[AI IN ACTION] Korea Inc. dashes into AI world with own platforms

Home > Business > Industry

[AI IN ACTION] Korea Inc. dashes into AI world with own platforms

Published: 31 Oct. 2023, 06:00 Updated: 01 Nov. 2023, 18:45

JIN EUN-SOO
jin.eunsoo@joongang.co.kr

Korea Inc. dashes into AI world with own platforms [SHUTTERSTOCK]

For LG, the launch Exaone embodies its vision of pushing the boundaries of the much-hyped generative AI technology. The large-scale AI platform can go from being a helping hand for individuals to a key tool for high-valued tasks such as developing new drugs and granting patents.

“Exaone helps humans make better decisions in a more effective way,” said Lim Woo-hyung, vice president of Applied AI Research Group at LG AI Research, in an interview with the Korea JoongAng Daily.

“There are limits when humans make decisions because humans cannot comprehend the multi layers of the situation and analyze data simultaneously. When AI steps in, they remove the possibility of missing details and making mistakes.”

Exaone is just one example, out of several, of Korea Inc.'s push to come up with their own large-scale generative AI. Portal giant Naver and other telecom companies have opted to develop their own language model tailored to an AI system in a similar way that OpenAI did with ChatGPT, while others are sourcing the system from third-party companies.

Regardless of the differences in their approach, the pursuit of generative AI is taking place across virtually all industries from electronics, telecommunications to pharmaceuticals. l

Lee Jin-sik, lab leader of LG's Exaone Lab, left and Lim Woo-hyung, vice president at Applied AI Research Group at LG AI Research, pose for the photo. [PARK SANG-MOON]

LG: expert-level AI
Instead of going after the masses, LG's Exaone specifically targets the so-called experts who are in search of professional data without having to worry about hallucination.

The recently revealed Exaone 2.0, for that reason, has been trained on 45 million professional documents including patents, books and theses. It also studied about 350 million images.

Under the group chief Koo Kwang-mo's leadership, LG established the LG AI Research, an AI-dedicated lab, in 2020, and pledged to invest 3.6 trillion won ($2.66 billion) in AI technology through 2027. It launched the first version, Exaone 1.0 in 2021, followed by an updated one this year.

Exaone's expertise shines in an area that needs to digest a vast amount of reading material in a short period of time.

That is why one of the lab's early clients include the Korea Intellectual Property Office where the Exaone platform will be deployed to significantly increase efficiency in distributing patents.

“The juries have to read hundreds of documents every day to determine whether a certain item is worth being patented, analyzing how it differentiates from similar patents from the past and the innovativeness of the technology,” said Lee Jin-sik, head of Exaone Lab in an interview with the Korea JoongAng Daily.

“The business is currently available in Korea, only, but we are at the forefront of deploying AI in the patent business globally and we plan to export the model.”

The focus on professional documents made sense for LG to make the model perfectly bilingual because a majority of the resources are in English, according to the lab. It made the platform to effortlessly go back and forth between Korean and English, and vice versa. It is the only generative AI platform that can do so in Korea, the company says.

“In order to expand the amount of collected data, we made the model bilingual from the get-go because most professional documents are only available in English,” Lee said.

“We are focusing on transferring the data obtained in English to Korean.”

Such specialization in expert knowledge has organically lead LG Exaone to cooperate with the group's chemical and bio affiliates.

Affiliates like LG Chem and LG Display have in fact already put Exaone into real use in their research into new drugs and materials.

The AI lab expects the deployment to decrease the number of synthesis trials from 10,000 to less than 100, and the average time spent on research from 40 months to five months.

Exaone Discovery, an Exaone-based AI service tailored to R&D, not only understands and generates text-based data but is able to understand graphic and image-based data such as molecule structures, tables and charts.

“Developing new materials and drugs takes a lot of time and effort because it requires a lot of testing,” said Lim.

“Such an analog work process is highly likely to get a lot of help from AI. And coincidentally LG Group is very strong in its chemical business. Each side needs to come together.”

The task-oriented design of the platform will differentiate it from rivals like OpenAI's ChatGPT.

“ChatGPT pursues a general language model for more common usage,” Lim said.

“Although not in all areas, Exaone is going to prove its worth in select areas in need of expertise.”

Lim Jung-hwa, leader of Hyperscale AI Planning, left, and Lee Sang-woo, leader of Hyperscale AI at Naver Cloud pose for the photo. [NAVER]

Naver: Tuned for Korea

Type in “punctual,” “meticulous,” and “adventurous” in the keyword box and jot down a couple of more sentences about yourself. Click “Generate.” Voilà!

A generative AI has just written up a cover letter introducing your coming-of-age story, blending in the characteristics you typed in.

The AI not only refined word choice and sentence structure but also added in plausible anecdotes that made the piece more complete. It also put catchy subheads for each section.

This AI has just saved you at least couple of hours that might have been spent looking for the right cover letter format, searching for the right expression and coming up with clever ways to make an average story look great.

Compare this result with OpenAI’s ChatGPT, inputting a similar request, and it becomes clear which service is more familiar with Korea’s job-seeking market and more fluent in Korean.

The former is a service called Letter Bee developed by DP Lab, a local startup.

Letter Bee is fluent in Korean and has insight into the country's job market because it processes data on HyperCLOVA, a generative AI platform developed by IT giant Naver.

Naver, which runs the country's biggest search engine, is a domestic leader in AI technology. It believes Korean data, collected through various services affiliated with the search engine including blogs and shopping sites, is its biggest strength.

In August, it unveiled the HyperCLOVA X, an upgraded version of HyperCLOVA revealed in 2021. It was a fruitful moment for Naver which has poured a total of 1 trillion won into the AI project for the past five years.

The exact amount of trained data for HyperCLOVA X is not disclosed, but it studied Korean language data equivalent to 50 times that of news searched on Naver annually. Comparatively, that is 6,500 times that of OpenAI's ChatGPT-3.

While data collected from blogs or online communities may not be as accurate as those collected from news or academic papers, Naver thinks they serve different and very important purposes.

“Many people think data collected from blogs or online communities will be low-quality,” said Lee Sang-woo, leader of Hyperscale AI at Naver in an interview with the Korea JoongAng Daily.

“But there is qualitative data there too. It is just a matter of how you want to approach the ‘diverse’ side by correctly training the model.”

Lim Jung-hwa, leader of Hyperscale AI Planning at Naver, echoed the view.

“We get a lot of hints about Korea-specific sentiment and how to read between the lines from those services,” Lim said.

The internet company launched HyperCLOVA X, an upgraded version of HyperCLOVA, in August and is now launching new services on it, starting with CLOVA X which is an AI chatbot service.

Connect X is another model based on HyperCLOVA X which is specifically trained to increase efficiency in the workplace by helping with coding, organizing schedules, summarizing meetings, and summoning the right documents. It is slated to launch later this month.

Naver also works with its affiliate LY Corporation which runs Japan’s most popular messaging app Line and search engine Yahoo Japan, positioning it in a more advantageous place to collect data. It is currently operating a Japanese version of HyperCLOVA which has been trained on Japanese data equivalent to 2,700 years of newspaper articles, according to the company.

“Japan is a viable option, as well as non-English areas like the Spanish-speaking world and the Middle East,” said Lee when asked about the AI platform's next destination..

“My ambition with HyperCLOVA X is to make it a global top-tier player. It sure is a difficult challenge but I think we have a chance.”

Telcos targeting telcos

Telecom service providers are bringing generative AI into sectors closely linked to their telecommunications service. They are developing large language models and AI systems tailored to customer centers and health care, and creating personal AI assistants based on the supreme amount of client data they've accumulated over the years.

SK Telecom, Korea's top mobile carrier, recently announced that it will establish a large language model with Deutsche Telekom to provide an AI system tailored to telecom companies by the first quarter of 2024.

The model will be multilingual, available in German, English and Korean.

Deutsche Telekom CEO Tim Höttges, left, and SK Telecom CEO Ryu Young-sang pose for the photo at the signing ceremony to jointly develop large language model for global telecom companies at SK Telecom's office building in central Seoul on Oct. 20. [SKT]

“Creating a large language model for the telecom companies means that the foundation model will be suitable to be deployed across the sectors in which such companies excel — in telecommunication, obviously,” an SK Telecom spokesperson told the Korea JoongAng Daily.

The large language model will help provide text summaries of calls and recordings. It could even advance as far as saving a phone number or organizing users' schedules as AI comprehends the context of the calls. It would be “your very own personal AI assistant,” as SK Telecom CEO Ryu Young-sang phrased it.

Within Korea, the company released its own AI assistant app “A.” on Sept. 26, which as of now is limited to completing simple tasks such as recording calls and creating text summaries of recorded calls. However, SK Telecom plans to upgrade the app to include functions like real-time translation services, updating a user’s agenda on the digital calendar based on recorded calls, and recommending and reserving the most adequate routes for a user to reach their destination.

The company does not hesitate to bring technologies from foreign companies in sectors they lack, forming multiple alliances with companies such as Anthropic, OpenAI and Konan Technology to develop its own large language model dubbed “A.X.”

KT, on the other hand, is pursuing a “full-stack” AI strategy to branch into other AI services with the development of its own large language model Mi:dm, which will be released at the end of this month.

“Full-stack” means the full, end-to-end development of applications.

The company is collaborating with Canada-based AI lab Vector Institute to develop a AI prompt engineering tool — the crafting of text or queries given to a language model to elicit accurate responses.

KT’s technology utilized in Mi:dm will be adapted to a Thai language model service as well, as part of its recent partnership formed with Thai telecommunication company Jasmine Group.

Jasmine Group will decipher the regional market to build a graphics processing unit (GPU) farm that can power the model. The GPU farm will begin construction some time in the first half of 2024, while the Thai language model will be developed in the second half of next year.

KT CEO Kim Young-shub, fourth from right, and Jasmine Group chief Pete Bodharamik, fourth from left, pose for a photo after holding a strategic partnership meeting for the generative AI business at the KT headquarters in central Seoul on Oct. 19. [KT]

It will provide a subscription-based service based on Mi:dm for business-to-business (B2B) and business-to-consumer (B2C) services. Mi:dm will be provided as a platform on KT Cloud and it will learn from client data, and will be adopted as a pay-as-you-go B2B business model.

For general consumers, Mi:dm will be adapted for caregiving services such as senior care and child care. Features may include an AI chatbot that can provide recommendations to improve a senior users’ health condition, or child counseling services for parents from data that AI learned from Oh Eun-young, the renowned child psychiatrist.

One sector that both telecom companies are striving to capitalize on is AICC, a service sector related to customer service or support centers using technologies.

The domestic AICC market is growing rapidly but its size hasn't been noticeably perceived due to the initial cost of installing AI software and their limited roles, which don’t go beyond merely assisting human counselors.

“The AICC solutions that companies are developing now generates more natural responses to customer requests based on vertical language models,” noted analyst Oh Dong-hwan of Samsung Securities. “Companies are also cutting costs by utilizing cloud-based AI software, which means there is no need to physically build a separate infrastructure on top of the companies’ existing system. We anticipate that existing workers at the contact center could be freed of repetitive tasks to focus more on value-added duties such as promoting subscription-based services. Small and mid-sized entrepreneurs who could not use the AICC system will now be able to do so with the cloud-based software as well.”

KT Cloud currently provides cloud-based AICC software as a subscription-based service to enterprises. The AI chatbot is able to recognize each customer’s own unique voice signature to verify their identities utilizing voice printing technologies, and automatically documents the content of the calls as well as summaries.

SK Telecom, on the other hand, derives the technology needed for AICC from a local AI-utilizing customer service developer called Persona AI. The telecom carrier became the startup’s third largest shareholder in August, although the size of the acquisition was not disclosed.

BY JIN EUN-SOO, LEE JAE-LIM [jin.eunsoo@joongang.co.kr]