SOUTH KOREA
South Korea has long been known for its innovative university-industry collaborations. Now, global developments in generative AI – most widely represented at the moment by ChatGPT – are prompting new ways for the country’s universities to work with South Korea’s own technology giants.
Rather than simply adopting ChatGPT, which was developed by California-based OpenAI, or other ‘universal’ generative AI technologies, which do not take into account local and regional languages, culture and political sensitivities, South Korea, like some other countries, is looking to produce its own versions of AI technology and generative AI in particular.
The search for ChatGPT alternatives is also motivated by reasons of ‘digital sovereignty’, to ensure a country has more control over its own data.
“We worry about technology dependency on big tech — on OpenAI and Microsoft and Google —because they own these large language models (LLMs) which will revolutionise the entire AI ecosystem,” Song Chong, head of the Kim Jaechul Graduate School of AI at Korea Advanced Institute of Science and Technology (KAIST), told University World News during an interview in Seoul.
KAIST – which ranks among the top five in Asia for producing machine learning conference papers in the past five years (behind only a few major universities in China) – and the Korean government research agency are putting funds into developing local LLMs in collaboration with big local tech companies such as: Naver, the Korean internet search engine; Samsung Electronics; KT (Korea Telecom); and LG – thereby spurring innovative forms of collaboration.
LLMs use statistical models to analyse vast amounts of data, from the internet and books, to understand patterns in language in order to generate AI-driven text. “We need to have our own LLMs, trained using more Korean language knowledge and Korean knowledge. But in addition, it will be useful worldwide,” Chong said.
Overseas markets
In April the South Korean government announced plans to invest about KRW563 billion (US$424 million) in the country’s AI industries and software development. The aim of the funding from the Ministry of Science and ICT is to strengthen the digital foundation of the country’s industries to facilitate their expansion overseas.
Some of the money will go into AI graduate schools, such as KAIST’s, to train 200,000 advanced software and AI professionals by 2027 to prepare for generative AI technologies such as ChatGPT to be rolled out in major industrial and professional sectors.
As well as improving Korean language generative AI applications, the aim is to build 200 types of specialised learning and language data models for non-English speaking markets.
Naver has already said it will be able to offer overseas versions of its latest LLM to foreign countries who may be concerned about Microsoft and Google’s focus on ‘universal’ AI language models.
If Naver is successful in this ambitious task, potential clients include countries in the Middle East with certain political sensitivities, and non-English speaking regions such as Latin America where governments are seeking AI systems based on their own cultural, language and political contexts.
Some countries are also concerned by new US government regulations under an executive order issued by President Joe Biden on 30 October which would require the vetting of AI models developed in the US before they are released.
For some governments, the vetting process represents a potential national security risk and confirms the need for alternatives to US-developed generative AI systems. But the order could also prevent Microsoft and Google from becoming even more dominant globally than they are now as they look to develop non-English language LLMS.
Naver’s CEO Choi Soo-yeon earlier this year expressed concern about the impact of Microsoft and Google’s AI tools on all aspects of life, describing it as “a threat, the likes of which we have never experienced before”.
While regarded as well behind the US and China in AI development, South Korea is hoping to ride these concerns towards becoming the world’s third AI power.
Home to two of the world’s largest memory chip makers (Samsung and SK Hynix) the country already has advanced memory chip capabilities required to fuel AI systems.
Korean search engine Naver has stepped up the development of generative AI services, initially focusing on South Korea and Japan, but as part of its bid to provide international services in languages other than Korean, it has also inked a preliminary agreement to provide AI tools for Saudi Arabia’s transition to a digital economy.
Joint research
A joint research centre was set up three years ago known as the KAIST-Naver Hypercreative AI Research Centre in response to developments in generative AI including LLMs.
The centre is based on the idea that huge advances in generative AI could lead to AI, at some point in the future, creating its own movie out of an existing book. “That’s the long-term goal but it is still LLM-based because it should be able to understand human language,” Chong explained.
“We call this ‘hyper creative’ AI. Around a hundred of our graduate students are working at Naver in this LLM area,” he said, noting they are working on almost a dozen different projects at the company, which has huge amounts of data from its online searches, online shopping, travel, videos and webtoons.
Like other universities in the country, KAIST does not have the computing power that big tech companies have, making partnerships essential.
“Computing [power] is important in AI training. But the university cannot afford that. That’s why we did this strategic partnership. We provide human resources for research and we can use their company data,” Chong explained.
“Some data cannot be taken out of their [Naver’s] campus. So, we are working within the company, using their computers, and using their data,” Chong said.
Intense collaboration
It is a new form of intense collaboration, and not just because of the large number of KAIST graduate students involved.
“Usually, when universities work with industry, students and professors work at the university and deliver the research results to the industry. But this is a different approach. We are working as a team at Naver’s campus,” he said.
According to Chong, a similar collaboration is in place with KT on LLMs, indicating how keen companies are to work with institutions such as KAIST.
“Tech companies have their own research. But unless they are big companies like Google or Meta (formerly Facebook), Microsoft or OpenAI, we still have a better research capability. Top tier universities have more high-quality researchers than industry research labs,” Chong said.
“Even at Google – most of the research is done by their own research people but they still collaborate with universities on some topics,” Chong said.
Rapid developments in AI have meant more ‘strategic’ partnerships with companies, including Samsung, rather than partnerships forged on a project-by-project basis, Chong said. For KAIST, these include research collaborations with Samsung semiconductor division’s business unit, and Samsung DSC (device solutions) which involves more than 15 products.
The strategic partnership with Samsung covers some 10 research projects, including how to apply semiconductor chip design or AI circuits to the AI system for such devices as Galaxy mobiles and appliances.
“The research topic is ‘AI for digital experience’, useful for mobile devices, appliances, TV and so on,” he said.
“KAIST has top-tier people. So eventually they [Samsung] will want to hire them, and they also want to have their own solutions.”
Universities’ need for data
Universities have their own motives for collaborating. “In this LLM era, institutions need the data,” Chong said. LLM’s accuracy relies on huge amounts of data.
“Universities don’t have the data, and some data cannot be made public. In many cases data must remain within the company. That’s why, for LLMs, this new form of collaboration at Naver makes sense,” Chong explained.
But universities like KAIST also need to conduct their own research.
“The AI area is fast-changing, particularly with the development of LLMs, but academia still has its own role. We cannot just follow their [big tech companies’] movements,” said Chong.
“We still need to work on fundamentals because LLMs are not perfect. They still have a lot of issues to overcome. That’s why we’re still running the basic and existing curriculum in AI here at KAIST, but with new classes on LLMs and generative AI,” Cong said.
Training students in the ethical requirements expected by society and research communities, as well as identifying bias, training in data privacy and protection, misuse of AI to harm people and society and other impacts on society, have been included in graduate courses.
Towards a global AI powerhouse
In its digital strategy the government says it hopes to make the country one of the world’s top three AI powerhouses by 2027. Though South Korea lags behind a number of countries in the size of its AI talent pool, the government wants to position the country globally as an alternative to US big tech.
“I cannot say that we are a leading country [in generative AI] right now, in the sense that we haven’t generated new knowledge; we’re using the knowledge generated by other leading countries. So, our challenge is to move from knowledge consumer to knowledge creator,” Chong noted.
Whether or not South Korea can achieve its ambitious goal, generative AI nonetheless presents a huge opportunity for research and university-industry research collaboration.
According to a report released in August by the Korean Chamber of Commerce and Industry, Microsoft, and Access Partnership (a London-based technology-focused public policy consultancy), generative AI has the potential to unlock South Korea’s productive capacity by up to US$476.6 million, equivalent to 29% of the country’s gross domestic product in 2022.
The study noted generative AI is likely to have a more significant potential impact in industries dealing with large-scale data or complex tasks. It also predicted that as it becomes widely applied across various industries, generative AI will affect working practices. Approximately 67% of all workers in South Korea will use generative AI in 5-20% of their job activities in five years, the study estimated.
These include the insurance, professional science and technology industries, information industry, wholesale, healthcare and welfare.
Chong said the government has an ambitious plan to roll out generative AI to “all industries” that is currently waiting to be approved by the National Assembly. Researching generative AI applications for the different industries will be key research topics for institutions like KAIST, he concluded.
