The first anniversary of ChatGPT ｜ How does AI application make money？ The big model waits for the ＂Watt moment＂
[Editor’s note] November 30, 2022 may become a day to change human history-OpenAI, an American artificial intelligence development agency, launched ChatGPT. It not only gave birth to another round of high-light period in the field of artificial intelligence, but also is not commonly known as "steam engine moment", "iPhone moment" or even "drilling for fire moment".
In the past year, the revolutionary technology called "generative artificial intelligence" has inspired the global scientific and technological community to "redo all software and hardware", which has caused the value of AI infrastructure providers with first-Mover advantage to skyrocket, making the scientific exploration from medical care to aerospace doubly empowered. The arrival of the legendary "singularity" has never been so possible.
Just like any technological change in history, ChatGPT has also brought us deep anxiety. There are both sci-fi fears that AI threatens human survival, and realistic concerns about smashing our jobs, defrauding our money, and manipulating our hearts. Even OpenAI itself has just experienced a crisis and almost escaped the fate of collapse overnight.
This year has given us more questions: what is the next evolution direction of the big language model? When will the shortage of AI chips be solved? Is the training data running out? How will the Battle of Hundreds of Modes in China evolve? Should AI technology development accelerate or decelerate? Will AGI (general artificial intelligence) exist in other forms? To this end, we invited people in the industry who were running on the AI track in 2023 to answer these questions and ask their own questions. If you have your own answers or questions, please let us know (www.thepaper.cn).
After a year of "noisy" generative artificial intelligence, how many companies have made money because of the application of this technology?
"Domestic applications are trying to generate value, but they rarely make profits. It is really profitable to be a’ small workshop’ for digital people." Chen Ran, founder and CEO of Beijing OpenCSG, told 澎湃 Technology. Chen Lei, vice president of Xinyi Technology and head of big data and AI, believes that among foreign companies, Microsoft, which has deep cooperation with OpenAI, "should have a relatively big improvement in revenue". Liang Jiaen, chairman and CTO of Yunzhisheng Intelligent Technology Co., Ltd., said that "the AI image generation company Midjourney is said to have achieved a revenue of 100 million US dollars", but "there is no real AGI-Native(AGI application".
"At present, it is still in the initial stage. If the cost input and income are calculated, the profitable industries are still rare." Xiao Yanghua, director of the Shanghai Key Laboratory of Data Science and a professor at Fudan University, said that the large model has achieved remarkable results in many industries, but the effect does not mean that it can be profitable. This involves many factors, especially the cost factor. The refining of large models requires huge costs.
If you put it another way: what are the most successful industries where the big model landed first or landed? Wang Xiaohang, vice president of Ant Group and head of financial model, said: "There are two types, one is creative industries, such as design, entertainment and games; The other category is highly knowledge-intensive industries, such as finance, medical care and law. " However, strong regulatory industries such as medical care and finance require extremely high reliability, which is not only an illusion, but also financial compliance and the value proposition of each industry. If these three problems are not solved, the potential of large models will not be brought into play when they land in rigorous industries.
Specific to the medical field, Liu Qingfeng, chairman of Iflytek, introduced to 澎湃 Technology that "the medical model can complete 90% of irrational drug use reminders", and Ren Feng, co-CEO and chief scientific officer of Yingshi Intelligent, bluntly said: "It seems that overnight, things that could not be done before can suddenly be done." For example, you can get relevant results by directly talking to the big model and asking "Help me find the target for treating a certain cancer".
In the financial field, Wang Xiaohang said that in many scenarios, the fields of intention understanding, qualitative analysis and quantitative analysis have surpassed manual work. "The future service will be completed through multimodal interaction and language interaction. How to integrate with the graphical interface of the mobile Internet to form a new interface and experience may happen in the next two years."
Respondents have different views on whether there will be phenomenal applications in the coming year. Wang Fengyang, vice president of Baidu Group and head of mobile eco-business system, believes that phenomenal generative AI applications will definitely appear and will exceed everyone’s expectations. Wang Jiaping, a chair scientist at the AI Security Inclusive System Research Center of IDEA (Guangdong-Hong Kong-Macao Greater Bay Area Institute of Digital Economy), said that "it will come faster than the phenomenal application of the original Internet". An Xiaopeng, director of Alibaba Cloud Science and Technology and Research Center, is cautiously optimistic, and points out that the premise of China’s argument that there are rich application scenarios is that there are enough data accumulated in specific industries. In addition, the underlying computing power, open source models and various tools are needed to build the whole ecology and reduce the technical threshold, application threshold and cost.
"The turning point of the industrial revolution brought about by any major technological change must come from the turning point of cost." An Xiaopeng said that the steam engine was first invented to pump the water from the mine. Its cost is high and it often breaks down. "What did Watt do? He greatly improved the stability, safety and reliability of the steam engine and reduced the cost, so the steam engine changed from a special machine that can only pump water to a general machine that can pull the train forward. "
The following is an interview record, which has been deleted due to space reasons:
澎湃 Technology: What are the enterprises or industries that have achieved profitability due to the application of generative AI?
Chen Lei (Vice President of Xinyi Technology, Head of Big Data and AI):We have seen many enterprises whose valuation, user volume or revenue have increased a lot. Because of the blessing of generative AI on the Office system, Microsoft’s revenue should have a relatively big improvement, and its market value has returned to the trillion scale. In the future, there should be companies that are profitable because of generative AI.
Zhang Peng: It’s in every scene. Customer service can be partly replaced by AI, or the quality can be improved by AI. In the scientific research scene, AI helps to read literature, find information and analyze data, improve the efficiency of scientific research, and even provide some key research ideas.
Chen Ran [Founder and CEO of Beijing OpenCSG]:Domestic applications are trying to generate value, but rarely make profits. Being a "small workshop" for digital people really makes money.Foreign applications have been very well adapted, such as Copilot’s auxiliary development code, and Tom Cat’s companion is also quite successful. In the future, applications whose data are ready, can be used immediately, are used more, and bring a lot of value to everyone will immediately form an industrial chain.
Liang Jiaen (Chairman and CTO of Yunzhisheng Intelligent Technology Co., Ltd.):Midjourney, an AI image generation company, is said to have achieved $100 million in revenue. In the past, UGC (note: user-generated content) was mainly text. With this multimodal tool, UGC will be more colorful. Therefore, the application of this piece may be faster, because it is equivalent to generating many candidates for us, and then people will work together to deal with them. If you directly let the machine generate a very good result, all AIGC (note: artificial intelligence generates content) can’t do it at present.
Xiao Yanghua (Director of Shanghai Key Laboratory of Data Science and Professor of Fudan University):Whether it is profitable or not depends on costs and benefits. At present, the cost problem of large-scale model application is very prominent, and the refining of large-scale model requires huge costs, and the application of large-scale model also has costs. After considering the cost factor, whether the income is significant enough and whether it is worthwhile relative to the cost is also the key issue in empowering thousands of industries in the big model industry.
At present, it is still in the initial stage. If the cost input and income are calculated, the profitable industries are still rare. If you ask another question, what are the industries and enterprises that have achieved the application effect of the big model, then I think the answer is very clear. The big model has achieved remarkable results in many industries. For example, a large number of virtual anchors are based on the big model, and most of the work of ordinary translators can be replaced by the big model to some extent. Still, it should be emphasized that the fact that a large model can achieve results does not mean that it can be profitable, which involves many factors, especially the cost factor. Although the big model can reach some abilities and levels of ordinary people, the cost may not be lower, so this is another problem completely.
澎湃 Technology: What are the most successful industries in which large models landed first or landed? What are the difficulties?
Wang Xiaohang (Vice President of Ant Group and Head of Financial Big Model):There are two types,One is creative industries, such as design, entertainment and games; The other is highly knowledge-intensive industries, such as finance, medical care and law.Large language model is a powerful container of world knowledge map, which can rival, empower and even replace some knowledge workers to some extent.
The big model is trained based on the public data available in the whole network, and it has a strong ability to understand and generate, but there are many challenges when it comes to professional and rigorous industries. The general big model is short in professional general knowledge, such as financial professionalism. The complexity and rigor of various financial decisions are very high, and the medical industry is more complex and rigorous. For example, making a family asset allocation protection plan is not a simple recommendation, but a calculation of deviation, risk concentration, risk level and affordability. Decisions like this are not good at by big models, and it cannot be learned. Even with enough data, its rigorous combing and calculation are far from meeting today’s industry requirements.
Strong regulatory industries such as medical care and finance have extremely high requirements for reliability. The reliability I am talking about is not just an illusion, but also financial compliance and the value proposition of each industry. If these three problems are not solved, the potential of large models will not be brought into play when they land in rigorous industries. How to solve these three problems? First, large models should be combined with small models in professional fields, such as asset allocation. After understanding the needs of users, asset allocation tools should be mobilized without recalculating how assets should be allocated. Second, a structured knowledge map is very important. For example, in order to do a good job in medical claims, we should build a very complete medical and insurance knowledge map, inject it into the training process of large models, reduce hallucinations, improve professionalism, and give priority to how to combine knowledge map with large models in the application process. The retrieval enhancement technology is also very effective. Answering after content positioning in the professional field is just like finding the answer from a professional book. These technologies can greatly improve the professionalism and factuality of the large model. This is a systematic project.
澎湃 Technology: What kind of big model application can really generate value?
Zhang Peng (CEO of Beijing Zhipu Huazhang Technology Co., Ltd.):A technology will produce several values in application, such as cost reduction, efficiency improvement, quality improvement and innovation. More specifically, for example, we can do things that people can’t do, help people do simple and repetitive work, free people from this kind of work, and improve the efficiency and effect of data flow, exchange and processing in the digital age.
Chen Ran:Applications should ultimately solve pain points and serve human beings, and human beings pay for the process of generating value. Many current applications will not immediately form the next generation of applications by leaps and bounds. The intermediate stage is to make the current applications AI and eventually become native applications, that is to say, they don’t even need to click on the webpage. These applications involve shopping, tourism, catering, clothing, etc.
Liang Jiaen:At present, there is no real AGI-Native(AGI application.Just like when the mobile Internet first came out, everyone’s idea was to make the PC application smaller and put it into the mobile phone. At present, the big model application is still at this stage, and the real AI-Native(AI application still needs iteration.
But no matter what the final form is, we still have to answer a question: what problem did we help users solve? For example, in a hospital, what users want most is to have a super doctor, all problems can be solved, and even the hospital is not needed. But in this serious scene, it is not realistic to do it in one step. We can now provide doctors with some tools to help them improve their efficiency and quality, let the basic work be done by machines, and better release more high-quality medical resources. In the process of cooperation between man and machine, man is adapting to the machine, and the machine is constantly improving in the process of learning from man. In the end, it may reach the upper-middle level of experts. But there may be some difficult problems left in the future, which need people to work together to solve.
Medical treatment is one of our important directions, which is a very knowledge-intensive industry. The accuracy and quality of medical-related data can’t reach the level of good medical experts just by grabbing them on the Internet. Therefore, we will also supplement relevant data to optimize related applications. In the end, it is still problem-oriented to find the final form of AI-Native.
澎湃 Technology: How does AI apply to intelligent medicine and life sciences, and how to consider privacy and ethical factors?
Liu Qingfeng (Chairman Iflytek):Over-the-counter drugs account for 45% of residents’ medication (the over-the-counter drug market in China was 195.17 billion yuan in 2022, accounting for 45.4% of the retail drug market), but there were more than 40 million suspected irrational drug use cases in more than 600 million follow-up cases at the grassroots level. Many people have basic diseases and are prone to improper medication. There is an 85-year-old man around us who suffers from Alzheimer’s disease. After taking related drugs, he had toothache. After taking a common anti-inflammatory drug, he stopped eating for nearly 20 days, and then he slowly recovered after seeing a doctor in a provincial hospital. Later, it was found that the reason was that I didn’t know that this anti-inflammatory drug had a reaction with the original drug. In addition, there were some contraindications for drug use. The people didn’t know it, nor did the sales staff in the pharmacy, and it was impossible to ask the doctor for every disease. So, how to meet this demand?
I think ourThe medical model can remind 90% of irrational drug use.Similarly, many people can’t understand whether there are deep-seated problems after getting the medical report, but not everyone has the opportunity to consult a doctor. Iflytek has tried thousands of internal sampling examples, and it can be seen that 40% of the samples should be given more reminders, some should seek medical treatment immediately, or pay attention to various taboos, and more than 3% should be reminded to deal with them immediately.
However, the best doctor can’t be 100% right, and even if the model is more professional than the doctor, it can’t be 100% right. Therefore, we hope that the society should be strict and cautious about such business on the one hand, and never recommend prescription drugs casually according to law, and never make a conclusion casually. Once problems are found, we must remind people to "go to the hospital for medical treatment". The task of the model is to make patients more clear, not to replace doctors, but to help patients better understand the situation in the future so as to better communicate with doctors.
Ren Feng (Co-CEO and Chief Science Officer of Intel):After the appearance of ChatGPT, we conducted secondary training based on it with internal data, so that the model can support professional and accurate information question and answer of biomedicine. In the past, the traditional way was to consult a lot of literature on the biological mechanism of each target. Now, we can talk directly with the big model, such as asking "Help me find a target for treating a certain cancer" to get relevant results. This impressed me deeply, as if overnight, things that could not be done before could be done suddenly.
We are currentlyAutomated laboratories assisted by AI are already being used.Combining artificial intelligence with automation, robotics and biological capabilities, it can not only perform a single task, such as Qualcomm screening, high-content imaging, second-generation sequencing, etc., but also realize an integrated series process, such as a fully automatic closed-loop wet and dry experiment that can complete target discovery and verification within 14 days. What is the effect? Take DMTA(Design, Make, Test, Analyze) in drug research and development as an example. In the past, it took a chemist about 3 to 6 weeks to manufacture, purify, quantify and identify the required compounds in each round of synthesis, and then a series of biological analysis was carried out. The automated laboratory can process it 24 hours, the reaction can be carried out at any time, and it is possible to shorten the synthesis time from 3 to 6 weeks to 3 to 10 days. In addition to the perspective of research and development, for hospitals, doctors and patients, the empowerment of AI is all-round, such as assisting diagnosis and tracking changes in patients’ conditions in an intelligent way.
Chen Ran:We see that multimodal molecular models and image models are being made. The fields and industries that need high IQ of human beings are the development direction of large models, and the directors in hospitals are scarce resources.
Liang Jiaen:The first thing we do is to write and review medical records. We will help doctors check whether there are hidden risks in diagnosis and treatment and which ones do not meet the medical insurance standards. The bottom line we have to keep is to protect users’ privacy, and our customer-related data are desensitized data.
澎湃 Technology: What specific educational changes will AI trigger? You can refer to your practice.
Xiao Yanghua:In the future, with the rapid development of general artificial intelligence technology, its impact on education will be very far-reaching. What to teach and learn in the future, and how to teach and learn will become problems.
Every progress of artificial intelligence seems to be marked by AI passing some human examinations, such as college entrance examination, registered doctor examination and registered judicial license examination. Then the progress of AI often reflects many problems in human education, such asSince AI has passed so many human exams, what is the significance of our exams?Advanced artificial intelligence seems to have been cultivating machines into adults, while backward education seems to be cultivating people into machines, and many excellent students have been cultivated into machines that brush questions. This is a problem that education should try to avoid, so the educational problems mapped out by the development of AI deserve our attention.
The second question is how to teach and learn. In the future, the significance of knowledge acquisition and skills learning will be reduced. Modern civilization has always taken knowledge discovery and acquisition as its main purpose. We used to take "reading poetry books and learning five cars" as our glory. Such a pursuit will become less dazzling in the era of big models. Because it is the big model that is the most knowledgeable, and he has learned almost all human knowledge, then the cheapness and depreciation of knowledge will be replaced by the wisdom of human survival and development.
Zhang Peng:Once I talked with the technicians of Good Future. AI is better at learning than human beings and faster than human beings. Do our children still need to learn these basic knowledge in the future? What will children in the future learn? I said in a joking tone, learn AI. In such an era, everyone should master the principles or basic knowledge of AI. Second, we still need to learn basic knowledge. Of course, we can consider how to use auxiliary means to improve learning efficiency.
The evolution of AI is forcing the evolution of human beings. If human beings don’t want to be replaced and enslaved by AI, their own intelligence level and learning ability need to be adjusted and evolved.
Chen Ran:School resources are limited. If the resources of good teachers are precipitated through AI and become teachers in all subjects, education will definitely change qualitatively.
Liu Cong (Dean of Iflytek Research Institute):The field of education belongs to a very important and deeply accumulated scene in iFLYTEK. Therefore, when we started the large-scale model research and established the "1+N" system in December 2022, we made it clear that education belongs to one of the important landing fields of "N". Facing parents and students, iFLYTEK AI learning machine empowered by iFLYTEK Spark Cognitive Model has realized the correction of Chinese and English composition, carried out companion dialogue practice in oral dialogue, and scored children’s pronunciation. Spark language companion App allows students to communicate face to face with virtual teachers. Facing teachers, Spark teacher assistants can innovate the planning unit teaching design, inspire the creation of situational teaching activities, and generate interactive teaching courseware with one click, thus improving the efficiency of teachers’ lesson preparation.
澎湃 Technology: With reference to your case, what progress has been made in the big financial model?
Wang Xiaohang:In the small model era before the big model era, we have been fully AI-oriented, involving digital guidance, risk management, claims, financial planners, consultants and sales teams. But the arrival of the big model has brought us great shock. It can easily refresh the best performance of the small model era in many financial scenarios.Intention understanding, qualitative analysis, quantitative analysis and other fields have surpassed manual work.In the era of small model, every application scenario needs to be deeply customized end-to-end. algorithm engineer and time cost are very expensive. The large model unifies the algorithm architecture. After simple fine-tuning and adaptation, a model can solve a large number of problems, do better than before, and improve the business efficiency of enterprises, so it can accelerate innovation and make people focus on professional and in-depth creative work. Specialized financial services will also have an alternative experience. At present, all mobile Internet interactions are mainly based on graphical interface (GUI).Future services will be completed through multi-modal interaction and language interaction. How to integrate with the graphical interface of the mobile Internet to form a new interface and experience may happen in the next two years.
Our big model has been tested in depth for half a year, which is the new version of Xiaobao. We hope to bring different generations of service experiences to mass customers, answer questions in finance, be responsive to hundreds of digital tools, customize personalized service plans, and change the interactive mode of graphical interfaces. We have high expectations and hope to collect user feedback in an orderly manner to provide a better experience. Internally, we provide "small assistance" for financial planners, analysts, sales staff, claims experts and content operation creation teams. On the basis of the original digital exhibition platform tools, AI brings new productivity and improves efficiency. As a big model room, financial technology has just started, but we have also seen some rapid iterations, which have opened our eyes under the impact of the big model.
Chen Lei:What we do is a vertical model, with one end assisting manual service users and improving customer service experience; One end serves internal professional staff, combined with our understanding of business and data in finance, and is used for code-aided generation and automatic data mining. Without optimization, the efficiency of verification stage is improved by about 20%.
澎湃 Technology: Will there be a phenomenal generative AI application in the coming year?
Wang Xiaohang:I’m sure it will, but I don’t know when it will appear The next application will not be GPT itself. It should be in the industry. There will be more than one application in the industry, and all industries will have it, such as e-commerce and finance. Digital financial services have been very rich and diverse, but it is still too complicated for the public. How to bring simple, professional and intelligent wealth management services to customers is of great value. The next service experience upgrade can only be achieved through AI. Every industry is waiting for its own AI super application. Whoever has more digital conditions in the industry and who combines industry and technology will be able to bring such products.
Chen Lei:Yes, especially for code generation and code-aided applications, it will be spread in the industry soon. Everyone is trying, and many attempts have shown initial results.
Wang Fengyang (Vice President of Baidu Group and Head of Mobile Eco-business System):We have seen some apps that can reach the top of the app store in 12 hours, and the number of users behind them has exceeded 1 million. I think the phenomenal generative AI application will definitely appear, and it will exceed everyone’s expectations.
Wang Jiaping [Chair scientist, AI Security Pratt & Whitney System Research Center, IDEA (Guangdong-Hong Kong-Macao Greater Bay Area Institute of Digital Economy)]:I thinkThis will be faster than the phenomenal application of the Internet.Generative AI directly produces content, and does not need to work hard to accumulate content producers. As long as it finds consumers, its iteration and growth will definitely be much faster than before. However, the existing technology needs to be improved, because the quality is not good enough now, and many people will say that it can be seen at a glance that it was created by AI. Because the existing content system has raised the threshold to a very high level, I think it will take time for AIGC to reach this height. Once it crosses this height, it will "kill" many industries dominated by content consumption, including online texts, short videos, short plays and so on. I think it will be subversive, but this subversion may not be new or a new source industry.
Chen Ran:Not for a while, any technological development has stages and accumulation. We developed by leaps and bounds, and it took us a year to walk for 10 years. Now I’m running blindfolded on the highway at a speed of 100 miles. How do you know where the next exit is? With the rapid development of technology, it must be difficult for you to judge the direction. When you slow down, you know where the scenery is the best. It is possible to get off the bus and open a supermarket and a homestay next to it. Stop and think a little while flying, and then there will be various business models and applications.
I think next year is an important time, how to let all kinds of forms settle down and form a closed-loop business. The United States has formed the ecology of thousands of companies, and now China is fighting alone, without forming an ecological platform and community, so our company is doing this. The business model comes from practice. It requires platforms to connect upstream and downstream. Some people produce big models, some people use big models, and some people serve big models. Finally, an ecology is formed. Company A and Company B can do business together. The element of platform and community ecology is open source, through which the ecology develops rapidly.
An Xiaopeng (Director of Alibaba Cloud Science and Technology and Research Center):I am cautiously optimistic about the forecast. It is the result of many factors going hand in hand, not a single factor. First of all, for China’s argument that there are rich application scenarios, we still have to go back to the previous stage.The premise of rich application scenarios is that there is enough data accumulated in a specific industry, which means that even if the application scenarios are very rich, if this premise is lost, the quality of the model will be discounted.Second, we need the basic computing power, open source model, model training, deployment and operation, and various tools to build the whole ecology, lower the whole technology threshold, application threshold and cost, and promote the whole commercial application, which is also the work that Alibaba Cloud has to do today.
The inflection point of industrial revolution brought by any major technological change must come from the inflection point of cost. When the input-output ratio has not changed, the industrial revolution will not come. The steam engine was first invented to pump up the water in the mine, and its cost was very high. So before Watt, the steam engine was invented long ago, but the steam opportunity often broke down. What did Watt do? He greatly improved the stability, safety and reliability of the steam engine and reduced the cost, so the steam engine changed from a special machine that can only pump water to a general machine that can pull the train forward.
Let’s look at the digital technology revolution, such as computers. When IBM’s mainframes and minicomputers appeared, its influence on the progress of the whole society was limited. Only when the PC (personal computer) arrived, especially when the Internet speed was highly improved and the cost was extremely reduced, did it promote the whole round of industrial revolution. soThe inflection point of technology and the arrival of the inflection point of technology cost, coupled with the lowering of the threshold of use, will promote the technological progress of an industry.This is very critical.
Everyone says that GPU is very important, and everyone can develop GPU, so what is the most important thing we can see from NVIDIA? As Andrew Ng said, before the emergence of CUDA (note: closed-source parallel computing platform and application programming interface developed by NVIDIA, which allowed software to use certain types of graphics processing units for general processing), there may not be more than 100 people who could program with GPU in the world, but at present, the number of CUDA developers in the world has reached several million. CUDA has extremely lowered the threshold of use, and when the threshold of use has been extremely lowered to build an ecology, this ecology is a powerful moat. It’s the same for us. Real commercialization needs technology and industrialization, but in fact, the use threshold of the general public and the sharp reduction of the use cost are the key points. But this is the result of iteration, not waiting for it to decrease one day, but under the continuous iteration of technological progress and industrial application, its cost will decrease rapidly, thus promoting the development of the whole ecology.
澎湃 Technology: What is the question you want to know the answer to about the application direction of the big model?
Wang Xiaohang:In which mainstream, rigorous and deep industry will major application innovations emerge and be recognized on a large scale?
Chen Lei:What is the commercialization path of the big model? At present, the charging mode of all large models is based on the number of tokens called (note: token usually refers to the smallest unit in the text processing process in the AI field), but the hardware and manpower input are uneven. In the future, it is a key problem that the application of big model can be related to the actual business value and fed back to its pricing, and we are also curious.
Zhang Peng:Will the phenomenal AI application exist, or where will it exist?
Liang Jiaen:How to solve the controllability and reliability problems of large models when entering vertical industries to solve practical problems? At present, it is difficult to fundamentally eliminate the "illusion" in the framework of the large model, and many back-end means are needed to help improve the large model. In the laboratory, 90% and 95% of the controllability and reliability are quite high, but it is difficult for users to use it with confidence in serious scenes.
Chen Ran:The application of big model serves people, but various applications may replace people, so how to identify people’s skills?
