Wang Xiaochuan: Slow by one step in ideals, fast by three in execution
Zhu Xiaohu looks at reality, Yang Zhilin looks at technology, and Wang Xiaochuan says both are only touching one leg of the elephant. What he wants to do is build a person, and healthcare is the towering, grounded scenario.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Slow by one step in ideals, fast by three in execution
Wang Xiaochuan positions himself as the third voice in China's AGI. Zhu Xiaohu talks applications and PMF, Yang Zhilin talks technology and unified modeling, and he thinks each sees one leg of the elephant. His strategy: the zero-to-one problem has already been solved in the US, and catching up is much less work than starting from zero—you can approach them with a tenth or even a hundredth of the consumption; but on applications you must have your own thinking. He repeatedly stresses 'slow by one step in ideals, fast by three in execution'—this is his positioning for doing large models in China.
— Wang XiaochuanThe application of AGI is building a person, not a tool
Wang Xiaochuan believes GPT has done two things that never happened before: first, it can use language—Marx said the difference between humans and animals is that humans use language; second, it can make and use tools. So he thinks the application of AGI is building a human-like new species, not a tool. He opposes talking only about productivity, and thinks the second thing is that it is like a person—an assistant, advisor, professional, like a doctor, lawyer, teacher. Any knowledge-intensive industry that needs supply is the greatest prospect brought by GPT.
— Wang XiaochuanThree model buildings: entertainment, healthcare, creation
Wang Xiaochuan divides the models to be built into three worlds. Entertainment, using the term from Dream of the Red Chamber, is the 'Illusory Realm'—the core is not Sora or Midjourney, but the story engine behind them; NPCs are just the connection between the story engine and the outside world. Healthcare is the life model, corresponding to the real world. The third is creation, helping people improve productivity. He stresses these are not vertical models, but part of the AGI model, called 'scenario models.' The three worlds may eventually merge, but he must know what capabilities the application scenarios ultimately need.
— Wang XiaochuanDoctors are a towering, grounded scenario, so dare to bet on GPT-4
Responding to Zhu Xiaohu's challenge 'Do Chinese companies dare to spend money on GPT-4 research?', Wang Xiaochuan says he is not at all intimidated. The reason is that the scenario is clear: in building a person, doctors are a towering, grounded scenario—grounded because it benefits service, towering because it demands a large model. He gives a judgment: today, when doing large models, small models cannot do well; the larger the model, the better, with no ceiling. So betting on large models has at least one pillar, ensuring that the model's technical capability will become the moat for the scenario.
— Wang XiaochuanResponding to Zhu Xiaohu point by point: before PMF, there is TPF
Wang Xiaochuan focuses on responding to Zhu Xiaohu's challenges. On 'no scenario, no data,' he says those companies probably do not include Baichuan. On the risk that if others open-source, betting on GPT-4 will lose everything, he says open source will not produce the best models, and healthcare scenarios cannot be done well with open source. On 'ten people cannot find PMF, a hundred people cannot either,' he thinks this is a deep prejudice, because the companies Zhu Xiaohu invested in found M with small compute, but healthcare's M cannot be achieved with small models. He proposes the core is TPF—what kind of product the technology fits, rather than talking about PMF first.
— Wang XiaochuanA technical colleague wanted to do Sora, and I shut it down
Wang Xiaochuan says he rarely gets angry at the company; the last time was because a technical colleague wanted to do Sora. His reason: Sora is not consistent with GPT; the AGI ideal must be built with language as the axis, language constructs the conceptual space, and it cannot be replaced by a model like Sora; video needs to incorporate language to become an engine toward AGI. From the AGI perspective, Sora is only a阶段性 product. He thinks the public is easily impressed by visual effects, but those in the tech circle know that the breakthrough in language is much greater than in vision.
— Wang XiaochuanAfter the New Year, let go of the inertia of following GPT
Wang Xiaochuan says last year was stressful overall, still in the stage of following 3.5 and following 4. After the New Year, his vision of the future and how to do things became much less conflicted, and he is more willing to communicate fully with everyone. He reflects on two inertias from last year: first, making following 3.5 to 4 the main task; second, outside the model, hoping to help the model through application thinking and external capabilities, rather than integrating model and application. After the New Year, the thinking of integrating model and application is stronger, putting differentiation into the model.
— Wang XiaochuanIn their own words · checked verbatim
We are actually building a person, building a person, right, we are not building a tool anymore.
我们其实市场 我们 在造人 造人 对 我们不是在造一个工具了
Wang Xiaochuan10:01
We dare to do entertainment, not because doing Sora means entertainment, or doing a Journey means entertainment, but because behind it, you can drive them with such a story engine—that is the core.
我们敢于做娱乐 并不是说 做了Sora就有娱乐了 做了一个名字Journey就有娱乐了 而是背后 你能够推动他们的 这样一个故事引擎 这是核心
Wang Xiaochuan27:56
Doctors are a very clear, towering and grounded scenario in the middle. So grounded means it benefits service, towering means it has demand for a large model.
医生是中间的一个很明确的一个顶天立地的场景 所以立地是它对服务有益 顶天是它对一个大模型有需求的
Wang Xiaochuan44:05
As for his words, most are right. As for us, what I just called 'slow by one step in ideals, fast by three in execution' actually echoes his topic.
我对他话来讲的话呢 大部分对 说对我们言的话呢 我刚才叫做理想上慢一步 落下快三步 其实就是暗和了他这样一个题目
Wang Xiaochuan1:06:15
Figures
| Baichuan Intelligence team size | 200+ people | 51:22 |
| Baichuan Intelligence technical algorithm proportion | 70% | 51:22 |
| Baichuan Intelligence sales headcount | 40+ people | 51:22 |
| Baichuan Intelligence former Sogou team proportion | 20%-30% | 49:21 |
| Baichuan Intelligence team size at founding | about 50 people | 11:03 |
| Number of healthcare companies Wang Xiaochuan personally invested in | more than 10 | 55:22 |
| Time for Wang Xiaochuan to return to Sogou CEO position | 18 months | 1:00:23 |
Glossary
- TPF / Technology-Problem Fit
- Technology first matches existing industry problems, then product-market fit is discussed.
- PMF / Product-Market Fit
- Product matches market demand, an indicator Zhu Xiaohu repeatedly emphasizes.
- RWS / Real World Study
- Collecting medical data for research in real-world scenarios outside hospitals.
- Illusory Realm / Virtual World Model
- Wang Xiaochuan borrows the term from Dream of the Red Chamber to refer to the virtual world model to be built for entertainment.
How to listen
Founders, investors, and engineers following China's large-model startup paths, especially those wanting to understand why Baichuan simultaneously does foundation models and healthcare and entertainment scenarios.
The opening pleasantries about Zhu Xiaohu's report and the 'blind men touching an elephant' metaphor can be fast-forwarded; the direct confrontation starts at 1:06:15.