On December 22, 2022, the Peking University Global Health and Development Forum 2022 was held with the main theme of Digital Transformation and Development Divides. Co-organized by the Beijing Forum, Asian Development Bank and PKU Institute for Global Health and Development, this Forum brought together world leading scholars, policy researchers and industry leaders from both China and international communities to share their insights and recommendations on the thematic topics, attracted over 10 thousands online viewers participated in the event. Erik Brynjolfsson, the Jerry Yang and Akiko Yamazaki Professor at Stanford University and Director of Stanford Digital Economy Lab delivered a keynote speech at the session of Digital Transformation in Healthcare.
The topic of my discussion today is what I term the "AI Awakening." This refers to the significant advancements in artificial intelligence that have occurred in recent years, particularly in the realm of large language models and foundational models. I will also discuss the implications of these advancements for the future of work, health, and development.
One way to chart this progress is through image recognition. In 2014, when I wrote my book "Second Machine Age," I mentioned that machines were not particularly proficient at recognizing faces and images. However, this was already beginning to change. The graph on the right illustrates the progress made in deep neural nets, as led by individuals such as Jeff Hinton, on image datasets such as ImageNet. On the left, you can see some examples from ImageNet, and the purple line illustrates the rapid growth in accuracy, nearing 100% on this database of approximately 14 million images.
Similar technologies are now being applied to medical images. For example, Sebastian Thrun has applied them to dermatology images, resulting in performance that surpasses that of human dermatologists. Radiologists are also finding that machines can assist with a wide range of tasks. When machines surpass human abilities in a particular area, it becomes a threshold effect. Entrepreneurs, hospital administrators, and consumers may choose the machine solution over the human solution once this threshold is crossed. It is akin to a phase change.
In addition, there has been considerable progress in the field of drug discovery. With the COVID pandemic, there was a surge of activity, which we tracked at the Artificial Intelligence Index, an organization I co-founded to track data on AI. If you visit the AI Index, you will find a wealth of data that you can use for your own research and analysis.
Perhaps the most notable advancements in recent years have occurred in the area of large language models and foundational models. For example, I wrote some comments on an NBER economic paper and then asked GPT-3, a large language model, to write comments in the style of the pop singer Taylor Swift. Within seconds, GPT-3 generated a song that was not only grammatically correct, but also demonstrated a level of sophistication in language that was comparable to, or even superior to, that of the average college student.
Similarly, there has been rapid progress in generative adversarial network scans that generate faces and other images. Recently, a popular podcast host, Joe Rogan, conducted an interview with Steve Jobs, even though Steve Jobs is no longer alive. The interview was highly plausible, with the same kind of voice intonations, comments, and insights that Steve Jobs might have made. Both the interviewer and the interviewee were completely artificial and generated by machines.
We are also witnessing rapid progress in the mastery of different types of strategic games, from games with complete information to games with incomplete information, where players do not know what the other players know. Even in games such as Diplomacy, where players must negotiate in English, make suggestions to others, and reach agreements, machines are performing exceptionally well in tournaments. This speaks to the level of progress that has been made.
In summary, we are witnessing a rapid improvement in the ability of machines to perform tasks that were previously only possible for humans. One way to track this progress is through a website called Metaculous, which gathers the predictions of experts from around the world on different AI-related topics. It is an exciting time as we continue to witness the rapid evolution of AI, and the potential it holds for shaping the future of work, health, and development.
The underlying factors driving the rapid advancements in artificial intelligence are the simultaneous dramatic increases in computer power, larger data sets, and more parameters. Each of these factors has grown by orders of magnitude and, interestingly and importantly, each time there is an order of magnitude improvement in any one of these factors, it leads to a comparable improvement in the quality of the system. This trend fits a power law quite well, which explains why we are seeing so much progress in this field.
To summarize the current state of technology, I would say that while the pace of change has never been this fast, it will never be this slow again in the future. The technology is continuing to accelerate, which has a lot of practical implications. For example, there is currently a "gold rush" in many fields, such as pharmaceuticals, R&D, medical imaging, and speech recognition, as businesses apply these underlying technologies. However, this also has important implications for work and the economy.
As an economist, my headline observation is that this digital progress is making the economic pie bigger by creating opportunities to solve many problems - economic, health, poverty, and environmental - that we were previously unable to solve. However, it is also important to note that there is no economic law that dictates that everyone will benefit evenly from these advancements. It is possible for the majority of gains to go to a small group of people, with many others not only failing to benefit, but actually being made worse off in absolute terms. Historically, progress has tended to lift most boats, but recently there are many people who are being left behind by these technologies, leading to dislocations and inequalities in various dimensions.
In terms of how technology is affecting work, one of the ways that it can have an impact is by substituting for human labor, which can drive down wages. However, technology can also compliment human capabilities, allowing them to do new things that they were previously unable to do. When technology complements human labor, it tends to raise wages. Historically, technology has primarily complimented human labor, but it is possible to design systems that can increase wages and create more widely shared prosperity, as well as systems that concentrate wealth and reduce wages for most people.
In order to understand which tasks are most likely to be affected by technology, we developed a Suitability for Machine Learning Rubric, which categorizes tasks based on whether they are well-suited for machine learning or not. By applying this rubric to 18,000 distinct tasks that are performed in the US economy, we were able to identify which parts of the economy were more likely to be affected and which were less likely to be affected.
For those interested in healthcare, as an example, let's take a closer look at radiology, one of the 950 occupations we studied. In radiology, there are 27 distinct tasks being performed by radiologists. We've listed the first seven tasks on the next slide. One important task that's likely to be affected by machine learning is the detection of abnormalities in medical images. We've also found that tasks that involve pattern recognition and data analysis are more likely to be affected than tasks that require medical knowledge, communication, and empathy.
In regards to the impact of artificial intelligence on the workforce, it is important to note that while advancements in technology have the potential to greatly affect the way work is done, it is not a case of mass unemployment or mass replacement of human labor. Rather, it is a case of mass restructuring and rearrangement of tasks within various occupations.
We have analyzed 950 different occupations and found that in each case, there are some tasks that are suitable for machines to perform and others that are not. This means that in the case of radiologists, for example, some tasks will increasingly be done by machines, while others will continue to be done by humans, resulting in a rearrangement of the workload. This is the case for all 950 occupations studied. In the United States alone, there is an estimated $713 billion worth of opportunity for these changes to occur, with similar effects expected in China.
Furthermore, our research has found that the tasks that are most likely to be affected by these technological advancements are those that involve relatively low wages and a high ability to be performed by machines. However, with the emergence of large language models, this trend may change, with higher-paying jobs becoming more susceptible to automation.
In order to more specifically understand which tasks will be affected, we have developed tools for analyzing hundreds of millions of job postings. These tools allow us to understand which tasks are most likely to be affected by organizations and the economy as a whole.
It is important to note that while these tools can greatly affect the way work is done, there will be uneven effects on the economy. Ultimately, our choices as to how we utilize these technologies will greatly impact the outcome. The same tools that have the potential to create more widely shared prosperity and help people more broadly, can also be used to concentrate wealth and control people in greater ways.
It is crucial that we consider our values and how we want to use these technologies to shape the future. The more powerful these tools become, the more choices we have about the kind of future we want to create. I encourage everyone to think deeply about the values they want to see expressed in society and the economy and to learn more about the research by visiting my website at the Digital Economy Lab at Stanford, as well as my personal website, where the papers I referred to are documented. Thank you for your attention.
