Large models are manufacturing, not magic: no lead lasts more than a month
The author of Death Knell for AI argues that large model makers have no protectable technical differences—data and compute are production inputs, just like in solar, phones, and EVs. No lead lasts more than a month.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI skills follow a seesaw: better coding, weaker prose
Xian Yu stopped using AI for writing in mid-2025—he finds it more satisfying to write himself, and the model simply can't translate original thought into personal voice. Language models can replace programmers but can't capture how a person thinks. The reason is a tradeoff in training: vendors pour resources into agentic programming datasets at the expense of writing datasets—‘it's like a whack-a-mole game’. By September 2026, coding ability kept improving while creative writing and the ability to express human thought fell, and noticeably so.
— Xian YuCollege students paid $10 to find the real ranking
In the past six months, ByteDance's Seedance series dominated video models globally. Xian Yu and his partner wanted a more interesting way to evaluate models than tech benchmarks, so they built wild prompts from memes popular on Bilibili and TikTok, had different models generate videos, then hired college students (about 300 of them) to rate the results for 10 yuan each. When H3, which MiniMax had just open-sourced, outranked Seedance 2.0 and 2.5 in their meme test—contradicting the narrative that ‘Seedance is comprehensively dominant’—Xian Yu seriously considered using technical methods to suppress H3's ranking but ultimately didn't. He thinks H3 won because it was optimized for tasks ordinary users care about (image-to-video, effects), while Seedance targets professional film studios.
— Xian YuWhen competitors complain, read it as insecurity
In July, Xian Yu published an article saying MiniMax's base model M3 had fallen behind competitors like Qwen and GLM because it hadn't been updated in four or five months. MiniMax, sensitive to ‘lagging behind’ getting out (it's public, stock price matters), filed a complaint. Xian Yu rewrote the piece as satire: ‘MiniMax is advancing so fast, we apologize to MiniMax’. Coincidentally, the day the apology went live, MiniMax's recruiter added his partner on WeChat asking about joining. This told Xian Yu the complaint was personal, not the company's official stance. Later MiniMax invited his partner to discuss technology—the relationship improved, and Xian Yu wore a MiniMax shirt to record this episode.
— Xian YuChina's AI narrative cycles every three months, copying Silicon Valley
Chinese AI startups rotate through themes on a three-month cycle, entirely following Silicon Valley's rhythm: whatever the Valley embraces, someone in China builds a clone project to raise money. After Manus popularized the autonomous agent narrative in early 2025, waves of people entered AI applications, but most lacked real user insight—pure self-serving entrepreneurship. The ‘users build agents’narrative had a moment too, but those products mostly died because product building is far harder than TikTok clips; ordinary users can't participate. By early this year, the torch passed to ‘world models’—all because Li Feifei and others are working on them, so dollar-based VCs want domestic copy-cats.
— Xian YuMost ‘world models’projects are repackaged video models
Xian Yu thinks the world model wave is largely storytelling: you know if an AI app works in one or two months, and if it doesn't retain users, it dies. But world models take three years to validate, and demos don't really matter—people are buying the narrative and upside. Most domestic ‘world models’projects take Alibaba's open-source video model One from a year ago, fine-tune it into digital humans or real-time generation, and wrap it in grander language. Hype has already cooled in the past few months, but after MiniMax open-sourced H3, these storytelling projects came back alive—H3 digital humans look much better than One, so they can show new demos and keep pitching.
— Xian YuChina's AI industry doesn't invent terms because it lacks confidence
There's an obvious pattern: Chinese AI barely invents terminology—the trendiest terms like RSI (recursive self-improvement) are all American. Company names are mostly English too (DeepSeek, MiniMax). It's not a technology problem, Xian Yu says; it's a confidence problem. The logic is ‘whatever Silicon Valley does, we build a replacement’, and funding sources matter: dollar-based LPs prefer the story of ‘cloning hot Valley trends, but cheaper and more efficient’. He points out that when investors see the same benchmark, they nearly always believe it if you tell them OpenAI did it over a Chinese company doing it. But he admits inventing terms isn't always good—the domestic model leap forward over the past few months might exist precisely because people were heads-down shipping, not debating grand concepts.
— Xian YuLarge models follow manufacturing logic: no technical secrets survive
Xian Yu sees the large model business as fundamentally manufacturing, with no truly proprietary technical secrets. How to build datasets and scale is a small circle—model makers hire from each other constantly, so there are no secrets. Compute isn't a secret either; anyone can stack cards. Give a model more data and compute (production inputs) and you get a stronger model. This chain needs no breakthrough innovation, just engineering to push scaling laws. It's the same logic as solar panels, phones, and electric cars—you can't name fundamental technical differences between EV makers, yet market share still shifts. Competition is brutal: any model's lead rarely lasts two months, usually just one. China's GLM 5.3, Qwen 3.8 Max, and K3 all ship within a month, landing in the same ability band.
— Xian YuMore money flowing in, crazier the people get
Xian Yu wraps up by noting that most people he knows are normal—just a bit unusual personality-wise. But AI is flooded with ‘so many liars, narcissists (NPD), and people doing who-knows-what’, and these people actually raise money and become industry stars. Half-joking, he suggests someone should study mental-health metrics in AI: rates of NPD or ADHD probably skew absurdly high. By contrast, ordinary AI app founders raise money at valuations under a few tens of millions of dollars; for model makers, Kimi's latest round values it at $5 billion. His conclusion: the industry has taken in so much hot money, yet the people are so odd—hot money and oddness are probably positively correlated.
— Xian YuIn their own words · checked verbatim
Dear AI hype dogs, stop making one-prompt-generates-article, one-prompt-generates-video, one-prompt-generates-podcast products. They're worthless, pure noise for the internet.
亲爱的AI炒作狗们, 别做一句话生成文章, 一句话生成视频,一句话生成播客的AI产品了, 没有价值,纯粹给互联网制造噪音,
Xian Yu3:06
It's like a whack-a-mole game: if you feed the model more data about programming, about agentic systems, then you can't feed it enough writing data. Its writing ability gets consumed by its coding ability.
它就像一个打地鼠样的事情 这是模型你给它为了更多的关于编程的关于Agentic的数据集 那么你没有该为更多的写作数据集 那它写作能力就会被编程能力给吞噬掉
Xian Yu7:29
When the ranking came out, I seriously thought about whether to use various technical means to suppress H3's ranking and put Seedance ahead. But of course I didn't in the end, and kept things fair.
当时排名出来之后我一度认真的思考过 要不要用各种各样的技术手段 包括条件 把H3的排名给降低一下 让CLS排名在前面 但当然最后没有这样做 维护了这个问题的公平性
Xian Yu19:49
This competition will be extremely brutal. Extremely brutal competition means any model—your lead is hard to maintain for two months. You can only hold on for one month.
这个竞争会非常残酷 非常残酷的竞争就是任何一个模型 你的领先很难持续两个月 你只能持续一个月
Xian Yu52:33
I asked him how he fixed the water leak. He said he asked Doubao (豆包) for help, and Doubao ended up finding him the phone number of a renovation contractor.
我问他怎么解决了漏水问题 他告诉我说 他问豆包应该如何解决 豆包最后给他搜出来的 这个装修公司老板的电话
Xian Yu55:37
But the AI industry is stuffed with so many liars, NPDs, and people doing who-knows-what, yet some people manage to raise money, some even become industry stars.
但AI行业里面充斥着如此之多的骗子, NPD, 和你不知道他在干嘛的人, 但这些人居然, 有的人居然还能拿到钱, 有的人居然还能成为行业明星
Xian Yu1:01:43
Figures
| H3 vs. Seedance 2.0 (meme benchmark) | approximately 80% of Seedance 2.0 | 14:38 |
| Meme benchmark pay | 10 yuan per person | 18:48 |
| Meme benchmark participants | approximately 300 | 19:49 |
| Instant funding | $2 billion | 29:58 |
Glossary
- benchmark
- A standardized test set used to compare the capabilities of different AI models side by side.
- RSI (recursive self-improvement)
- A proposed technical path where a model iteratively improves itself to achieve capability jumps.
- scaling law
- The principle that trading more data, compute, and parameters yields corresponding improvements in model capability.
- product-market fit
- The state where a product is genuinely needed by its target users and they keep choosing to use it.
How to listen
Founders, investors, and operators tracking China's AI model competition; anyone trying to see past corporate PR to what's really happening in the market.
9:32–11:00, a riff comparing American late-night comedy to philosophy; tangential and can be skipped.