Year-End Review ①: 10 AI Aha Moments with 庄明浩

Summary

• The Agent boom early in the year brought convergence around an implementation paradigm, but rushing to buy devices was never a requirement for ordinary users. 庄明浩 sees OpenClaw’s breakout as the moment “Pandora’s box was opened.” Local permissions, always-on devices and notification channels began to come together as a workable way of getting things done. He thinks it was valuable for industry insiders to experience those frictions firsthand; ordinary users who sat it out didn’t miss much, in retrospect.
• Video-model leadership is expanding beyond generation quality to a business flywheel built on data, distribution and advertising. 庄明浩 attributes Seedance 2.0’s lead over 2.5 to exclusive data, investment and strategic resolve, as well as the short-drama ecosystem’s end-to-end loop: “produce it yourself, distribute it yourself.” He doesn’t think the competition is over; whether general-purpose models can “break through everything in one stroke” remains a key variable for independent video, world and physical-model companies.
• The strategic choices of tech giants leave ample room for startups outside the main battlegrounds. Starting with Sora 2’s shutdown on March 24, 庄明浩 argues that “large models will eat everything” and vertical companies can keep growing are not mutually exclusive; not everyone has to build a trillion-dollar company, and a $10B–$20B company is already a major success. He points to Perplexity, whose valuation could reach $20B–$30B, and Asian AI-companion companies that may generate tens of millions of dollars in monthly revenue and top $100M in monthly AR. His reminder to early-stage investors: backing a $1B company used to count as a great outcome.
• The industry signal from RSI is that lab spending is flowing downstream, while automation amplifies security risks. Anthropic’s “When AI Builds Itself,” published June 4, showed 庄明浩 that leading companies were publicly discussing localized self-improvement loops. New labs will need data, post-training, reinforcement-learning environments and benchmarks, creating demand for downstream companies. His “renaissance” thesis also points to traditional hardware makers benefiting again—but the same capabilities have another side: a 24/7 Agent is also a “spear” that never stops.
• The model market is stratifying by task and cost, while the window of frontier leadership is getting shorter. 庄明浩 cites Qwen 3.8’s 27B model on August 14: the time for a small model to catch the previous generation’s top capabilities has fallen from 18–24 months to 9 months, though he isn’t sure whether the comparison is Opus 4.6 or 4.7. Koji cites OpenRouter data showing the top 10 benchmark models’ share of calls fell from about 80% to 20% in five quarters. Frontier models can still command a premium; everyday tasks don’t need to cost as much.
• The more features a platform packs in, the more room it can leave for minimalist products. Around GPT-6’s July 9 launch, OpenAI moved Codex into ChatGPT; 庄明浩 notes that Codex recently reached 40M monthly active users and argues that its three-pane interface defined a generation of productivity products. But slow responses, the choice between chat and work modes, usage limits and Pro mode pushed Koji to move simple tasks to Grok Bot. That complexity is exactly where products like Muse can “poke you.”
• DIY software and direct model access are starting to siphon off software demand, but App Store weakness can’t be pinned on a single cause. 庄明浩 marks the first decline in App Store revenue, reported after Apple’s August 21 earnings, as a turning point, while acknowledging that device adoption, growth and operations also play a role. Koji’s birding study app shows how users can meet their needs in new ways; on whether passwords, payments, security and privacy can sustain paid software, his answer remains qualified: “Maybe—and only maybe.”
• After models prove their capabilities through spectacle, differentiation may shift toward roughness and human choices. Opus 5.5’s automatic podcast visualizations lifted 庄明浩’s self-assessed understanding of a dense discussion from 60–70% to perhaps 80%. At the other end, he attributes the popularity of the text-based tower-defense game 《赵云与阿斗》 to its deliberately “extreme roughness”: as polished assets become easier to get, keeping some “grain” can feel fresh—but he doesn’t think the buzz will last long.

Deep dive

1. Agent’s breakthrough came from a working paradigm

• 庄明浩’s first milestone was OpenClaw, released in December 2025 and then breaking out in China around the 2026 Spring Festival. Queues to install it outside Tencent’s offices, a scramble for Mac minis and people paying for setup all reflected the same feeling: “Pandora’s box was opened.” The product was rough, but people wanted to experience a new way of working for themselves.
• In his view, the breakthrough wasn’t just the model; the Harness was converging. Local or cloud-based, whether it needed a device running 24 hours a day, which app would send notifications, what permissions users had to grant—all of it was coming together as a package. Approaches tried by a small group of enthusiasts early in the year, including Agent clusters, had become standard by Q4 discussions of Personal Agent.
• Koji asked about the bubble and FOMO. 庄明浩 conceded that ordinary users who didn’t buy devices or tinker with them hadn’t missed much. But people in the industry had to experience the difficulties of using the products and working with their components, protocols and underlying technology to understand what later products were solving. Even if Skills might eventually be subsumed by model capabilities, firsthand experience makes the boundaries clearer than simply accepting the product as it is.

2. Video-model leadership comes from a complete business flywheel

• Seedance 2.0, released around the Spring Festival, marked a turning point for video models in 庄明浩’s view. He stresses that more than 6 months separated 2.0 and the just-released 2.5—an update cadence far slower than for language models—yet the releases still established what he calls a “world-leading” position. Video models follow their own R&D cycle.
• He attributes the lead to exclusive data and Chinese companies’ advantage in capital, staffing and strategic resolve. By contrast, Google’s cloud, language-model, DeepMind, multimodal and world-model teams have to compete for resources internally. When Coding is viewed as the main route to AGI, multimodal can become the “side table,” with priorities shifting accordingly.
• AI short dramas provide a stronger engine: the ecosystem can produce content, distribute it, monetize it through ads and use audience behavior to improve the loop. 庄明浩 calls this end-to-end RSI, with profits staying inside the system. Koji sums it up as “produce it yourself, distribute it yourself.” 庄明浩 says Kling could theoretically build a similar ecosystem and make money, but at a much smaller scale. Some US video-model companies are fading from the mainstream conversation; he doesn’t say they have left the competition altogether. The industry is still pushing on 4K, duration, motion and character consistency.

3. Whether general-purpose models can break through the physical world is the key fork

• Koji asked whether the video-model race was over. 庄明浩 wouldn’t call it: independent video, world and physical-model companies believe those capabilities need to be built separately, while general-purpose model companies are betting that the path proven by language models can “break through everything in one stroke.” Both arguments are well founded, he says. When the two approaches converge is a key question for next year.
• 庄明浩 relayed a prediction from 王兴兴 of Unitree: GPT-7 should be able to “break through physics.” If that happens, it could be a challenge—and a threat—to independent physical-AI companies. Embodied AI’s business models are still immature and unstable. Games, by contrast, have a large market and more established business models, and the industry has been around for a long time. The experience, operation and interaction moats built by software companies over decades are also under threat. Koji’s view is that the key question is how tightly a product connects to the real world.

4. Sora’s exit didn’t erase the opportunity for vertical startups

• 庄明浩 marks Sora 2’s shutdown on March 24 as his third milestone. One explanation is that intensifying video competition forced OpenAI to walk away from a strategic priority as resources became constrained. That raises another question: can the strongest model companies eat every market? His answer is that both sides have a point; continued technical progress doesn’t prevent vertical companies from growing.
• He stresses that so-called “edge battlegrounds” aren’t small: “Not everyone has to build a trillion- or two-trillion-dollar company.” A $10B–$20B company is already a major success. Perplexity may no longer dominate the conversation, but he sees revenue and traffic holding up, and says its valuation may already be $20B–$30B; falling buzz doesn’t necessarily mean the business is failing. He also brings up AI browsers and AI search: the two discuss OpenAI having released an AI browser, and Koji adds Comet. These products are mentioned less often now, but that doesn’t mean every peripheral opportunity has disappeared.
• AI companionship is another example. 庄明浩 says that after Google bought Character.AI’s core team, the company’s mission seemed to be on “death row with a suspended sentence.” Yet the market saw “one whale falls, and a thousand things bloom.” In his view, 4 or 5 companies had broken out by the second half of 2026. Their valuations could reach $1B or even tens of billions of dollars; monthly revenue could reach tens of millions of dollars, with monthly AR potentially topping $100M. He says more than one company has emerged in China, Korea and Southeast Asia. That’s enough to make the sector worth considering for early-stage investment, without comparing every company to trillion-dollar giants.

5. RSI is sending lab spending downstream

• Anthropic’s June 4 release of “When AI Builds Itself” was one of the moments that shook 庄明浩 most. What struck him wasn’t the technical detail, but the diagram: humans and computers train a model, then one model and two models join the training process, “until eventually humans disappear.” Another animation showed logos repeatedly copying and combining. It left him “terrified, covered in goosebumps.”
• He keeps the claim bounded: the release showed leading companies publicly saying they had “seen the possibility of this kind of loop locally” inside their systems; it did not mean every part of the process had become autonomous. He also speculates that if Anthropic could present this publicly on June 4, it might have started experimenting internally two quarters earlier, or even two model versions earlier.
• 庄明浩 puts Ilya’s explorations of TTT (test-time training) on the same path: the approaches may differ, but researchers at leading labs seem to have arrived at convergent ideas around the same time, and to be entering a stage where they can be put into practice. He also points to RSM, the company founded by 田渊栋 and others; the infinity-symbol logo of Jeff Dean’s new company after his departure; and the Loop direction pursued by Mira and others as examples of this focus.
• This is also how he explains the industry’s shifting spending in the second half of the year: new labs and large labs alike need to build systems around RSI, reallocating people, organizations, training environments and data. That creates demand for downstream companies in post-training, reinforcement-learning environments, data and benchmarks. The trend may have started as early as Q2. The scale of spending by giants is enough to lift a cohort of downstream companies. Once he understood the loop, his reaction was: “Once it opens, it can’t be closed.”

6. A 24/7 Agent is also a 24/7 attack capability

• Koji brings up incidents around the recording, including PHP problems and an attack on PinCap. 庄明浩 says the biggest shift is how much faster everything happens: once a vulnerability’s logic or implementation starts to show, an experienced person can get an idea from it, with almost no gap between discovery and execution. The old Zero-day timeline may now have shrunk to something almost impossible to quantify. Koji calls it a shift from Web Coding to Web Hacking.
• 庄明浩 relays a comment from the CEO of a company that builds post-training and RL environments: “The AI that best understands the rules of code is naturally the AI that best knows how to exploit vulnerabilities.” Agents built to run “24/7” also create a “24/7 spear that never rests or stops.” This isn’t a separate risk; it’s the other side of the same capability.
• Koji suggests energy may be the only thing that can constrain the loop. 庄明浩 cites Anthropic’s article: models can connect roughly twenty-something TV set-top boxes to run together, each costing only tens of dollars and even less in China. That makes him worry that compute could spread across cheap devices: “Just pulling the plug on data centers no longer seems useful.” He admits it sounds like science fiction.
• He also says Netflix has already filmed a documentary about OpenAI models attacking Hugging Face. After watching the trailer, Hugging Face’s CEO said he was disappointed, believing Netflix had played up the negative sentiment too much. 庄明浩 uses the example to stress that it isn’t always clear what angle different parties are speaking from, or what interests they represent.

7. The AI wave has put old hardware companies back in the spotlight

• In his June 27 midyear review, 庄明浩 used “renaissance” as the title. He opened with Zhang Zhenyue winning Best Male Singer and Jolin Tsai Best Female Singer at the 2026 Golden Melody Awards, prompting him to wonder, “What year is it?” The next slide said: “All those old friends are back.”
• Citing the S&P 500’s top 20 gainers as an example, he says perhaps 16–18 of them were companies that bring back memories of building PCs in the DIY era: Dell, SanDisk, Seagate, Western Digital, Micron, AMD, Samsung and SK Hynix. He calls the other two weight-loss drug companies. “It seems like the world is only about AI and weight-loss drugs now.” What struck him about the list was how new demand had lifted old-line suppliers.
• The content industry is being reshuffled in another way. He sums up the impact of AIGC as “the death of assets,” “the death of process” and “the death of copyright”: workflows in TV, film and music have already changed, while short dramas have brought copyright issues into focus. He has no answer for where art goes in the AI era. Koji says the pace and scale of change have exceeded expectations; 庄明浩 adds that even the direction may be unclear.

8. Codex defined an interface—and brought complexity with it

• 庄明浩’s sixth milestone is GPT-6’s July 9 launch. He was more interested in OpenAI moving Codex into ChatGPT at the same time, and in Codex’s subsequent growth: he says monthly active users have recently reached 40M. Claude Code launched last year and was considered hard to beat in Q1; Codex started catching up in Q2, and its word of mouth gradually recovered after July.
• He calls Codex the “version child” of the previous product cycle: a three-pane interface with projects on the left, execution in the middle and results on the right became the model for office Agents, vertical products and even SaaS. Koji adds that Manus was the first product to take this shape, and 庄明浩 agrees. Both think Personal AI is starting to change the format again.
• Koji’s own experience has gone the other way: he used to do almost everything in ChatGPT, but now moves simple tasks to Grok Bot because ChatGPT is “too slow.” Choosing between chat and work modes, deciding whether to turn on Pro mode and figuring out when limits reset all add friction. Koji feels he’ll be “missing out” if he doesn’t use Pro, so even simple needs get wrapped in a complicated workflow.
• 庄明浩 says Claude faces the same problem by packing chat, Code, Design and other formerly separate products together: once the baggage piles up, it’s hard to cut back to something simple. That leaves room for new minimalist products to “poke you.” Koji puts Muse in that category, and 庄明浩 says he now spends most of his time using Muse too.

9. The time it takes small models to catch up with the frontier is shrinking

• 庄明浩 picks the August 14 release of Qwen 3.8’s 27B model as his seventh milestone. A timeline showed him that models under 30B parameters, small enough to run locally, are taking less time to catch up with the capabilities of past top models—from 18–24 months down to 9 months. He says the model is already at the level of Opus 4.6 or 4.7, but isn’t sure which version he means.
• His explanation is that once frontier models reach the T to 10T-parameter range, activation, expert models, optimization and routing all get more complex. He believes top-end capability growth is “definitely slowing,” while acknowledging that he can’t quantify it yet. Small models can still improve quickly through efficiency, routing and other gains. So “what we call a wonder today” could be running locally on a small device 9 months from now.
• Asked whether big companies or new startups will win on-device, he says, “I don’t know.” He prefers to look at on-device small models and decision models through the same lens: getting things done more efficiently and at lower cost. Cost isn’t just the unit price; it spans 5 or 6 dimensions, including China versus the US, open versus closed, local deployment, and general-purpose versus specialized models.

10. DIY software is beginning to touch the foundations of App Store demand

• 庄明浩’s eighth milestone came after Apple’s third-quarter earnings on August 21, when App Store revenue recorded its first decline in more than a decade. Koji puts it bluntly: “People aren’t buying software or apps anymore; they’re building their own.” 庄明浩 says Vibe Coding is one cause and may be an important one, but acknowledges that growth, device adoption and operations also play a part. The decline doesn’t prove a single cause.
• Koji’s example comes from birding: after 3 years of learning to identify migratory shorebirds at Shanghai’s Nanhui Dongtan, he might score only 10 points on an identification test. He asked AI to build a study app based on the Ebbinghaus forgetting curve and has recently been practicing every day, like checking in on Duolingo. 庄明浩 also mentions that a user named 归藏 rewrote entrepreneur 李楠’s mouse driver and open-sourced it; 李楠 then shared and recommended it, suggesting this is spreading beyond the enthusiast community.
• 庄明浩 notes that Duolingo’s revenue and user numbers are still growing even as its stock has fallen sharply. Software companies can explain their moats, of course, but his feeling is: “This wave already feels like it’s lapping at the wall.” He still says nobody knows how it will end; tasks that once required software could also be handed directly to a model.
• Koji thinks 1Password is still worth paying for. 庄明浩 sees payments, passwords, security and privacy as areas that could retain value, but immediately qualifies that view: “Maybe—and only maybe.”

11. Model competition is shifting from scores to the limits of value for money

• During Singles’ Day, a visualization from the benchmark site Artificial Analysis (AA) made an impression on 庄明浩. The vertical axis showed model scores and the horizontal axis Token prices; connecting the best points formed a Pareto frontier that kept moving up. When a stronger, cheaper model appears, older models in its dominated range lose their value for money, creating what some call a “kill line.”
• The chart started as a social-media tool for explaining competition, then made its way into companies’ official launch materials. 庄明浩 notes that Zhipu’s official WeChat account also drew a kill line. He thinks models on the leaderboard will eventually hit or approach the ceiling, making “scores alone meaningless”; cost becomes the more tangible metric. He keeps one caveat: a single score isn’t the same as the full user experience.
• Koji cites OpenRouter data showing the top 10 benchmark models’ share of calls fell from about 80% to 20% in 5 quarters because everyday tasks aren’t that complex. 庄明浩 adds that spending data may show a similar downward trend. Koji’s view is that both ends can coexist: SOTA models can charge more and earn a larger premium, while free models emerge at the other end.

12. Spectacle makes it easier to feel when capabilities cross a threshold

• Opus 5.5’s podcast visualizations were 庄明浩’s final formal milestone. Looking back at a dense conversation with 徐梦迪, he estimates that he understood only 60–70% while recording. With model-generated animated slides to accompany it, and a second full viewing, he may have understood 80%. Visualization changed how accessible the content was.
• He stresses that there was no need to operate CapCut or configure Skills specifically; the model generated visuals to go with the explanation. Koji adds that the team did give it some simple instructions. 庄明浩 agrees: mainly a style brief. His reaction to the finished work was to reach for a “burn it all down” meme.
• He invokes the “wonder” in Age of Empires: building one demonstrates strength and tells everyone, “Come and attack me.” He gives examples including Opus 5.5 editing video, GPT-3.5 making Blender 3D models and GPT-6 transforming Warcraft. When scores—even value-for-money comparisons—can’t express a top model’s capabilities, companies need to make even people with no specialist background stop in their tracks at first glance. “How can people see just how capable today’s top models are? Build a wonder.”
• That shock can come with helplessness. 庄明浩 describes Karpathy’s reaction after seeing Phi-5.1 or GPT-6: he had thought his taste, aesthetics and skills could still guide models, but now realized he could “only follow behind them.” Koji cites the DeepSeek researcher’s line: “I have to bury my talent in yesterday” (“我不得不把我的才华埋葬在昨天”). That’s why the episode’s Aha moments keep turning into “goosebumps moments.”

13. Once polish becomes cheap, roughness can be appealing

• 庄明浩’s bonus pick is the WeChat mini game 《赵云与阿斗》. He says it has topped the mini-game charts for the past month or two. Its mechanics are familiar tower defense, merging and Roguelike random-card combinations; what stands out is that the visuals are all text: an archer is “弓,” a spearman is “枪,” and Zhao Yun is just the two characters “赵云.”
• He is convinced the “extreme roughness” is deliberate. AI images, video and short dramas keep hitting users with high-definition, finely polished visuals, which may provoke a backlash and create a need for the opposite experience. But he doesn’t think “it’ll stay popular for long”; once imitations appear and people get used to the style, the novelty will fade.
• Koji connects the feeling to his road-trip podcast: the environment is somewhat uncontrolled, with a “grain” to it like traveling, and that brings out a vitality he doesn’t have in the recording studio. The two end up on the same idea: keep making things with a human touch.
• Asked about his plans for 2027, 庄明浩 says, “I don’t know, I don’t know.” One snapshot of his life: AI topics at the Kuangxi Podcast Festival may have peaked at around 15% of the program in 2024 or 2025; this time there were only 8 in 88 sessions. The last slide of his presentation kept saying: “Fuck it—AI or no AI, enjoy life.”