loading
loading
Xianyu (咸鱼), the writer behind 葬AI — “Funeral AI” — produces some of my favorite satire about China’s AI industry. ChinaTalk readers may remember his essay “Open Models Are the United Front of AI”. In this conversation, we talk about: - Why AI writing isn’t funny; - How Xianyu made a website for ordinary people to rank video model by meme-ability; - Why Five Year Plans are a kind of benchmark; - And why there’s so much Chinglish in AI. We also disagree about where the hype is worse: China or America. The conversation is in Mandarin. Listen to the original audio here, or read a lightly abridged English translation of the transcript below. On Writing Humor Irene: Xianyu, welcome to ChinaTalk. I’ve been a devoted fan since your essay on Juche AI last September, so it’s a real honor to have you here. Could you introduce yourself? When did you start writing, and how did you get into independent publishing? Xianyu: Before we begin, I just need to get it on record: Juche is a glorious ideology. I started writing Funeral AI last July or August, approaching AI through subculture. A lot of people in the AI industry maybe didn’t quite understand that. If you have a purely technical background, you may prefer very serious, professional articles. But from the beginning, I thought that even the most advanced technology — the models and APIs that frontier labs spend tens of billions of dollars developing — is available at a relatively affordable consumer price, perhaps $20 or $100. So any ordinary person has the right to describe how it feels to use AI. That was my starting point. And it has proved true: even with the most advanced AI, including GPT-6, which came out a few days ago, or the hottest venture-backed projects, an ordinary person without a technical background can use them and describe the experience. You can offer original observations that are different from the slop and garbage already on the internet. Irene: You mentioned that you previously worked in Chinese tech media, specifically at 36Kr and Huxiu. Did those experiences make you consider this more subcultural approach? Xianyu: No. Institutional media are very buttoned-up. They try to explain what the industry’s most important people and institutions are doing in a relatively professional voice. You have to network with all sorts of people, reach the big names, and persuade them to give interviews. I really wasn’t a fan of that song and dance. I never lasted more than six months in any of my previous jobs. Those experiences convinced me that I’m not suited to a conventional job. I’m better off at home doing whatever I feel like. … Irene: How do you think about the issue of novelty in AI creativity? Xianyu: I started out writing with AI. The article you just read was written with AI: it did at least 80% of the main structure, and I polished it afterward. But sometime in the second half of 2025, I stopped using it entirely. Everything I write now is written by me. On one hand, I’ve become more interested in this kind of subcultural writing. Writing it by myself feels better than having AI do it. But it’s also because models simply don’t do writing well. They can program; they’ve completely replaced human programmers. Yet they can’t express your original ideas well in your own language. This is a complicated problem. For example, some friends and I are making a writing benchmark. How do you evaluate whether an AI-written article is good? You have to extract indicators: how strong the AI flavor is, how original it is, etc. One obvious example is connective language. AI loves “not X, but Y,” and em dashes. Those patterns make the writing feel very AI-generated. A comprehensive evaluation might need hundreds of such indicators. It becomes a complicated engineering task. Meanwhile, as models get better at programming, their writing is getting worse. It’s like whack-a-mole: you feed the model more coding and agentic data without adding more writing data, and the coding capability swallows up writing. So from 2025 to now, the mainstream models have improved at programming while their writing has deteriorated. Here we are in September 2026: models are more advanced, but their originality and ability to express human thought through creative writing, rather than an engineering task, have declined substantially. That’s why I don’t use AI to write anymore. Irene: Although I’ve lived abroad for a long time, I read Chinese and regularly follow WeChat blogs. Over the past two years, I feel like their writing has become harder and harder to read. There used to be accounts and publications that I appreciated for their beautiful prose or using phrases I wanted to remember. Now it feels like I’m reading Chinese entirely for work and information gathering, rather than for the pleasure of the language itself. Your humor genuinely moved me at that language level. It’s so fun, and it’s about a subject I’m interested in. I saw your essay about going back from AI writing to writing by hand, where you talked about entertainment as something AI can’t do. How were you thinking about that? Xianyu: Humor requires conviction. I sometimes watch American political stand-up. There’s a famous comedian — I’ve forgotten his name, a bald older white guy. He may have died; I can’t remember. People like that are funny because they really believe something. A right-wing comedian might genuinely not believe in progressivism or the narratives about diversity over the past few years, so he makes those jokes without holding back. It works the same way in reverse for someone on the left. Humor requires some original thinking and a real belief in something outside the consensus. For me, that non-consensus belief is that AI isn’t that important, and you don’t have to believe in it. You may see this more clearly in America, where perhaps most ordinary people oppose AI. In China, most ordinary people are technooptimists — or perhaps they aren’t even that, but just simply have an instinctive openness to a new technology. Most people I meet don’t think of AI as something bad, something that harms their lives, or something that can be irrelevant to them. They instinctively see it as the most important thing, something that will multiply the world’s productivity. There’s nothing wrong with that view; it’s just a different view. I believe ordinary people’s lives matter more, at least compared with the model companies’ grand narratives and the trillion-dollar IPO feast they’re heading toward. Your own experience of these AI products still matters enormously. And if ordinary people’s experiences can be expressed well, in an engineering form, they’re something model companies particularly need. One key is benchmarks. I mentioned them earlier: benchmarks actually do something very similar to journalism. Take the Wall Street Journal‘s parent company—I’ve forgotten its name [editor’s note: Dow Jones]. What it originally did was establish standards. Financial institutions adopted those standards and the information platform behind them, and that’s how it developed into what it is today. Our Fifteenth Five-Year Plan is a benchmark too: the Chinese government sets a development goal, then breaks it into quantitative indicators. If you meet the indicators, you can say you’ve achieved the plan. AGI, by contrast, is a very bad term, because it has no quantitative indicators. It isn’t broken down like the Five-Year Plan: when model capabilities or adoption reach a particular level, we will have achieved AGI. There’s no definition like that. The same applies to any specific aspect of AI. Writing ability is poor, but if you break it down into indicators, quantify them, and run different models through repeated tests, you can obtain both trajectory data and an objective conclusion. You can use this engineering approach to examine any aspect of it. In fact, that’s what the model companies have done over the past year in programming and agentic capabilities. Most people I meet lack a concrete path for pushing back against AI, or recovering confidence in their own lives. They ask: I’m not a leading researcher at one of these labs, or one of the hundred people at the heart of AI, so what’s the point of my own life? Do the things I spend my days thinking about matter? I think they do. First, you have to believe that your own life and your real thoughts matter. Second, there’s a path forward. You don’t necessarily have to devote all that effort yourself to constructing a benchmark and quantitatively defining what you’re good at. But that path exists. In their own field of expertise, ordinary people can be more convincing than the hottest AI researcher. A Benchmark for Memes Irene: You’ve built a few benchmarks yourselves. Could you talk about them? I saw that there was a general benchmark and a meme benchmark. Xianyu: Let me start with MemeBench. Over the past six months or so, Chinese companies have been absolutely dominating video generation. ByteDance’s Seedance, from 2.0 to 2.5, has essentially taken over the international market. The biggest video-agent platform in America is Higgsfield, and in China it’s LibTV. On both, much of what users generate runs on Seedance. MiniMax then open-sourced H3. Judged across a comprehensive set of metrics, it might perform at about 80 percent of Seedance 2.0’s level, but it’s open source, so anyone can deploy it locally. Video models have been sweeping everything aside over the past six months. When we were building this benchmark, we knew H3 was coming, but it hadn’t been released yet. ByteDance’s Seedance dominated everything. It was a bleak situation. Imagine a world without OpenAI, where only Anthropic exists. That’s terrible and frightening — my Anthropic account has already been banned. ByteDance hasn’t gone that far, of course. But Seedance is still an extremely expensive API. If you’re seriously building an AI application with Seedance 2.0, the cost is unacceptable, probably even if you’re doing it in America. We wanted a way to compare these models. Apart from Seedance 2.0, hardly anyone had discussed the other models over the previous six months. Kling still hadn’t released its new model, though it was supposed to be coming soon. We thought: beyond technical metrics, could we find more entertaining ways to assess capabilities? We collected meme templates from Bilibili and Douyin comments — things people doing serious AI work wouldn’t normally think about, but might encounter while scrolling in their spare time. We thought the model companies probably hadn’t trained on this material: Nailoong, Zhang Xuefeng — we didn’t actually include Zhang Xuefeng, to avoid potential copyright issues — and these ridiculous memes. … We put these images in and gave the AI wild prompts, such as Nailoong shooting lasers and blasting apart all the buildings in front of him. That’s a rather tame example off the top of my head; the actual prompts were probably funnier. We turned it into an arena: a website. What matters most for models is data. We initially needed people to label the outputs and evaluate them, so we’d have an initial ranking before opening it to the public and increasing participation. My partner Muqiu found university student groups on QQ and posted a part-time job: complete this set of questions and we’ll pay you ten yuan [around US$1.5]. Lots of students took part. I went to university from 2018 to 2022. The little jobs I used to see in student groups paid three, four, or five yuan. They weren’t complicated tasks, so ten yuan was a very generous rate. Then something hilarious happened: people kept asking Muqiu whether it was a scam, because ten yuan seemed too generous. My first reaction was: wow, this really is China’s demographic dividend. Where else in the world can you find an educated young person, who can use a basic AI product, willing to spend twenty minutes completing a test for a few yuan? Probably nowhere. We initially recruited about three hundred people to establish a basic ranking. H3 had come out by then, so we added it. At first we were in despair: after the students rated these AI slop meme videos, H3 ranked above Seedance 2.0 and Seedance 2.5. It was the exact opposite of the dominant industry narrative: surely 2.5 was the most advanced, then 2.0, then H3. I seriously considered using technical adjustments, including changing the weights, to push H3 down and put Seedance first. Of course, we didn’t do that. We preserved the benchmark’s fairness. What it shows is that ordinary people voting in an entertainment context prefer MiniMax H3. From a comprehensive technical perspective, across a broader range of indicators, Seedance is definitely the better model overall. But that doesn’t matter. You can drive a Tesla; I ride my electric scooter everywhere and get more fun out of it. After publishing the article, we talked with people at MiniMax and ByteDance. ByteDance questioned the result: surely Seedance was better at instruction-following? But this is a question of perspective rather than a purely technical question. As I explained in the article, comparing Seedance 2.5 and H3 reveals whether you’ve made the model itself into a product. H3 has had more product work done on it: its data was specifically optimized for image-to-video generation and video editing, such as adding effects the way you would in Premiere. That’s why it does better in those situations. Seedance 2.5 is aimed more at professional studios. ByteDance probably worked with directors to build substantial professional datasets. Google’s video models may have fallen behind partly because they haven’t invested that much effort. Seedance 2.5 is better suited to professionals, but an ordinary person using it directly may get worse results than with H3. Even the most advanced models, built at a cost of billions of dollars, can still be judged by ordinary people — and should be. Otherwise, the companies will keep pursuing coding, increasingly specialized capabilities, and whatever makes more money, until ordinary people discover they’ve created a monster they can’t even talk to. Anthropic already has a little of that quality. Having a conversation with Fable 5 is difficult because it produces all sorts of garbage language. … Irene: While preparing for this interview, I spent half an hour on MemeBench and discovered I’m a total H3-head. Every video I liked was from MiniMax. When the big companies contact you, is it because they care about their MemeBench ranking? Xianyu: Of course they care about evaluations. Hardly anyone had publicly said anything bad about Seedance. We approached it from a specific, narrow angle. It’s similar to your work in journalism: if you’re writing a long piece about US interest-rate hikes or cuts, would you write a feature, or an inverted-pyramid report? You’d find an angle for the story. We did the same thing, using memes as a very particular way to assess Seedance. Suddenly we were offering data to explain why H3 was better from that perspective, so naturally Seedance’s team cared. But it was a healthy interaction. They didn’t say we’d done a terrible job; they came over to ask why the result had turned out that way. It was quite funny. Irene: But MiniMax did file a complaint about you in July, didn’t it? Could you tell that story? Xianyu: It’s a microcosm of how fierce the competition is. MiniMax’s latest base model is M3. I forget exactly when it came out — early this year, anyway. At launch it wasn’t exactly a laggard; it was probably second tier, behind Zhipu’s GLM or Alibaba’s Qwen. But Chinese models have made an enormous leap over the past six months. At the start of the year, their coding models could mostly handle relatively simple work. Professional programmers would probably pay for a Claude account. That’s still true today, but less so. Now, I think most people’s work, including that of programmers and other professionals, can be handled by Chinese models such as Kimi K3, Qwen 3.8 Max, or GLM 5.3. … When we wrote about MiniMax two months ago, we were in the middle of that leap in Chinese model capabilities, and MiniMax had clearly fallen behind. M3 was four or five months old and hadn’t been updated, though a new version was coming soon. I don’t remember all the details of the article, but the basic point was that MiniMax had fallen behind and we wanted to discuss it. MiniMax was very sensitive about that. It’s a Hong Kong-listed company that needed to support its share price, so it complained about the article. Our response was: fine, we’ll apologize. We rewrote everything we’d said. Where we’d previously said MiniMax was behind, we now said it was incredibly advanced. We immediately apologized: “Sorry, Fat Cat. Sorry, MiniMax. You’re China’s most advanced model, the model most likely to achieve AGI.” We later had some conversations with them. … The person who complained wasn’t someone with real decision-making authority at MiniMax; they had acted individually. Another interesting point is that the ability to explain a model clearly—what works and what doesn’t—is extremely rare. Most people can only recite technical metrics: the model ran SWE-bench and the other standard benchmarks, and here’s its ranking. Repeating that just adds noise to the internet. Being able to explain actual problems in terms a technical person or researcher can understand is valuable. Model companies around the world are doing this. GPT-6, for example, strongly emphasizes 3D and computer use. Clearly, someone with product-definition skills explained those needs to the researchers, and they built the relevant data together. Improving a model isn’t solely researcher-driven; product and technical people define the requirements together. But I’m on very good terms with MiniMax now. In fact, I’m wearing a MiniMax shirt right now. Irene: I’ve only just noticed! Congratulations on repairing the relationship, and may Fat Cat keep succeeding. Since I don’t live in China, a lot of my information is secondhand. I learned about two products through your writing: LibTV and Plaud. Today I actually saw a Plaud advertisement when I opened the New York Times app. We’ve been discussing how a product expresses a need. What kinds of demand do you think these newer Chinese AI startups are anticipating? What does the general public actually want from AI? Xianyu: That’s a very good question. Chinese AI startups go through a cycle roughly every three months, rotating from one theme to the next and following Silicon Valley’s rhythm exactly. Right now, a personal-agent project called Instinct is very hot in Silicon Valley. It recently raised money at a valuation of about $2 billion. China has several personal-agent companies too, and they consequently get more attention in the venture market and more articles written about them. The themes rotate every three months, following Silicon Valley’s narratives. Around this time last year, AI applications were the hot thing. Manus launched in early 2025 and prompted a discussion about general-purpose agents. Manus was a China original, which was quite interesting. It initially borrowed from Devin, an American product, but it was more consumer-oriented. From the first half of 2025 through roughly October, investment in AI applications was booming. Suppose you were a young ByteDance employee, very good at bullshitting, with a decent cofounder and team, claiming to build some kind of application: you could probably raise money. It was just a question of whether your valuation was $10 million, $20 million, or $30 million. But that created a problem: AI applications require an understanding of users. You have to identify a real need and build for it. That isn’t something you can simply scale up. When everyone piles in, including boring people who’ve never done anything interesting in their lives, most of the resulting applications have no user insight. They exist to satisfy the need to raise money. That described most AI applications last year, and still does today. But investment in applications is now in a winter, so the people still building them probably have at least some real enthusiasm; otherwise they wouldn’t keep going. Compared with a year ago, we haven’t lost much, and the surviving applications may actually be more interesting. One thing particularly stuck with me: when general-purpose agents became popular, founders looking for funding immediately wanted to build OpenManus — to remake Manus in every possible way. That’s a viable category. Genspark is the obvious example and is doing well; I think it’s valued at over $3 billion now. Whenever a product has genuine product-market fit, copying it quickly enough can get you somewhere. … After AI applications cooled off, the hottest category in China in the first half of this year was world models. That follows Silicon Valley’s narrative: Fei-Fei Li and Yann LeCun are working on world models, so dollar-denominated funds want Chinese counterparts to invest in. My particularly uncharitable explanation is that you can find out whether an AI application works within a month or two. If users can’t figure it out, the usage numbers are bad, and paid advertising doesn’t work, your application is probably finished. A world model, by contrast, can be difficult to validate even over three years. Fei-Fei Li released a demo recently: does anyone care whether it was good? Apparently not. Once you’ve announced that you’re building a model based on a new paradigm, what you’ve actually built doesn’t matter. People are buying a story and an expectation. Over the past six months, Chinese VCs, especially dollar funds, have backed all sorts of world-model projects. The terminology gets wilder and wilder, as if they’re combining words at random: first good ol’ world models, then “causal” world models, then “social” world models. But most of these world models are actually video models. What they’re really doing is post-training Wan, Alibaba’s open-source video model released a year ago. Before H3, Wan was the strongest open-source video model, but it’s now outdated. They turn it into a digital avatar or a real-time generator, call a terrible AI video a world model, and tell an enormous story around it. It’s an even worse story than the metaverse. Yet investors and founders have spent the past few months frantically telling these stories to raise money. Perhaps this is where their vision and understanding are ahead of mine: venture capital rotates between themes. Video models were the theme for the past few months; I don’t know what’s next. But as long as themes rotate, attention and hot money can come back to yours. World models have cooled off over the past month or two, but now H3 is out, and people post-training it can tell another round of world-model stories. H3 is more capable than Wan, so their avatars and real-time videos will look better than the demos they showed a few months ago. H3 has given these world-model projects another lease on life, which is very funny. H3 itself is a serious, valuable contribution: it’s the only usable open-source video model that can compete with Seedance. It occupies a unique niche, yet it has rescued a bunch of scammers who do nothing but tell stories. Irene: I went through your older posts and found a stand-up routine you did last year. You said the biggest, most certain trend in AI was its “Web3-ification”: almost nobody makes money except people selling courses and influencers. That seems somewhat distinctive to China. Could you say more? Xianyu: Isn’t America the same? Cluely’s success suggests that American AI is even more Web3. Irene: I’ve been thinking about this over the past few days because of what’s happening with Anthropic. America’s weaker technological optimism may partly reflect how earnest Americans are. They don’t approach these arguments with that basic expectation of hype. I think Silicon Valley’s earlier intellectual movements, such as effective altruism, have influenced this. People now discuss technology very earnestly, whether they’re alarmist doomers or accelerationists. In China, people seem more willing to accept the inherently promotional aspect, or at least recognize it. Xianyu: I think it’s the opposite. What we’ve been discussing takes place in a tiny circle, perhaps fewer than 10,000 people in China. Following AI startup products is a very niche interest in China, and probably in America too. Most people haven’t heard of these products. Take Fable 5: I don’t know, have a million people used it? It’s niche and very expensive. The central point is that China’s stories about AI products and startups come from America. Dollar-denominated VCs in particular invest earlier, and their first instinct is to find a Chinese counterpart. If models are hot in America, find Chinese counterparts. If personal agents are hot, find Chinese counterparts. The same goes for AI applications. It’s not absolute; China has some original startup narratives. But broadly, the stories come from America, so American storytelling must be even crazier. The hype here is actually less extreme. I think America’s core competitive strength is its ability to tell stories and coin terms. Take benchmarks: they’re essentially an engineering problem. Define users’ needs and the relevant metrics, and building one isn’t particularly complicated. It’s certainly easier than training a model. Yet the mainstream benchmarks that everyone runs when releasing a model, whether SWE or Terminal-Bench, are mostly American. Americans instinctively think that when a technology is young, they shouldn’t just work on the technology or its biggest component: they should also build industry organizations and define standards. Watching Anthropic and OpenAI argue about AI safety over the past few days, my first thought was: here comes another standards body or industry gatekeeper. We’ll inspect you, and if you fail, you don’t get a license to enter. That’s a classic American approach: establish an association or standards organization, define the rules, and exclude anyone who doesn’t meet them. Coining terms works the same way. RSI, recursive self-improvement, is one of the hottest terms in large models right now: the idea of using it to achieve a leap in model performance. That’s another American coinage. My immediate reaction is that Americans are far better at coining terms. It may be one of America’s most important capabilities. Irene: Why doesn’t the Chinese industry coin its own terms? Xianyu: Coining terms is part of the superstructure. You need real confidence to do it. This very small circle of Chinese AI startups largely operates on the logic of building a cheaper alternative. I very much want genuine innovation, and there is plenty of genuine innovation. But broadly, a project becomes popular in Silicon Valley, and a few returnees or Tsinghua students build something similar to ride that theme. That probably has something to do with where the money comes from. Dollar funds have to explain their investments to their limited partners. To older businesspeople and investors from the Middle East or Europe, you can tell two stories: this is an original idea from young Chinese founders, or this is an established, hot Silicon Valley category that a Chinese team can execute more cheaply, better, and faster. Those old guys, whether in the Middle East, Europe, or America, are more likely to buy the second story. The source of capital matters, and that brings us back to who gets to define the conversation. Chinese renminbi funds can think more independently and invest according to China’s circumstances, rather than following Silicon Valley. Of course, that’s an idealized picture. It’s all still early, and we can’t make such a strict comparison. Broadly, I think the Chinese AI industry lacks confidence. You can find many reasons for that, but confidence is the essential thing. You need it before you can propose benchmarks and establish standards. That takes us to a much bigger question. For example, one Chinese AI application tells investors it has a million American users. Another has a million Indian users but only 10,000 American users. Which one gets funded? There’s no question: investors everywhere would choose the one with a million American users. That’s an expression of America’s overall national power. Similarly, if one benchmark comes from a Chinese company and another from OpenAI, most companies will choose OpenAI’s. Perhaps OpenAI’s models really are more advanced and its benchmark more representative. But broadly, people instinctively place greater trust in American standards, metrics, and definitions. That’s a fact. Irene: I understand the first part. Purchasing power is lower in developing countries, so investors put more faith in business models that work with American consumers. But on a technical level, my observation is that Western researchers understand and respect Chinese labs’ capabilities, especially since DeepSeek. When Kimi K3 came out, for instance, many people read its paper carefully. If Chinese companies wanted to influence technical terminology, I think there would absolutely be an audience. Xianyu: It’s a question of mindset and confidence. It simply doesn’t occur to Chinese people to do this. Irene: It’s fascinating. I joked with friends last year that AI had revived Chinglish. Chinese media writing about AI technology now mixes in lots of English words. We finally have a Chinese translation for “token,” ciyuan 词元, but there’s still “agent,” “benchmark,” and even the companies themselves have English names. I’ve always found that hard to understand. Doesn’t calling yourself DeepSeek or MiniMax affect your marketing in China? Xianyu: Kimi’s company has a Chinese name, Yue Zhi An Mian 月之暗面, “the dark side of the moon,” which sounds quite nice. MiniMax’s Chinese name is Xiyu Technology 稀宇科技, which is indeed rather obscure.
Xianyu (咸鱼), the writer behind 葬AI — “Funeral AI” — produces some of my favorite satire about China’s AI industry. ChinaTalk readers may remember his essay “Open Models Are the United Front of AI”. In this conversation, we talk about: - Why AI writing isn’t funny; - How Xianyu made a website for ordinary people to rank video model by meme-ability; - Why Five Year Plans are a kind of benchmark; - And why there’s so much Chinglish in AI. We also disagree about where the hype is worse: China or America. The conversation is in Mandarin. Listen to the original audio here, or read a lightly abridged English translation of the transcript below. On Writing Humor Irene: Xianyu, welcome to ChinaTalk. I’ve been a devoted fan since your essay on Juche AI last September, so it’s a real honor to have you here. Could you introduce yourself? When did you start writing, and how did you get into independent publishing? Xianyu: Before we begin, I just need to get it on record: Juche is a glorious ideology. I started writing Funeral AI last July or August, approaching AI through subculture. A lot of people in the AI industry maybe didn’t quite understand that. If you have a purely technical background, you may prefer very serious, professional articles. But from the beginning, I thought that even the most advanced technology — the models and APIs that frontier labs spend tens of billions of dollars developing — is available at a relatively affordable consumer price, perhaps $20 or $100. So any ordinary person has the right to describe how it feels to use AI. That was my starting point. And it has proved true: even with the most advanced AI, including GPT-6, which came out a few days ago, or the hottest venture-backed projects, an ordinary person without a technical background can use them and describe the experience. You can offer original observations that are different from the slop and garbage already on the internet. Irene: You mentioned that you previously worked in Chinese tech media, specifically at 36Kr and Huxiu. Did those experiences make you consider this more subcultural approach? Xianyu: No. Institutional media are very buttoned-up. They try to explain what the industry’s most important people and institutions are doing in a relatively professional voice. You have to network with all sorts of people, reach the big names, and persuade them to give interviews. I really wasn’t a fan of that song and dance. I never lasted more than six months in any of my previous jobs. Those experiences convinced me that I’m not suited to a conventional job. I’m better off at home doing whatever I feel like. … Irene: How do you think about the issue of novelty in AI creativity? Xianyu: I started out writing with AI. The article you just read was written with AI: it did at least 80% of the main structure, and I polished it afterward. But sometime in the second half of 2025, I stopped using it entirely. Everything I write now is written by me. On one hand, I’ve become more interested in this kind of subcultural writing. Writing it by myself feels better than having AI do it. But it’s also because models simply don’t do writing well. They can program; they’ve completely replaced human programmers. Yet they can’t express your original ideas well in your own language. This is a complicated problem. For example, some friends and I are making a writing benchmark. How do you evaluate whether an AI-written article is good? You have to extract indicators: how strong the AI flavor is, how original it is, etc. One obvious example is connective language. AI loves “not X, but Y,” and em dashes. Those patterns make the writing feel very AI-generated. A comprehensive evaluation might need hundreds of such indicators. It becomes a complicated engineering task. Meanwhile, as models get better at programming, their writing is getting worse. It’s like whack-a-mole: you feed the model more coding and agentic data without adding more writing data, and the coding capability swallows up writing. So from 2025 to now, the mainstream models have improved at programming while their writing has deteriorated. Here we are in September 2026: models are more advanced, but their originality and ability to express human thought through creative writing, rather than an engineering task, have declined substantially. That’s why I don’t use AI to write anymore. Irene: Although I’ve lived abroad for a long time, I read Chinese and regularly follow WeChat blogs. Over the past two years, I feel like their writing has become harder and harder to read. There used to be accounts and publications that I appreciated for their beautiful prose or using phrases I wanted to remember. Now it feels like I’m reading Chinese entirely for work and information gathering, rather than for the pleasure of the language itself. Your humor genuinely moved me at that language level. It’s so fun, and it’s about a subject I’m interested in. I saw your essay about going back from AI writing to writing by hand, where you talked about entertainment as something AI can’t do. How were you thinking about that? Xianyu: Humor requires conviction. I sometimes watch American political stand-up. There’s a famous comedian — I’ve forgotten his name, a bald older white guy. He may have died; I can’t remember. People like that are funny because they really believe something. A right-wing comedian might genuinely not believe in progressivism or the narratives about diversity over the past few years, so he makes those jokes without holding back. It works the same way in reverse for someone on the left. Humor requires some original thinking and a real belief in something outside the consensus. For me, that non-consensus belief is that AI isn’t that important, and you don’t have to believe in it. You may see this more clearly in America, where perhaps most ordinary people oppose AI. In China, most ordinary people are technooptimists — or perhaps they aren’t even that, but just simply have an instinctive openness to a new technology. Most people I meet don’t think of AI as something bad, something that harms their lives, or something that can be irrelevant to them. They instinctively see it as the most important thing, something that will multiply the world’s productivity. There’s nothing wrong with that view; it’s just a different view. I believe ordinary people’s lives matter more, at least compared with the model companies’ grand narratives and the trillion-dollar IPO feast they’re heading toward. Your own experience of these AI products still matters enormously. And if ordinary people’s experiences can be expressed well, in an engineering form, they’re something model companies particularly need. One key is benchmarks. I mentioned them earlier: benchmarks actually do something very similar to journalism. Take the Wall Street Journal‘s parent company—I’ve forgotten its name [editor’s note: Dow Jones]. What it originally did was establish standards. Financial institutions adopted those standards and the information platform behind them, and that’s how it developed into what it is today. Our Fifteenth Five-Year Plan is a benchmark too: the Chinese government sets a development goal, then breaks it into quantitative indicators. If you meet the indicators, you can say you’ve achieved the plan. AGI, by contrast, is a very bad term, because it has no quantitative indicators. It isn’t broken down like the Five-Year Plan: when model capabilities or adoption reach a particular level, we will have achieved AGI. There’s no definition like that. The same applies to any specific aspect of AI. Writing ability is poor, but if you break it down into indicators, quantify them, and run different models through repeated tests, you can obtain both trajectory data and an objective conclusion. You can use this engineering approach to examine any aspect of it. In fact, that’s what the model companies have done over the past year in programming and agentic capabilities. Most people I meet lack a concrete path for pushing back against AI, or recovering confidence in their own lives. They ask: I’m not a leading researcher at one of these labs, or one of the hundred people at the heart of AI, so what’s the point of my own life? Do the things I spend my days thinking about matter? I think they do. First, you have to believe that your own life and your real thoughts matter. Second, there’s a path forward. You don’t necessarily have to devote all that effort yourself to constructing a benchmark and quantitatively defining what you’re good at. But that path exists. In their own field of expertise, ordinary people can be more convincing than the hottest AI researcher. A Benchmark for Memes Irene: You’ve built a few benchmarks yourselves. Could you talk about them? I saw that there was a general benchmark and a meme benchmark. Xianyu: Let me start with MemeBench. Over the past six months or so, Chinese companies have been absolutely dominating video generation. ByteDance’s Seedance, from 2.0 to 2.5, has essentially taken over the international market. The biggest video-agent platform in America is Higgsfield, and in China it’s LibTV. On both, much of what users generate runs on Seedance. MiniMax then open-sourced H3. Judged across a comprehensive set of metrics, it might perform at about 80 percent of Seedance 2.0’s level, but it’s open source, so anyone can deploy it locally. Video models have been sweeping everything aside over the past six months. When we were building this benchmark, we knew H3 was coming, but it hadn’t been released yet. ByteDance’s Seedance dominated everything. It was a bleak situation. Imagine a world without OpenAI, where only Anthropic exists. That’s terrible and frightening — my Anthropic account has already been banned. ByteDance hasn’t gone that far, of course. But Seedance is still an extremely expensive API. If you’re seriously building an AI application with Seedance 2.0, the cost is unacceptable, probably even if you’re doing it in America. We wanted a way to compare these models. Apart from Seedance 2.0, hardly anyone had discussed the other models over the previous six months. Kling still hadn’t released its new model, though it was supposed to be coming soon. We thought: beyond technical metrics, could we find more entertaining ways to assess capabilities? We collected meme templates from Bilibili and Douyin comments — things people doing serious AI work wouldn’t normally think about, but might encounter while scrolling in their spare time. We thought the model companies probably hadn’t trained on this material: Nailoong, Zhang Xuefeng — we didn’t actually include Zhang Xuefeng, to avoid potential copyright issues — and these ridiculous memes. … We put these images in and gave the AI wild prompts, such as Nailoong shooting lasers and blasting apart all the buildings in front of him. That’s a rather tame example off the top of my head; the actual prompts were probably funnier. We turned it into an arena: a website. What matters most for models is data. We initially needed people to label the outputs and evaluate them, so we’d have an initial ranking before opening it to the public and increasing participation. My partner Muqiu found university student groups on QQ and posted a part-time job: complete this set of questions and we’ll pay you ten yuan [around US$1.5]. Lots of students took part. I went to university from 2018 to 2022. The little jobs I used to see in student groups paid three, four, or five yuan. They weren’t complicated tasks, so ten yuan was a very generous rate. Then something hilarious happened: people kept asking Muqiu whether it was a scam, because ten yuan seemed too generous. My first reaction was: wow, this really is China’s demographic dividend. Where else in the world can you find an educated young person, who can use a basic AI product, willing to spend twenty minutes completing a test for a few yuan? Probably nowhere. We initially recruited about t
document fetched 2026-10-30 02:55 · canonical https://chinatalk.media/p/how-to-make-fun-of-chinese-ai