Dialogue with Luo Ye of the University of Hong Kong: There are three "disjoints" in the field of artificial intelligence, and OpenAI is an "imperfect answer"

AI is undoubtedly one of the most exciting areas in the tech world and the most concerned area in the investment community in 2023. From the rapid iteration of ChatGPT at the beginning of the year to the OpenAI "infighting drama" at the end of the year, artificial intelligence is advancing rapidly at a pace that is significantly ahead of theoretical research, both in terms of scientific innovation and organizational management.

"People tend to overestimate the progress of AI in the short term, but underestimate the progress of AI in the long term. Luo Ye, associate professor at the School of Economics and Management of the University of Hong Kong and deputy director of the Institute of Digital Economy and Innovation, has been researching artificial intelligence for nearly a decade, and in his opinion, the progress of artificial intelligence in the industry in the past decade has far exceeded expectations. Including large language models, the arrival time is also much faster than expected.

However, the rapid development of industry has also brought many new issues. In Luo Ye's view, there are three "disjoints" behind this that need to be resolved. One is the disconnect between application and theory, the second is the disconnect between industry and academia, and the third is the disconnect between the entire field of artificial intelligence and other social fields.

As an AI scientist and economist, Luo Ye not only pays attention to the progress of artificial intelligence in the field of scientific innovation, but also pays attention to the management problems and power redistribution problems brought about by new organizational forms. Recently, the first financial reporter had an in-depth dialogue with Luo Ye on the cutting-edge issues in the field of artificial intelligence and the huge controversy caused by OpenAI. In his view, OpenAI's "infighting" turmoil is a huge warning - power is increasingly concentrated in the hands of a few, and traditional corporate governance theory and practical experience are no longer enough to face an organization that manages "geniuses".

There are three "disjoints" in the field of artificial intelligence

CBN: Looking back on the year of artificial intelligence, what disruptive progress do you think has been made?

Luo Ye: The launch of ChatGPT has indeed had a disruptive impact on the entire industry. But people tend to overestimate the progress of AI in the short term and underestimate the progress of AI in the long term.

I've been working on artificial intelligence for almost a decade. Over the past decade, AI has advanced far more in industry than I estimated. At first, I thought that a lot of problems were very difficult in terms of mathematical principles and machine Xi theory, and they might not be able to solve it in ten years, but they often did it in two or three years.

I started tracking Transformer models as early as 2017, and the early technology was rough, and I thought it would be impossible to use it in a real sense for a decade. But it has only been five or six years since major breakthroughs have been made, such as the sparse reinforcement Xi problem studied by Deepmind. So in fact, industry is always moving at least twice as fast as I expected. Of course, there are still some insurmountable technical problems such as logical reasoning ahead, and it remains to be seen whether they can be solved quickly.

In the long run, we always seem to be underestimating the progress and development of artificial intelligence, but in the short term, what is the difference between each generation of ChatGPT and the previous generation of technology, there is no real change in terms of technical principles. But it should also be noted that this model breaks through a tipping point at the application level.

CBN: What is the tipping point?

Luo Ye: In the past, when we did all the natural language processing problems, we were faced with very complex engineering problems. With the advent of large language models, it is now possible to handle all problems from the same base, which solves a big challenge at the application level.

The GPT model is essentially a compiler. Its input is "code" written in data and natural language, also known as prompts. It's just that the content it outputs is not accurate enough, and many times it is in the opposite direction, or made up. But in any case, it makes the programming process greatly simplified. Using ChatGPT as a compiler to develop a wide variety of applications is very promising, and the resources, manpower, and time invested in many scenarios may be one-tenth of the traditional programming method, and the effect may be better. And because of the creation of this base, there will be a huge breakthrough in the progress of the application level. Assuming that AI technology does not move forward in the future, the existing technologies and potential applications are enough to set off a huge wave of change in various industrial fields.

CBN: You have repeatedly mentioned that the progress at the application level has exceeded your expectations every time. Have you ever wondered why you always underestimate the speed of industry?

Luo Ye: That's a very good question. This is because there are several disjoints in the current development of artificial intelligence.

The first is the disconnect between application and theory. At the application level, both AlphaGo and ChatGPT are empirical applications, and there is no complete theoretical guidance, and experience is far ahead of theory. When we don't understand many principles, the industry is moving fast at the application level and has achieved many breakthroughs through exploration. Just like early alchemy and chemistry, alchemists developed many complex compounds early on, but chemists were still in a state where they could not even understand basic chemical reactions.

It also makes me wonder and hold a conservative attitude about whether this kind of rapid development at the experimental level can be sustained without major theoretical innovations in mathematics, statistics, machine science, Xi.

Yicai: The exploratory way produces empirical applications, which is by "luck"?

Luo Ye: Not exactly, it is likely that the industry has grasped the insights of certain optimization areas, but why this is the case, we still can't fully understand it. There are many experts in machine science Xi theory who are conducting research and exploration in this area, but so far there has been no basic progress in completely convincing and complete mathematical theories.

CBN: Can we understand?

Luo Ye: That's a good question. At least for now, it's not easy.

Human cognitive abilities are also very limited. Experiment-based scientists have their explanations for why these large language models are effective and why attention models are better in the field of deep learning Xi, but their explanations lack basic theoretical support and appear very empty. It's hard to see scientists like Albert Einstein and Newton come out and fully reveal the underlying mechanism. At present, the ability of applied mathematics to describe deep learning Xi is still relatively limited. In the process of practice, experience occupies the main guiding significance, so the accumulation of practical experience is very important in the application of GPT.

CBN: What are the other "disjoints"?

Luo Ye: The second disconnect is the disconnect between industry and academia. The industry has the advantages of more computing power, more resources, more problems, more data, and can see more direct economic benefits, so many scientists are active in the industry, and a considerable number of the best artificial intelligence scientists have jumped from academia to industry. Due to their limited resources, colleges and universities are relatively limited in what they can do.

The third disconnect is the disconnect between the entire AI field and the rest of society. In the past, there was a metaphor for the rise of the IT industry, which is called the "generalization effect" in economics. It is the industry that comes into contact with IT technology earlier, and the productivity of the industry is affected by IT faster, and the speed is faster than that of the industry that comes into contact with it later. Therefore, it will be seen that the financial industry is an industry that has been affected by IT relatively early and relatively largely, while agriculture is relatively late and slow.

The disconnection of the industry will inevitably lead to the problem of wealth distribution, because the higher the productivity, the faster the wealth will be gathered, so the employees in the industry will have relatively higher wages and faster wealth accumulation. Over the past 30 years, research on labor economics in the United States has found that the new high-paying jobs are mainly in the IT and financial industries, while traditional industries have been left behind, and many jobs have even been moved to China, Vietnam, and other places. In the future, whether the rise of artificial intelligence technology will once again affect the redistribution of wealth in the whole society is also a question that needs to be faced.

CBN: This still needs to ask "yes"? Isn't the answer "definitely"?

Luo Ye: Not necessarily, it depends on how the regulator designs the redistribution mechanism. In fact, this is also a deep reason for OpenAI's infighting.

Government oversight is more or less necessary in any field, and perhaps even more so in the field of artificial intelligence. Although there is an opinion that scientists should have the freedom to do what they want. It is true that there should be freedom in scientific innovation, but the results of scientific innovation and the various consequences that lead to it need to be supervised.

An example can be seen. In the 1960s, when General Motors had 1 million employees, members of the U.S. Congress declared that whatever was good for GM was good for America. Looking at IBM in the 1970s, IBM had 5~100,000 employees, which was one of the most influential companies in the world at that time, and its influence on the U.S. Congress was also very large. IBM is followed by Microsoft, with about 10,000~20,000 employees, Microsoft is followed by Facebook, with only 3,000 employees at the headquarters, and today, OpenAI's core personnel are only about sixty or seventy before the release of ChatGPT. As you can see, the number of the world's most influential organizations is declining dramatically, which essentially means that power is being concentrated in the hands of fewer people. While we are all talking about the business value of AI and the redistribution of wealth, we must also pay attention to the redistribution of power it brings. This rebalance of power, brought about by technological advances, could have far-reaching implications for the world.

Judging from the OpenAI infighting, it is a question whether the traditional corporate governance system is still applicable to enterprises in an intellectually concentrated industry such as artificial intelligence.

OpenAI是"不完美的答案"

CBN: Corporate governance in technology companies has become an increasingly problematic issue. If traditional architectures don't work anymore, what do we use instead?

Luo Ye: Obviously, we need a new governance structure, but we don't have a complete answer to what it is, even in the theoretical system of corporate architecture and governance around the world.

From the perspective of the smallest "atom", an artificial intelligence enterprise or organization, there is a clear opposition. For example, the chief scientist of OpenAI is Ilya Sutskever, and some people say that Ilya is 1 at OpenAI, and everyone else is 0, and nothing else makes sense without him. Traditionally, technical staff have obeyed management's decisions, but that doesn't seem to be the case at OpenAI.

Capital would have feared being abandoned by Ireya, but for Ierya, he clearly faced more choices than capital. Although Sam Altman has returned to OpenAI, what is certain is that it is only a temporary state and the real problem has not been solved.

CBN: What is the real problem?

Luo Ye: The real question is how to manage "geniuses".

Traditional corporate governance theories do not address how to manage genius. We tend to think that genius management is difficult.

For example, watching NBA games, Rockets general manager Morey is known for his data-based management. Data management is effective and has a high probability of bringing a team to the playoffs, but it is difficult to become a champion. If you go to produce champions, it's a very difficult thing. Champions are difficult to be formulaic, mass manufacturing, and champion temperament is difficult to explain with conventional management theories, which is why in various sports events, champion coaches, champion players often have a great advantage over first-class coaches in terms of income.

Genius does not like to be managed, and the work he does has three typical characteristics: the consumption of resources and time is highly uncertain, the outcome is highly uncertain, and the process is highly uncertain and unsupervised. He may have spent 99 days lying on the beach basking in the sun, but the last day solved the world's problems. For top AI organizations, traditional business incentives and equity incentives may be difficult to be effective, and the KPI management system will drive out good money and eliminate real geniuses.

CBN: You threw out a question, but we want to hear an answer.

Luo Ye: We have thought about this question before, and we have made some progress, but we don't have a complete answer yet. This is a difficult question all over the world.

In fact, you can think of it the other way around, OpenAI is a question, but it can also be an answer. It's just that its answers are not perfect, so it exposes some problems that are worth thinking about from the perspective of management and economics.

One of the reasons why OpenAI can succeed is precisely because it has a different management model from many traditional enterprises. OpenAI is a non-profit organization that initially received an initial amount of money and then brought together a group of idealistic scientists who did not define what to do or how much money they were going to make, but simply asked these people to do research on an ethereal ultimate proposition: how to regulate artificial intelligence. This is rare in human history where scientists have so many resources and so much autonomy at the same time.

Microsoft, Google, and China's leading technology companies actually have their own research institutes and their own top scientists. But they are still subject to the board of directors, to the company's business, and in many cases to serve the company's business needs.

OpenAI, on the other hand, is an organization that doesn't need to serve the business at all. If a for-profit company says that it needs to invest $1 billion to make a large language model, and it is likely to produce nothing or fail in the competition, the proposal will most likely be killed by the board of directors, or the CEO will be killed outright.

Yicai: OpenAI not only shows us the disconnect between the theory and application of artificial intelligence, but also shows the disconnect between theory and application in the field of management. OpenAI is a management experiment in its own right.

Luo Ye: Yes, OpenAI is not only at the forefront of technology, providing us with a lot of opportunities to observe, but also giving us a huge warning in terms of corporate governance and organizational management.

It is a huge challenge to manage a technology that has the potential to change the world, on the one hand, we want scientists to have a lot of enthusiasm and space to experiment, and on the other hand, we want to limit the unequal impact of scientific innovation, and at the same time balance capital with the interest of scientists.

I don't think we have any radical solutions in management, economics, we don't even have formal seminars. There is still a gap in the whole management science on how to manage genius and how to balance the efficiency and risk of disruptive innovation.

Every field has an ultimate problem, and the ultimate problem leads to the ultimate ideal. Geniuses tend to think about very original problems, and even change the traditional view of original problems. Although some people think they are paranoid, I think they are all very respectable, because the pursuit of the ultimate ideal itself requires a great deal of courage.

GPT-for-Finance"理论上完全可行"

CBN: Finance is a highly data-dependent field, and it is also an important application field of artificial intelligence. We know that the team from the University of Hong Kong is currently building a deep industrial-grade automated decision-making platform in the financial field based on natural language by combining generative AI, knowledge of the field of finance, and a general financial solver - GPT-For-Finance.

What is the stage of GPT's development in the financial field?What are the core breakthroughs?What are the biggest challenges?Some people believe that financial GPT still faces many difficulties, such as many reports of listed companies are tables and graphs, how should large language models process these data?

Luo Ye: Bloomberg was the first to develop BloombergGPT, which specifically uses financial computing, but more applications in the financial field are biased towards a Q&A assistant. We're trying to drive decision-making, not Q&A, and we're trying to build tools that can replace part of the production process in financial institutions.

Based on limited resources, we mainly try to solve the problem of finance that is relatively core. The core breakthrough that has been made so far is that it has been proven to be feasible in principle on certain issues, such as report writing in certain areas of finance. Our model has been able to generate certain chapters of a research report to a relatively professional level.

Tables and images are also fully processable and can be generated, but instead of using the multimodality of large language models, a lot of work is still processed in the traditional way, using AI-tools to integrate into the core of large language models. If there is relatively standardized data in the financial field, such as financial statements, as long as you design prompts for the data you want to obtain, and then combine it with appropriate data to extract. When it comes to calculations, special modules can be designed to complete them, which can improve the overall accuracy and get the calculation results faster.

The accuracy of the model is very important. We've done a lot of work to improve accuracy. For example, by dividing the production process into many small steps, and then distinguishing which links the large language model is most suitable for, the accuracy can be greatly improved after disassembling the layers of the process. This method of dismantling the task and performing traditional machine programming and prompt word engineering is collectively referred to as natural language programming. The nature of this approach is that there is no one-size-fits-all solution, but rather that it varies from application to application.

In this way, there is a great hope that it will be able to achieve a state that can be used in the industry. In terms of the process of decision-making, experiments have shown that it is completely feasible. However, the biggest challenge is that we have limited resources, so many subsequent engineering problems need to be completed by the industry.

CBN: Which parts of the financial sector will be the first to be replaced?

Luo Ye: It may be a substitution, or it may be an enhancement. A more likely way is for the AI to do some of the work first, and then the humans do some of the processing.

For example, part of the work of industry analysts and quantitative analysts. Many analysts do work that is highly repetitive in the program, such as extracting data, analyzing data, etc., which can be achieved using standardized AI-agent. In the policy optimization stage, you can also explore the use of AI-agent.

Finance is very much a game of probability. For example, banks do not pay attention to which loan is risky among hundreds or tens of millions of loans, but pay more attention to the overall non-performing rate. There is also a lot of room for AI to be applied in such fields.

Whether it is in the field of banking, insurance or investment, artificial intelligence can not only play a role in serving customers, but also make relatively large changes in the work process, which may have a great impact on the operation of the financial industry.