The Open-Source Kill Line Has Risen; Passing Fable Is Only a Matter of Time

“The corpus a model eats is in fact its shackles.” — Wang Tiezhen, on why AI can’t create anything new

In July 2026, Raymond went back and listened to the Q1 large-model episode recorded three months earlier, and it felt like another lifetime — “the me from last episode sounds like an old fart, out of date, and not especially knowledgeable about the technology.” In three months, Chinese models went from a big three to a big four to a big five: GLM-5.2 displaced Opus 4.8 in real workflows; Meituan, in the gaps between the food-delivery wars, trained a 1.6T model called LongCat on fifty thousand domestic chips; Tencent skipped the clever tricks, held the horse stance for six months, and delivered Hunyuan 3. Across the ocean, Anthropic turned Fable’s release into a piece of national-security theater, and Musk and Tang Jie made a long-distance bet on the timetable for Chinese models catching up — six to nine months, or “it won’t take that long”?

In this episode, Wang Tiezhen, Hugging Face’s China lead and the guest whose last appearance drew nearly two million plays, returns to hand in a midterm paper: what RSI actually is, and why 100% self-evolution is a physically unreachable problem; why wish-driven programming is a lottery scratch card; why Anthropic had to train “a corrupted Einstein”; and, as open source pushes the kill line ever higher — where the first place is Moutai and the second-place baijiu can’t raise prices — where an investor’s hundred dollars should go.

What follows is the full conversation, edited and condensed.

1. “Fable leads, GLM-5.2 does the work”: RSI is letting the horse build its own saddle

Raymond: It’s July 2026. Three months ago we recorded the Q1 2026 large-model episode, and listening back last night it felt like another lifetime — this industry changes far too fast, and the me from last episode sounds like an old fart and out of date, but that was three months ago. Chinese models seem to have caught up with America, going from a big three to a big four to a big five. Today we have season one’s guest, Wang Tiezhen, back to discuss the 2026 half-year report, the midterm on large models. That episode was very popular; video plus audio came to nearly two million plays. The opening topic is RSI. Last season, when I asked about the breakthrough that most impressed you, you talked about the AutoResearch that Karpathy described. Over the past few months, RSI has become the most widely discussed term. What is RSI? And how does it differ from AutoResearch?

Wang Tiezhen: Let me step back and cover what happened over these months. The thing that shook me most was the release of GLM-5.2 — in real-world use it essentially replaced Opus 4.8. My workflow now is close to self-evolving: Fable does the division and decomposition of work and dispatches tasks to GLM-5.2, and I barely use other models. Opus 4.8 is like a very senior employee at a company, always haggling with me: your requirement isn’t clear enough, can you confirm it again; I don’t think what you said is right, the scope should change. But on the things I’m doing, I understand them better than it does — it’s an employee who talks back. GLM-5.2 may be marginally less clever than Fable, but it’s extremely diligent and conscientious and can carry a job through in one push. For long-horizon tasks where I don’t want to intervene, it’s very well suited.

The long-horizon task I’m working on is getting AI to “bootstrap.” That’s a common term in compilers: a compiler turns Python or C++ code into machine code a machine can read, and the most challenging thing is compiling itself — the first version of a C compiler was written in an earlier language, and after that C is written in C, compiling itself, iterating continuously. RSI is very similar: first you stack up a sufficiently clever model using human intelligence, and then it needs less and less human intervention, obtaining information from the outside world itself, optimizing itself, iterating itself, working toward some goal. AutoResearch is me optimizing something else — lowering the loss, say, having the model run various trials, but the act of running trials isn’t itself optimized. RSI is me optimizing myself — building on AutoResearch and threading yourself through it so it becomes a loop.

Raymond: I ask because RSI may be the most important keyword of Q2: Recursive raised a large round with Yuandong Tian joining as co-founder; Anthropic also published that RSI piece — 80% of the code in Claude Code is now written by Claude Code itself, and the human share of Opus’s progress keeps shrinking. The worry the article describes is like raising a child: in first or second grade he can learn on his own, by third through sixth grade you don’t need to supervise, and by middle school you find he’s picked up bad habits and isn’t fully aligned with you — recursive drift has already happened — but he’s hit adolescence and you can’t control him. So the piece calls for the world’s governments to jointly regulate AI. Setting the apocalyptic scenario aside: has any company achieved RSI? Once you have it, the pace of evolution is infinitely fast and nobody can ever catch up.

Wang Tiezhen: I don’t think 100% RSI is achievable. From my own exploration in this area, it feels a bit like entropy increase in physics: making something fully autonomous, looping and iterating with no external information at all, has a bottleneck. The spaghetti code it writes gets bigger and bigger until the complexity far exceeds the model’s context-handling capacity and its intelligence itself, and then it gets worse and worse. It has to keep ingesting low-entropy information from outside during the process, extremely valuable human input. A potential metric for RSI progress is the human participation share gradually falling — always approaching 100%, but very hard to reach.

And you have to separate two layers. One is the ratio of purely human-written code to machine code — machines write fast, so human code gets diluted quickly. If you look at impact, you have to look at how the machine code was generated: even at Anthropic, I believe there are still large numbers of engineers prompting Claude Code, and what it generates is code under human direction. Of that 80% machine code, perhaps 60% is directly directed by a human.

Raymond: As long as it’s a human’s direction issued through a prompt, it’s still fine.

Wang Tiezhen: But traditional software engineering is a white box inside a company, and as machines participate ever more deeply it will gradually become a grey box and then a black box: humans define the architecture and define what good and bad code are — do you want a black box or a black sphere, humans can set that frame; but internal predictability and interpretability are basically beyond human grasp. I’m not that pessimistic though. Decades ago it was possible to write an operating system entirely in assembly, and later humans moved from writing assembly to writing higher-level languages. Apply that today: C++ is already too far from human perception, so let machines write it; humans write the higher-level “code” — prompts, design, the actual business logic.

2. Wish-driven programming is a lottery scratch card: the model is trapped in its corpus, and a human’s value is walking over and picking it up

Wang Tiezhen: Let me describe my own path using AI to write code. At first I used it to finish a function: what goes in, what comes out, write me a line or two. Later it wrote tests automatically. Later still I’d split modules up and throw them at it. Then it got close to wish-driven development — I say I want to build something and it does it according to its own understanding. And then I found the project had nothing to do with me anymore, like a project I’d picked at random on GitHub. I had no feeling for it and no idea what it was doing. Once a project gets complex, ask it to solve one problem and it robs Peter to pay Paul: solved here, and a new problem created over there.

Raymond: You wouldn’t necessarily mind, would you? It works, and it isn’t something meant for many people to use.

Wang Tiezhen: I mind a lot — past a certain level of complexity it basically stops working: I can’t read the code, and I don’t know what the AI is changing. People say wishing is very useful, but wishes always have consequences: someone wishes for money, genuinely gets a lot of money, and the consequence is their health deteriorates. Collaborating with AI feels exactly like that; it can always satisfy your immediate request, and as for anything else, it doesn’t care — that’s the extent of its boundary and its range of understanding. My original workaround was: two months in, the project becomes unmanageable, and then a new AI version descends from heaven — fine, Fable is out, help me clean up this spaghetti. After a few rounds of that I concluded it doesn’t work: I’d expended a great deal of effort essentially pressing Next, Next, Yes, and once the code came out I couldn’t predict its behavior at all.

Raymond: What kind of task makes it unpredictable? “Make me a coffee every morning” — if there’s no coffee the next morning, it got it wrong, and both the feedback and the correction path are clear.

Wang Tiezhen: Let me offer a metric: the size of project it can get right in one push. A normal project starts with a prototype and then adds things layer by layer. The prototype might be twenty thousand lines, which current models handle with no problem and get running correctly; but when you add something new, it changes the prototype’s code beyond recognition. My current approach is to prescribe its design: as the project evolves, I only need to change a particular module, and I tell it you may only touch modules A, B and C, and the module boundaries stay fixed.

Behind this is that humans and large models understand the world differently. What a large model learns is the world’s average: feed in all the corpus, memorize it, record the world’s information in the weights — it doesn’t necessarily understand the world. Humans are creatures capable of reducing complexity: we can abstract Newton’s second law out of all the things happening on Earth, and abstract three hundred thousand or a million lines of code into a few invariants and a few well-jointed modules. What I can do is, each time I think of a good abstraction, tell it to implement accordingly, and keep moving forward through collaborative iteration.

Raymond: So when does wish-driven programming hold up?

Wang Tiezhen: When you have hopes for the result but no expectation of reproducibility or stable output — you finish this one thing today and won’t use it again, so go ahead and wish. That’s scratching a lottery ticket: you might one day scratch a billion. But if something breaks in production and right now you have to close a billion-yuan funding gap, and you have no command of the code, no established trust, no understanding of the whole company’s balance sheet, you can’t instantaneously scratch a winning ticket.

Raymond: That article also described three states. One, models stop getting stronger and human civilization stops here — which honestly wouldn’t be a big deal. Two, models strengthen exponentially and need better abstractions and instructions to do better, at which point people with research taste are the bottleneck — which is also the core premise of Recursive’s fundraise, that researchers with taste are rare and valuable — and we appear to be in this stage. Three, RSI keeps accelerating. My challenge is: everything you describe as abstraction is stuff 5.2 can’t do, and in three months Fable will have shipped ten more versions — will it all be solved? What we’re discussing may all be the past — listen to this episode three months from now and it’ll probably be entirely irrelevant.

Wang Tiezhen: Let me describe my project and you’ll see why I stress abstraction and command. I want GLM-5.2 to start by generating a harness it can itself use — effectively writing a GLM-5.2-specific version of Claude Code, letting the horse build its own saddle. I first used a big horse with a big saddle — Fable plus Claude Code — to build a small saddle for the small horse, GLM-5.2, and the initial version exists. The next step is seeing which links in the chain I don’t yet command: inference uses off-the-shelf frameworks like vLLM and SGLang, so can I use the small horse with the small saddle to build GLM-5.2’s own inference? It already runs, just slowly. The step after that is training, starting from LoRA, where the computational complexity won’t be too large. Get that chain running and my RSI loop is closed, and I only need to give it my preferences. The current goal is only to bootstrap; the next goal is improving some performance metric.

But during this I found a very serious phenomenon: anything the model has seen even once in the training corpus, it can do — ask it to write a CUDA kernel and it does; ask it to optimize toward a target and it goes round and round and can’t get out. It can apply a method humans used in domain A to a domain B where humans have never used it; but ask it to create an entirely new domain C, using a completely independent method C, and it can’t. That isn’t what it was trained for — its objective is to predict the next token, to predict what humans already know. Creating something nobody knows is right or not is, from the standpoint of its training, a hallucination.

Raymond: That’s interesting — any scientific or technical invention looked no different from a hallucination at the time. When there was electricity and there were trains, Empress Dowager Cixi may well have thought it was sorcery; and for the past twenty years everyone thought Musk was hallucinating.

Wang Tiezhen: Which is exactly why a researcher’s taste matters — it directs hundreds of billions of investment. Ask the model and it may quickly tell you that recovering rockets isn’t realistic, because nobody has done it that way, and the weights of the existing corpus will tell it very clearly: here are 137 reasons it can’t be done. A model gets trapped in the corpus it was trained on; the corpus it eats is in fact its shackles. Humans are different; a human can, at a very slow pace — very slow compared with machine learning — explore the boundary of something bit by bit, and that’s a human’s native value. When I was building the inference engine, I looked at the execution graph and knew some spot could definitely be optimized — I can’t write CUDA, but I knew. And it simply couldn’t write it, going round in circles; put Fable on it, and Fable’s proposal looks beautiful but likewise can’t be written. It cannot lift itself. It’s actually very light; a human walks over and picks it up. But left to itself, it just can’t lift itself.

Raymond: Like a robot vacuum that can never make itself fly. So current RSI can only push the efficiency frontier and can’t really go beyond human cognition?

Wang Tiezhen: The efficiency frontier is certainly one part. Let me throw in another point: the distribution of intelligence — intelligence itself has network effects. We all know Anthropic is distilling all of us: one user writes a very useful skill and workflow for some scenario; another person in another scenario finds the current version can’t solve his problem; Anthropic distills both of them in, the next version comes out, and the previously unsolvable problem is solved. Where did that intelligence come from? Not from the model itself — some people on the other side of the planet explored a new path, and it happened to get distilled in. In any unit of time, the newest model is the aggregate of all knowledge workers’ intelligence at that moment — it looks very clever, but ask it to do something nobody in the world has ever done, an entirely independent new domain, and it can’t manage. Which is also why I think traditional handcraft still has great value — otherwise we lock ourselves in.

Raymond: It feels like there’s still hope for humans, that we can still walk over and pick it up, that I still have some value myself. “Every person is born with value” — I’ll cut that kind of inspirational line.

3. The price of ubiquitous intelligence: “what’s being sold is me”

Raymond: Last episode you had a core view: once past the threshold of usability, Chinese models can collect large volumes of work data, which is a huge boost to model progress. That was very instructive for me. This time OpenAI shipped a feature called Record and Replay: I want to revise a résumé, it starts recording and watches how I revise; I want to pick candidates out of a hundred résumés, it watches how I pick, and then the recording goes to ChatGPT — and next it repeats the same actions and finishes the rest. A complete set of worker data. What do you make of this interaction innovation?

Wang Tiezhen: This direction was inevitable. Building products in future is two things: first, collect as much human intelligence interaction data as possible in the most efficient way, not missing a single piece of data that might help improve the model; second, push the model’s intelligence down into every scenario as fast as possible. Pushing down means making intelligence ubiquitous: Claude Code and Codex both started from the TUI, the command-line interface, and both then have to build apps, because more people use apps. Anthropic has a feature I dislike a great deal but which is genuinely practical: press a shortcut and a question box pops up anywhere on screen, so you can ask any time, anywhere. In future there will even be applications on permanent standby: while you’re chatting or working, it pops up prompts, information, things you did before that might interest you — pre-empting every fragment of thought you need to have.

Raymond: While you were talking I was thinking my recording device may be listening to our conversation, and an enormous context has already been sent over.

Wang Tiezhen: Use open source where you can. Although it can’t be avoided — you live in this society. ChatGPT has a switch: the enterprise version doesn’t train on your data by default; the subscription version joins training by default and you have to go opt out yourself. Think about why the two prices differ by a factor of a hundred — what’s being sold is me.

Raymond: Let’s talk harnesses. This quarter OpenAI’s harness also improved substantially, and they merged Codex into ChatGPT. You’re building a small saddle for a small horse — is a harness hard to build?

Wang Tiezhen: Two observations. First, once enough people have done something, the model has seen it in the training corpus. Two years ago a harness was hard to build; today it’s easy: I say I have a model interface that looks like this, build me a harness that can do A, B, C and D, and out it comes. The papers exist, the Reddit threads exist, and every component on the saddle has been pulled apart and analyzed — it’s like walking into Huaqiangbei, Shenzhen’s electronics market, snapping a few parts together and assembling a new phone. But before the phone was invented, you couldn’t have found those parts in Huaqiangbei either. Second, what’s hard is user experience. Why does Claude Code ship daily? A simple harness can’t deliver an outstanding experience; you have to do the grinding, unglamorous work. For instance, one direction harnesses are evolving in is not letting the model be capable at every moment of deleting every file on your hard drive: Claude Code has a small model dedicated to checking whether a command you’re about to run will harm the computer; Codex doesn’t have that but has a sandbox; my own approach is very simple, have the model check the danger before each command and then run it inside Docker. Each vendor solves it differently, and it’s a matter of taste — which is why, beyond Claude Code and Codex, there’s also OpenCode and Cline, each with its own distinctive experience and design.

4. Fable was good marketing: Anthropic trained “a corrupted Einstein”

Raymond: Let’s talk about Anthropic, which has been in the eye of the storm all year. The most melodramatic part was Fable’s whole release process. What did you make of it?

Wang Tiezhen: Very good marketing.

Raymond: Do you think it was deliberate?

Wang Tiezhen: Sure. National security is a basket you can put anything into.

Raymond: When Fable came out I hadn’t used it yet, and my assumption then was: GPT arriving was tier one, it can chat; then the agent tier, it can do work — both of those are already history, and anyone can access them for twenty or two hundred dollars. At the Fable tier I did think it might genuinely be a national-security-grade thing — like the RSI we discussed, where once someone holds the key and climbs exponentially, their three months is nothing like your three months. Say it can crack banks’ security holes — I honestly can’t say whether defense or banking is more secure: defense is certainly secure, but fewer people attack it; banks have more money, so more people attack them. So when I saw that most of the first batch of “trusted users” allowed to use Fable were financial institutions — I don’t remember whether the Pentagon was among them — my reaction was: wow, that’s fairly extreme. Fable’s headline claim was catching a lot of zero-day vulnerabilities, and that struck me.

Wang Tiezhen: I ran a few experiments: using Kimi K2.5 and K2.6 to dig for the vulnerabilities it named, like the FreeBSD ones. Give it some hints — tell it which file might have a problem and let it look — and it finds the same issue. Once the problem is localized, the ability to find it isn’t the issue; the bigger problem is whether it can traverse an enormous surface and spot it immediately. So my guess at the difference between open-source models and Fable has two parts. First, open-source models haven’t been trained on offensive security corpus — Anthropic certainly pays close attention to this and must have collected large volumes of hacking and security material that open-source models don’t have. Second, intelligence level still relates to scaling law and model size: open-source models are around 1T, and a few months ago they could already do “tell it which file and it finds the problem.” MiniMax says it will do 2T and 3T models, and once size goes up, can fewer hints get you straight to the problem? I think that’s very likely. So Fable, in terms of paradigm, probably has no fundamental gap with existing models. But Fable had to make an extremely dramatic release — the open-source kill line has risen, so it has to open up a full body length over open-source models, and it needs a good story. Otherwise open source is far cheaper with all sorts of advantages, and people won’t use it.

Raymond: Let me confirm the breakdown: you think the marketing component of Fable’s release far exceeded any enormous product leap?

Wang Tiezhen: Yes. Its capability may be much stronger, but the paradigm hasn’t changed enormously.

Raymond: Then a second question follows. It wants to be the world’s only god-like “good side,” and to do that it has to learn how every bad actor does bad things — collecting, directly or indirectly, the corpus of every hacker in the world. Ask Kimi today and Kimi doesn’t have a department dedicated to researching how hackers hack, nor a vendor able to supply hacker corpus. Which is to say, the party claiming the highest moral standard is instead the aggregator of all malefactors’ corpus?

Wang Tiezhen: One part is right: without this much attack corpus, improvements in model intelligence alone shouldn’t produce attack capability this strong.

Raymond: In other words, intelligence itself doesn’t equal insecurity — the greater a person’s wisdom, the safer they may be. Einstein wouldn’t know how to hack NASA, whereas a sufficiently well-trained 19-year-old high schooler might — indeed Einstein would consider hacking NASA beneath him. So Dario’s situation is: he has collected every method by which 19-year-old high schoolers might want to hack NASA. You could say he has trained a corrupted Einstein.

Wang Tiezhen: Right, he trained a corrupted Einstein.

Raymond: So who is manufacturing insecurity? For his own purposes — his core logic is: I must prevent a corrupted Einstein from appearing, so I must first learn and first create a corrupted Einstein.

Wang Tiezhen: I think this is the same current of thought as the China threat narrative the Pentagon has always promoted: if there’s nothing there, manufacture something anyway.

Raymond: I feel slightly more relaxed now. When Fable first came out it felt like Chinese models had been left very far behind and I was anxious; then 5.2 came out, and after hearing you out — this isn’t undoable.

5. Musk framed it as six to nine months: “first place is still Moutai; second-place baijiu can’t raise prices”

Raymond: The most entertaining Anthropic item is also the 5.2 and Musk exchange. In late June someone asked when Chinese models would reach Fable’s level. Musk answered: Q1 2027. Zhipu founder Tang Jie replied: it won’t take that long, it’ll be quick. Musk immediately responded: not benchmark-gaming — a real large model has to be judged on whether it’s useful intelligence, and the only standard is revenue: high revenue means users are buying, and users buying means your intelligence is useful. That exchange was what capital-market investors were most focused on at the time. It settled a few things: the highest standard of intelligence is Fable, indisputably; Musk thinks China’s gap is at most nine months and possibly six; Tang Jie says it won’t take nearly that long. The next day Zhipu’s share price rose 20% or 30%, a recent high. From which listeners can also tell that Zhipu hasn’t bought advertising on this show. What do you make of it?

Wang Tiezhen: This was Q2’s biggest piece of gossip. We discussed last time that Chinese models’ iteration speed might be slower than America’s — but look at these past months, it isn’t; Chinese models’ iteration speed has come up. I’m not clear on the reason: it may be that China has put a lot of effort into infrastructure, or that the model architectures are more efficient than America’s. If iteration speed has come up, and every Chinese model about to launch will be better than GLM-5.2 — if it weren’t better, it probably couldn’t be launched — then surpassing Fable really is only a matter of time. Because there’s no paradigm change; it isn’t like ChatGPT 3.5 appearing out of nowhere with nobody knowing where it came from. You already know what the source of intelligence is, and doing preference tuning and RL on top of it, with more compute, more people and more corpus, produces higher intelligence. Give me enough time and enough money and it can be done. So we particularly need investors like you to buy some Chinese large-model company stock and give them a bit of support.

Raymond: I’ve already bought that stock. Next question: how do you judge that a company has left the table? Last episode we joked about Meta, that nobody knew where Llama was, and it promptly shipped a new model claiming rough parity with 5.5. xAI looked like it was about to fade out and then started running again after acquiring Cursor, and Grok’s newest model performs decently too. It seems everyone can flip the table every few months and return to it in a blaze of glory. How long without iterating counts as leaving the table? Is there a framework?

Wang Tiezhen: It’s hard to say anyone has left the table — from a capital-market contention standpoint they’ll always exist; the question is only whether they deserve today’s heat. If Chinese models genuinely reach Fable’s magnitude, Fable certainly won’t have today’s heat. There’s another very interesting point about the open-source world: progress in open source is progress for everyone. It pulls the whole checkpoint forward enormously — technical papers, model weights, exploration of new architectures, all of which benefit everyone. Your own model isn’t as good as the open-source one? No problem, take the open-source checkpoint and keep fine-tuning, put your own data on it. Airbnb and Cursor are both exploring this path. The better open source is, the better the ecosystem.

Raymond: But here’s the part investors don’t like: too many people at one table. If I want to invest in Zhipu, the question is simple: what does Zhipu’s future look like? You tell me anyone can do open-source models — Anthropic has the highest ARR today, and yet its service isn’t that much better than anyone else’s. If that’s true of the industry’s number one, the pressure on number two and three is greater. In the end you may have a dozen or twenty people at one table, and nobody makes much money.

Wang Tiezhen: Setting Zhipu aside, let me describe how open source helps investing. Back in 2023 there was China’s hundred-model war, with a hundred models all claiming ChatGPT-level capability — and we later learned many were just wrappers, and the Cyberspace Administration issued specific regulations to check whether you were a wrapper. Without open source, all the intelligence behind those models would be opaque and you’d have even less basis to invest: look at growth? Everyone piles into advertising. Look at who government industrial funds back? A crowd chases that. Nobody would put the money they raised into genuine research.

Raymond: So open source performed a “hundred down to ten” function in private markets. Is it six down to four now?

Wang Tiezhen: I don’t even know how many China should be counted as now — at least ten that are capable, including Meituan and Xiaomi.

Raymond: From a public-market standpoint it’s simple: a sector with twenty or thirty players still capable, each roughly equivalent, and some giving it away free — that’s a bad sector, and the number one can hardly raise prices. If you think there’s a wall between China and America and Americans won’t use Chinese models, then investing in each side’s number one is also fine. But on the American side xAI and Meta are chasing hard, and GPT will catch up soon too. In Q2 this year Anthropic led by a clear gap and could name any price and any valuation; now it seems that’s no longer the case, and whether it will be again, nobody knows. American investors wanting to rank the American players and Chinese investors wanting to rank the Chinese ones are the same problem.

Wang Tiezhen: I think it’s simple. If everyone keeps playing the open-source game, each model is a commodity on a supply chain, and their positions are roughly the same: the model is capable, so someone buys the API and applications are willing to connect. Then the question becomes whether you have enough chips, enough data and enough researchers to train the newest and strongest model. Open source enables you to keep playing this game; it doesn’t mean you play it well. But if lots of people can play, prices won’t be especially high and it becomes a fairly commoditized thing. First place will certainly still be Moutai, still selling dear; but the second-place baijiu will find it very hard to raise prices. That isn’t a bad thing — models becoming a general commodity is good for the country, for society, for research and for education. It’s just that when investing, you’ll find that backing the eighth-place baijiu is a somewhat foolish thing to do.

6. LongCat, Hunyuan 3 and data islands: kill lines everywhere, one large model per capita

Raymond: Speaking of kill lines, let’s talk about Meituan. Meituan has been beaten ragged by the food-delivery wars for two years, wave after wave, and yet found time to build a model: LongCat 2.0, 1.6T, trained entirely on a fifty-thousand-chip cluster of domestic chips. I was fairly shocked seeing it. What was your reaction?

Wang Tiezhen: Two impressions. First, the engineering is well done; they’re exploring the model’s frontier. Previously Ant Group’s Ling tried training models in a mixed-chip data center with both Nvidia and other chips; LongCat trained the model with no Nvidia chips at all, which is a very large breakthrough.

Raymond: Educate me: what’s the difference between domestic chips and Nvidia chips in training?

Wang Tiezhen: Inference is a far simpler problem than training; producing tokens requires relatively little energy and compute. Training is a very long-horizon task that has to run enormous volumes of corpus: how fast the GPUs run, communication between GPUs, the stability of arithmetic precision, how well the software platform is adapted — the challenges are all high, and there’s a great deal of work on both the engineering and model sides. I greatly admire the Meituan people. The second impression is that the model is decently optimized and runs fast.

Raymond: My thought at the time was: China has kill lines everywhere. People once thought DeepSeek was the kill line, and then found lots of people pushing the line forward — non-model companies like Meituan and Xiaomi keep raising it. If someone’s casual side project reaches this level, then to put it another way: Meituan has its own 1.6T model and won’t casually buy your API service unless you’re genuinely much better. Second thing: one day I was in a mall and wanted to ask Dianping a question — is there a coffee shop in this mall doing especially poor business, with free seats, especially quiet? It should search every coffee shop within three hundred meters of the mall and rank them. It turned out Dianping doesn’t have that feature. Why? Because Chinese apps’ data isn’t connected at all: Doubao can’t get that data, and nobody can. So will it end up that every Chinese tech giant entrenches the fragmentation of information and each ships its own model — Xiaohongshu will ship its own model tomorrow — one large model per capita?

Wang Tiezhen: Besides brand effect, attracting technical talent and flexing technical muscle, data islands genuinely are another reason each has to ship a model. Same in America: Cursor won’t sell its own data — programming data touched by hand by users that good is extremely valuable, and why did xAI want to buy it? I think that’s inseparable from data.

Raymond: Let me run the menu: GLM-5.2 is already this year’s Q2 champion, no need to discuss it; MiniMax and Kimi have both been in the news lately, about to ship fairly large models.

Wang Tiezhen: MiniMax’s 2.7T model — if they can train it at all it should be pretty good; being able to launch it means the model is good, since launching a model far behind competitors badly damages your own brand. Separately, the recent release of Hunyuan 3 is a very interesting event.

Raymond: Please. A great many listeners want to hear this part — everyone’s stuck holding Tencent. Tencent’s share price has been very poor over the past six months, partly because of the capital environment and partly because of its AI progress. Everyone’s mentality is: you’re Tencent after all, you can’t fall behind, so should I buy the dip? This certainly isn’t investment advice, but let’s hear Mr. Wang’s view on Hunyuan.

Wang Tiezhen: The previous preview version of Hunyuan 3 came out in February or March and drew little attention, because there genuinely was a performance gap. I tested Hunyuan 3 heavily; it’s very good. What’s remarkable about it? Its model architecture is unchanged from the preview — they just changed the data and the training method, all using the most classical architecture, and yet achieved an enormous performance leap. Right now every major domestic player is frantically working on new model architectures — whether to adopt linear attention, whether to do sparse attention — and Hunyuan 3 shows that people research model architecture mainly to lower costs, not to raise performance; architecture may not help much in raising the capability frontier, or we’re nowhere near that bottleneck. That’s my reasonable inference as an outsider; researchers certainly have firmer evidence from ablation experiments. Conversely, this suggests data’s role may be very large. And Tencent has a lot of data: WorkBuddy is decently built and used by plenty of people inside and outside Tencent, so it can collect a great deal of data; overseas Google has Gmail, and domestically everyone uses WeChat — I have no secrets from Tencent. If the data from WeChat official accounts can be reasonably used by Hunyuan, the power would be very large, and Hunyuan 3.5 and Hunyuan 4 may make even bigger qualitative leaps. One more point: its architecture is smaller than GLM-5.2’s, and it runs on chips you can buy domestically, on a single eight-card machine.

Raymond: LatePost wrote an exclusive in-depth piece on Hunyuan 3 mentioning two things. First, over the past year Tencent gave Hunyuan a lot of organizational room to run — a personnel adjustment many large companies can’t manage. Second, Shunyu Yao trained this model without pursuing clever tricks, without innovating on model architecture, without gaming benchmarks — practicing martial arts without first seeking the Nine Yin Manual, instead holding the horse stance, holding it continuously, for six months. That’s real wisdom. China genuinely isn’t short of data and Tencent least of all; but how to clean the data well, by what standards to define it, how the model gets feedback, needs someone running the operation. Take it one layer further: how many companies in the world can actually “produce data”? Data is a living thing. Buying things on Taobao, scrolling Douyin, chatting on WeChat, editing in Tencent Docs — all of that is producing data. Some data is valuable — a programmer writing a complex task, giving feedback, fixing his own bugs, that’s very expensive data; me chatting with my mother isn’t worth much. Tencent happens to sit at China’s largest data port. Whether and how that data can be used may be the most important topic ahead.

Wang Tiezhen: One addition: it needs to be used lawfully and legitimately. Your chats with your mother contain a lot of privacy. How to use the intelligence in collective wisdom while protecting user privacy as much as possible is a very interesting problem, and I’m curious how Tencent handles it going forward.

7. Coding is a misunderstood word: what it’s really about is automation

Raymond: After Anthropic took off, every model vendor made coding priority one — before that some were chasing multimodality, chasing audio, chasing image and video. Is that shift reasonable?

Wang Tiezhen: Very natural. Previously everyone competed on chat and on role-play, each with its own use case, but you have to look at the share of users willing to pay directly for it, and the hourly wage of the human work it can replace. A Google engineer’s annual salary might be three or four hundred thousand dollars, and thirty or forty percent of his work can be replaced; the social value of that is many times greater than a role-play model. And generating code is the substrate for a lot of things: getting a machine to produce a research report and present it as a beautiful HTML page is a code-writing task. Code-writing also naturally has a data loop: does the code compile, is the result correct, how far from expectation, and the feedback goes back in, making the whole thing faster and faster — a role-play model can’t get data that good. Finally, nobody can define AGI clearly, but it looks as though writing code helps us move a bit closer to AGI, makes models cleverer, and accelerates how fast everyone does everything.

Raymond: I have a thought I want to put to you: it’s July 2026 and I think coding is still non-consensus, and the capital market’s misunderstanding of the word runs deep — the word is wrong; it shouldn’t be called coding, it should be called automation. Investors will say, how many engineers does Google even have, a hundred or two hundred thousand dollars each, and that’s the size of the TAM. But who wants to code? Nobody. Yet we live in a fully digitized world where every interaction with another person is expressed through some digital behavior, and all of those digits can ultimately be implemented in the form of code. An accountant’s job is turning one plus one into two in Excel — code can do that. A designer has to draw a cat and a dog on the front end and put them together — still an HTML page, still implemented as code. So what it’s really about is automation: everything can be automated. You see me “coding” every day; do you think I’m coding? Which is why companies named CodeBuddy and WorkBuddy strike me as problematic — Buddy is fine, but there shouldn’t be a Code in front of it. Sam Altman may have listened to our last episode and renamed Codex to ChatGPT.

Wang Tiezhen: Yes.

Raymond: But today it’s still non-consensus: many friends of mine use Claude Code and their first tab is still Chat — in which case you may as well use Doubao. Chat, cowork and code will ultimately merge into one place, but it shouldn’t be called code — the word doesn’t sound like a tool a normal person would use, and people aren’t programmers. Whereas if you say make a wish, a great many people will do that.

Wang Tiezhen: This is really the natural evolution of a word’s semantics. Many words have completely changed meaning over the past few decades, and maybe code will too.

Raymond: Search for images of “code” today and you get black screens with green text. Maybe when my son grows up, his understanding of code will be saying a few words and making a wish.

Wang Tiezhen: I burn incense at a temple and make a wish too — is that also code? His understanding of the word will be completely different.

Raymond: Aladdin’s lamp, make a wish. I think that’s the core reason every company is pivoting to coding — coding has the biggest data flywheel. Back to Cursor, purely as a thought experiment: suppose OpenAI had the chance, $60 billion in its own stock — should it buy Cursor? If you had one coffee with Sam Altman, would you advise him to buy?

Wang Tiezhen: Sam’s motivation to buy it isn’t as strong as Musk’s. Musk has compute — xAI’s compute was idle at one point and he even sold the idle capacity to others — so what he needs more is a company with data and a product to make things happen with him. And Musk is a person of great imagination.

Raymond: That transaction should have happened in March or April, and I completely failed to understand it at the time. It isn’t the logic of the Devin and Windsurf era — buying a coding agent already happened in 2025. This time the situation was: Anthropic had already built Claude Code, Cursor was under great pressure and had launched Composer itself, and buying it even then was entirely about data; data’s weight in that transaction is far higher. My thought at the time was: if I were rich enough I’d certainly buy as much compute as possible and own as many research teams as possible — if pursuing intelligence itself is the only thing to do, then just pursue intelligence itself. So I didn’t understand it.

All right, many thanks, Mr. Wang. We need to meet once a quarter; AI iterates far too fast — in these three months Chinese models have closed a big stretch of ground, genuinely an enormous leap. And the next three months? Can Anthropic, today’s number one, hand in an even more impressive paper? Uncertain. In Q3 very likely Anthropic and OpenAI will already have listed — what will their financials look like? Unknown. Will their listings perform well? Unknown. Will the whole AI sector be repriced? Unknown. Will more companies join the large-model race, and does China still have many vendors about to ship? Unknown. And where does the kill line finally settle? Unknown. All of these are Q3’s questions.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.