Stanford University Institute for Human-Centered AI's AI Index project director Sha Shaqadee recently stated in an interview that for South Korea to achieve its goal of becoming one of the world's top three AI powers, it should not directly compete with the United States and China in the realm of ultra-large general-purpose models. Instead, it should focus on developing industry-specific models.
Sha Shaqadee pointed out that South Korea has clear competitive advantages in per capita AI patents, talent cultivation, semiconductors, robotics engineering, and hardware, meeting the conditions required to lead in the AI ecosystem. However, he also noted that South Korea has yet to translate these strengths into top-tier model competitiveness globally. While Korean-specific models are being launched, they have not yet reached the upper echelons of global benchmark tests.
South Korea Needs to Focus on Developing Industry-Specific Models
Sha Shaqadee believes that the biggest obstacle South Korea needs to overcome is scale. The United States and China dominate the competition in ultra-large general-purpose models thanks to their vast domestic markets and massive investments, making it difficult for South Korea to follow the same path. He suggests that South Korea leverage its strengths in software, semiconductors, and hardware to develop specialized models that excel in specific industries such as manufacturing, science, and public services.
Regarding the South Korean government's push for national AI foundation model projects, as well as projects for universal AI and cybersecurity-specific AI models, Sha Shaqadee offered a positive assessment. He emphasized that the concept of sovereign AI is broader than simply owning a domestic foundation model, and should encompass infrastructure, data, models, applications, and talent. South Korea has already met many of the conditions for achieving this, and the next step is to integrate domestic AI models with infrastructure, data, services, and talent into a complete ecosystem.
On the issue of model evaluation, Sha Shaqadee noted that current benchmarks have limitations, including the lack of transparency in training data affecting the credibility of evaluations, and the rapid loss of discriminative power in benchmarks. He suggests adding predictive assessments to the evaluation metrics, which would involve predicting the performance and risks of a model before it is deployed in real-world scenarios, while also balancing the transparency of evaluations and mitigating the risk of companies gaming the test questions.
Sha Shaqadee also revealed that the AI Index plans to strengthen indicators in areas such as physical AI, world models, AI education literacy, and economic and employment effects. He stated that in the field of physical AI, it is necessary to measure the ability of robots to perform tasks in real environments, and to examine accident liability and regulatory systems, while also considering the inclusion of hardware and robotics education levels in the assessment scope. Additionally, the ability of ordinary citizens to use AI effectively and responsibly should become a new object of evaluation.