Baek Gyeonghun, Head of LG AI Research Institute, "New Version of ExaOne to be Released in August"

37th Hankyung CEO Jeju Summer Forum Lecture
Combining Exaone and LLM Solves 3-Year Task in 3 Months
LG Subsidiary CEOs Compete to Study AI
"CEOs Must Personally Oversee AI Before It's Too Late"

Baek Hwang-hoon, head of LG AI Research Institute, revealed on the 11th that "next month, we will upgrade the LG Exaone model and announce a new platformized version." LG AI Research Institute had unveiled the group's ultra-large multimodal Exaone 2.0 last year, and now they have announced plans to release a new version next month.

Baek Gung-hoon, Director of LG AI Research Institute, attended the '2024 Hankyung CEO Jeju Summer Forum' held at Lotte Hotel Jeju on the 11th and gave a presentation on the topic of 'Current Status and Response Directions of the Generative AI Ecosystem.'
[Photo by Hankyung]

On the same day, Baek delivered a lecture at the "2024 Korea Economic Association CEO Jeju Summer Forum" held by the Korea Economic Association at Lotte Hotel Jeju in Seogwipo, Jeju Island, stating, "There will be an announcement next month regarding the new generative AI 'Exaone' (new version)." The lecture topic was "Current Status and Response Directions of the Generative AI Ecosystem."

LG released Exaone 1.0 in 2021 and Exaone 2.0 last year. Exaone is not a consumer-oriented ultra-large AI like ChatGPT but an AI designed for experts in each affiliate's business.

Baek admitted that LG Group also faced difficulties in gaining the response of each affiliate during the 'AI transformation' process. Initially, few CEOs of affiliates applied AI to specific businesses, but now the entire group, including LG Display, is introducing Exaone into their operations. In the case of LG Display, about 85% of the workload is transformed by AI, and 15% is 'post-processed' by humans.

He said, "CEOs should not delay attempts to adopt AI," adding, "LG has been making AI transformation efforts across all affiliates for four years, and now CEOs of each affiliate are eager to adopt AI and are studying it themselves."

Baek explained that the reason companies find it difficult to apply AI is because securing data is not easy. During the data acquisition process, AI deep learning must quickly verify accuracy to grow AI models. However, this process risks generating false information. It is also difficult to purchase data 'stitch by stitch' at a fair price. To solve this, LG AI Research Institute is collaborating with Elsevier, the world's leading academic publisher, to acquire data worth 2 trillion won.

He emphasized that properly applying AI to business operations can yield definite results. In molecular design prediction experiments by chemical affiliates such as LG Chem and LG Household & Health Care, applying LG Exaone showed performance 10 to 30 times better than when performed by traditional chemists. Tasks that used to take three years were completed in just one month.

However, as LG AI Research Institute receives many project requests for new drug development, anticancer vaccines, cathode materials, electrolytes, and display luminescent material development, there are limitations with the limited pool of AI scientists. Baek said he plans to solve this problem by combining expert-use Exaone with general-purpose large language models (LLMs).

Baek stated, "Connecting AI models for each specialized field with a massive generative AI knowledge system will enable rapid resolution of detailed business problems," adding, "The era has arrived where combining the two produces results more than 85% faster than traditional expert work." He further noted, "Initially, even when business experts and AI experts worked together, the process took three years, but now it can be done in three months."

Baek advised that simply purchasing AI from global big tech companies like Google and Microsoft at high prices is not always the best approach. Companies need to clearly organize what problems they need to solve, how difficult they are, and how much cost they should spend.

He said, "OpenAI's GPT-4 is a connection of 20 models each with 200 billion parameters, and even Anthropic's Claude 3 is a very large model with 100 billion parameters, so a significant cost must be borne," adding, "It is important to accurately assess whether the problem you want to solve is that difficult."

This content was produced with the assistance of AI translation services.