'Multimodal AI' Similar to Human Learning
Supports Custom Design for Purpose and Budget
ExaOne 2.0 Supports Various Applications
“Will artificial intelligence (AI) replace human labor?”
When this question was input, AI provided an answer spanning more than two pages within 5 seconds. From an automation perspective, it said AI could partially replace human labor, but more importantly, it could help enhance the unique abilities humans possess. Within seconds, it explained not only expert knowledge but also the rationale behind it.
On the 19th, LG AI Research held the 'LG AI Talk Concert 2023' at the Convergence Hall of LG Science Park in Magok, Gangseo-gu, Seoul, demonstrating the super-large multimodal AI 'Exaone 2.0'. Multimodal AI refers to AI that learns in the same diverse ways humans perceive objects. While generative AI like ChatGPT mainly focuses on language, multimodal AI can process multiple input types such as text, images, and voice to think.
On the 19th, at LG Science Park in Magok, Gangseo-gu, Seoul, Bae Gyeonghun, head of LG AI Research Institute, unveiled the super-large multimodal AI 'EXAONE 2.0' at the 'LG AI Talk Concert 2023'. Photo by Kim Hyunmin kimhyun81@
On that day, LG AI Research sequentially demonstrated Exaone 2.0, composed of three major platforms: Universe, Discovery, and Atelier. Exaone Universe is a generative AI expert conversational platform. Exaone Discovery is a knowledge exploration platform that can be used in fields such as chemistry and bio. Exaone Atelier is a multimodal AI platform used to express images in language and visualize language as images.
Using Exaone Discovery, it was possible to quickly explore how to utilize new battery materials. While related tasks previously took 40 months, it was explained that with Exaone Discovery, this was reduced to 5 months. Exaone Atelier, when given a single product image and the request 'Generate marketing phrases for product advertising,' immediately provided several advertising phrases consisting of three to four lines each.
The first version of Exaone was released in December 2021. Since then, LG AI Research has continuously conducted research and development (R&D). The goal is to enable domestic and international partners to use expert AI specialized in each field through Exaone. In this process, about 45 million professional documents such as patents and papers, and 350 million images were trained to produce Exaone 2.0. The amount of training data was increased more than fourfold compared to the previous model to improve performance. It also supports customized design according to purpose, usage, and budget.
Noting that most expert knowledge data is primarily in English, Exaone 2.0 was developed as a bilingual model that understands and responds in both Korean and English. To solve the problem of excessively high construction costs, efforts were made to lighten and optimize the large language model (LLM) and multimodal model. As a result, for the LLM, at the same performance level, inference processing time was reduced by 25% and memory usage by 70% compared to the existing model. This led to an overall cost reduction of about 78%. The multimodal model uses twice the memory of the existing model but reduces inference processing time by 83%, saving about 66% of costs.
On the 19th, at LG Science Park in Magok, Gangseo-gu, Seoul, Bae Kyung-hoon, head of LG AI Research Institute, unveiled the super-large multimodal AI 'EXAONE 2.0' at the 'LG AI Talk Concert 2023'. Photo by Hyunmin Kim kimhyun81@
From the 31st, LG AI Research will first open Exaone Universe to AI researchers within the LG Group and partner universities. In September, formal services will be launched for LG employees researching and studying AI. Specialized services by domain such as chemistry, bio, and pharmaceuticals will be introduced in the future. For researchers in the chemistry and bio fields within the group, Exaone Discovery services will start in the fourth quarter. This aims to pursue innovation in R&D related to new materials, new substances, and new drugs. Prior to this, in the third quarter, Exaone Atelier services will be provided to professional designers inside and outside the group.
Baek Gwang-hoon, head of LG AI Research, said, "LG is the only company in Korea to commercialize both bilingual models and bidirectional multimodal models," adding, "We will develop into an AI company that creates customer value differentiated from other generative AI." Since Koo Kwang-mo became LG’s CEO in 2018, LG has been focusing on AI as a future growth engine. In 2020, LG AI Research was established to serve as the group’s AI research think tank.
Meanwhile, LG introduced LG Electronics’ AI Contact Center (AICC) as a representative use case of Exaone. AICC analyzes and summarizes customer consultation content in real time. When a consultant talks with a customer, AI generates answers or content that perfectly matches the consultation and displays them on the monitor. After the consultation ends, it summarizes the content in one line for storage, helping quick processing in future after-sales service (AS).
LG AI Research plans to convert the pilot-operated AICC into a formal service in the second half of the year. Next year, AICC will also be used for customer service in English-speaking countries. Although customers may not realize it, AI has begun to influence human life.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
