Author｜Dong Mei

Interview guests｜Yu Youping, President of Zhongguancun Science and Technology; Kenneth Stanley, former distinguished researcher of OpenAI

In the past two years, we It is clearly felt that large models have become an important force in promoting industry change. These models with powerful natural language processing capabilities and deep learning capabilities, such as the GPT series, BERT, etc., are showing their unique value in various fields, but the endless introduction of large models is far from meeting people's expectations for artificial intelligence. What people want to see more is that large models can actually solve problems in life and bring substantial changes to human society.

To realize this expectation, the application of large models is particularly important. Only when large models are truly integrated into people's daily lives and become a powerful tool for solving problems can they truly realize their potential. However, the implementation of large model applications is not easy. Although some leading enterprises and scientific research institutions have successfully applied large models in actual business, for most enterprises, the implementation of large model applications is still a huge challenge. challenge. How to overcome these difficulties and promote the implementation of large model applications has become the focus of current industry and academic circles.

To this end, this issue of "Geeks Meet" specially invited Kenneth Stanley, a former outstanding researcher at OpenAI and the author of "Why Greatness Cannot Be Planned: The Myth of the Objective" Professor, and Mr. Yu Youping, President of Zhongguancun Science and Technology, discussed the challenges and opportunities of large-scale model application.

The following is an edited transcript of the interview:

InfoQ: I am very happy to discuss the implementation of large model applications with two teachers, which is currently a topic that is of great concern to both the industry and academia. First, please briefly introduce yourself.

Yu Youping: I am Yu Youping, President of Zhongguancun Science and Technology. I have worked at Baidu for the past 17 years and have been responsible for Internet advertising products, big data, AI He Yun has been engaged in management for many years. After joining Zhongguancun Science and Technology, I focused on leading the team to combine large-scale model technology with the actual needs of enterprises, provide enterprise customers with end-to-end high-quality products and services, and help enterprises realize their goals. Really gain the business value brought by large models

Kenneth. Stanley: Thank you for having me here. My experience is quite diverse, mostly in the field of artificial intelligence research. I was a professor and I was an entrepreneur. I started a company that became a company. Uber acquired it and turned it into the Uber Artificial Intelligence Lab. I led basic research there for a few years, and then I went to OpenAI, where I led a team called OpenAI.team.

I then founded a company called Maven, whose purpose was to create a new social network based on open artificial intelligence technology, which I spent a lot of time developing. In fact, I recently left Maven and was thinking about what to do next.

InfoQ: Can you share with us some of the technology trends you are paying attention to recently?

Kenneth Stanley: One of the big trends I am paying attention to is the development of large models and artificial intelligence. This is very interesting to me. I have been paying attention to the progress of artificial intelligence for a long time. At that time, There aren't any big models. A hot topic recently is scaling law. People are speculating whether large models will continue to scale with more data and calculations. If so, does it mean that we are moving towards what people call artificial general intelligence (AGI). If the scaling law fails, will the large model stagnate? These are hotly discussed topics, and this is also a topic that interests me. I'm a researcher and I think about the underlying nature of things and how they actually work.

Yu Youping: Zhongguancun Science and Technology focuses on large model applications. On the To C side, the applications of large models are more eye-catching in AI search and question and answer, followed by AI creative creation. Zhongguancun Science and Technology focuses on the application of large models on the To B side, helping companies improve their external customer service quality and marketing conversion rate, and internally helping companies improve their knowledge management and operational services. I can see many useful and interesting cases in these areas. Zhongguancun Science and Technology is involved in both general application fields and industrial application fields of large models, and has made some good progress.

I am also paying attention to the topic of scaling law mentioned by Professor Stanley. Whether scaling law will continue is uncertain and may take three to five years or more to be proven, so perhaps it’s important to be patient.

Challenges encountered in the implementation of large model applications

InfoQ: If 2023 is the first year of the explosion of large model applications, then 2024 will be It is known as the first year of implementation of large model applications. What are the large model applications that impressed the two teachers this year? What scenarios do these large model applications belong to? Why do you think their presence is important?

Kenneth Stanley: Yes, I think programming is obviously a big application area, and there are many applications that have emerged in this area.

Making the professional experience of programming easier is an important thing, because in some extreme cases, it may even require fewer programmers or no programmers to do some things. I don't think we have it yet. Not quite to that extent.

But this is already amazing. If I could just describe an application and let a computer program it, that would definitely be revolutionary in the industry, so the advances in automated programming are very fascinating.

At the same time, creative applications based on large models are also very interesting. You mentioned that image generation and video generation have a significant impact on the industry. Music generation is also interesting. The focus of this process is not just that the big model can make good music or good art, but that as a human being, I can have my own ideas and then filter these ideas through the model to ultimately generate interesting music. The big model and I think it’s cool that humans work together to do this.

A very potential partnership is a perfect match between a person with a good idea but no skills and a model with a bad idea but good skills. I think it could really change the landscape in a lot of industries where there was once only a small elite group that could do things.

For example, in a professional field like musicians, only professional musicians can really participate in it. Professional musicians are the only people who really participate in actual discussions related to modern music today, but if suddenly an amateur What could a hobbyist make that was just as good? They had musical ideas but no skills, but now they can actually implement their ideas, which means they can now participate in discussions they couldn't before.

So it has greatly broadened the scope of people who can participate in the forefront of many fields. I think this has a very significant impact on many industries.

Yu Youping: Yes, I completely agree. Large models are used in creative fields, such as image and music generation. At the same time, large models are also making great strides in the field of programming. When I was at Baidu, there was also a very good programming tool called Baidu Comate (Wenxin Quick Code). About 30% of the code was generated by Comate. The effect is impressive.

InfoQ: I would also like to ask Professor Kenneth Stanley, based on your research experience at OpenAI, can you tell us which cases of large model application have left the deepest impression on you? Why?

Kenneth Stanley: When I was working at OpenAI, coding was a very important programming topic there. When I joined OpenAI in 2020, it was very impressive that computers could actually code, but also astounding how creative these models were at finding new ideas or searching for ideas.

In terms of creativity, they are not close to human levels, but these models can provide a lot of help in creative conception. For example, having a large model comment on whether a story is well told, it Being able to give your own ideas, before the advent of big models, this was not the case. There's absolutely no hope for a computer to say something like that. You can't input a story.Go into your computer and ask it if this is a good idea.

So I was very interested at the time, and now with the development of technology, this has changed, and I can even ask the computer if my idea is interesting, although (the computer) is not necessarily at the human level, but It can at least comment on these things. So, the question it raises is, what can humans do with it?

They can do work such as creativity, opinions, art and business ideas, as well as work with subjective and qualitative judgments like programming. These are completely new and I am full of expectations for them.

InfoQ: You mentioned that large models will be codeable by that time in 2020, which should be a very early stage. In your opinion, what is the difference between the large model programming of the past and the automatic programming of today?

Kenneth Stanley: I think there have been rapid improvements, so it is now more reliable, can do more, make fewer mistakes, and the tools are more mature for operation The interface is more mature and works more seamlessly with your coding setup. But I still think it hasn't crossed that critical mass where I want it to build an app for me if I don't know how to program.

If we can reach that level, it will change the entire industry. But it's evolving step by step and getting better every year. I have to say that this is very relevant to scaling laws, because if the development of scaling laws slows down, then this progress will also slow down. In fact, we are still far from the point where we really don’t need programmers.

InfoQ: So, can you predict how long it will take to reach the goal of no programming skills? Is it that you only need to have an idea and an app can be automatically generated?

Kenneth Stanley: That's an interesting question, I think it depends a lot on scaling laws, and because of that it's hard for me to give an exact number because we don't know whether scaling laws will last. If the scaling law ceases it does not mean that progress has stopped, it just means that progress will be more difficult and slower.

The important thing about scaling law is that you don’t need any new ideas, all you need to do is add more computing resources and more data, so you don’t need to use any more brains. . We just have to wait until we get more GPUs, get more data from the internet, and then it will get smarter and smarter. If scaling laws continue as a trend, maybe two or three years from now we'll get the answer we want, which is to describe an idea to a large model and get an application directly out of it.

IfThis won't happen due to stagnant scaling laws. But it’s also interesting to think about it conversely, because once the scaling law stagnates, it means that new ideas are needed, including new architectural-level ideas, to overcome the difficulties we are observing now. The current architecture is not adequate to continue to drive progress, but that doesn’t mean progress is impossible, it just means we need new ideas, and humans do always have new ideas.

The big model is an idea that we have had, and we will continue to have new ideas, so I think we will continue to see progress. It’s just that ideas don’t appear according to a predictable timeline, just like I can’t tell you when the next Einstein will be born. This doesn’t follow a timeline, and it’s not a rule, so it’s hard to predict 10 Years or 20 years or more. My general view is that it will be either a few years from now or ten to twenty years from now, depending on the scaling law.

InfoQ: So let’s go back to the implementation of large models. What do the two teachers think is the biggest challenge encountered in the implementation of large model applications? Is it a technical limitation, market acceptance, or other factors such as scenarios (including algorithm optimization, data processing, model training, etc.)? How do you think these technical challenges affect the implementation of large model applications?

Kenneth Stanley: I think the currently known limitations of the model are the biggest obstacles to its practical application, including model illusion. Hallucination means that they tend to say something that is not true and treat it as if it is true. Presented as a fact, this is obviously a big obstacle to many applications because it brings greater risks. Whether in the financial or health fields, the more impact the model has on the application of specific scenarios, the more likely it is that hallucinations will occur. The more dangerous it is.

Because once the model is wrong, a lot of money may be lost, so people will be hesitant to apply the model. If it may be life-threatening, they will be even more hesitant, such as in medical applications, In application scenarios such as legal applications, if the model makes up laws that do not exist, it may cause trouble in court. So currently this type of application requires a lot of manual supervision.

Some fields such as image generation do not have this problem, and I think this is why applications in those fields are more popular. In the field of image generation, even if it generates a bad image, no one will die and no one will lose money.

Yu Youping: Yes, if we can solve the illusion, this industry will have greater room for development. But there are other challenges, such as creativity bottlenecks with large models and difficulties in generating novel content.

Kenneth Stanley: Indeed, what you mentioned may still be closely related to hallucinations. For example, we ask the large model to give a new recipe or a new invention, many times it will tell you something that already exists, I call it creative illusion, it's like an illusion, you think you are saying something new, but in fact you are not.

It takes credit for things that are actually in its training data, which is a big problem for creativity. Because of this creative illusion, it is also difficult to use it for creative applications, because not only does this illusion show poor creativity, it also poses risks because you can't be sure that what comes out of the system is a truly original idea.

I mean you can make it create art and music and all kinds of stuff on its own, but there are a lot of key limitations that need to be addressed in order to fully implement it, limitations like this It is a real basic research problem in artificial intelligence. I think it can be solved, but it will not be as simple as improving scaling law.

Yu Youping: Zhongguancun Science and Technology mainly serves corporate customers, so we have encountered more problems. First of all, there is still a lot of room for improvement in the capabilities of basic large models. In enterprise application scenarios, we need to provide very accurate answers. As Kenneth said, general large models are far from being able to directly meet the needs of scenario applications. Secondly, when dealing with complex problems under the Agent architecture, the response speed of the model is not good enough. ; Finally, corporate customers’ awareness of large models is also a very important aspect. Enterprise customers often have high expectations for large models and hope that they can immediately replace part of the labor force. This is still a huge technical challenge.

Zhongguancun Technology focuses on integrating traditional software to provide end-to-end solutions, such as full-media contact centers. We strive to balance model accuracy and responsiveness so that our customers can enjoy current technology and gain real value from it. From a business perspective, this is crucial.

Data security and privacy

InfoQ: This is indeed the case. We face many challenges in the process of implementing these applications. I would also like to know what the two teachers think about how we can strike a balance between technological innovation and data security and privacy protection when implementing large model applications? Does OpenAI have any experience in this area to share?

Kenneth Stanley: OpenAI has an entire department specifically responsible for APIs, that is, commercial applications. They are really worried about these issues, so they will try to establish some security, privacy and other potential risks. safeguards, such as potential misuse of the model by customers. I feel like this is an evolving capability, just like a lot of companies are figuring out how to solve problems like security and privacy when it comes to large models.

This is different from other types of business applications or computer applications. I think there is still a lot to learn. Security and privacy are important to OpenAI because they have many commercial customers.

InfoQ: I would also like to ask Mr. Yu, what has Zhongguancun Kejin done in this regard?

Yu Youping: The core of large model applications is high-quality data. Data security and privacy protection are indeed the keys to large model applications. In this regard, Zhongguancun Science and Technology has formulated very strict data management systems and policies to protect the privacy of corporate customers and consumers.

InfoQ: In addition to doing a good job in data security and privacy, the promotion of large models is also very important. Just because a large model application has been developed, it does not mean that it is accepted and loved by users. Can the two teachers use their own experience to talk about how to improve users' awareness and trust of large model applications, so as to promote their implementation?

Kenneth Stanley: I think this is closely related to the challenges mentioned earlier about these models. When we talk about things like hallucinations, creative illusions and even safety issues, all of these issues relate to the customer's acceptance of these models. To be honest, I think there is a limit to what we can do. From a customer service perspective, models are only a small step in improving customer service work, and model technology ultimately still has a lot of room for effort.

For example, when we were using large models for customer service, I would feel uncomfortable if I called a company and had a recorded conversation. The idea of ​​​​improving this situation sounds very attractive, but the key lies in issues such as model illusion. The risks it brings are too great, and it is not clear whether it is ready (application) yet.

So the fundamental thing is to solve the basic problems of large models themselves, so that they can reach a level where we can accept their mistakes, that is, a level that can compete with humans.

InfoQ: I have seen many companies using large models to engage in customer service automation. The valuations of these companies are very high, but the quality of their handling of actual scenario problems is not satisfactory. What do you think? What do you think about this phenomenon?

Kenneth Stanley: That's interesting. I think the reason why they're highly valued is probably just because customer service is a huge industry, and if someone can actually solve this problem, it will unlock huge value. , this is obvious.

Because all industries require customer service, I think the situation you mentioned does exist in reality. But I don't think we're there yet, because the current models are not good enough to do that, and they're doing a pretty good job, and five or 10 years ago it would have been hard to imagine that we could actually replace all of these customer service agents with computers. , but suddenly it's possible because the big models are already so good.

I think this kind of imagination is so attractive that people are willing to take risks and bet on it.

Yu Youping: In our practice, clients hope to get very accurate answers and minimize illusions. Zhongguancun combines science, technology and financeUsing very high-quality data and training the model in some specific fields, you can get a vertical large model application that serves specific scenarios.

In the field of telemarketing, we have cooperated with some customers in the fields of finance, manufacturing and e-commerce, and the results are much better than expected. In actual marketing scenarios, the marketing conversion rate has been significantly improved, and the results are very good, but there is still room for improvement.

Will low-skilled workers be replaced by AI?

InfoQ: As more and more artificial intelligence models become smarter and smarter, some people may worry that large models may lead to social inequality. For example, automation will replace some low-skilled jobs. How do teachers view this concern?

Kenneth Stanley: I do think the development of artificial intelligence is a serious problem because it may have a range of broad economic impacts. The range of possibilities, or continuum, is very broad, starting with some jobs being replaced by automation, and ending with the potential replacement of all jobs on the planet.

At first, only a few jobs were replaced. It is not clear whether this is really a serious situation. Maybe some jobs were lost, but some jobs were also added. Maybe efficiency was improved. Improvement, the situation is not bad.

Overall, everyone can benefit. But as things get more extreme and technology gets smarter, things start to get weirder because entire industries could be displaced. For example, you may no longer need programmers in the future, just like you no longer need drivers, with robotics you may no longer need domestic workers.

Even the creative side is even stranger. Because I think before when we thought about automation, we thought it was going to replace people like factory workers. About 10 years ago, we might have been talking about factory workers being replaced by robots. We would not have thought at the time that people like artists would also be replaced. The current development of artificial intelligence is leading in strange directions.

So what we are seeing now is that in extreme cases almost everything has been replaced, which is extremely profound impact, and this is not just a small economic direction topic, this is like A radical reconstruction of society, so will this actually happen? When will it happen? We have no way of knowing. But because it could happen, it's worth at least thinking about what that would mean and how we would respond to social problems in that scenario.

I think one thing is clear, the economy is going to need a complete restructuring, the way the economy is run now is not going to work, and I'm not sure if it will still be capitalism then, because a lot of people will have nothing to contribute. . Of course, there are still some people who can still contribute something. For example, when it comes to making very important decisions, humans may still be needed. For example, political decisions and decisions involving wars still need humans to complete.

But bigMost things can probably be replaced by computers, so how will the world work? How will it be built? No one knows, but it’s something to think about. But there’s no need to panic because it won’t happen tomorrow, so don’t be overly anxious.

InfoQ: You said that many industries and jobs will be replaced by artificial intelligence in the near future. For programmers and developers, what do they need to do if they do not want to be replaced by AI? How to stay competitive at the skill level?

Kenneth Stanley: I don’t want to imply that it’s in the near future. My point is I don’t know how soon it will happen. It’s not that it’s close. It could happen soon, but it could also be close. We are far away.

This is a good question, and many people have asked me this question. Especially young people, people in college, they will say if everything is going to be replaced, what major should I study? This is a tough question. For those who are really concerned about this issue, I think it's important to note that too much planning may not be necessary. Because there are some things in life that you simply cannot plan for, just like if an asteroid is going to hit the earth, we cannot prepare in advance.

I would advise you to keep doing what you are interested in, because there is nothing you can do about something like an asteroid hitting the earth, so I would not advise you to give up on your life, you should not spend all your time On panic. Continue to live your life, explore your talents, and don’t worry about catastrophic events that are beyond your control. But there are some things to worry about, such as artificial intelligence that has made significant progress to a certain extent. This is different from the previous situation where everyone is unemployed. , that is a more extreme situation, and artificial intelligence is developing step by step. You can see which fields will be affected only when artificial intelligence develops to a very advanced level. In my opinion, highly creative fields are short of It will not be replaced by AI in a short time, and those who do programming may be in danger in the short term.

So I don’t think you need to prepare for this in advance. I say to young people, don’t give up on your interests and don’t give up on the world.

We don’t know when or if these things will happen, the best thing to do is to continue to maintain your enthusiasm and try to put yourself in a position where you can provide creative ideas.

Yu Youping: In my opinion, only some very simple and repetitive tasks will be replaced by large models. Because we've seen that large models can increase productivity and creativity. Large models will create many new things, and the emergence of some new jobs also benefits from the development of AI technology.

At Zhongguancun Kejin, we provide enterprises with some intelligent tools based on large models, such as text robots and voice robots, as well as some tools to help employees improve work efficiency, such as large model sparring, large model assistants, etc. . I'm not worried about this issue. I think it's a good thing and the world will continue to change. There's no need to worry too much.Be careful and do what you can.

What is the field most likely to be disrupted by AI?

InfoQ: Artificial intelligence models have greatly changed our lives in recent years. Which industries or fields do the two teachers think are most likely to be disrupted by large models in the next few years? Why are these areas leading change?

Kenneth Stanley: I think low-risk creative fields, such as visual artists, have been affected now. If you need to make a marketing poster, you can basically do it for free now, you don't have to pay anyone, whereas in the past that was the designer's job, so I think a lot of industries are going to change dramatically.

We see that as the capabilities of large models improve, this impact will extend to fields such as video, music and audio. These fields are relatively low-risk, which means that even if they encounter problems, they will not be affected. It poses a threat to people's life safety and will not cause major economic losses.

Programming is also a field that is relatively easy to be subverted. Although it is very dangerous for programs to go wrong in some places, I think it may still affect the fields of programming and engineering.

Because humans will collaborate with AI to achieve huge improvements in work efficiency, this will have an impact on the entire industry. Although there are concerns that AI may displace some jobs, I firmly believe that this risk is not as severe as people think. In fact, those jobs that are at risk of being replaced are often just a link in the entire process, and the missions of these people are often more critical and indispensable. At the same time, as AI becomes more ubiquitous, fewer workers may be needed, but those who remain will be able to leverage the power of AI to complete more tasks with greater efficiency and greater productivity.

Yu Youping: I think those data- and knowledge-intensive industries are most likely to be changed, such as law, medical care, education, etc. Because there is a huge amount of data in these fields, large models can provide more automated services and tools to help everyone improve work efficiency.

At the same time, I think professions such as lawyers, doctors, and teachers will not be completely replaced. After all, these professions have extremely high requirements for professionalism and accuracy of output results. But it is very certain that large models can help people work more efficiently.

InfoQ: I would like to ask Mr. Yu, based on your observations, what are the differences between the implementation of large model applications at home and abroad? Why this difference?

Yu Youping: Regarding the situation abroad, Professor Stanley has just shared a lot. Domestic applications are mainly concentrated in finance, office, education, medical and other fields, as well as in employee efficiency improvement and customer service inside and outside the enterprise.

As far as customer service is concerned, domestic use of telephone services or marketing is more common, while foreign countries prefer online text services. This is related to the habits of domestic and foreign users. Foreign large model clothingPublic cloud services are the main services, while domestic services are mainly privatization. This is because there is a gap in business models between China and foreign countries, but there is not much difference at the technical level.

InfoQ: Two teachers think that in the current technological environment, are we ready for the full implementation of large model applications? If so, please expand on what preparations we have made? If not, what else do we need to do?

Yu Youping: To embrace new technologies, there is no need to be demanding about completeness. Nowadays, large model technology can be implemented in many fields. Zhongguancun Science and Technology mainly focuses on two aspects: 1. Help corporate customers to better apply internal knowledge and data, create knowledge assistants for enterprises or industries, and improve employees’ Work efficiency; 2. Help corporate customers, provide better services and marketing to their customers, and improve corporate marketing and service capabilities.

Chinese customers prefer to receive end-to-end services. We combine our product capabilities in all-media contact centers and audio and video platforms to provide customers with end-to-end products and solutions. At the same time, we are also actively exploring the export of products overseas to serve global customers. In terms of large model technology, we also keep up with cutting-edge technological progress and constantly optimize our product experience and services. We believe that as large model technology continues to improve, the products and services we provide to customers will become more and more sophisticated. Intelligent.

Kenneth Stanley: I think it depends on the industry, some industries are better prepared than others, and the differences in different areas are very interesting. There are some industries where it's obvious, like in customer service, they're trying to do something with big models, but other industries haven't even realized that this could impact their fields, and I think what's really interesting is that it can impact every industry, even It can be said that no industry will not be affected by large models.

I'm reminded of a recent incident where some bugs flew into my house. I didn't recognize the bugs. It was very hot at that time, so we opened the windows and these bugs started flying in. They were red, had long wings and seemed to die when they hit the ground, I thought maybe the heat killed the bugs.

So I was curious about what was going on. I had never seen them before, but I wasn't too worried.

Later, my ten-year-old son said why don't you ask ChatGPT what it is, and I said that's a good idea, we should ask ChatGPT.

So I pointed at the bug and asked what it was, and ChatGPT answered that it was a winged termite. I had never heard that termites were wood-eating insects, but I never knew they had wings.

So I asked ChatGPT further, and it explained that when a termite colony becomes too large, it needs to create a new colony, and they will grow wings so that they can create new colonies. I learned something new and also learned that my house was in danger due to termites getting into the house so I called a termite inspector.

The reason I tell this story is because it’s incredible to me that this termite inspector got a business because of ChatGPT. You wouldn’t think that people inspecting termites and AI have anything to do with each other, but that’s exactly what happened. You see, anyone can benefit from this,

which shows how widespread its potential impact is, not to mention that my house might be safer because of AI. I think everyone should learn about AI and think about what it can do for their industry.