In the AI era every company ends up a generative system company, and the middlemen get killed off
Internet platforms exploit the production side through their power of distribution: the same content gets different reach depending on who posts it, which shows the distribution mechanism has failed. Generative systems internalise distribution, connecting demand and production directly — production and consumption become one.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Data determines the boundary of intelligence; algorithms only determine how much emerges
Wang Guan imagines intelligence as a circle: the boundary is set by data, compute is the speed at which you approach that boundary, and algorithms are the part that breaks through the boundary and produces emergence. He quotes the industry's self-mockery that "as much artificial as intelligent" — when algorithms stop moving the needle, it always ends up being the product manager who goes and gets data, because data represents an understanding of business knowhow. Technology can trade RL for data, but only on the premise that closed domains like code and maths can automatically verify quality; in a great many open scenarios the knowhow can only be turned into data by people.
— Wang GuanThe public-data stage is over; the product startup's opportunity is the third data
Wang Guan splits data into three stages. The public-data stage is a race to reach a fixed boundary first, suited to organisations with high talent density, plenty of compute and fast decisions. The domain-data stage is a race of "I have it and you don't", favouring the big platforms and traditional industries with good IT. The third is product-endogenous data: data that did not exist before and only came into being because you designed the product form. ChatGPT is the sample — the data of conversations solving all kinds of problems never existed in history, and it built its moat by creating new data. An application company has to design, from day one, a new kind of data that can be trained back into the model.
— Wang GuanThe data flywheel is the wrong term; you have to filter for data above the model's level
Wang Guan thinks "data flywheel" is an inaccurate term, because not all data is useful. If you train all user data back without filtering, the model's capability converges on the average intelligence level of all users, and can even get dumber. He notes that the data FSD can use for training today may already be less than one thousandth of what it collects — it certainly wasn't like that at the start; the filtering bar was raised continuously. The core principle is to find, within existing data, the part that adds to wisdom: you have to learn from people smarter than you.
— Wang GuanThe boundary between model companies and application companies is disappearing
Wang Guan thinks there is no longer any point distinguishing model companies from application companies: model companies have all started doing products, and application companies should not be seen as shells either, but as companies that have not yet started building their own models. Cursor has been building its own models all along; only by holding the foundation model in your own hands are costs controllable and profits actually generated. Application companies will run into the capability boundary of the foundation model, because the foundation is general, its parameter count is fixed, and capability is allocated across different tasks — some promote each other, some are mutually exclusive.
— Wang GuanA generative system is a method, not a specific product
Wang Guan likens generative systems to recommender systems: a recommender system is a technical method that, applied to articles, becomes Jinri Toutiao, and applied to video, becomes Douyin. A generative system has three parts: the DSL defines how the problem is expressed precisely; the Context layer lowers the entropy of user intent and the entropy of agent action; the Environment is the interface where humans and agents act together, and needs a reward function to filter valid data. He stresses that the Agent-versus-Workflow debate does not exist internally — both are essentially about producing more valid Context.
— Wang GuanGenerative systems kill the middleman: there is no distribution step
Wang Guan says the essential difference between a generative system and a recommender system is that "there is no middleman taking a cut" — and all internet platforms are middlemen. He offers a thought experiment: suppose two people have identical user profiles and post the same content at the same time and place; the reach will still differ, because their identity weights differ, which shows the distribution mechanism has failed. A generative system internalises distribution: the user's demand for content goes straight to the production side, and the production side generates content straight to the user — from an inventory-allocation logic to a direct-order logic.
— Wang GuanThe north star is degree of intelligence, not user count
Wang Guan says there is no doubt that building a product has commercial value, but above that you have to measure by the intelligence of the whole system. Intelligence has two dimensions: producing higher-scoring content from the same input; and consuming fewer tokens for content of the same quality. He uses a test-taking analogy: one person reads the problem once and gives the answer, another works at it for ages, gets it wrong, starts over, and finally both are right — the former is smarter. So he values a small number of users generating enormous revenue more than a huge number of users generating the same revenue.
— Wang GuanThe future produces IP; the currency shifts from attention to trust
Wang Guan reasons that future producers will no longer produce single pieces of content but continuous content with shared attributes abstracted into a Recipe, which he calls IP. The currency of the economic system changes too: it used to be attention, which is why you get content that leaves you unable to see clearly for two or three seconds, looping to game the recommendation algorithm. But when content supply is excessive and attention is limited, value shifts to trust. He cites "Blizzard出品必属精品" — Blizzard is not what it was today, but in its early years that phrase represented trust.
— Wang GuanIn their own words · checked verbatim
There is no middleman taking a cut. So who is the middleman now? The middleman now — actually all internet platforms are middlemen.
没有中间商赚差价 现在中间商是谁 现在中间商 其实所有的互联网平台都是中间商
Wang Guan0:00
The capability of your model will converge on the average intelligence level of all the users it is used by.
你的这个模型的能力 就会趋同于所有用户使用 所代表的那个就是就是那个用户的平均 平均智能水平
Wang Guan49:43
Today's application companies — we shouldn't just see them as so-called shells, or software companies, app companies, but as companies that haven't yet started building their own models.
今天的应用公司 我们也不应该只把它们看成所谓的壳 或者是软件的公司 app的公司 而是还没有开始去做自己模型的公司
Wang Guan1:42:28
The same goods, priced differently — so why can a generative system do better? Because that step is gone. There is no distribution step.
有同样的商品 它的价格不一样 那生成系统为什么可以做得更好呢 因为没有这个环节了 没有分配这个环节了
Wang Guan2:40:09
With a small number of users — I use three users to generate 1 million in revenue — versus using 100,000 users to generate 1 million in revenue, which do you think, for your company, right, we would value the former more.
用少量的用户产生 我用三个用户产生 100万的这个营收 和用10万的用户 去产生100万的营收 你说对于你们公司来说对吧 对我们会更看重前者
Wang Guan2:55:13
Figures
| 12X team size | 30 people | 3:05:21 |
| Views on the WeChat Channels video Wang Guan made with Mideal | over two million | 3:17:24 |
| Revenue share from the WeChat Channels video Wang Guan made with Mideal | a few hundred yuan | 3:18:24 |
| Share of the 12X team who have been a founder or co-founder | about half | 3:19:25 |
| Share of FSD data usable for training today | less than one thousandth | 50:46 |
| Number of companies in the US market doing video-processing SaaS products | 20 to 30 | 1:07:01 |
| ARR of US video-processing SaaS products | tens of millions of dollars | 1:07:01 |
| Share of the global population that can read and write | 70% | 2:46:13 |
Glossary
- DSL / domain-specific language
- A structured language designed for a specific problem, sitting between natural language and code.
- System 1 / System 2
- Wang Guan borrows from cognitive science: System 1 is the model's instinctive reaction, System 2 is the construction of Context outside the model.
- Context Engineering
- Designing effective Context outside the foundation model so the model outputs better results.
- prosumer
- Producing while consuming; the act of production itself has value for the producer.
- reward function
- The mechanism that evaluates and filters valid data, determining what the system reinforces.
How to listen
AI application-layer founders, product managers, investors watching the shift in the platform economy — especially anyone thinking about the boundary between application companies and model companies.
2:00-28:00 personal history and early product attempts, skippable.