linux in a very short time, making it what he called the most-starred project in GitHub history. He also warned that the concentrated consensus contained an element of “deification.”

• Giving an Agent a computer is not new: Manus, cloud-based virtual machines and multiple local PC-control products had already tried similar approaches. NotebookLM summarized OpenClaw as “crossing the engineering chasm, from chat to getting work done,” but 庄明浩 pushed back: “That can’t be right—wasn’t that exactly the story Agents were telling all along?”

• The real difference is the engineering stack: memory is written to Markdown files, tool calls execute actions, and the system reads local files directly; elevated system permissions give AI “hands and eyes.” It also plugs into existing messaging channels—WeCom, QQ, Feishu and Telegram—so sending a message feels materially different from opening a chatbot.

7. A Single “Are You Still There?” Gives a Local Agent a Companion-Like Quality

• 庄明浩 first used MiniMax’s cloud-hosted version, then installed OpenClaw on an idle Windows laptop, running it with the Kimi model and purchased tokens. He wanted it to run continuously for 72 hours, but the system settings triggered an automatic reboot every day; after coming home 2 nights in a row, he found the Agent offline.

• On the third day, he asked OpenClaw to change the reboot settings and ensure that the software would auto-start after a restart. At 12:30 a.m., he sent “Are you still there?” on Feishu. It replied that the settings were complete and there was no need to worry, then added: “Go to bed early.”

• The experience led him to define OpenClaw as “essentially an AI companion product.” Its value is not just the task result, but leaving the office with something on your mind—a continuously working presence you know is still there.

8. OpenClaw Exposes a Token Shortfall, an Ecosystem Bubble and a Product Opportunity

• 庄明浩 maps the application breakout onto Anthropic’s product timeline: Claude Code opened up coding in February 2025; Claude released “16” in October; Opus 4.5 arrived in November; Claude Code Work in January 2026; and Opus 4.6 in February. Once technology crosses a threshold and unlocks a use case, use-case growth feeds back into the model.

• OpenClaw’s founder quickly aligned with OpenAI, while the side project Moltbook became “Reddit for lobsters.” Entrepreneurs began building “something for Agents,” Web3 users recast it as Web 4.0, and the author publicly refused to issue a token; 孙宇晨 nevertheless announced “All in Web4.0” on February 23, after which the weirdness multiplied.

• Statistics show token demand after the breakout growing more than 10% per week, even 15%; Opus 4.6 can execute tasks continuously for more than 14 hours. Chinese OpenClaw deployments may number only several hundred thousand and are unlikely to reach 1M, yet have already exposed compute shortages at cloud providers, model companies and API platforms of all sizes.

• The product is still like an EV that is fully autonomous on a left turn but makes you tighten the screws yourself on a right turn: capability is wildly uneven. 庄明浩 therefore holds both views in play—products that lower the installation barrier will be everywhere over the next 1–2 months, but “if I learn slowly, I can simply avoid learning”; mastering complex operations too early may become obsolete as quickly as learning Stable Diffusion prompt keywords did back then.

9. Software Stocks Have Already Priced in “Agents Eating the Incremental Growth”

• Over the past several quarters, software, SaaS and cloud ETFs have fallen roughly 20–30%, while data-center hardware and storage stocks have risen sharply. SAP, Salesforce and Workday continue to grow revenue and even profit, but their share prices keep falling, reflecting long-term market-size expectations rather than current earnings.

• The concern is not that existing demand will disappear, but that individuals and enterprises can use Web Coding and Agents to build personalized tools, leaving traditional software with part of its installed base while Agents capture new demand. “Software ate the world, but AI will eat software” has become the core expression of this repricing.

• Coding has been first to prove the path, while product design, finance, data analysis, marketing and legal have continued to produce high-adoption use cases. 庄明浩 does not equate short-term share prices with a certain outcome, but “replacement, augmentation—or something else, who knows?” The capability trajectory is clearly moving toward more bearish software scenarios.

10. The “2028 Doomsday Theory” Has a Coherent Transmission Chain, but the End Point Hasn’t Arrived

• The first chain is: AI replaces white-collar tasks, forcing companies to lay off workers; layoffs then force companies to adopt AI more aggressively, creating a “the more layoffs, the more AI” loop that ultimately hits consumption funded by white-collar income.

• The second chain targets intermediaries whose profits depend on friction: insurance, real estate, travel, two-sided platforms and even Visa and Mastercard could lose profits as Agents eliminate friction. Corporate debt and white-collar personal debt then deteriorate, ultimately becoming a financial and even economic crisis.

• On the day the article was published, Visa, Mastercard, DoorDash, ServiceNow and Blackstone all fell by anywhere from several percentage points to the low teens. The next day, Block announced layoffs affecting 4,500 of roughly 10,000 employees, and the stock jumped more than 20% on the spot. “All the narratives collided.”

• 庄明浩 retains 2 objections: the article underestimates government intervention, and technology diffusion need not happen instantly. After ATMs appeared, bank tellers took nearly 40 years to enter a sustained downtrend. If AI compresses 40 years into 10, the pain will be more concentrated. What has already happened is overinvestment, rising debt, AI penetration, falling software stocks and a higher unemployment rate; what has not happened is ROI collapse, the disappearance of intermediary profits, a white-collar consumption collapse or large-scale debt blowups.

11. OpenAI’s $110B Financing Has Only $35B in True Upfront Funding

• The financing consists of $50B pledged by Amazon and $30B each from NVIDIA and SoftBank, for a total of $110B. Amazon deepening its ties after Microsoft was no longer exclusive is not surprising; SoftBank secured an allocation but still has to find funding for later commitments.

• NVIDIA had previously promised up to $100B of investment, structured at $10B for every 1GW built. As construction progressed, OpenAI’s valuation kept changing and each tranche would represent a different share count, making the mechanics too complicated. 黄仁勋 ultimately chose to earmark a $30B allocation at the current round price, but under the $10B-per-GW formula, the actual first tranche is $10B.

• Of the $110B on paper, only $35B is upfront: Amazon pays $15B first, with the remaining $35B triggered if OpenAI achieves AGI or launches an IPO; NVIDIA’s first $10B remains tied to compute, while SoftBank’s first $10B and subsequent funding also depend on IPO milestones.

• OpenAI said its weekly active users had risen from 800M to 900M. It forecasts 2030 revenue of $284B, but cumulative burn could reach $218B. Prediction-market odds put the chance of an IPO by the end of 2026, before January 1, 2027, at roughly 55%, and by June 1, 2027, at roughly 72%; the highest-odds range for first-day market cap is $750B–$1T, but even that is only around 10%.

12. Anthropic Is Closing on, and in Some Pockets Overtaking, OpenAI in Enterprise

• 庄明浩 cites Anthropic ARR of roughly $11.9B versus about $25B for OpenAI, with the gap narrowing rapidly; some forecasts even put Anthropic revenue ahead of OpenAI by mid-2026. This is not a model leaderboard story, but a shift in enterprise payment mix.

• Enterprise payment platform Ramp’s data is more extreme: Anthropic has more than 60% of AI Chat for Business, and nearly 90% of API-related enterprise spending. Because the data comes from actual corporate payments, 庄明浩 considers it more alarming than ordinary surveys or third-party traffic estimates: “What happens if Anthropic’s revenue exceeds OpenAI’s in 2026?”

13. Chinese Models Have Entered a Dense Major-Release Cycle Every 6 Months

• The cycle began with ERNIE 5.0 on January 22, followed by Kimi K2.5 on January 27, Step 3.5 on February 2, GLM-5 on February 11, MiniMax M2.5 on February 13, the full Doubao Seedance 2.0 lineup on February 14 and Qwen 3.5 Max on February 16. Most were major releases carrying “5.0” or “.5” version labels.

• 庄明浩 believes the Lunar New Year clustering was not accidental, but it was not an isolated event either: a similar window appeared in July 2025, and a cadence of major releases every 6 months is taking shape. GPT has reached 5.4, with 3 versions released in a few months; frequent launches have themselves become normal.

• In OpenRouter’s full-February token statistics, MiniMax M2.5 ranked first with 5.44T, Kimi K2.5 second with 4.2T, DeepSeek V3.2 fourth, GLM-5 eighth, MiniMax M2.1 twelfth and Step 3.5 fifteenth. OpenClaw contributed incremental demand, and whether the data represents the entire internet remains debatable, but the trend is clear.

• 8 of the top 10 models on Arena’s open-source text leaderboard were Chinese, while the top 6 on the Coding leaderboard were all Chinese: GLM-5, Kimi K2.5, MiniMax M2.5, Qwen 3.5, DeepSeek V3.2 and Xiaomi MiMo V2. Anthropic later said DeepSeek, Kimi and MiniMax had distilled Claude through large numbers of accounts, but whether distillation counts as theft remains unanswered.

14. 林俊旸’s Departure Puts Model Competition Back on Organizational Capability

• At an AI event whose audience had already been heavily filtered for interest, 庄明浩 asked who had known 林俊旸 before the news; only around 3 people out of more than 100 raised a hand. After his departure, Twitter filled with expressions of respect and recycled the line from the OpenAI controversy: “Qwen is nothing without its people.”

• 庄明浩 is more interested in the conflict between technical-team goals and organizational requirements from Alibaba Cloud and Alibaba Group. 林俊旸 had previously proposed “three in, three out” and native multimodality; if pretraining, post-training, reinforcement learning, text, image and speech are split across multiple organizations, whether they can still support a unified model has no standard answer.

• Subsequent reporting also pointed to compute constraints at the general-model team, which 庄明浩 compares with early OpenAI team conflicts. The questions left behind include what role a model team should play inside a large company, how core personnel should govern an open-source community, whether the community’s KOMs should speak publicly, and whether “fit” matters more than abstract right and wrong.

• A listener who had not yet joined Alibaba to work on AI strategy used an NBA analogy: “I thought I was going to work for Popovich, but then GDP left.” The joke captured the shock that sudden organizational change can deliver to talent expectations.

15. Seedance 2.0 Moves Video Generation from Demo Effects to Industrial Restructuring

• The samples 庄明浩 showed included a woman skateboarding through mountain roads as the seasons changed naturally, a consistent driving sequence generated from a given character and car, and a God of War character fighting Thor. Even without extensive post-production, the samples still had rough edges, but the model was “extremely strong—frighteningly strong.”

• From community distribution, evaluations and overseas feedback, he believes this is the first time a Chinese video model has truly reached the front rank. 冯骥 saw “Kill the game” at the end of the user manual, called it “quite objective” and wrote: “AIGC’s childhood is over.”

• 冯骥’s industrial judgment is that multimodal information understanding and integration have taken a leap. The cost of general video production will no longer follow the traditional logic of the film industry, but gradually approach marginal compute cost; content supply will experience “unprecedented inflation,” while organizational structures and production workflows are rebuilt.

• When a shot can be generated in hundreds or thousands of versions, the scarce skill shifts from making the content to identifying “the one.” Drawing on 郑民’s projection in “When the Cost of Dreaming Hits Zero: The Next 5 Years of Film and Television,” 庄明浩 puts future competition in taste and selection rather than raw generation volume.

16. Interactive Video, Copyright and Real Voices Bring the Risks to the Fore

• Seedance 2.0 is not an isolated leap: Kling 3.0 continues to improve shot control, consistency, 4K, multilingual output and lip synchronization; the Google DeepMind team also released Veo 3 test footage showing interactive video. After the Veo 3 beta launched, Unity fell more than 20% and Roblox about 16%, although 庄明浩 thinks “the game industry is over” is mostly a gimmick.

• Video going mainstream means e-commerce ads, pre-shot content and interactive film could all be remade, but it also means deepfakes and a trust crisis. 冯骥 warned friends and relatives that videos containing personal likenesses and voices but lacking authoritative-channel backing should be verified through multiple channels.

• Hollywood quickly sent letters, while overseas discussion focused on copyright. 庄明浩’s view is that the existing copyright system may not fit the AI era, but the replacement regime is still unclear. Tim from Hurricane uploaded his own face, and the model automatically matched his original voice; the platform then banned uploads of real faces, using a one-size-fits-all guardrail for now, but the capability itself “can’t be stopped.”

17. The Mood Music for DeepSeek V4 Is in Place; the Release Is Still Missing

• From DeepSeek V3.2 in December 2025, its new architecture and papers, and results from partner institutions, to gray testing in late January or early February, every signal was read as evidence that V4 was near. Media reports also said its coding ability would catch OpenAI and Anthropic, that it would support native multimodality for the first time, and that testing used fewer NVIDIA GPUs and shifted toward domestic chips.

• 庄明浩 did not turn rumors into a conclusion: “The atmosphere has been built up to this point.” But as of the recording date, March 9, V4 still had not launched and no specific date was known. He kept only 2 pages of content, waiting for V4 to provide actual user experience, an architectural inflection point and an industry milestone.

18. The Lunar New Year Red-Envelope Campaigns Created a Peak, but Did Not Rewrite Chatbot Retention

• Doubao stayed relatively quiet, relying mainly on Spring Festival Gala placements and 3 rounds of “technology gift” lotteries to generate a Lunar New Year’s Eve peak. Qwen centered on ordering and free-order cards, officially saying it received 4.1B “Qwen, help me…” requests in 6 days and completed 120M AI orders, including 52.2M milk teas, 35.16M fruit teas and 11.38M coffees.

• Yuanbao first spread through sharing and “Yuanbao Pai” social mechanics, some of which WeChat quickly blocked; the gold-colored Moments posts sent through Yuanbao on New Year’s Eve became highly visible. Its February 18 recap claimed 50M DAU, 114M MAU and more than 1B pieces of content created.

• Data showed Doubao’s Lunar New Year’s Eve DAU peaking at roughly 140M; Qwen surpassed 70M when it announced the RMB3B free-order promotion; Yuanbao exceeded 40M on New Year’s Eve. Once the campaigns ended, all of them fell quickly, with average daily usage returning to around 5 sessions and even dropping to 3.

• 庄明浩 saw “almost no surprise” in the result: the chatbot format struggles to support extremely long sessions, ultra-high retention and high-frequency use. On February 14, the State Administration for Market Regulation summoned Alibaba, ByteDance, Baidu, Tencent, JD and Meituan, among others, and demanded an end to “involution-style competition”; the rapid collapse in attention after Lunar New Year’s Day 1 and Day 2 was also tied to this backdrop.

19. Zhipu and MiniMax’s Fivefold Rally Has Spilled Over to Kimi and StepFun

• Before listing, Zhipu and MiniMax were valued at roughly HK$40B–50B; at the Lunar New Year peak, both neared HK$300B, a gain of roughly 5–6x. Using 1% of OpenAI’s roughly $840B valuation in this round—about HK$65B—as a unit of account, both companies moved from 1 unit to nearly 5.

• 庄明浩 cautioned that “we see the new names’ smiles, not the old names’ tears.” During the same period, SenseTime, Unisound and Kuaishou were in clear downcycles, and Hong Kong tech overall was not strong. New model-company listings contain substantial emotion and do not mean that all AI assets have been re-rated together.

• The secondary-market premium has nevertheless provided a real-money anchor for private markets. Kimi raised $500M at a $4.3B valuation on December 31, 2025, then raised $700M at a $10B–$12B valuation on February 24; the Kimi K2.5 launch and leaderboard performance coincided with the financing, giving it room to raise its price almost immediately.

• Market reports also said Kimi K2.5 generated more revenue in 20 days than in the entire previous year, although the comparison may reflect a low base. StepFun announced RMB5B in financing on January 26, with 印奇 as chairman, and was reported to be preparing a Hong Kong listing; the stock-price effect from Zhipu and MiniMax spilled over to Kimi and StepFun.

20. 樊麾 Ran the Full Human Psychological Curve of Facing a Machine

• 庄明浩 uses the 10th anniversary of AlphaGo to address the collective anxiety over “Will I be replaced?” 樊麾 was a professional Go player, head coach of France’s national Go team and European champion in 2012, 2013 and 2015. In 2015, he accepted DeepMind’s invitation to play an early AlphaGo in a 5-day, 10-game test.

• In the first game, he attributed the loss to a small mistake; in the second, he used a roundabout strategy to level the match and thought it was “nothing special.” By the second day, the patterns stopped working, and the third and fourth days brought consecutive crushing losses. The machine had no emotions; every doubt, hesitation and loss of confidence “hit a wall and came straight back,” leaving him with only 2 wins in 10 games.

• 樊麾 realized that the matter was not so simple and joined the AlphaGo team. After Nature disclosed the games, however, outsiders accused him of being weak, playing terrible Go and even taking payment. He did not defend himself, because he could not reveal the truth at the time.

21. 李世石’s Move 78 Did Not Reverse the Trend, but Rebuilt Humanity’s Place

• Against 李世石, the AlphaGo team internally predicted a 5–0 score. After losing the first game, 李世石 still felt the gap was limited; when move 37 appeared in the second game, professional players were puzzled and even laughed, but 李世石 stared at it for 12 minutes, his expression shifting from relaxed to serious as his confidence began to wobble.

• 李世石 lost 3 games in a row, and the third game did not even resemble his normal level because the machine offered no feedback, sending every suspicion back into himself. In the fourth game, down 0–3 and under the world’s gaze, he played move 78, later called “God’s move,” said by some to have had odds of only 1 in 10,000, and eventually won the game.

• In 庄明浩’s telling, this was humanity’s last victory over AlphaGo. It did not alter the machine’s continued progress, but it took 李世石 from arrogance through doubt and shattered confidence to rebuilt confidence: “It’s a miracle, but humans are capable of creating miracles.”

• 10 years ago, viewers did not know what the match meant; today, everyone seems caught in a similar vortex—“the first time you hear the song, you do not know its meaning; by the second listen, you are already inside it.” A global population chart also shows that high-frequency or paid AI users remain only a tiny sliver: anxiety is already at our throats, but true mass adoption has yet to begin.