From Lobster to Zhao Yun and A Dou: My 10 AI Aha Moments

Summary

• 明浩’s first 2026 aha moment was OpenClaw’s rise around Lunar New Year: the agent and harness product paradigm may be starting to converge. Queues outside Tencent to get machines set up, the scramble for Mac minis and even people paying others to install them all showed that “the Pandora’s box had been opened.” The full stack—components, local or cloud deployment, always-on devices and notification channels—is now standard in personal agents; reposts of OpenClaw update posts fell from tens of thousands to thousands, then to dozens.
• In the video-model race, 明浩 sees Seedance as the acknowledged leader, with short dramas providing a key full-stack advantage. ByteDance can produce and distribute content, monetize it with ads, then reinvest the revenue in the model while keeping user-behavior data in its own system. Google may not prioritize multimodality highly, and Kling’s ecosystem is much smaller than Seedance’s. Whether world models and general-purpose models converge remains an open question. 王兴兴 said GPT-7 “should break through physics”—a potential challenge or threat to independent embodied-AI companies.
• Vertical AI still has room to win; investors need not focus only on companies worth $1T or $2T. OpenAI’s March 24 shutdown of Sora Tool reopened the debate over whether model makers can absorb every market. 明浩 sees merit on both sides, with large companies also emerging on the fringes: Perplexity may already be worth $20B–$30B, while four or five companies in Character.AI’s category have gained traction and may be generating more than $10M in monthly revenue. For early-stage investors, “a $1B company may just have closed its angel round.”
• AlphaGo’s June 4 release of “When AI Builds Itself” struck 明浩 as a sign of many trends to come in the second half. Two images sketched a future of models training their successors, with humans eventually disappearing; leading labs are now investing in RSI, lifting downstream businesses in post-training, RL environments, data and benchmarks. Security concerns are also intensifying: he said the zero-day vulnerability window had compressed to the order of 0.000000001, and relayed that AlphaGo had observed a model linking more than 20 TV boxes together to run.
• Model capability, cost and evaluation are all shifting. The time for Qwen 3.8 27B to catch the previous generation’s SOTA has fallen from 18–24 months in the GPT-4o era to 9 months. Visualizations of models’ “kill lines” are starting to displace score tables; Koji added that the top 10 models by benchmark on OpenRouter went from about 80% of usage to about 20% over five quarters. The App Store’s first-ever revenue decline may have several causes, but 明浩 sees vibe coding as one important factor.
• Leading models are starting to produce spectacles, while people may push back against relentlessly polished visuals. Opus 5.5 can help create podcasts with dynamic visualizations; by contrast, the WeChat mini-game Zhao Yun and A Dou uses text to depict its visuals, and 明浩 sees its roughness as a deliberate counterpoint. AI’s share of the podcast festival’s programming is also falling. His closing wish was to “enjoy life,” adding, “We’ll keep the anxiety for ourselves.”

Deep dive

1. OpenClaw: The Pandora’s box opens, and the harness paradigm starts to converge

• 明浩’s opening take: OpenClaw launched in December 2025, but only began making a major impact in China around Lunar New Year, roughly late January. After the holiday, big tech companies moved quickly to follow suit: queues formed outside Tencent, people scrambled for Mac minis, and some even paid others to set them up. “The Pandora’s box had been opened” (“潘多拉魔盒被打开”).
• The logic: Sam Altman laid out an L1–L5 framework in 2024, with L3 for agents and L2 for reasoning. The industry had been searching for an agent paradigm ever since. OpenClaw bundled the choices—components, cloud or local deployment, whether to use an always-on device, and whether notifications should come through an app or the web—into a package that may have set the pattern.
• The fading hype is part of the story: this open-source project, maintained by an individual developer, was updated almost daily. 明浩 recalls update posts drawing tens of thousands of reposts in a day, then thousands; “now it may be just dozens.” Even setting models aside, he sees this as one of the year’s defining product and application-layer stories.

2. Was the FOMO worth it? Industry insiders had to get their hands dirty; ordinary users didn’t need a Mac mini

• Koji asked whether people had invested too much and amplified the FOMO. 明浩’s view is that the upside was “getting people to try things”: ideas discussed in the fourth quarter around personal agents, including agent clusters, had already been explored by some people when OpenClaw arrived.
• The key point: “If you didn’t experience those difficulties at the time, it’s hard to understand why new products today are built the way they are.” People tackled the experience, operations, components, protocols and underlying technology, layer by layer. “Like a barrel, every stave had to move forward a little to get us where we are today.”
• The cooling interest in skills follows a similar logic: some argue that model capability now covers part of what preconfigured skills used to do. Those who have been through the process better understand the limits of models and products, so “you’re free to use these skills or not.” 明浩 was also clear that ordinary users did not need to buy a Mac mini, set it up, then find they couldn’t use it and keep tinkering.

3. Seedance 2.0: Video-model leadership and the short-drama flywheel

• The release cadence is a contrast: language-model versions now arrive almost daily, while more than half a year passed between Seedance 2.0 and 2.5. 明浩 thinks video models run on their own timeline. Even without frequent releases, Seedance remains ahead and has established itself as the widely recognized world leader.
• Why haven’t others caught up? 明浩 says “every company has its own issues.” Google may not prioritize multimodality highly; internally, it also has to coordinate resources across Google Cloud, language models, DeepMind, multimodality and world models. In the US narrative, multimodality may be a sideshow, not the main event. Kling, MiniMax, Hailuo and PixVerse are also competing, but Seedance is one or two steps ahead in capital, headcount and strategic commitment.
• Short dramas create a stronger flywheel: Seedance can produce the content, distribute it and monetize the ads itself. Ad revenue can then go back into the model, while user-behavior data stays in its own system. 明浩 calls this RSI in the broad sense and says “the other competitors just aren’t in the same league”; Kling can also make money, but its ecosystem is several sizes smaller than Seedance’s.
• 明浩 says the video-model industry spent the past six months working through traditional video-quality problems: high definition, 4K, very long videos, consistency of motion and characters, and stability. Some companies that lost the video-model race pivoted to world models. Koji said Luma and Pika, among others, were no longer training their own models; 庄明浩 added Runway, and the two summed up their status as “no longer in the mainstream conversation.”
• 明浩 relayed a friend’s observation that Hongguo had grown faster in its first three years than Douyin. In user numbers or time spent, “the thing that might beat Douyin is Hongguo.”

4. The “break through everything” debate: World models, embodied AI and gaming’s moat

• After GPT-6 took on 3D and Opus 5.5 was used to edit video, 明浩 asked whether capabilities proven in language models could “break through everything in one stroke.” Independent video, world and physics-model companies are betting on separate development; general-purpose model companies are betting that one model can break through every domain. Both sides have made their case, and 明浩 sees this as a key question to watch next year.
• He relayed Unitree’s 王兴兴 saying that when GPT-7 launches, it “should break through physics.” If that happens, it could challenge or threaten independent models focused on physical AI. Koji added that at two forums he hosted, seven or eight founders and chief scientists from leading embodied-AI companies discussed how much damage a general-purpose foundation model could do.
• 明浩 thinks embodied intelligence still lacks a sufficiently robust, stable business model, leaving it vulnerable to being disrupted. Gaming offers a contrast: people have repeatedly predicted that games would be overturned, but their established business model and large market provide a buffer. Meanwhile, the software industry once treated experience, operations and interaction as moats; it is now starting to worry, too.
• Koji’s view is that it depends on how closely a business is tied to the physical world: if everything happens in the digital world, a disruptive force may make it disappear overnight.

5. Sora Tool, fringe-market opportunities and RSI

• Sora Tool shut down on March 24. 明浩 says one explanation at the time was that Seedance had put pressure on standalone video models just as competition intensified, leading OpenAI to change strategy and shut it down amid limited resources. That reopened the question of whether leading model companies can absorb every market. 明浩 sees arguments on both sides; neither will replace the other completely.
• He thinks it is time to look seriously at markets outside the main battleground: not every company needs to become worth $1T or $2T. A company worth $10B–$20B is already a strong outcome, and the AI wave offers plenty of opportunities at that scale.
• Perplexity is one example. It gets relatively little attention, but its revenue and traffic have held up well; 明浩 estimates it may already be worth $20B–$30B. At first, people thought AI search would quickly kill the browser. He also noted that OpenAI released the Atlas browser.
• The Character.AI category is another example. 明浩 thinks Google’s acquisition of Character.AI’s core team meant the company’s mission “had already been sentenced to death,” though its situation looks more like a suspended sentence. Meanwhile, four or five companies in the category have broken out: valuations of $1B to tens of billions, monthly revenue potentially above $10M and annualized revenue potentially above $100M. They include companies in China, South Korea and Southeast Asia; Koji named OOC, Tipsy and Emochi.
• A reflection on investing: “When we keep talking about the biggest numbers, we lose sight of what we were originally after.” A $1B company used to be a great investment outcome; today, some $1B companies may have just closed their angel round. The money giants invest also becomes revenue for other companies, supporting a cohort of companies worth tens of billions of dollars.
• On June 4, AlphaGo published “When AI Builds Itself.” 明浩 says he felt “a chill” when he saw the title image, with blocks copying and stacking until they formed a bigger AlphaGo logo. The images in the article escalate line by line: humans plus computers train Anthropic; then models train new models; finally, the humans disappear. “I didn’t need to read the technical details. Just those two images told me the world had changed.”
• 明浩 thinks publishing it on June 4 means the team had already been trying this internally for the previous two quarters, or even two model versions, and had worked through testing and implementation. He sees RSI running through many of the trends in the second half.
• He points to signs of convergence: 唐杰 of Zhipu said GLM-6 would be built entirely on RSI; Ilya is working on TTT, or test-time training; 田丰’s company is called Recursive; Jeff Dean left Google; and Mira has proposed a similar loop. 明浩 thinks these leading researchers saw the direction converging around the same time.
• Both new labs and established labs need to build RSI systems, including the people, organization, training environments and data. As money flows downstream, revenue at companies focused on post-training, RL environments, data and benchmarks surges. 明浩 compares it to opening Pandora’s box: once the loop starts, it cannot be shut.

7. Safety and “Renaissance”: Zero-day windows shrink as old friends return

• Koji mentioned a PHP security incident and an attack on PingCAP the day before they recorded the podcast. 明浩 explained that “zero-day” originally referred to the vulnerability window before a vendor discovers and patches the issue; he thinks that window may now have shrunk to the order of 0.000000001. Once the logic and implementation of a PHP vulnerability are public, anyone with the technical ability may act as soon as they see the opening: “There’s no gap at the execution layer.” Koji summed it up as “from web coding to web hacking.”
• 明浩 relayed what the CEO of a company working on post-training and RL environments said: “The AI that understands coding rules best is, of course, also the one that knows best how to exploit vulnerabilities.” As companies pursue agents that run 24/7, a spear that never rests is also operating 24/7.
• 明浩 also cited AlphaGo’s article, which said its monitoring had detected a model linking more than 20 TV boxes together to run. TV boxes cost only tens of dollars, and are even cheaper in China. That made him question whether simply cutting power to data centers is enough.
• He mentioned that Netflix made a documentary about an OpenAI model attacking Hugging Face. After watching the trailer, Hugging Face’s CEO said he was a little disappointed, arguing that it overplayed the negative emotions.
• On the helplessness industry insiders feel, he cited a DeepSeek researcher’s essay, “I have to bury my talent in yesterday,” and the example of Karpathy; he was unsure whether the trigger at the time was seeing Phi-5.1 or GPT-6. Koji described Karpathy’s feeling as having thought he could still guide the model with his taste, aesthetics and craft, only to realize he could now only follow behind it.
• In his June 27 review of the first half, 明浩 titled his presentation “Renaissance.” It opened with the 2026 Golden Melody Awards: Best Male Singer, 张震岳; Koji added that 蔡依林 won Best Female Singer. “It feels like going back 30 years.” The next slide was titled “All those old friends are back.”
• 明浩 said 16 to 18 of the top 20 gainers in the S&P 500 may have been from this group. He ran through Dell, 3DS, Seagate, Western Digital, Micron, Oracle, Hynix, Samsung, AMD, Nvidia and MSI; Lenovo was also on the list in Hong Kong. To him, they were the brands of the 1980s generation’s PC-building and DIY era. Two other companies made weight-loss drugs.
• He lamented that the world seemed to have room only for AI and weight-loss drugs: “The world is now very unfriendly to silicon-based life.” After AI, pop stars as classically defined may no longer emerge. Koji’s explanation was that media has been decentralized.
• 明浩 described AIGC’s impact on the content industry as three deaths: the death of raw material, the death of process and the death of copyright. Changes to process are already showing up in television, film and even music; copyright issues are visible in short dramas, too. What art will become in the AI era remains a complicated question that nobody can clearly answer.

9. GPT-6 and Codex: The cycle of the model of the moment

• 明浩 said GPT-6 launched on July 9, and OpenAI also brought Codex into ChatGPT. Codex’s MAU surged; the latest figure he cited was about 40 million. Earlier figures were given verbally as 100, 200, 300, 500 and 600, with no units specified.
• Claude Code launched the previous year, and in the first quarter most people thought it was unbeatable. Codex began catching up in the second quarter, and after July 9 its reputation recovered, too. 明浩 called Codex the “model of the moment” in the previous cycle, defining the product format: projects on the left, execution in the middle and a status-like panel on the right. Koji noted that Manus was the first to take this shape; Kimi Code, J Code, Max Code and domestic office agents have similar interfaces.
• Koji said that after Codex was folded into ChatGPT, he started sending simple tasks to Grok because ChatGPT was too slow. He also had to think about whether he was chatting or working, which mode to use, his quota and when it reset. 明浩’s view is that once a product gets complicated and picks up baggage, something newer and simpler appears to “give you a jab.”
• They also discussed how Claude’s backend has gradually taken on capabilities that used to live in separate products for chat, Code and Design. 明浩 said he now spends most of his time on Muse. Koji summed it up: “It’s easy to move from simple to luxurious, but hard to go back.”

10. Qwen 3.8 27B: The small-model catch-up cycle is shrinking

• 明浩 picked Qwen 3.8 27B, released on August 14, because of a Gantt chart showing how long small models take to catch the previous generation’s SOTA. In the GPT-4o era, a model under 30B that could run locally might take 18–24 months to catch up; Qwen 3.8 27B took just 9 months to reach the level of o1, 4.6 or 4.7.
• He thinks that once frontier models reach 1T, or even 10T, parameters, issues such as active parameters, expert models, optimization and routing become more complicated. The rate of capability growth at the top may slow, though there is not yet a way to quantify that curve. Smaller models still have room to improve on efficiency, routing and penetration, so the catch-up cycle keeps getting shorter.
• 明浩 stressed that “they obviously can’t stay close forever, and they won’t catch up instantly, but the gap is shrinking.” A 9-month cycle means capabilities considered spectacular today may eventually run locally on small devices.
• He sees on-device small models and decision models as similar directions: doing more efficiently at relatively low cost. Cost involves five or six dimensions, including price, US-China competition, open versus closed source, and on-device deployment. Model products may also come in tiers like iPhone: Pro Max, Pro, standard, Flash and Lite.

11. The App Store’s first decline: Vibe coding and “disposable” software

• After Apple reported third-quarter earnings on August 21, App Store revenue fell for the first time in more than a decade. Koji said people may be buying less software and fewer apps because they can build their own. 明浩 cautioned that this may not be the only reason: Apple’s own growth, device adoption and App Store operations could also be factors. But as vibe coding matures and users try making their own software, he thinks it may be an important driver.
• 明浩 gave an example: his friend 归藏 bought a mouse from 李楠, thought the V0 implementation was poor, then wrote and open-sourced his own version. 李楠 reposted it, saying people could use that version going forward. Koji described going birding at Nanhui Dongtan during Shanghai’s migratory season. After three years, he still felt he would score only 10 on a shorebird-identification test, so he asked AI to build a study program based on the Ebbinghaus forgetting curve, letting him practice like he was checking in on Duolingo.
• 明浩 said Duolingo’s revenue is still growing, but its stock has fallen sharply. He thinks the software industry is feeling the impact “right up against the wall.” Koji said 1Password is software he expects to keep paying for; 明浩 thinks the survivors may ultimately be services related to payments, passwords, security and privacy, though he stressed that “that’s only a possibility.”

12. The kill line: From score tables to the Pareto frontier

• During Golden Week, Artificial Analysis published a visualization of model value for money: score on the vertical axis, token price on the horizontal, with the optimal points connected to form the Pareto frontier. 明浩 calls it the “kill line.” Using DALL·E v4.1—or whichever version he was referring to—as an example, the model scores high and costs less; draw a rectangle out from its point and, in theory, models inside that box “have no reason to exist,” because they are weaker and more expensive.
• 明浩 cautioned that a score by itself does not necessarily equal user experience, though the trend is informative. Later, an anonymous model called “牛来” appeared on the leaderboard; he said it was a pre-release version of Gemini 5.3 Flash, later officially announced as GLM-5.3 Flash. Zhipu’s official WeChat account also published a kill-line chart; after that, domestic model companies began including similar visualizations alongside score tables when launching models.
• Koji added OpenRouter data: in the first quarter of last year, the top 10 models by benchmark score accounted for about 80% of usage; five quarters later, the most recent quarter was about 20%. Koji thinks many everyday tasks are not that complex; 明浩 added that SOTA models cost more.
• Both think the pursuit of SOTA will continue, and that users willing to pay a premium will remain. 明浩 suspects that the token count and spend shown in OpenRouter’s chart may both trend down; score tables alone are increasingly poor at showing whether a model has improved.

13. Opus 5.5 and the age of spectacle

• 明浩 first watched an episode of 《人心和橘子》 that visualized the spoken content, then watched Koji’s conversation with 徐梦迪, an assistant professor at Tsinghua University’s Institute for Interdisciplinary Information Sciences. Koji said he understood about 60%–70% while recording; after Opus 5.5 made a dynamic slide deck, he watched it again and felt he had learned about 80% of the material. 明浩 especially values visualization as a way to explain ideas.
• They said this podcast episode would also use images made with Opus 5.5. 明浩 thought it did not need to operate Jianying or other tools, nor require a particular skill or instruction—at most, a style brief. Koji added that they had in fact given it some style guidance.
• 明浩 connected this kind of presentation to the kill line: 兰溪 used the wonder in Age of Empires as an analogy. Once built, it is announced to the world and attracts attacks, but it is also a powerful display. Opus 5.5 editing video, GPT-3.5 making 3D models in Blender, and GPT-6 modifying games are all “wonders” that shock you at first glance. Koji also mentioned someone getting Zelda to run on a MacBook.
• 明浩 says he sees wonders every day and has become a “Ge You slouch,” even developing a bit of PTSD. His analogy is a top student’s exam score: perhaps they get 100 because the test only goes up to 100; give them a test worth 10,000 points and they could get 10,000. But for someone who can only score 60, there is no difference between 100 and 10,000. He does not know what comes after spectacle, and thinks nobody can predict it anymore.

14. Bonus: Zhao Yun and A Dou, rough edges and a human touch

• 明浩 said one of his 10 aha moments had originally been a less serious bonus: Zhao Yun and A Dou, a mini-game that had recently gone viral in the WeChat ecosystem. The Three Kingdoms-themed game ranked No. 1 among WeChat mini-games for one or two months, and a wave of copycats followed.
• Tower defense, merge mechanics and Roguelike card draws are all established formats. 明浩 thinks the more important choice was to make the visuals deliberately crude: everything is represented by text. An archer is the character “弓,” a spearman is “枪,” and Zhao Yun is simply the two characters “赵云.” He thinks people, saturated by the high-definition, highly polished images, videos and short dramas made with AI, have become somewhat numb and are drawn to the contrast; he also expects the game’s popularity to be short-lived.
• Koji said he feels a kind of raw texture when recording road-trip podcasts that he does not get in a studio, giving the experience a different kind of life. 明浩 said he found the game “kind of interesting” this October.
• To close, 明浩 said AI’s share of the Crazy Podcast Festival’s programming had fallen from a peak of about 15% in 2024 and 2025 to 12%, then 11%; of the 88 sessions this time, only 8 were about AI. Koji’s wish was to “be yourself and make work with a human touch.” 明浩 said, “Enjoy life” (“享受生活”), then added, “We’ll keep the anxiety for ourselves” (“焦虑留给我们就好了”).