Model companies will eventually disappear; product companies are just getting started
Qi Fanchao (齐帆超) argues that models themselves cannot directly create value, and that model companies will not exist in the future; he raised nearly 400 million yet voluntarily gave up training larger models, turning instead to solving the uncertainty of large-model output — the biggest blocker to productization.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The value of training bigger models shrank in H2 2023
In H1 2023 Qi Fanchao was still talking about technical differentiation, such as adding instruction-following ability early in the pre-train stage and handling long structured content. But by August-September of H2 2023, he found that all players were flooding into the same track: the differentiation he had thought of was either seen by others too, or others were already doing it, and coming up with new differentiation was getting harder and harder. More critically, he judged that directly training large models was contributing less and less to achieving his product goals, so he made a strategic pivot: toward problems that cannot be solved well by off-the-shelf large models alone, even with continuous scaling.
— Qi FanchaoModel companies will not exist in the future
Qi Fanchao does not accept "the model is the product," nor does he believe AGI will arrive within three to five years. His judgment: models themselves cannot directly create value, so model companies must find a specific scenario or product direction to convert into value, and will ultimately become product companies. He uses Li Auto as an analogy — pure-electric technology will keep improving, but Li Auto chose extended-range hybrid, adding a fuel tank and generator on top of the battery as backup power, partially solving range anxiety, so it could productize faster and ultimately win user recognition.
— Qi FanchaoCompress the large model's error probability into a limited range
Qi Fanchao believes the biggest challenge of large models is that they are probabilistic models, and combined with the need for diversity, two sides of the same coin are hallucination and unexpected errors. His solution is to keep the large model within a relatively limited space: do forward operations on the input, using more deterministic methods to narrow the room for play; do backward operations on the output, using knowledge bases and other means to verify results. He mentions that Lu Qi (陆奇) also believes insufficient model robustness and stability is one of the most important reasons blocking large-model productization. He does not accept that RAG alone can solve hallucination — the core of AI search is RAG, but an answer can still retrieve A and answer B.
— Qi FanchaoContent is only the carrier of information
Qi Fanchao defines information as "reducing your uncertainty." The same thing can be obtained through different content carriers — an article, audio, video, a WeChat Moments post — and what you get after consuming it is the same. From this he judges: what was called information distribution in the past was actually content distribution, with the smallest granularity being an article or a video. But for users, consuming a complete piece of content is not necessarily the most efficient way to obtain information, because it inevitably contains low-value time. His product idea is to deconstruct content into the smallest information units, then reassemble them into new content according to user interest.
— Qi FanchaoThe center of gravity in the content chain keeps moving toward consumers
Qi Fanchao traces the content industry: before the internet, the production side had strong say and creators had high social status; after the internet and mobile internet, self-media emerged, the barrier to creation dropped, and distribution platforms gained stronger say; he judges that in the future content consumers will have even stronger say, because AI makes content supply increasingly excessive and production costs keep falling. He even believes that even if he did not do it, this change would definitely happen — it is a fundamental law of social development. For creators, he believes upstream source information and content remain very valuable, but what is delivered will change, and what may be provided is corpus for AI processing.
— Qi FanchaoStartups are like sailing, with no nautical chart
Qi Fanchao says that when you have no nautical chart, the difficulty and uncertainty may be several orders of magnitude lower than when you have one. He uses Columbus as an analogy: Columbus believed that by continuing west across the Atlantic he would find India, but no one had verified it — it was only a belief in his heart, and there were enormous challenges along the way; in the end he did not reach India but North America. Qi Fanchao says that in startups, especially when you want to do something differentiated that no one has done before, you cannot follow a map; you can only believe in your inner conviction and solve whatever problem arises.
— Qi FanchaoI did a very stupid thing
Qi Fanchao recounts: he is very J, willing to make everything organized, and one day he changed everyone's email names in the company to a uniform format with numbers appended. In hindsight the benefit did not exceed the cost — not the technical cost, but the high switching cost for everyone, since previously registered accounts would all be affected. From this he extrapolates: entropy increase in the world is an inevitable trend, and wanting to reduce entropy requires paying a cost, so you have to consider whether the ROI is positive, and whether the cost paid to reduce entropy in certain links is worth it.
— Qi FanchaoModel uncertainty cascades through the organization layer by layer
Qi Fanchao believes the biggest difference between the organizational form of AI companies and internet companies comes from underlying uncertainty: internet technology is deterministic, goals and rhythms are clearer, and division of labor and collaboration is less difficult; but large-model output uncertainty is extremely high, and even algorithm engineers cannot say clearly whether it can be done or when it will be done. This cascades layer by layer into how the organization sets goals, how different teams coordinate, and whether information can be aligned faster. His conclusion is that AI companies must have higher information transmission efficiency and consistency, and individuals must cover longer chains, because knowing only your own link makes it hard to understand how upstream and downstream affect you.
— Qi FanchaoIn their own words · checked verbatim
I think there may be no such thing as a model-based product in the future, and maybe no such thing as a so-called model company either, because the model itself cannot directly create value.
我认为可能将来是不存在模型系产品 可能也不存在所谓的模型公司的 因为模型本身它是没法直接产生价值的
Qi Fanchao0:00
I think this statement is both right and wrong. I can say the model is always getting better — but on what timescale? A few years, three or four years, or ten or eight years, or decades?
我觉得这句话既对也不对 我可以说模型永远都在变好 比如说它是以多长时间的尺度 几年三四年 还是十年八年 还是几十年的尺度
Qi Fanchao30:24
Content is only the carrier of information — content is the carrier of information. Right. How do you define information? Information is actually reducing your uncertainty.
内容只是信息的载体 内容是信息的载体 对 你怎么定义信息 信息其实就是把你的 不确定性降低
Qi Fanchao44:52
When you don't have a nautical chart, your difficulty and uncertainty, compared with having a nautical chart — especially when someone tells you how you can always reach the other shore — the drop in uncertainty may be beyond your imagination, a drop of several orders of magnitude.
就是当你没有一个航海图的时候 你的难度和不确定性 比你有了一个航海图 尤其是有人告诉你 怎么总能到达一个彼岸的时候 这个不确定性下降的 可能是你难以想象的 是有好几个数量级的下降的
Qi Fanchao1:13:02
I fundamentally believe that as long as this product truly brings irreplaceable value to users, it will definitely be able to make money through various means.
我是底层相信说 只要这个产品真的给用户带来了不可替代的价值的话 它是一定能够去通过各种方式赚到钱的
Qi Fanchao1:39:19
Figures
| Company valuation | 1.2 billion RMB | 1:00 |
| Company founding date | March 17, 2022 | 1:00 |
| Funding round | Series A+ | 1:00 |
| Founder's age | Born in 1994, just turned 30 last month | 1:00 |
| Number of investors met from H2 2021 to mid-2022 | No fewer than 100 investment institutions | 17:19 |
| Investors talked to after the MiraclePlus roadshow | Sixty to seventy over time | 17:19 |
| Model scale he wanted to train at the time | 10B to 20B | 18:19 |
| Estimated investment to train the model | Tens of millions | 18:19 |
| Company headcount | 60-plus | 1:22:07 |
Glossary
- conditioned generation
- Given existing content or information, generate content that conforms to a specified input.
- RAG
- Retrieve relevant documents first, then have the model answer based on those documents; can alleviate the knowledge-update problem.
- PMF
- The state in which a product is genuinely accepted by the market and users are willing to keep using it.
- MVP
- The first version of a product that validates the product direction with the fewest features.
- Foundation Model
- A concept proposed by Stanford, referring to a model that after large-scale pretraining can be transferred to many tasks.
How to listen
Founders and investors watching AI application-layer startups, content distribution products, and endgame judgments about large-model companies.
1:23:18 to 1:27:11, the Q&A on post-90s CEO personality, drawing big pictures, and information-asymmetry management — low information density.