43. “26Q2” Large-Model Midyear Report: RSI, Fable, 5.2, Hunyuan 3, LongCat—Kill Lines Everywhere — 王铁震

Summary

• 王铁震的期中考判断：中国模型迭代速度已经上来；在后续中国模型持续超过 GLM-5.2 的前提下，超过 Fable 只是时间问题。 GLM-5.2 has “basically replaced Opus 4.8” in his live workflow: Fable handles task decomposition, while GLM-5.2 does the work. Wang sees no paradigm shift in Fable; given enough compute, data, researchers and time, it can be built. The market conversation relayed by Raymond: Elon Musk said Chinese models would reach Fable’s level in Q1 2027, and 唐杰 replied, “It won’t take that long.” Zhipu shares rose “20% or 30%” the next day, hitting a recent high.
• 100% RSI (recursive self-improvement) is considered extremely difficult to reach: like entropy increasing, a fully autonomous iterative loop will produce an ever larger mess of code until complexity exceeds the model’s context and intelligence. The system must continuously ingest low-entropy, high-value human input from the outside world. Raymond paraphrased an Anthropic paper as saying that 80% of the code in Claude Code is written by Claude Code itself; Wang estimates that “perhaps 60%” of that still reflects direct human instructions. One potential RSI metric is a gradual decline in the share of human participation, although the original wording later says the ratio will “forever approach 100%”; the direction of the ratio is ambiguous.
• The model’s fundamental ceiling is that “the corpus a model consumes is, in practice, its shackle.” A model can transfer methods from field A to field B, but creating an entirely new field C with a completely independent method looks like a hallucination under its training objective. Humans can compress complex systems into invariants and modules through abstraction; that is why researcher taste is valuable and why Recursive’s large financing round makes sense. For now, RSI is at least easier to push along the efficiency frontier.
• Wang characterized Fable’s national-security-style launch as “very good marketing,” not a paradigm-level leap. Using Kimi K2.5 and K2.6, he found the FreeBSD vulnerability promoted by Fable after being given a hint. His guess is that the gap lies mainly in offensive-security data and model size, not in the paradigm. “The kill line for open-source models has moved up,” and Fable now has to create meaningful separation and tell the story well. Raymond therefore compared Anthropic to having trained “an Einstein gone dark.”
• Open source is commoditizing models, with a direct investment implication: “Moutai will still be No. 1 and command a premium; but it will be hard for the baijiu in second place to raise prices.” China has at least 10 companies capable of building models, including Meituan and Xiaomi. With too many players and free models, even the leader may not be able to keep raising prices over time. Open source served as a “100 into 10” filter in the primary market, but “investing in the eighth-place baijiu is a bit stupid.”
• China has “kill lines everywhere,” while data silos could push the industry toward “one big model per capita.” Meituan trained the 1.6T-parameter LongCat 2.0 on a 50K domestic-chip cluster, and non-model companies are continuing to raise the floor. Because data across Chinese apps is not connected, Raymond worries that every major company will build its own model. Wang agrees that data silos are one reason each company needs its own model; the US has similar dynamics, with xAI’s acquisition of Cursor tied to its programming data.
• Hunyuan 3 is the quarter’s key signal: its architecture is unchanged from the preview, yet swapping in new data and training methods delivered a huge performance jump. Wang’s reasonable outside-in inference is that architecture research is mainly about lowering costs, not necessarily directly improving capabilities. That in turn suggests data may matter enormously. Tencent has data channels through WorkBuddy, WeChat and Official Accounts; if it can use them appropriately and legally, Hunyuan 3.5 and Hunyuan 4 could still deliver qualitative leaps. Hunyuan 3 is smaller than GLM-5.2 and can run on a single domestic machine with 8 GPUs.
• Raymond’s non-consensus view: “Coding” is the wrong word; it should be called “Automation.” Every model company is moving into coding because it covers work with relatively high, replaceable hourly value, underpins many digital tasks, and offers a natural feedback loop through compilation and execution results. The TAM is not programmer compensation but every automatable interaction in the digital world. On whether OpenAI should spend $60B to acquire Cursor with its own stock, Wang says Sam’s motivation is weaker than Musk’s, given Musk’s idle compute capacity; Raymond thinks the deal is mainly about data. The Q3 questions remain whether Anthropic and OpenAI will “very likely” go public, what their financials look like, whether AI stocks will be re-rated, and where the kill line ultimately settles.

Deep dive

1. GLM-5.2 Displaces Opus 4.8: RSI Is a Horse Making Its Own Saddle

• Raymond named RSI a Q2 keyword, alongside Recursive’s large financing round, 田渊栋 joining as a co-founder, and the Anthropic RSI paper he relayed. Wang opened with his biggest surprise of the past 3 months: the release of GLM-5.2. “In real-world use, it has basically replaced Opus 4.8.” His workflow is already approaching self-improvement: Fable breaks down the work and assigns it to GLM-5.2. Opus 4.8 is “like a very senior employee at a company who is always bargaining with me… a backtalking employee”; GLM-5.2 is slightly less intelligent than Fable, “but extremely diligent and hardworking—it can finish the job in one sustained push.”
• The distinction between RSI and AutoResearch can be explained through a compiler bootstrapping itself: the first C compiler was written in an earlier language, and later versions compiled themselves. AutoResearch is “I optimize something else”; the experiment itself is not being optimized. RSI is “I optimize myself—building on AutoResearch, I string myself together into a loop.”
• Anthropic’s claim that “80% of the code in Claude Code is written by Claude Code itself,” and the regulatory concerns around it, need to be split into 2 layers: the ratio of purely hand-written code to machine-generated code, and whether humans are still directly dictating the machine-generated output through prompts. Wang estimates that of the 80% machine-generated code, “perhaps 60% is the direct expression of human intent.” Software engineering will move from a white box toward gray and black boxes, but he is “not that pessimistic”: just as humans moved from assembly language to high-level languages, people will write a higher-level form of “code” in the future—prompts, design and business logic.

2. 100% RSI Is Hard to Reach: This Is an Entropy Problem

• Wang believes 100% RSI will be very difficult to achieve. His physical intuition is that a system running completely autonomously, with no external information, eventually hits a bottleneck: “The pile of garbage code it writes will keep getting bigger until its complexity far exceeds the model’s context-processing capacity and intelligence, and then it will get worse and worse. It has to keep ingesting low-entropy information and extremely valuable human input from the outside world.”
• A potential metric for RSI progress is a gradual decline in the share of human participation. But Wang also said that the ratio would “forever approach 100%, but be difficult to reach”; the direction of the ratio is ambiguous, and no additional interpretation has been imposed.
• His own RSI experiment chain is worth recording. He is having GLM-5.2 first generate its own dedicated Harness—“let the horse make its own saddle”—using “a big horse with a big saddle” (Fable plus Claude Code) to build a small saddle for the small horse. Next comes GLM-5.2’s own inference engine, which already runs but slowly; after that, training beginning with LoRA. “Once this chain works, my RSI loop works.” The current objective is simply bootstrapping; improving a specific capability comes later.

3. Wish-Driven Programming Is a Lottery Ticket; Abstraction and Corpus Are Model Constraints

• Wang described his own progression from filling in a function to “wish-driven development,” before realizing, “This project has nothing to do with me anymore… I spent a lot of effort just clicking Next, Next, Yes.” His metric is the size of the project the model can complete correctly in one pass: a 20K-line prototype is fine, but when new features are added, “it changes the prototype’s code beyond recognition.” The solution is for humans to define the abstractions, lock the module boundaries, and let the model modify only the designated A, B and C modules.
• The underlying difference is that a model stores the world as an average in its weights, while humans can abstract invariants and modules from a mass of observations. Humans derived Newton’s second law from countless events on Earth, and can compress 300K-line or 1M-line codebases into a few well-connected modules. Wang’s role is to give the model the right abstraction and have it implement accordingly.
• When does wish-driven programming work? “When you have an expectation for the result but no expectation of reproducibility or stable output… that’s scratching a lottery ticket.” When a $1B funding gap appears in production, “you can’t instantly scratch a lottery ticket and make it appear.”
• Experiments show that if the model has seen something once in its corpus, it can do it—writing a CUDA kernel is no problem. But “ask it to optimize against an objective, and it just circles around and around without finding a way out.” Creating something whose correctness nobody yet knows how to judge “is a hallucination under its training.” Raymond added that every scientific discovery looked like a hallucination at the time: “For the past 20 years, everyone thought Elon Musk was hallucinating too.” Faced with reusable rockets, a model might list “137 reasons why it can’t be done” from existing data.
• That leads to the network effects of intelligence: “We all know Anthropic is distilling all of us.” A new version solves an old problem not because intelligence appears from nowhere inside the model, but because “some people on the other side of the planet discovered a new path, and it happened to be distilled into the model.” So “old-school craftsmanship is still extremely valuable—otherwise we get locked in.”

4. The Cost of Intelligence Everywhere: “Isn’t It Just Selling Me?”

• Wang sees OpenAI’s Record and Replay—recording how you work, learning from it and reproducing it—as an inevitable direction. Product building will ultimately involve 2 things: collecting human intelligence-interaction data as efficiently as possible, and deploying intelligence into every scenario as quickly as possible, “claiming every fragment of your day that requires you to think.”
• The deployment surface is also expanding. Claude Code and Codex started in TUI, or terminal user interfaces, and will move into apps because apps have more users. Anthropic also offers a shortcut that summons a question box, allowing users to ask questions anywhere on the screen. A further step is for applications to proactively surface information the user may care about, things they did in the past and suggestions for what to do next.
• The privacy warning comes with a barb: ChatGPT Enterprise does not use data for training by default, while the subscription version opts users into training by default and requires them to opt out. “Think about why these 2 cost 100 times as much—the product being sold is me.” Wang’s advice is to “use open source whenever you can,” though he also believes people living in society cannot completely avoid this kind of data collection.

5. Fable’s Launch Was a National-Security Spectacle: Anthropic Trained a Dark Einstein

• Wang’s assessment of Fable’s launch was a single phrase: “very good marketing.” “National security is a basket you can throw anything into.” Raymond added that he initially assumed Fable represented national-security-grade capabilities because it claimed to find large numbers of zero-day vulnerabilities. Most of the first trusted users were financial institutions, by Raymond’s recollection; he could not remember whether the Pentagon was among them.
• Wang’s empirical test was to use Kimi K2.5 and Kimi K2.6 to hunt for vulnerabilities highlighted by Fable, including the FreeBSD vulnerability. With a hint, “they can find the same problem.” He sees 2 main sources for the gap: open-source models have not been trained on offensive-security data, and there is still a size gap. Several months ago, open-source models at the 1T scale could already “find the problem if you tell them which file to look at”; if MiniMax reaches 2T or 3T, perhaps they will need less prompting to find the issue directly.
• The conclusion is that there is no fundamental paradigm gap, but “the kill line for open-source models has moved up.” Fable must pull a full step ahead of open-source models and sell the story; open source is cheaper and has other advantages. Raymond asked why people claiming the highest ethical standards would assemble a corpus of malicious actors. Wang agreed that Dario had effectively “trained an Einstein gone dark,” tying it to the mindset that says, “If there’s no problem, create one—there has to be trouble.” Raymond’s initial post-launch anxiety had shifted to, “This isn’t something that can’t be done.”

6. Musk’s Six-to-Nine-Month Wager, and Baijiu Beyond Moutai

• Q2’s biggest piece of gossip was Musk saying Chinese models would reach Fable’s level in Q1 2027. 唐杰 replied, “It won’t take that long.” Musk responded, “It’s not about leaderboard scores”—the only standard for useful intelligence is revenue. Raymond interpreted Q1 2027 as 6 to 9 months from the time of the conversation. Zhipu shares rose “20% or 30%” the next day to a recent high, and Raymond disclosed that he had already bought Chinese-model stocks.
• Wang revised his previous quarter’s judgment. He had thought Chinese-model iteration might be slower than in the US, “but look at the past few months—that’s not what happened.” The reason is unclear: China may be investing heavily in infrastructure, or its model architectures may be more efficient. If iteration has in fact accelerated, and every Chinese model still to be released has to beat GLM-5.2—“if it’s not as good, it may not be released”—then surpassing Fable is only a matter of time.
• He refused to give a firm answer on who gets cut: “It’s hard to say who gets pushed off the table.” Meta’s new model is said to be approaching 5.5, and xAI has come back into the game after acquiring Cursor. “It seems like everyone can turn the tables every few months.” The significance of open source is that “open-source progress is everyone’s progress”: without it, the wrapped models in the 2023 model war were opaque and impossible to underwrite. Open source served as a “100 into 10” filter. Wang says China has at least 10 companies that can build models, including Meituan and Xiaomi.
• Raymond’s investor-style cold water is that there are too many players at the table. Anthropic still had an overwhelming lead in Q2—“it could name any price and any valuation”—but that no longer seems to be the case. Wang’s conclusion: open source turns models into supply-chain commodities. “Moutai will still be No. 1 and command a premium; but it will be hard for the baijiu in second place to raise prices. That’s not a bad thing… it’s just that investing in the eighth-place baijiu is a bit stupid.”

7. LongCat, Hunyuan 3 and Data Silos: Perhaps One Big Model Per Capita

• Meituan’s LongCat 2.0 is a 1.6T-parameter model trained on a 50K domestic-chip cluster, entirely without Nvidia GPUs—an advance beyond Ant’s hybrid-GPU attempt. Wang calls it “a very major breakthrough.” Training is much harder than inference: GPU speed, inter-GPU communication, numerical-precision stability and software adaptation are all hard problems. The model itself is also well optimized and runs quickly.
• Raymond’s broader takeaway is that “China’s kill lines are everywhere.” If a non-model company can develop a side business to this level, it will not casually buy someone else’s API. With data across Chinese apps completely siloed—Dianping cannot even answer a question such as “what is the quietest coffee shop within 300 meters of the mall?”—he worries the industry could end up with every major company building its own model, effectively “one big model per capita.” Wang added that the US has a similar dynamic: the programming data personally generated by Cursor users is extremely valuable. “Why does xAI want to buy it? I think it has everything to do with the data.”
• For Tencent shareholders stuck in the stock, Hunyuan 3 is the key passage. Its architecture was unchanged from the preview released in February and March; it uses a classic architecture and changed only the data and training methods, “yet achieved a huge leap in performance.” Wang stressed that this was a reasonable outside-in inference: “People research model architecture mainly to reduce costs, not to improve performance.” The reverse implication is that data may be far more important than expected.
• Tencent’s ammunition includes WorkBuddy, WeChat and Official Accounts data. Raymond said, “I have no secrets from Tencent.” If Official Accounts data can be used appropriately by Hunyuan, “Hunyuan 3.5 and Hunyuan 4 could still deliver an even bigger qualitative leap.” Hunyuan 3 is smaller than GLM-5.2 and can run on a single domestic machine with 8 GPUs.
• Raymond cited an exclusive, in-depth LatePost report: Tencent gave Hunyuan substantial organizational latitude over the past year. 姚顺雨 did not chase clever tricks, architecture innovation or benchmark scores, but instead said, “First, stand in horse stance; hold it for 6 months. That is real wisdom.” Data is alive: programmers writing complex tasks, providing feedback and fixing their own bugs is expensive data. “Chatting with my mother isn’t very valuable.” Wang added that data must be used reasonably and legally; using collective intelligence while protecting privacy is a major issue. A brief aside on MiniMax’s 2.7T model: the fact that it can be released shows the model is good, but falling too far behind peers would damage the brand.

8. “Coding” Is the Wrong Word: This Is Automation

• Wang put the logic behind every model company’s shift into coding plainly: it covers work with relatively high, replaceable hourly value—Google engineers may earn $300K-$400K a year, and 30% to 40% of their work may be replaced; code underpins many tasks; and “writing code naturally creates a data loop.” Compilation, execution results and the gap between those results and expectations all feed back into the system. Role-play models cannot access data of comparable quality.
• Raymond says this remains a non-consensus view as of July 2026: “The word is wrong. It shouldn’t be called Coding; it should be called Automation.” Investors who calculate TAM from the number of programmers are misunderstanding the market. An accountant’s Excel work and a designer’s frontend work ultimately become coding implementations. “Buddy” is fine as a company name; just don’t put “Code” in front of it. The 2 also imagined how the meaning might evolve: “Maybe when my son grows up, Code will mean saying a sentence or making a wish”—an Aladdin’s lamp, make a wish.
• On the thought experiment of whether OpenAI should use $60B and its own stock to acquire Cursor, Wang believes Sam’s motivation was weaker than Musk’s. xAI once had idle compute capacity, Musk needed a company with data and a product, and “Musk is an extremely imaginative person.” Raymond admitted he did not understand the deal at the time: Claude Code already existed and Cursor had launched Composer, so buying it anyway means “the thing being valued is completely and entirely the data.”
• The close is a string of open Q3 questions, with Raymond repeatedly saying “I don’t know”: Can Anthropic deliver an even more impressive answer? He thinks it is “very likely” that Anthropic and OpenAI will go public in Q3. What will the financials look like, and how will the IPOs trade? Will they trigger a re-rating of AI stocks? Will more Chinese companies release models? And where will the kill line ultimately settle? “Those are all Q3 questions.”

Verification Notes

• The original uses both “the share of human participation gradually declines” and “forever approaches 100%” in its RSI metric discussion. The direction of the ratio is ambiguous; the wording has been preserved without attempting to correct it.