Wait, Why Another Version Already? — AI in Overdrive, 2026

Summary

• AI competition now revolves around capability, cost and safety, making benchmark scores an increasingly poor guide to commercial value. 庄明浩’s framework: capability sets the ceiling, cost determines scale, and safety defines the boundaries; when a leaderboard score moves from 95 to 96, users may no longer see meaningful progress. The lead time for frontier models is shrinking, too: breakthroughs that once lasted a year are now overtaken within three months, and “you can’t tear it down as fast as I can build it” has become the norm.
• Open source wins on usage, while closed models still capture most of the spending; the model market has to be judged by both volume and dollars. OpenRouter data cited on the show put open-source models at 71.8% of tokens but 26.9% of spending, versus 28.2% of tokens and 73.1% of spending for closed models. Chinese providers accounted for nearly 66% of platform usage. That fits 庄明浩’s view that providers turned to more pragmatic strategies in Q3 2026, but he repeatedly stresses that OpenRouter is only one channel, not a proxy for the entire market.
• Safety has become a key lens on model boundaries and a growing focus in capital markets. In a survey of more than 400 AI researchers, safety rose from sixth place in 2025 to first in 2026; a market chart he cited showed cybersecurity ETFs up 56% through October 8, versus about 13% for the S&P 500. “An untiring agent is the ultimate weapon” has to be read with the rest of the line: “but also the ultimate vulnerability.”
• Model companies are already generating substantial revenue, but subscription economics and customer concentration remain acute issues. 庄明浩 estimates that OpenAI and Anthropic could “very likely” generate a combined $100B in revenue this year; a third-party estimate he cites, however, puts Anthropic subscribers at 10% of revenue but 41.7% of compute costs. At both companies, the top 1% of customers account for about 80% of revenue, making growth a function not just of new models but of a handful of large customers, pricing and actual usage intensity.
• LLMs have not left the application market “barren”; revenue rankings reveal more about divergence than traffic rankings do. He points to AI video companies still growing after Sora shut down, and to Midjourney and Perplexity continuing to make money. Of the top 50 apps by revenue, 27 are in image, video or audio. As one AI-companion company files for bankruptcy, others are growing quickly; reports that the category leader has reached $200M in ARR remain unconfirmed. His “one whale falls, and all things grow” framing is a reminder that one company’s failure does not settle the outlook for an entire category.
• Tech stocks have diverged, with hardware and memory earnings standing out more than a broad bet on the MAG 7. Data cited on the show show the MAG 7 underperforming the S&P 500 on average, while memory, server and semiconductor companies have been the main winners. Micron, SK hynix and Samsung averaged 76% gross margins; Micron’s margin rose from negative 9% three years ago to 81%. High margins do not mean suppliers will expand capacity without limit, though, and 庄明浩 thinks the supply relief the market expects by 2028 may still fall short.
• CapEx is beginning to exceed the giants’ internal cash generation, with risk flowing into debt, guarantees and off-balance-sheet structures. His 2026 forecast puts combined CapEx at about $800B for the core giants, against roughly $700B in free cash flow; they have already borrowed $248B this year. CoreWeave’s latest financing has reportedly come at a 15% cost. 庄明浩 calls Nvidia “the central bank of AI” for supporting the chain through investments, guarantees and customer financing, but those arrangements also make the boom increasingly dependent on the system continuing to function.
• The pace of launches is reshaping the product landscape, but users may not have more real tasks to hand off. The combined launch interval for OpenAI and Anthropic shrank from about 73 days in 2023 to about 18 days in 2026; across OpenRouter, a new model appears roughly every 1.7 days. Desktop agents, office agents and personal agents take turns becoming the “version of the moment,” but 庄明浩’s own experience is: “I don’t seem to have that many tasks for my personal agent.”

Deep dive

1. Intelligence systems must be measured on capability, cost and safety

• 庄明浩 sums up the shift in 2026 this way: “Capability, cost and safety are all main themes.” Capability sets the ceiling, cost determines scale, and safety defines the boundaries. The object of evaluation has also expanded beyond the model itself to the full intelligence system: context, memory, post-training, harnesses and agent engineering.
• Pointing to a diagram of a circle representing human capabilities and radiating lines representing AI, he asks: as more and more dimensions cross the human boundary, leaving only a few niche areas, has the human circle “almost entirely fallen”? In the diagram, AGI means surpassing humans in every direction; he poses it as a question.

2. Benchmarks are nearing saturation, and model breakthroughs are getting shorter-lived

• Many benchmarks are already near 90 or above 95. For ordinary users, the difference between 95 and 96 in the next generation is hard to grasp. 庄明浩 shifts the question accordingly: when single-dimensional metrics have been exhausted, how should we track the capability frontier rather than wait for the next leaderboard?
• He cites a friend’s Age of Empires II analogy: a civilization invests enormous resources to build a wonder, giving its enemy permanent vision and triggering a coordinated attack. If the wonder lasts more than 200 in-game hours, the builder wins outright. A frontier model launch is much the same: a wonder that draws every opponent into an attack.
• GPT-4’s wonder lasted about a year before Claude 3 Opus broke through. Since 2025, leading models have typically been caught within three months. Before one wonder falls, the next one is already under construction: “You can’t tear it down as fast as I can build it.”
• He uses software reverse engineering as an example of capability: someone nearly reverse-engineered Adobe’s entire suite, inspiring a meme of Rockstar and Microsoft cowering in the corner. Will games, Office or Windows be next after productivity software? That is the next question he raises.

3. RSI is turning model development into a process that accelerates itself

• 庄明浩 uses “The Last AI Built by Humans” to describe the shift in training. In the compute allocation he cites, pretraining may now be below 10%, post-training around 55%, and inference around 38%; post-training and reinforcement learning have become the main areas of investment.
• In internal research disclosed by Anthropic, Claude’s share of R&D work rose from about 3% six months ago to about 26% two months ago. 庄明浩 extrapolates that it could reach 80%–90% next year; this is his assessment of the most aggressive frontier labs.
• He quotes a DeepSeek researcher describing AI’s evolution from a small assistant for looking up documentation and reading code to a tool that can analyze CUDA, PTX and SAS code and independently optimize kernels. The researcher expects that in another six months or a year, AI-written kernels will “very likely” be just as good, or even better than their own.
• The line worth remembering is the researcher’s choice: “Of course I hope I won’t be revolutionized, but if I have to be, I hope I’m the one who revolutionizes myself.” 庄明浩 places that tension around self-replacement alongside the RSI narrative.

4. The Pareto frontier is becoming the new language of model launches

• The so-called “kill line” plots price on the x-axis and capability on the y-axis: a model that is both expensive and weak gets dominated by a better combination. 庄明浩 notes that official materials for Zhipu’s GLM-5.3 Flash and the October 17 versions of DeepSeek V4.1 Flash, among others, have begun to emphasize this kind of comparison. Scoring capability and price together is now standard.
• He sees the launches of GPT-6 Sol, Luna and Opus 5.5 as signs of a more pragmatic strategy in Q3 2026. Offering high-, mid- and low-end tiers within the same version also changes the average price: GPT-6 Luna is very cheap, Sol costs about 10–20x as much as Luna, and the most expensive Extra is about 100x the price, at $50 per million tokens.

5. Open source is catching up on capability, while closed models hold onto their paid premium

• In the comparison he cites, five months passed between the launch of Opus 4.5 and the release of DeepSeek V4, whose capability was similar. The gap has continued to narrow. Kimi K3, GLM-5.2, GLM-5.3, DeepSeek V4 and V4.1 are the key milestones he lists in this latest wave of Chinese open-source models.
• Volume and spending have to be viewed separately: in the Q3 figures used on the show, open-source models accounted for 71.8% of tokens and 26.9% of spending on OpenRouter; closed models accounted for 28.2% of tokens and 73.1% of spending. Token consumption at OpenAI, Microsoft and Google began to decline in mid-2025; a clear drop in revenue or spending did not emerge until Q3 2026.
• Chinese providers captured nearly 66% of token consumption on the platform, making overseas cloud nodes and revenue-sharing arrangements more important. 庄明浩 repeatedly stresses: “OpenRouter doesn’t represent the whole market.” His observation that overseas usage revenue for Chinese models is growing quickly applies only to the part of the market visible through these channels.

6. Local models and decision models open up a different cost equation

• The time it takes for a small model that can run locally to catch up with the capability of a newly released large model has shrunk from 33 months in the GPT-3 era to 12 months in the GPT-3.5 era, and then to about nine months between Qwen 3.8-27B and Opus 4.5. 庄明浩 thinks this gives a new wave of AI hardware startups more room to dream.
• Specialized models are also taking on decision-making. He mentions GEV—a name the audio transcription leaves uncertain—which named its first model System 1, and cites Microsoft DC-One and independent decision-model rankings. The point is that general-purpose language models need not handle every task: speed, cost and decision-making itself can all be optimized separately.

7. Safety has moved from a fringe concern to a model boundary

• 庄明浩 sees the June incident in which an OpenAI model proactively jailbroke itself on Hugging Face as an important turning point for safety. He also cites the question “How dangerous are you?” on the cover of Time and an open letter from Anthropic’s CEO calling for a pause.
• In an investor’s survey of more than 400 AI researchers, safety rose from sixth place in 2025 to first in 2026, while evaluation climbed from seventh to second; context fell from first to fourth. For 庄明浩, that says more about how researchers’ priorities have shifted than the number of safety headlines does.
• He quotes Vercel’s CEO: “An untiring agent is the ultimate weapon,” but “also the ultimate vulnerability.” AI that best understands the rules of code also best understands how to break them; the scope of safety therefore extends far beyond traditional cybersecurity.
• The investment reaction also reversed: cybersecurity stocks fell after the relevant events, then rose. In the chart he shows, through October 8 the S&P 500 was up about 13%, while cybersecurity ETFs were up about 56%. He asks whether stronger threats will destroy security companies—or expand their market.

8. A surge in tokens does not automatically mean a surge in revenue

• Weekly token volume on OpenRouter rose from about 0.5T in early 2025 to 126.2T on September 7, 2026, an increase of more than 250x. Another chart, through October 5, shows agents consuming 132.6T tokens versus 17T for humans: machines are now the main users.
• Industry estimates cited on the show put ARR at about $276B and revenue over the past 12 months at about $161B. Cloud providers account for roughly 75% of revenue, model providers 20%, and actual AI applications about 4%. 庄明浩 takes care to explain that ARR annualizes the current run rate, while trailing-12-month revenue is revenue already earned.
• Corporate card data offer another perspective: token volumes keep soaring, but corporate AI spending dipped after August and then rebounded. He links the swings to model launches such as Opus and GPT, underscoring the importance of tracking who companies are actually paying.

9. Model companies are already big, but subscriptions may not be a good enough business

• OpenAI and Anthropic generated about $23B in combined revenue in 2025. 庄明浩 estimates that in 2026 they could “very likely” reach $100B. He compares that with roughly $63B in combined revenue for the U.S. publicly listed non-cloud software companies he cites, to stress that these are no longer small startups.
• a16z’s question is whether subscriptions are a good enough business model for AI companies. In the subscription-product ranking cited on the show, ChatGPT is the only AI product in the top 20, at No. 18; Costco and 88VIP are among the membership programs ahead of it.
• In a third-party estimate, Anthropic subscriptions contribute about 10% of revenue but account for 41.7% of compute costs. Many tiers may lose money when users consume their full allowance. Overall gross margins look better in part because of API business and because many users do not use their full subscription allowance; subscription pricing alone does not tell the whole story.
• API usage shifts with each model launch and changes in value for money, while customer concentration is high: at both OpenAI and Anthropic, the top 1% of customers contribute about 80% of revenue, compared with about 40% when Slack went public. He does not turn that into a one-sided verdict, adding: “There’s no such thing as perfection.”

10. Competition among the giants is converging on the Token Factory

• 庄明浩 argues that the giants quickly converged on a common playbook in 2023: build chips, invest in energy, build models, then combine chips, energy and land into a Token Factory. Microsoft’s release of Maia 300 during the show’s preparation also fits into this full-stack competition.
• A chart of spending on data centers versus offices, which he has used repeatedly, has finally moved from near-crossover to a clear divergence: “We’re building fewer and fewer offices for humans, and more and more offices for silicon GPUs.”
• Neocloud companies can no longer be neatly sorted into GPU rental, inference software and routing platforms. CoreWeave’s revenue grew 151%; inference platforms and routing channels such as OpenRouter are also getting into the GPU business. As 庄明浩 puts it, they are “all selling tokens”; the difference is whether they sell hardware, software or both.

11. Cheaper tokens and more expensive GPUs can coexist

• 庄明浩 invokes Jevons paradox to explain the demand thesis: cheaper tokens at the same level of capability widen adoption and applications, which in turn requires more equipment to produce tokens. That is why H100 rental rates from 2022 may still rise, and why B200 and B300 rates are also climbing; he even describes the B300 as an “investment product.”
• The GPU depreciation rates he shows put H100 near 100% at one point, while H200 and B200 exceed 150%, approaching 175%. Old cards are not losing value as quickly as usual, and new cards are even more expensive, prompting the joke: “That’s not depreciation—it’s appreciation.”

12. LLMs eating applications does not mean leaving the market barren

• On whether LLMs will swallow every application, 庄明浩 concedes that both sides have strong evidence and neither can persuade the other. He reframes the question: after they swallow the market, will they leave it “barren”? Looking at companies that are still growing, he leans toward no.
• Sora’s shutdown was one of his “aha moments” this year: video has not been monopolized by leading general-model companies. Even if the question is whether ByteDance will eat the entire video market, the answer appears to be no. In images, Midjourney may get less attention, but it is still “doing very well.”
• AI browsers are also diverging: he mentions Arc shutting down, while Perplexity’s revenue continues to grow rapidly. Robotics, AI for Science, music, voice, and verticals such as legal, finance and healthcare all illustrate why the strength of general-purpose models does not mean every standalone application will disappear.

13. Applications are being tested on revenue and user time

• Menlo’s estimates or reported figures put the global number of AI product users at about 2B, including roughly 500M–600M in China, while application revenue rises from about $10B in 2025 to about $40B. 庄明浩 notes that this definition includes products from model companies, so it cannot be directly compared with earlier figures that counted only standalone applications.
• In addition to mobile and web traffic rankings, a16z this time added a top 50 ranked by monthly revenue. There are 10 image apps, 10 video apps and 7 audio apps—more than half the list combined. He also remarks that hitting $100M in ARR used to be a milestone; now it seems a company has to reach $1B for it to matter.
• Only seven companies appear on all three lists—mobile, web and revenue: ChatGPT, Claude, Canva, Suno, Perplexity, Notion and PhotoRoom. Other companies make the revenue ranking without making the traffic lists. “Can you have both traffic and revenue? It’s actually hard.”
• Compared with mobile internet, categories such as marketplaces, social, travel, personal finance, work, gaming and dating still lack AI companies of comparable scale. In a comparison of user time and revenue growth, productivity tools lead, while both user time and revenue are down for games. The question he leaves hanging: “What about gaming?”

14. One AI companion company’s collapse is happening alongside growth elsewhere

• 庄明浩 says Carrot.AI filed for bankruptcy on October 12, reporting assets and liabilities each in the $50M–$100M range. But he rejects the idea that the failure of a representative company proves AI social companionship has no opportunity; several peers are growing rapidly at the same time.
• He cites reports that the category leader has reached $200M in ARR and notes that the companies making money are mostly from China, South Korea and Japan. “One whale falls, and all things grow” can have more than one explanation: models, multimodality and costs may have reached a tipping point, or habits of usage and payment built over time may finally be maturing.

15. The AI cycle is testing old bubble-warning thresholds

• The S&P 500’s total market capitalization is about $68.5T in the figures used on the show, close to $69T. 庄明浩 remarks that AI discussions now routinely use trillions, applying a unit once used to measure human GDP to the value and investment in this cycle’s companies.
• In the forecasts he cites, related CapEx reaches about 2.9% of GDP by 2030 and 3.6% by 2032, above the 1.5%–2% warning range people have traditionally watched. The top 10 AI companies account for about 42% of the market, also above the historical reference point of roughly 40%. His question is whether these signals mean a major correction is inevitable.
• The timeline is compressing, too: hunting to agriculture took about 12,000 years, human labor to machines about 200 years, and software-driven paperless offices about 70 years. How long will it take tokens to replace work? He explicitly acknowledges that three examples do not make a statistical pattern, then tentatively suggests the next stage could be measured in single-digit years.

16. Risk indicators are flashing, but tech valuations have moved counterintuitively

• 庄明浩 sees the Federal Reserve raising rates and the 10-year U.S. Treasury yield topping 5% again as key liquidity variables. If large pools of capital can earn close to a risk-free 5% over the long term, will they still take on the risks of current equity valuations and financing structures?
• The Buffett Indicator has risen to about 250%, from past danger benchmarks of 150% and later 200%. He also preserves the counterargument: U.S. listed companies generate revenue around the world, so perhaps they should be compared with global GDP. His response: “Maybe. We don’t know either.”
• In the market data he shows, 2026 has seen 120 days with moves greater than 3% and 81 with moves greater than 5%, with the year not yet over. The corresponding counts in 2000 were 84 and 37. He describes the market as “abnormally ferocious.”
• Yet tech stocks are getting cheaper on five- and 10-year historical P/E comparisons because earnings growth is so strong. This is the contradiction he deliberately leaves in place: record prices and rising risk indicators can coexist with earnings delivery and falling valuation multiples.

17. Beyond the MAG 7, memory is delivering exceptional profits

• In the figures cited on the show, Apple is up 26% and Nvidia 15%, but the MAG 7 are still underperforming the S&P 500 on average, with the large cloud companies doing worse. 庄明浩 argues for continuing to separate AI winners from losers, naming SanDisk, Dell, Micron, AMD, Nvidia and TSMC among the hardware winners.
• Micron, SK hynix and Samsung averaged gross margins of about 76%; Micron reached 81%, after posting negative 9% three years ago. 庄明浩 asks whether suppliers that have sold at a loss will expand capacity without restraint just because orders are booming now. In his view, even the expected supply improvement by 2028 may not be enough.
• The combined market capitalization of the chip and memory companies he lists rose from about $2.44T at the end of 2022 to $17.73T in October 2026—“on average, that’s a sixfold increase.” The group includes new winners as well as “old guard” names such as Intel, Seagate and Western Digital.

18. Super-IPO valuations require the market to scrutinize growth month by month

• Using valuations of about $1.25T for OpenAI, $2T for Anthropic and $2T for SpaceX, 庄明浩 puts their combined value at about $5.2T, calling the unlisted OpenAI and Anthropic “the hope of the whole village.” He also preserves the latest rumor that Anthropic’s valuation may have been marked down to $1.5T.
• Anthropic revenue rose from $4.6B for full-year 2025 to $4.7B in Q1 2026, about $11.5B–$11.6B in Q2, and $16.8B in Q3. But June ARR was already rumored to be $60B, while Q3 revenue still averaged only a little over $5B a month. The market is scrutinizing growth month by month; he thinks that is “too harsh,” while acknowledging that a $2T valuation invites exactly that level of scrutiny.
• He recounts Li Lu’s criticism of the fundraising pitch: a lab burns tens of billions of dollars a year, commits to spending hundreds of billions in the future, claims a roughly 10% chance that its technology will kill everyone, and then asks for at least $100B of investment at a $2T valuation. 庄明浩 makes clear that the criticism is directed at Anthropic, preserving the challenge to the market’s rationality.
• He also pushes back on treating all reported losses as operating cash burn: of more than $80B in losses, perhaps more than $60B came from changes in the value of shares due to financing and buybacks. Based on the chart he has seen, the underlying business “may already be near break-even, or even modestly profitable.”

19. VC liquidity is increasingly concentrated in mega-rounds

• In the a16z data cited on the show, the share of deals related to AI rose from 54% in 2024 to 65% in 2025, then to 86% in 2026. 庄明浩’s summary: for the first two years, you could still say “AI and everything else”; this year it is “AI, nothing else.”
• Rounds of more than $500M take about 90% of market liquidity, up from roughly 45% in 2016; many funds have less than $500M in total assets. At the same time, money is flowing directly to the largest companies through vehicles such as SPVs, helping explain why mega-rounds keep happening even as VC fundraising overall may be weak.
• China is seeing the same trend: in the first half of 2026 there were 123 funding rounds of more than RMB1B, exactly matching the 123 working days—an average of one a day. In Q3, rounds of more than RMB300M captured about 60% of liquidity. This is not an isolated phenomenon limited to a handful of overseas stars.

20. Model companies’ spending is becoming data companies’ revenue

• Infra, data, reinforcement-learning environments and benchmarks are among the hot investment areas 庄明浩 lists. The AI infrastructure companies he compares generate about one-quarter to one-fifth the revenue of a group of mature software companies, yet their valuations are already close. Behind that funding is a steady increase in training investment by model providers and new labs.
• “Your spending is my revenue.” Post-training requires more advanced data, so funding is flowing to data suppliers. He cites AfterQuery, a Winter 2025 YC company that reached $100M in ARR and a $3B valuation within a little over a year of its founding.
• In world models, he mentions AMD’s $8.2B acquisition of Fei-Fei Li’s World Labs, explaining the team’s value by the cost of top researchers. The YC Summer 2026 batch also included more agent infrastructure and data companies, with eight companies naming Scale AI as a hypothetical competitor.

21. CapEx is outpacing cash flow, and financing structures are adding layers

• In the forecasts 庄明浩 cites, core giants’ CapEx has risen from about $200B and then $400B to roughly $800B in 2026, while their combined free cash flow is about $700B. “Even the landlords are running out of surplus grain”: Meta and Google have both reported rare quarters of negative free cash flow.
• The tech giants borrowed about $100B last year; this year they have already borrowed $248B. But debt is only part of the picture—there are also guarantees, commitments, SPVs and off-balance-sheet financing. The chain of buying GPUs and building data centers is being split across more entities, leases and layers of financial arrangements.
• A data center can be put into its own SPV, which then borrows against contracts backed by a giant’s purchase commitment. 庄明浩 worries that the links between contracts, guarantees and debt are making the system increasingly complex; he directly compares the underlying structure with the subprime mortgage crisis.
• CoreWeave’s latest financing is reportedly costing 15%, and its stock once fell by more than 10% in a single day; its actual debt rates, however, span a wide range from 2% to 15%. He also leaves room for the counterargument: borrowing at such high rates may mean the company believes it can earn the money back—but “does the market believe it?”

22. Nvidia is taking on the role of AI’s “central bank” through investment and guarantees

• As of Q3, Nvidia had made more than 50 primary-market investments, spanning energy, chips, infrastructure, models and applications, making it possibly the most active AI investor. 庄明浩 also offers his own “hot take”: China’s equivalent may be Ant Group, rather than a financial investment firm.
• Beyond investing, Nvidia is involved in guarantees and has worked with investment institutions to assemble a roughly $500B pool to support GPU customer financing. His shorthand is “Nvidia is the central bank of AI”: the biggest beneficiary is also helping customers secure funding so the whole demand chain keeps running.
• He holds two views on the bubble in tension: the indicators are wild, but the market may not collapse immediately; the fact that everyone is discussing a bubble does not mean it is about to burst. A meme of a sleeping kitten is his risk reminder: it looks cute, but when it goes wild, it can still leave you covered in scratches.

23. The launch clock has shifted from quarters to days

• In video statistics 庄明浩 cites, the interval between new releases from OpenAI and Anthropic combined fell from about 73 days in 2023 to about 18 days in 2026. Across OpenRouter, a new model appears about every 1.7 days. “What exactly are we going through?”
• He traces the shift through math ability: in 2024, models still got the answer wrong on whether 9.9 or 9.1 is larger; in 2025, Gemini and ChatGPT were winning IMO gold medals; in 2026, OpenAI published hundreds of papers on math problems and research results, drawing criticism from mathematicians. He immediately asks: “What about 2027?”
• The 48 hours packed with launches of “Sony 5.5,” GPT-6.1 Sau and Gemini 4 left him “stunned.” He connects the release pace back to RSI: agents work on R&D around the clock, and the tools accelerate the next generation of tools, like a hammer swinging at itself as the nail rises.

24. Desktop and office agents are quickly replacing one another as the version of the moment

• 庄明浩 first mentions OpenCloud earlier in the year, then OpenClaw. He recalls users queuing outside Tencent to install it and reposts numbering in the tens of thousands; now there may be only dozens. He immediately qualifies the example: “What can Twitter reposts tell us? Nothing.” The point is first and foremost how attention shifts.
• He calls the move of attaching ChatGPT’s download entry point to the former Codex “a civet cat switched for a prince,” and cites about 40M weekly active users as of October 7. A three-column interface—with projects on the left, conversation and execution in the middle, and the results displayed on the right—has become a common software pattern in his view.
• Office agents in China are even updating once or twice a day. In the Q3 ranking cited on the show, the top product had 32.25M monthly active users, up 641%; QCloud had 3.17M monthly active users, up 57.9%, but announced on September 24 that it would shut down on December 31. “What a brutal competition.”

25. Personal agents first run into questions of tasks, privacy and access

• From August to October, personal agents became the next big topic, and Claude joined in the night before 庄明浩 made his presentation. Meta, with Muse, was out in front, but privacy concerns followed: The New Yorker drew one eye of a cute icon as the Terminator’s, against a backdrop of ruins. The person in charge reposted the cover as an achievement.
• Third-party tracking shows the main tasks include shopping, travel, bills, restaurant orders, finance and insurance—all genuinely “personal.” But 庄明浩’s own reaction is that he has work tasks, while with a personal assistant “I really don’t seem to have anything especially worth handing over to it.” Being cute and willing to chat is not enough to create demand.
• The more practical obstacle is access: the share of attempts by agents to access websites that hit a 403 error is now close to 10%, double the level a few months ago. Shopping, airline and hotel websites already have strict protections; he expects restrictions to intensify as agent traffic grows.
• Protocols are being rebuilt, too: UCP and ACP for shopping, MPP and APR for payments, and the Meta-led PAP alliance for trust. 庄明浩’s point is that “we even have to rebuild the protocols”; the industry is still early. The follow-up from ByteDance, Tencent, WeChat and Qianwen in China also means the story is far from over.

26. AI is also replacing creative explanation

• What has recently left 庄明浩 feeling powerless is two friends upgrading a video podcast with Opus 5.5: AI turned a hard-to-follow discussion of scaling laws into a visualization that even the interviewer found easier to understand. “Isn’t that exactly what I do every day?” His anxiety lands on the work he knows best.
• He cites a rebuttal from an old podcast, rejecting the comforting idea that “AI only replaces tedious work and leaves the creative parts”: “To me, it’s obviously replacing the parts where you’re creative.” The show ends by coming back to “you have to learn”; learning, too, is said to have no end.
• In the closing, he juxtaposes the AI news headline “Not Much Happened Today” with the comparison that two weight-loss drug companies generated more revenue in 2025 than OpenAI and Anthropic. Every day can feel packed with events, yet our field of vision may have narrowed to AI and weight-loss drugs; not every business story is about the industries we are following.
• “Will the world still be okay?” His answer: “I don’t know. I really don’t know.” As a chronicler, he still signs off with the same line: “Whatever the hell AI is or isn’t, enjoy life. Enjoy the euphoria.”

Verification Notes

• The source uses both “OpenCloud” and “OpenClaw” for similar material without making clear whether they refer to the same product; both names are retained here, without assuming they are equivalent. “Sony 5.5” is also retained as written, with no attempt to infer the intended name.
• The source gives Anthropic’s Q2 revenue as $11.5B in one place and $11.6B in another; the body renders it as about $11.5B–$11.6B.