Lack of Korean model to combat global AI raises industry concerns

[Courtesy of Naver]

As global big tech companies deepen their monopoly over artificial intelligence (AI), there are growing concerns that South Korea’s information and platform industry could be subordinated to them unless the government, companies, and academia join forces to focus on investment and research capabilities in this field.

According to sources on Sunday, only five South Korean companies have succeeded in developing super-scale AI foundation models. They are Naver Corp., Kakao Corp., KT Corp., SK telecom Co. and LG Electronics Inc. A foundation model is the basis of AI that learns from inputs such as text, images, voice and video and generates new data.

“The cost of developing super-scale AI is astronomical,” said Kim Jin-hyung, a professor emeritus at Korea Advanced Institute of Science & Technology’s computer Science department. “The key is the language model and it is not easy to compete with global big techs, which have secured economies of scale based on their overwhelming capital and technology.”

Naver is the most advanced Korean company in the development of super-scale AI. In 2021, the company developed and unveiled HyperCLOVA, a super-sized AI, for the first time in Korea, utilizing its own supercomputer and data in the local partner ecosystem.

Naver built an infrastructure for processing large amounts of data by using a supercomputer with 700 petaflops (1 petaflop is a unit of computer performance capable of processing 10 trillion operations per second). HyperCLOVA X, which will be unveiled in July, was developed with 240 billion parameters, exceeding 175 billion of GPT-3. Naver is also planning to introduce ‘SearchGPT’, an AI chatbot on HyperCLOVA, in the first half of this year.

Kakao plans to release an upgraded version of Ko-GPT, a super-sized AI language model, through its subsidiary Kakao Brain in the second half of this year and introduce an AI chatbot service that responds to OpenAI’s ChatGPT as early as the third quarter of this year.

When ChatGPT, based on GPT-3.5, was first released for public use, the Korean AI industry and authorities were patting themselves on the back, saying that a homegrown AI model would be better than that.

However, the overall perception changed when GPT-4 released in March was found to outperform ChatGPT even in Korean language. According to data released by OpenAI, the Korean accuracy rate of GPT-4 is 77.0 percent, which is higher than the English accuracy rate of GPT-3.5.

This effectively shattered the perception that the Korean language can be a shield for domestic companies to compete in the super-sized AI competition.

In addition, there is a growing concern that Korean companies will become dependent on overseas AI technology if they do not have a foundation model with differentiated technology.

OpenAI already started monetizing ChatGPT by launching a $20 per month subscription product, and recently applied for trademark rights for its service.

A bigger problem is that by relying on overseas AI, the data produced by domestic users may be absorbed or subordinated to overseas AI source technology companies.