"don't think within the paradigm of the past"

interview｜"chinese entrepreneur" reporter he yifan and zhao dongshan

arts｜"chinese entrepreneur" reporter zhao dongshan

edit|yao yun

image source｜respondent

"using large models to create doctors" is what wang xiaochuan, founder of baichuan intelligence, is currently focusing on.

on august 28, 2024, baichuan intelligent signed a strategic cooperation agreement with beijing children's hospital. the two parties plan to jointly launch five ai medical products, "one large and four small", using ai to empower the expansion and sinking of high-quality pediatric medical resources, as well as a balanced regional layout .

wang xiaochuan interpreted the significance of this cooperation this way: "we believe that there is an opportunity to create an ai pediatrician with the level of an attending physician in a tertiary hospital within 3 years, which is equivalent to creating 1 million attending physicians, enough to cover township-level clinics across the country. ”

in the c-end application market where large model manufacturers are fiercely competing, wang xiaochuan and baichuan intelligence are not seen. even before this, baichuan intelligence also launched "bai xiaoying" to compete with kimi, bean bags and other products.he believes this is dangerous. once major internet companies concentrate their resources and energy on competition, they have little chance of winning. "getting out of the range of big manufacturers is a problem faced by ai startups." wang xiaochuan told "chinese entrepreneur".

after a year and a half of exploration, when the evolution curve of large model technology has turned from steep to flat, ai large model manufacturers have begun to become more pragmatic, exploring the boundaries of large model performance while focusing on the implementation of commercial applications.

but in wang xiaochuan's view, too many people regard large models as efficiency tools, calculators, and computers. the new era has arrived, but people still think about problems within the old paradigm.he believes that the implementation of large-scale models should be to find industries with the highest knowledge density and insufficient supply, and find shortages to increase supply, rather than everyone working together to figure out a new unmet demand.

wang xiaochuan also emphasized,he is not using large models for medical treatment, he is doing large models because of medical treatment. he found that it was too easy for people to look up to technology, but they also despised the achievements of technology. he felt that it was too divisive to have so-called "slaves" serving me.

wang xiaochuan knows that doctors using large-scale ai models are still facing a series of problems: for example, medical treatment is a matter of life and death. is the technology reliable? are national policies and laws approved? how to build a business model? after being asked many times, wang xiaochuan found a very simple answer: "then what do you think about driverless driving? if there is a car accident, driverless driving is also a matter of life and death, whether it can be on the road or not."traffic laws, don’t you have all these worries? so why are we doing autonomous driving? "

two months ago, baichuan intelligence completed series a financing, with a total financing amount of rmb 5 billion. currently, alibaba, tencent,millettechnology giants and many top investment institutions are its shareholders.

as one of the first domestic ai large model companies established in april 2023, what is the current mentality of wang xiaochuan, the helmsman of baichuan intelligence? what thoughts does he have on implementing large model applications? how do you view the competition between startups and major internet companies? with these questions in mind, we also found some answers in our conversation with wang xiaochuan.

the following are the core points:

1. everyone was worried about not being able to buy a card last year, but this year they are worried about having a card and not knowing what to do.

2. i make large models because of medical treatment, not to use large models for medical treatment. in my world view, the biggest development of this generation (technology) is that language has become mathematics, not multi-modality.

3. too many people regard large models as efficiency tools, calculators, and computers. large model companies, investors, and application parties all think this way. the new era has arrived, but we still think about problems within the old paradigm.

4. there are two types of phenomena in application implementation. we call the blind man touching the elephant and the pony crossing the river. those who fail to see the world clearly are like blind people touching an elephant. these people do not know what a model is, whether it is a tool or a partner. the pony crossing the river means that they do not know their own positioning and what role they play in it.

5. it is too easy for people to look up to technology, but they also despise the achievements of technology. they feel that it is too divisive to have so-called "slaves" to serve me.

6. getting out of the range of big manufacturers is a problem faced by these (startup) companies. therefore, large-scale model entrepreneurship must go beyond consensus, otherwise it is a very dangerous thing. if a company is seen very clearly, it is not a startup company.

the following is the transcript of the exclusive interview with wang xiaochuan by "chinese entrepreneur" (with abridgements):

talking about joining the game: i have already thought about making a doctor from a big model

"chinese entrepreneur":what was your feeling and experience like in terms of landing this year?

wang xiaochuan:at the end of april last year, i mentioned "super model + super application". i have already figured out how to understand the relationship between models and applications, as well as the relationship between big manufacturers and startups. along the way, my remarks in the technical field have now begun to become a consensus. for example, when it comes to language being more related to intelligence than multimodality, openai has begun to practice it, weakening the direction of multimodality, and strengthening reinforcement learning.

on the application-driven side, i think medical care is the biggest, or even the only, application in this wave of change. everyone couldn't find it after searching for a long time, but even if i found it, no one would believe it. this is our own internal status and position in the industry.

"chinese entrepreneur":will you focus on medical care right from the start?

wang xiaochuan:yeah, no shouting last year. i do large-scale models because of medical treatment, not to use large-scale models for medical treatment. actually,in my world view, the biggest development of this generation (technology) is that language has become mathematics, not multi-modality.

the transformation of language into mathematics is a major progress in human civilization. the last time physics was turned into mathematics, it was newton who wrote "mathematical principles of natural philosophy", so this became a mathematical problem, which was an amazing progress. this time, the knowledge, thinking, and communication behind language become mathematical problems. after language changed to mathematics, knowledge, thinking, and communication changed to mathematics. this is the biggest change.

so i think the future thing is doing mathematical models of life.creating a doctor is not a vertical language problem, it is solved using agi and the life model. in my mind, this is not a vertical thing. maybe the application field is vertical, but the model technology used leads to the future. this is my bottom-level thinking. it does not mean that everyone talks about medical treatment as vertical application.

of course, medical care is still facing many doubts: is it technically reliable? it is a matter of life and death that is very serious. will the national policies and laws approve it? how to build a business model? a bunch of questions. i later found a very simple way to answer this question: what do you think of autonomous driving? if the above problems are the reasons for not doing it, then why do we do autonomous driving?

"chinese entrepreneur":perhaps everyone naturally thinks that the medical scene solves more complex problems than the driving scene.

wang xiaochuan:(making) a doctor is indeed more difficult than this, but when you talk about life issues and driving law issues, they are both in the same category. medical care is the only industry that requires services and scientific research. this is because the understanding of medicine as a whole is insufficient and must be found in practice. so don’t think about how strong a single doctor is - it’s a good deal. a single doctor will see 30,000 patients in a lifetime. patients, but large models today can put all the data together and even see 10 million patients like a doctor.

talking about the path: he who has doctors can conquer the world

"chinese entrepreneur":what is the combination of large models with doctors in medical scenarios, or what is the empowerment of doctors?

wang xiaochuan:in the short term, it is empowered with the experience of mature doctors, and in the long term it will surpass doctors. today we are still in a position of contempt for technology. many entrepreneurs have explored this field before, which was more model-oriented and connected doctors, but the problem of insufficient supply of doctors has not been solved.

in 2016, i learned a saying, "he who has doctors wins the world."the internet is a matching logic. take one side to leverage the other side, or discover the product thinking model of hidden needs. all day long, you are wondering about what potential unmet needs there are.but medical care is not like this. medical care is supply-driven. your needs are very clear, which is to treat diseases. the supply of doctors is insufficient.

"chinese entrepreneur":is the hospital footing the bill for what you do now?

wang xiaochuan:in the short term, the hospital may pay the bill, but the final payer should be insurance, the government, or individuals. in my logic, we are not helping doctors, but creating doctors. we are not in the hospital, we are going outside the hospital.doctors are friends to us, but they are not our clients. they are our partners.

we have followedbeijing children's hospitalcooperation, president ni xin clearly proposed to create 1 million pediatricians who can "stay at home". we have proposed a big concept to change everyone's medical path: it turns out that the technology is behind the doctors, providing cdss to doctors, and making intelligent image reading systems in hospitals, but we are moving from inside the hospital to outside the hospital, and even from online go down to the line. creating doctors is a must.

"chinese entrepreneur":why pediatrics?

wang xiaochuan:pediatrics has the most urgent need for path change. children get sick every now and then, and worried parents ask for leave to go to the hospital, or even go to the emergency room in the middle of the night. cross-infection may occur, especially during a major influenza outbreak. in fact, 80% of pediatric diseases do not require going to the hospital. this is what the director said. going to the hospital for treatment is very costly to the patient, and the hospital experience is not good, so how can we stop him from coming? similar to the hierarchical medical care advocated by the country, this path can achieve the universalization of medical supply.

"chinese entrepreneur":just now you repeatedly mentioned the word "creating a doctor", which means that it needs to have the judgment ability of a doctor?

wang xiaochuan:communication skills and judgment skills. especially communication skills, it does not mean that you can help me make a judgment when the materials are ready. on the contrary, consultation is its core ability. today, consultation is still a slash-and-burn method, and it still relies on the doctor's mouth. it has not been standardized. if you ask the wrong questions, you will not get good answers. our target is general practitioners in the united states. we should be able to do things they can do. don’t expect me to do things they can’t do.

"chinese entrepreneur":was it a gradual process for you, or did you figure out how to do it from the first day?

wang xiaochuan:i said two things in 2018: first, "when machines master language, strong artificial intelligence will arrive." second, let digital family doctors empower the grassroots. what i'm talking about here is not empowerment in the hospital. later, i also published an article saying that we can see this thing being done in the next 5 to 10 years, that is, by 2028.

talking about thinking: don’t think in the past paradigm

"chinese entrepreneur":we chatted around and found that many general model companies are looking for scene-based experience. everyone will encounter various problems. how about the hospital scene?

wang xiaochuan:too many people in the industry regard large models as efficiency tools. calculators and computers are thinking about this. large model companies, investors, and application parties all think this way. the new era has arrived, but we still think about problems within the old paradigm.

to give a typical example, a participant in a core project of a leading company came to me and asked me, is the large model reliable? i said why are you asking this question? he said he couldn't even do 7-digit multiplication. i then asked, can you? you sitting in front of me as an executive can't do 7-digit multiplication, why should you expect it?

there is something wrong with everyone’s thinking. we are “creating people”, not tools.

in fact, what i mentioned about "creating people" is from the perspective of creating partners. the concept that everyone started to mention is called assistant. assistants have a wrong understanding. we always think that people are superior and inferior. they do miscellaneous tasks for you. i use my brain and you do it. the assistant is a role that helps you do things. someone joked that we wanted machines to sweep the floor and wash dishes, and let people write poems and draw pictures. but today, the machines are writing poems and drawings, while people are still sweeping the floors and washing dishes.

today we think that the more advanced ones are not making assistants, but making consultants, who are more knowledgeable than you. which industry is the most knowledge-intensive and undersupplied? this is the correct path to open.when a shortage is found, supply must be increased, rather than everyone figuring out a new unmet demand.

"chinese entrepreneur":is the idea of ​​using it as a productivity tool fundamentally wrong?

wang xiaochuan:right. everyone always says (what you have to do) is to either kill time or save time. killing time is for entertainment, saving time is for improving efficiency. do teachers, doctors, and lawyers kill time, or save time? no, don't think in the past paradigm.

"chinese entrepreneur":looking beyond baichuan intelligence, what detours have you taken or are taking on the path to commercialization?

wang xiaochuan:one is, i think, "there is no future now". for example, if you want to make quick money immediately, you can write an advertising copy or a ppt and the effect will be very good, no problem. but this is just a small application and will not become the main theme in the process of technology iteration.

another way is to think that the scene will come naturally after training the model. i don’t quite agree with this. we need to think about the relationship between large models and search, and how they relate to reinforcement learning. we need to think about it completely. we don’t just publish a paper on a single point and call it technology.

"chinese entrepreneur":for companies that want to apply models, what misunderstandings do you see about models?

wang xiaochuan:everyone is a little too ambitious. it turns out that many vertical industry internet companies immediately think that i want to train models and turn from the demand side to the supply side. many people come to us and ask if they can help me train a model? i asked him what he needed? he said that i don’t know what i need, anyway, i will train a model for others to use.

there are two types of phenomena in application implementation. we call the blind man feeling the elephant and the pony crossing the river. those who fail to see the world clearly are like blind people touching an elephant. these people do not know what a model is, whether it is a tool or a partner. the pony crossing the river means that they do not know their own positioning and what role they play in it.

talk about ability: have more understanding and belief in medical treatment

"chinese entrepreneur":what is the need to match or switch between intermediate capabilities of a large medical vertical model?

wang xiaochuan:i never thought of it as a vertical model, we called it medical augmentation.a universal model cannot solve medical problems. to me, the core is the need for the team to have more understanding and belief in medical care. faith and understanding require events and processes.some time ago, i happened to tell our technical classmates, what would you think if i asked you to think about medical treatment half a year ago? he said i might quit my job, but what now? it feels so good.

the core of our cooperation with beijing children's hospital is to use their data to create "doctors" and build large medical models so that "doctors" can have direct contact with patients and parents.

"chinese entrepreneur":in this case, the team's ability requirements are very high. they must be both ai experts and understand medical care.

wang xiaochuan:we have a team of doctors, and some even write code. in the past, the biggest characteristic of people in the medical field was that they were "injured" and were not excited about anything. however, when talking to them about the large-scale model of building doctors, a small number of them get excited, and even want to participate and contribute. those are particularly good candidates.

"chinese entrepreneur":in the past, hospital data was in data silos. in this case, is the model company involved in this issue?

wang xiaochuan:i think there is a problem, but i know that healthcare is a place with more data than other industries. this industry already has a certain amount of data accumulation and has done a good job in digitalization. many industries have not yet achieved digitalization.

"chinese entrepreneur":now everyone has a saying that the inference speed of large models will slow down.

wang xiaochuan:yes, one end will go down dozens of times, but the other end will also go up ten times or even a thousand times. because the machine starts to learn to think slowly, it will not give you a flash of inspiration after reasoning, it needs to think slowly. at this time, the requirements for reasoning will become higher. for example, openai will be slower, and it will take at least 10 to 20 seconds to think, so the unit cost goes down, but in order to get a good answer, other costs go up.

talking about anxiety: getting out of the range of major internet companies

"chinese entrepreneur":what is the industry practice like this year? what is the difference between b-side and c-side?

wang xiaochuan:this year everyone is even more anxious. if you want to take the c-end route directly, the risk is quite high.when working on the c side, if there is no depth and we are just doing the model as service thing, according to the consensus, there is no chance for a startup company. you have to have industrial depth. you have to go out of your way to do things that others can't do. you have to think further than everyone else, and not only further in technology, but also in business and industry. if you still think within the capabilities of large manufacturers, there is nothing easy to do about this matter.

i feel that it is too easy for people to look up to technology, but they also despise the achievements of technology. i feel that it is too divisive to have so-called "slaves" to serve me. it is not a mistake to create an assistant, we have also done a lot of things. but i don’t think it represents the cutting edge of big models.

"chinese entrepreneur":after moving from the mobile internet era to the ai ​​era, i found that many of my original entrepreneurial logics were not applicable. what do you think?

wang xiaochuan:there may be several changes: first, the core change is that the logic of the product has changed from pmf to tpf. pmf is the fit of the product and the market, but we are supply-side reform, and the market demand is already there. what needs to be considered now is how can we build it and solve the problem of insufficient supply, so we changed from pmf to tpf; secondly, it changed from thinking about user needs to thinking about whether the technology can realize the various capabilities of the product, and product development changed from making tools to create "people".

"chinese entrepreneur":have your anxieties changed this year and last year?

wang xiaochuan:last year i was worried about not being able to buy a card, but this year i am worried about having a card and not knowing what to do.

"chinese entrepreneur":is not being able to make money one of the causes of anxiety?

wang xiaochuan:we have bypassed the main battlefield where everyone competes, and there is no such anxiety. my anxiety is how everyone can balance their beliefs about medical issues and technical issues: don’t be a techie who doesn’t care about medical care, and don’t be a medical practitioner who doesn’t pursue technology. raising money, setting direction, and finding people are areas where we need to continuously upgrade internally, but i think the money and direction are clear.

"chinese entrepreneur":according to your vision of creating a doctor, what aspects of medical care will you be involved in in the future?

wang xiaochuan:in the short term, we must first let pediatricians and general practitioners go to the grassroots level to change the way patients seek treatment. going further down will allow doctors to have time to do more things, which will also be helpful for scientific research. when he can deal with patients in real time, he can help better collect and organize full-link data, so that he can have enough data. ultimately, your medical model is to know what kind of diagnosis and intervention methods should be given to what kind of people, and to find a personalized paradigm in precision medicine.

"chinese entrepreneur":from last year to now, what is the question you have been asked the most by investors?

wang xiaochuan:how to catch up with gpt-4 in the first place is the most frequently asked question. everyone has released applications this year. what do you think? they don’t ask this question anymore, they think medical treatment is good.

"chinese entrepreneur":between big manufacturers and model startups, will big model startups be under pressure from big manufacturers next?

wang xiaochuan:certainly, the flow of talents is now beginning to flow to large factories. getting out of the range of big manufacturers is a problem faced by these (startup) companies. therefore, large-scale model entrepreneurship must go beyond consensus, otherwise it is a very dangerous thing. if a company is seen very clearly, it is not a startup company. in the early years, bytedance was not optimistic, and many major companies gave up on it after talking to it - non-consensus is the meaning. getting out of the range of big manufacturers was the first reason why we ended up in the first place.

news hotline & submission email: [email protected]