"This Year’s AI Focus Is 'Agents'... Major Transformations Expected"
Industrial Applications... 45,000 LG Group Employees Currently Using It
Moon Tae Lee, Senior Executive Director and Head of the Superintelligence Lab at LG AI Research, is giving a lecture at the Nuclear Industry Breakfast Meeting held on the 9th. Korea Nuclear Industry Association
"The power consumption of artificial intelligence (AI) is closely related to the number of parameters. We are developing technologies that improve efficiency while reducing the number of parameters."
Moon Tae Lee, Senior Vice President and Head of the Superintelligence Lab at LG AI Research, introduced these recent AI development trends during a lecture titled "From Text to Reality" at the Nuclear Industry Breakfast Meeting hosted by the Korea Nuclear Industry Association at the Westin Chosun Hotel in Seoul on the 9th.
Moon Tae Lee explained, "Big tech companies developing large language models (LLMs) previously competed to increase the number of parameters. However, they faced issues of inefficiency and a sharp rise in power consumption. As a result, the focus has now shifted toward improving efficiency and cost-effectiveness."
In fact, ChatGPT-3 developed by OpenAI increased its number of parameters to 175 billion. While OpenAI has not disclosed the number of parameters for ChatGPT-4, experts estimate it to be over 1 trillion. Google's PaLM, which has been made public, features 540 billion parameters. As the number of parameters increases, resource consumption rises, leading to inefficiency and significantly higher costs. Power consumption also increases accordingly.
For this reason, recent AI companies are focusing on developing technologies that achieve performance similar to large-scale models without increasing the number of parameters. For example, LG AI Research's "Exaone Deep 32B," released in March this year, has been evaluated as outperforming China's DeepSeek "R1" despite having only 3.2 billion parameters. In comparison, R1 has 67.1 billion parameters.
Moon Tae Lee stated, "In developing Exaone 3.0 and 3.5, we prioritized improving efficiency rather than simply increasing the number of parameters," and added, "We are focusing on whether we can maintain emergent abilities without maximizing parameters."
Exaone is an LLM developed by LG AI Research. In August last year, Exaone 3.0, an open-source version, was released for the first time in Korea. In December last year, LG Group began the official service of "Chat Exaone," an enterprise AI agent based on Exaone 3.5, for LG Group employees. "Exaone Deep," released in March this year, is an inference-type AI.
Whereas traditional AI required a process of repeated questions and answers between humans and AI to solve problems, inference-type AI does not simply present retrieved information but instead draws new conclusions or predictions through inference. This is expected to solve complex challenges and potentially replace human intelligence.
Moon Tae Lee commented, "In 2023, concerns about the side effects of AI were prominent. In 2024, AI sovereignty was a major topic. This year will be the year of 'AI agents,'" adding, "Many transformations are taking place." He compared the emergence of various AI-powered services to the launch of the smartphone app store.
In particular, a variety of AI services applicable to industrial sectors are expected to emerge. In fact, LG Group is already utilizing Exaone, developed by LG AI Research, across its affiliates. Moon Tae Lee explained, "With AI, it is possible to create materials with desired properties or predict the characteristics of new molecular structures," and added, "We are collaborating with other affiliates, including those in the pharmaceutical sector." He further noted, "Currently, 45,000 people within LG Group are using Exaone."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
