Whereas the ChatGPT — a free chatbot system capable of generating conversational dialogue — displays a surprising level of sophistication in response to English language prompts, the Korean version of its service is lacking. And that might leave an opening for a niche market for local competitors.
On Tuesday, Microsoft said that it will incorporate Open AI’s GPT technology into its Bing search engine.
“This technology is going to reshape pretty much every software category that we know,” said Microsoft CEO Satya Nadella.
The updated version of Bing will have a conversational interface similar to ChatGPT. The goal is to have the unpopular search engine compete against Google.
ChatGPT, which first made debut in November last year, is a chat system of San Francisco-based Open AI’s language model GPT-3, or Generative Pre-training Transformer 3.
Two months into operation, ChatGPT garnered more than 100 million users worldwide. And the explosive popularity of ChatGPT has nudged Google to launch its rival service.
Google CEO Sundar Pichai Monday introduced the company’s upcoming “experimental conversational AI service,” named Bard, on the company’s blog.
Bard will run on Google’s LaMDA, which stands for Language Model for Dialogue Applications.
The sudden surge of interest in generative AI models fueled competition among big techs, and Korea’s major IT companies are also hoping to jump on the bandwagon.
What is noteworthy is that Korea is the only country in the world that Google failed to dominate the search engine market, save for China and Russia.
Naver operates a search engine website, and Kakao runs a messenger app. Korea’s two largest tech companies have been gathering data while running the country’s leading online services with overwhelming dominance in the market.
In 2021, Naver became the third company to introduce its own large-scale language model with HyperClova, following OpenAI’s GPT-3 and Huawei’s PanGu-α.
HyperClova processed 50 years of Naver new articles, according to the company. The amount of Korean language-based dataset that HyperClova learned is 6,500 times larger than that of GPT-3, said Naver.
In 2021, Kakao Brain, Kakao’s research subsidiary, introduced two large-scale AI systems based on Open-AI’s GPT-3 and Dall-E: KoGPT and minDALL-E.
Many IT companies share their algorithms and source code for computing models as an open source for others to use. Therefore, what makes a real difference in developing an AI-infused service is often the dataset that the software learns.
That is how Kakao, which has an extensive experience in managing large-scale datasets through their online services, has been quick to come up with AI-based services.
“Google could develop large-scale language model LaMDA and the Bard service because the company runs a huge search engine,” said Zhang Byoung-tak, director of Seoul National University’s AI Institute.
“Naver and Kakao have a strength in that sense,” said Zhang.
Collaboration in the tech industry
Data, algorithms and computing and network capabilities are considered the three factors that enable artificial intelligence technology. And that is why telecom companies, chipmakers and IT service companies are launching joint projects.
SK Telecom runs a GPT-3-powered virtual assistant app named A., while working with SK hynix and Sapeon to develop AI chips. KT is developing its own large-scale language model Mi:Deum and plans to release it in the market in the first half of the year in collaboration with Rebellion, a local chip designer.
Naver invest in Furiosa AI, an AI chip company, and also established a task force with Samsung Electronics for a AI chip project in December last year, to develop a chip that can power its HyperClova.
Local start-ups are coming up with new services based on computing models developed by big tech companies.
Wrtn Technologies developed Wrtn Training, a writing assistant tool, based on Naver’s HyperClova. Scatter Lab, known for Lee Lu-da chatbot, is using GPT-2 to develop a conversational chatbot.
“What matters the most for hyper-scale AIs is the quantity of data, as it needs overwhelming amount of data for machine learning,” said Kim Jin-hyung, a professor emeritus of computer science at KAIST. “Therefore it is important to be part of the ecosystem after speedily developing the service using open source.”
Korean companies believe non-English speaking countries may be a promising market for their services.
Kakao Brain hopes to release Vietnamese and Malaysian versions of its KoGPT targeting the Southeast Asian market, as well as English and Japanese versions. Wrtn Technologies plans to develop Japanese, French and Spanish versions of its service.
Experience in developing and running digital services can also be an asset for the companies.
“Companies need to secure the competitive edge with their own service idea that goes beyond the limitation of the previous models, instead of just emulating ChatGPT,” said Yoo Jae-yeon, AI researcher of Yellowdog, which is a local venture investment company.
The technology rivalry between the United States and China may also create a new opportunity for Korea.
“Korea has competitiveness in price compared to the United States, and in security compared to China,” said a founder of an AI start-up which operates globally.
Cost challenges
Korean companies are significantly lagging behind global big tech, in terms of technological level, financing capability and infrastructure.
GPT-3, which was released in 2020, had 175 billion parameters, and Microsoft and Nvidia’s MT-NLG has 530 billion parameters. Meanwhile, Naver’s HyperClova has 204 billion parameters and LG’s Exaone has 300 billion.
Heavy costs of cloud computing infrastructure remain a major obstacle as well.
Naver CEO Choi Soo-yeon said that Naver will introduce “SearchGPT” that specializes in queries in Korean within the first half of this year, yet admitted that “there are many challenges that need to be reviewed in terms of cost-effectiveness.”
“Hyper-scale AI development requires a massive amount of investment and human resources, and therefore is done mainly by companies rather than academic institutions,” said Ha Jung-woo, head of Naver’s AI Lab, and called for an active support from the government for research and development projects.
Evolving from information to insight
Computing capabilities of AI models improve at a steep pace, doubling by every six months.
This even outpaces Moore’s Law, which states that the number of transistors on a microchip, hence computing capability, doubles every two years.
Generative AIs such as ChatGPT are expected to bring a huge change in how people work by automating writing, summarizing and translating texts, especially since Microsoft will use OpenAI’s technology in its office software.
Bill Gates, founder of Microsoft, believes AI will lead to yet another paradigm shift, saying that “this is every bit as important as the PC, as the internet,” according to a Forbes report.
Gates also forecast that the emergence of generative AI would inevitably result in the fall of some white-collar jobs.
As AI will be a crucial part of everyday life in near future, the ability to utilize the technology is likely to be a must on both personal and professional levels.
“The future generation will grow up as an ‘AI native,’ that learns, converses, and plays using AI technologies,” according to a report from National Information Society Agency(NIA)’s AI Future Strategy Center.
But it may discourage people from trying to learn, warned Kim Tae-won, an executive principal researcher at NIA.
“Humans accumulate knowledge by learning from their own experience or someone else’s through media,” said Kim. “But AI service changes how a knowledge is defined, as it just skips the experiencing part of the learning process.”
Issues of students using ChatGPT in writing their essays or reports already became a hot topic.
Kim emphasized that digital education will gain importance in the era of AI, as “students need to be aware of data source of the generative AI system and the limitation of automatically generated texts.”
BY KIM KYUNG-MI, SHIM SEO-HYUN, KIM NAM-YOUNG [shin.hanee@joongang.co.kr]
