On ＂Reflection on the Development Status of AI＂ ｜ Rational View on the Short-term and Long-term Effects of AI
[Moderator: Liu Yongmou (Professor Wu Yuzhang of China Renmin University)]In recent years, the AI technology represented by the big model has developed rapidly, setting off a wave of AI development upsurge sweeping the world. People who pay attention to the development of AI are not limited to AI developers, promoters, critics of AI development, humanities and social science researchers, but also include the general public who deeply feel that life will be deeply affected by AI. The issue of AI development is no longer a purely technical issue, but a public issue in a certain sense. This is very clear in a series of related events such as the recent release of Sora by OpenAI and the open source Grok by Musk. In various related public discussions, the current situation of AI development is particularly concerned. The basic question is: Is there any problem in the general direction of AI development at present, and what direction should it go in the future? To this end, eight scholars in the fields of computer, economics, Marxist theory and philosophy were organized to reflect on the current situation of AI development in an interdisciplinary way, in order to attract the attention of the experts.
There are 8 articles in this series, reprinted from science economy society, No.2, 2024, and this article "Rational View of the Short-term and Long-term Effects of AI" is the fifth. In the article, Chen Yongwei thinks that AI is a new general-purpose technology, and it can be predicted that AI will have a significant impact on social and economic development in the long run, but the development of AI is still in the "sowing stage".
Since OpenAI released ChatGPT in November 2022, the AI field has ushered in a new round of development boom. In just over a year, various AI models have sprung up, and their capabilities have been rapidly upgraded and iterated. According to the evaluation report released by Microsoft, the GPT-4 model developed by OpenAI has been similar to human beings in many fields, even higher than human intelligence level, thus showing the "spark of general artificial intelligence". Recently, Claude-3 model developed by Anthropic showed better performance than GPT-4 in evaluation, and reached 101 points in IQ test, which exceeded the average level of human beings. In addition, in terms of creative work, AI also performed very well. For example, the Sora model newly released by OpenAI has been able to generate a video with a quality comparable to CG animation in one minute according to simple prompts.
Facing the rapid development of AI technology, how much impact AI will have on economic and social development has become a concern. At present, there are two completely different views on this issue. One view is that the influence of AI will be subversive, which will not only revolutionize productivity, but also completely change the organizational form of social economy. Of course, as a force of "creative destruction", it will also bring many negative effects. For example, the large-scale replacement of labor by AI may cause serious technical unemployment. Another view is that the influence of AI may not be as great as imagined. People who hold this view point out that there have been several rounds of AI development upsurge in history. In each round of upsurge, people have made high estimates of the impact of AI, but the results afterwards prove that these expectations have not been realized in the end. In the recent round of AI craze, although many AI models have shown excellent performance, their impact seems to be limited to a few narrow fields. From the perspective of the whole society, their positive and negative effects are relatively limited.
So, which of the above two views is correct? To answer this question, we need to have a clear understanding of the nature of AI technology.
1. AI is a general-purpose technology.
Economists who study the impact of technology on economic development often divide technology into two categories: "specific purpose technologies" and "general purpose technologies". The former category of technologies can only be applied to a specific field and generate value in this field; The latter technology can be used in many fields and produce value in these different fields.
Historical experience shows that compared with special purpose technology, the impact of general purpose technology on economic development is often more essential. For example, in the book "Economic Transformation: General Purpose Technology and Long-term Economic Growth", Lipsey and others have investigated the role played by general purpose technology in history. They analyzed 24 kinds of general-purpose technologies, including printing, steam engine, electric power and internal combustion engine. The results show that all these technologies have had a great impact on long-term economic development and changes in economic structure. From this, it can be inferred that if AI is also a general-purpose technology, it is likely to have a very significant impact on the whole economy and society in the long run.
Bresnahan and Tretenberg once put forward a standard for judging general-purpose technology. According to this standard, a technology can only be called general-purpose technology if it meets the three characteristics of universality, improvement and innovation spawning. Universal applicability means that this technology can be used as an input and applied to many different departments; Progressiveness means that the performance of this technology will continue to improve over time; Innovation gestation means that this technology can stimulate related innovations in many different departments, thus driving the technological development of these departments. By applying this criterion, we can test whether AI belongs to general-purpose technology. Next, we will verify the three features in the standard one by one.
(1) Universal applicability. As a technology, AI obviously meets this standard. After the "deep learning revolution", AI technology has entered many industries. Whether it is factories, shops or financial institutions, people widely use AI to assist decision-making. Especially after the rise of generative AI represented by GPT, the threshold for people to learn and use AI is further lowered, which makes the application of AI have more imagination. In reality, many individuals and enterprises have applied generative AI to their work and life. In August 2023, McKinsey conducted a survey on the AI usage rate of enterprises, and the results showed that 55% of the interviewed enterprises had already used AI. It can be seen that AI is indeed a universally applicable technology.
(2) progressiveness. The advancement of AI technology is manifested in three different aspects. The first is the rapid improvement of the task processing ability of the AI model. Take the GPT model as an example: when GPT-1 was launched in 2018, its ability was very limited, and its ability to analyze words logically was very weak. However, only five years later, GPT-4 has shown great ability in semantic recognition, logical analysis and problem solving. Microsoft researchers once asked GPT-4 to take the American Bar Examination. As a result, it got a high score of 298 points, exceeding 90% of human candidates. The second is the breakthrough of multi-modal ability of AI model. More than a year ago, AI models usually could only handle single-mode information, such as GPT-3.5, which could only handle text information. However, in just over a year, multi-mode has become the standard of AI models, and mainstream AI models such as GPT-4, Claude-3 and Gemini have the ability to handle multi-mode information such as text, images and videos. The third is the decline in the cost of the AI model. With the continuous optimization of AI model, the unit cost of training and using the model has also dropped significantly. Take the API calling GPT model as an example, the cost per Token has dropped by 90% compared with the past. It can be seen that in the past period of time, the technological progress of AI model is very obvious. It can be predicted that this progress will continue as long as there is sufficient computing power to support it.
(3) innovation is pregnant. Now, AI has become an important working tool in some fields, which has played a significant role in optimizing workflow and promoting technological progress in these fields. In turn, technological progress in these fields has also promoted the progress of AI, thus forming a virtuous circle. At present, the case that best embodies this virtuous circle is the design of AI chip. Because AI chip needs to arrange hundreds of millions of electronic components on a small silicon wafer, its design accuracy requirements have far exceeded the ability of traditional human design, so many companies including Google and NVIDIA have used AI to assist chip design, and have achieved a lot. Conversely, these chips designed with AI can provide more powerful computing power and provide a good foundation for the training and application of AI models.
To sum up, as a technology, AI fully meets the three characteristics of universal applicability, progressiveness and innovation, so it can indeed be recognized as a general-purpose technology. According to historical experience, it can be speculated that it will have a great impact on the whole economy and society in the long run.
Second, why is the influence of AI not as great as expected?
The question now is: since theoretically, the impact of AI will be enormous, why does it still seem to be very limited so far? To understand this phenomenon, we need to start with the characteristics of general purpose technology.
General purpose technology may have a great impact on the whole economy and society because it can penetrate into different fields. Therefore, the extent to which its influence can be exerted depends on the degree of its integration with various specific fields, which usually requires the support of relevant infrastructure and the efforts to combine general technology with professional applications. In their research, Helpmann and Trettenberg once divided the influence of general purpose into two stages: "sowing" and "harvesting". In the "sowing stage", the related infrastructure supporting general-purpose technology has not been popularized, and its integration with related industries is relatively shallow. At this stage, general-purpose technology will not immediately lead to a significant increase in productivity, and its impact on society is relatively small. In the "harvest stage", with the completion of relevant infrastructure and the deep integration of technology and related industries, the power of general-purpose technology will emerge in generate.
Because the influence of general-purpose technology has such a stage, there is usually a long time interval from the appearance of a new general-purpose technology to its real impact. This point has been proved many times in history. Watt completed the improvement of the steam engine in 1765, but by about 1830, the total power of the steam engine in Britain was still only 166,000 horsepower, accounting for only 1.5% of the total power in the country at that time. It was not until 1850 that the appearance of turbine steam engine made the steam engine truly become an efficient and cheap power source, and thus it was widely used. Similarly, at the end of 19th century, people mastered the generation and transmission principle of alternating current. However, for a long time, it had little influence on the United States, the country where it was invented. It was not until 1915 that the national power grid was built one after another and various electrical appliances were developed one after another that the influence of electricity on the American economy was reflected.
By comparing the current development stage of AI with historical experience, we can easily find that, as a new general-purpose technology, the current development of AI is still in the "sowing stage" as a whole. Although on the surface, AI has penetrated into all walks of life, on the whole, their integration with the industry itself is not deep. For example, although some industrial enterprises claim to have used AI models in their work, in fact, these AI models are only used to help solve marginal work such as customer service, and have no substantial impact on their core business. At the same time, the infrastructure related to AI, such as computing center and vector database, is still very inadequate, which restricts the influence of AI to a certain extent. It is under the combined effect of these factors that the impact of AI on the economy and society is far below people’s expectations.
Third, policy thinking to promote the potential of AI
Based on the general-purpose technical attributes of AI, in order to give full play to its potential influence on the economy and society, it is necessary to help this technology quickly realize the transfer from "sowing stage" to "harvesting stage" through policy means. Specifically, the following aspects are the most worthy of attention.
(A) Need to speed up the construction of infrastructure related to AI
For AI, there are two types of infrastructure that are the most important.
The first type of infrastructure is computing power. Whether it is training or using AI model, it needs to call huge computing power. In deep learning, even if there is no progress in the algorithm level, the performance of the model will be greatly improved with the increase of parameters and training data, all of which require computing power as the bottom support. When OpenAI was training GPT-3 model, it used tens of thousands of NVIDIA A-100 GPUs. It is under this huge amount of computing power input that the performance of GPT model has made a qualitative leap, and even the phenomenon of "emergence" of intelligence has appeared. However, in the market, companies that can mobilize huge computing resources like OpenAI are very rare. Most enterprises and individual developers can only rely on computing resources in the cloud to train AI models. Although many cloud service providers can provide AI computing power, the accessibility of AI computing power is still low and the use cost is still high in the absence of basic hardware such as GPU. In this case, it is still an urgent problem to strengthen the research and development of hardware such as GPU and increase the supply of AI computing power in the cloud.
The second type of infrastructure is storage facilities. The great development of AI will drive huge storage demand. A considerable part of the data to be stored is unstructured data. In order to better store, organize and retrieve these data, it is necessary to have new storage facilities, such as the support of vector database. At present, there is still a great shortage of such storage facilities, so in order to promote the development of AI, it is still necessary to support their construction to a certain extent.
(B) should do a good job of coordination in the process of AI technology diffusion
In the process of developing and popularizing general purpose technology, many different subsystems are usually involved. However, the industry standards between these subsystems are usually inconsistent, and their development speed is also uncoordinated, which is likely to have a negative impact on the diffusion of technology. In this case, it may play a significant role in promoting the diffusion of technology by coordinating related subsystems with policies.
Specific to the current development of AI, the basic principles of the models are all different, and they lack interoperability with each other, which limits the coordinated development of different models to a considerable extent. In this case, we should actively cultivate an open source atmosphere and encourage developers to open source some basic technologies. At the same time, policies should be used to ensure the unification of some key technical standards and data interaction standards. In this way, when a developer makes a breakthrough in an AI technology field, this technological breakthrough can quickly spread.
(C) We should actively promote the deep integration of AI technology with specific industries.
Agrawal and others put forward two concepts when analyzing the integration of AI and industry: "single point solution" and "system solution". Among them, "single point solution" refers to using AI to solve a specific problem, but keeping the overall workflow unchanged. For example, replacing part of customer service or word processing with AI in reality is a single-point solution to the problem. In contrast, "system solution" refers to redesigning the whole process with AI. For example, some logistics enterprises began to abandon the original human-centered scheduling, and instead adopted AI as the scheduling center, and rearranged the relevant positions accordingly, which is a manifestation of systematic solution.
It is easy to see that in the "single point solution", the degree of integration between AI and the industry is shallow. Although it can bring some efficiency improvements, the range is usually not too large; The "system solution" requires deep integration between AI and industry, which can greatly improve production efficiency. Therefore, in order to give full play to the role of AI, it is necessary to actively promote the integration of AI and industry from "point solution" to "system solution".
It should be noted that when enterprises use AI to systematically change the production process, they not only need to invest a lot of costs, but also may face considerable transformation risks. Under this circumstance, we can consider using policies to provide some support, provide some subsidies and help for the AI transformation of enterprises, and make their transformation smoother.
If the above points can be achieved, then AI, a general-purpose technology, can complete the leap from "sowing stage" to "harvesting stage" faster, and its potential can be better stimulated.
Fourth, on the problem of "technical unemployment" caused by AI
It is worth noting that historically, the development of any kind of general-purpose technology has promoted the development of social productive forces, brought people a lot of benefits, but also brought many related problems. This is no exception for AI. In fact, at this stage, many AI-related problems have begun to emerge, and among these problems, "technical unemployment" is undoubtedly the most noteworthy.
As early as 2013, scholars at Oxford University made an estimate that 47% of jobs in the United States will be replaced by AI within 20 years, involving tens of millions of employed people. After the explosion of ChatGPT, researchers of OpenAI made a similar estimate, and thought that 10% of the 80% labor force’s work tasks in the United States might be affected by ChatGPT, and 19% of the labor force’s work tasks might be affected by ChatGPT. These research conclusions have aroused great concern in the society and caused a certain degree of AI panic.
In my opinion, we should treat the above research rationally. Historical experience tells us that although technology may destroy certain jobs, at the same time, it usually creates more jobs. For example, the development of automobiles eliminated the position of coachman, but created more positions of drivers. Therefore, from a long-term perspective, we don’t need to be too afraid of "technical unemployment". But in the short term, because the employment skills between the old and new jobs do not match, a short-term unemployment shock is indeed possible. For example, the development of AI can make people no longer need a large number of illustrators, but it will give birth to the demand of more high-end AI engineers. On the whole, the increase or decrease of positions may be balanced. However, it is obviously impossible for illustrators to become AI engineers immediately. At this stage, there may be some unemployment. Especially when AI and the industry begin to deeply integrate and make systematic changes to the industry, the unemployment rate may be higher.
At present, technical unemployment caused by AI has begun to appear in some industries. In order to cope with this impact, relevant policies must be actively involved. These policies include but are not limited to: strengthening the construction of vocational training and re-education system, encouraging service industry, encouraging the development of odd-job economy and sharing economy, etc. Only by doing these jobs well can we not panic in the face of unemployment caused by AI.
V. Conclusion
Bill Gates, the founder of Microsoft, famously said, "We always overestimate what we can do in one or two years and underestimate what we can do in five or ten years." This sentence is quite appropriate to understand the influence of AI.
As a general-purpose technology, the technical attributes of AI determine that it may have a great impact in the long term, but its impact may not be obvious in the short term due to various factors. In this case, we should actively use relevant policy means to promote the positive impact of AI technology; At the same time, using the current time window, various coping mechanisms are established for its possible negative effects. If we do this, we can foster strengths and avoid weaknesses, so that AI technology can better benefit people.