Troy Stangarone is the senior director of congressional affairs and trade at the Korea Economic Institute.
Integrating AI into Korea's economy
Since the debut of ChatGPT in late 2022, generative AI has been viewed as a key component for economic growth and technological development. How best to integrate this new technology, however, is the key challenge faced by companies and countries alike.
Despite the hype behind AI, companies have yet to fully adopt generative AI. With any new technology, it takes time for companies to implement and best utilize the technology. However, generative AI comes with additional challenges. It has a tendency to “hallucinate,” meaning to make up data, raising questions about its reliability. In a New York legal case, for example, AI created a fictional legal precedent in the brief it generated.
For a country like Korea, the challenge could be even more difficult. The large language models that underpin generative AI models such as ChatGPT require enormous amounts of data to train. While ChatGPT has used publicly available data on the internet, it and other leading AI companies have also worked to integrate proprietary data into their models. It is unclear whether Kakao or Naver can compete with Google, Meta or Microsoft, which supports ChatGPT, in a competition for data.
It is also unclear if building ever-larger large language models is a viable path forward for AI. The world’s most advanced generative AI models have already absorbed much of the world’s available data. Firms are now increasingly turning to synthetic data — data created by AI itself — to fill the data gap. However, a recent study in Nature found that because of the compounding of errors from using synthetic data, AI systems risk collapsing if trained on increasingly larger amounts of synthetic data. Without a new technological breakthrough, generative AI based on large language models will face limits on its development.
On a more basic level, companies face a cost-benefit calculation. Generative AI is generally presented as a way to address a wide range of tasks. However, as Erik Hoel, an American neuroscientist, has pointed out, because large language models need immense amounts of data to train, they are best at skills where data is plentiful. These also tend to be low value-added areas. From a corporate perspective does it make sense to replace low-wage workers with higher-cost technology for skills that are widely available? This suggests for generative AI to be successful it needs to begin providing higher-value skills or improve the productivity of workers in higher-level tasks. This basic challenge could provide Korean firms and Korea more broadly a path forward for integrating AI into the economy.
Rather than viewing generative AI as a means to replace workers, a better approach is asking how AI can help workers become more productive and avoid the hallucination and potential collapse issues related to large language models. This involves focusing on more specialized models.
Specialized AI models tend to have lower levels of hallucinations. The data used to train the models is higher quality and often possesses fewer errors than consuming publicly available data on the internet.
Google’s DeepMind provides an example of how specialized models can produce higher value data. Last year, it discovered 2.2 million potential new crystalline structures, of which 380,000 are considered the most stable. By traditional methods, this would have taken 800 years. These new crystalline structures could provide breakthroughs to improve semiconductors or EV batteries. This approach would also help Korean firms maintain a competitive edge in these industries. More broadly utilizing AI to improve or develop new products would enhance productivity in ways that would justify significant investments in AI.
Korea’s own demographic challenges may also give it an advantage in developing advanced AI models. One promising area for developing AI is robotics. Because robots learn from their interaction with the real world, they produce their own data for learning and have fewer limits on consumable data. They are only constrained by their interactions.
Declining birthrates have pushed Korean firms to integrate more industrial robots into production. Integrating AI with industrial robots could provide Korea with a way to develop high-value advanced AI, but also aid in improving the productivity of small and medium-sized firms (SMEs). Based on data from 2019, SMEs are only 32 percent as productive as large firms in Korea. Boosting productivity in SMEs would significantly improve Korea’s overall economy.
Korea already plays an important role in the advancement of AI due to the development of high-bandwidth memory. Without these chips, Nvidia’s graphics processors would not be able to power AI models to the degree they have. However, for Korea to move beyond being a hardware partner for AI and to integrate the technology more broadly into its economy, it will need a different approach focused on specialized models for research and development, along with integration into robotics.
Troy Stangarone is the director of the Hyundai Motor-Korea Foundation Center for Korean History and Public Policy and the deputy director of the Indo-Pacific Program at the Woodrow Wilson Center.
Troy Stangarone