LG Unveils 'Smart Super Large AI That Understands Human Speech' X1 (Comprehensive)

The metaverse space decorated with a Christmas concept by 'Exaone,' the super-large AI released by LG AI Research on the 14th. (Photo by LG)

[Asia Economy Reporter Jeong Hyunjin] "Decorate the room atmosphere like the photo I saw on social networking services (SNS)." To decorate a metaverse space as a place for a Christmas party, instructions are given to artificial intelligence (AI). The AI recommends designs including a candy-shaped sofa, dazzling lights, Christmas stockings, and wallpaper. When asked to change the view outside the window to a Christmas atmosphere, it presents a snowy landscape. The AI acts as a kind of 'designer,' understanding the user's intent and decorating the indoor space. All of this is the work of LG AI Research Institute's ultra-large AI 'EXAONE.'

LG AI Research Institute Celebrates 1st Anniversary, Unveils ‘EXAONE’... The Most Studied AI in the World

LG AI Research Institute, the 'core weapon' of LG Chairman Koo Kwang-mo's future new business, has unveiled the ultra-large artificial intelligence (AI) 'EXAONE' and announced plans to expand the global ultra-large AI ecosystem. Starting with actively applying the 'top 1% expert-level AI' across all LG affiliates' businesses, it aims to expand its applications through alliances with global partners and ultimately popularize ultra-large AI.

On the 14th, marking its first anniversary, LG AI Research Institute held an online 'LG AI Talk Concert' where it unveiled EXAONE and announced major research achievements and future plans. Ultra-large AI is designed similarly to the human brain structure, enabling it to think, learn, and judge like a human. EXAONE stands for 'EXpert Ai for everyONE.' LG AI Research Institute has been researching ultra-large AI since May.

Parameters refer to where the data learned by AI through deep learning is stored, and theoretically, the more parameters, the more sophisticated the AI's learning can be. EXAONE possesses about 300 billion parameters, the largest in Korea, and has multi-modality capabilities that allow it to acquire and handle various information related to human communication, including language, images, and videos.

As multi-modal AI technology advances, AI can go beyond acquiring and understanding data to reasoning and creatively generating across sensory domains such as vision and hearing. Decorating a metaverse space for Christmas is a representative example. While existing AI could analyze text to find images, EXAONE can independently judge and create images based on learned information. For example, if you say, "Make me a pumpkin-shaped hat," it draws a new image directly.

To develop such multi-modal AI, LG AI Research Institute explained that it utilized the world's largest scale of training data. EXAONE learned from 600 billion corpora and over 250 million high-resolution images combining language and images. Including specialized data held by LG affiliates, as well as refined corpora such as papers and patents, it is increasing its potential to act as an expert in various industrial fields.

EXAONE is a bilingual AI. Unlike OpenAI's GPT-3, an ultra-large AI developed in the U.S. that focuses on English, and other ultra-large AIs developed domestically that concentrate on Korean, EXAONE understands and uses both Korean and English at a native speaker level.

LG Announces Expansion of Ultra-Large AI Ecosystem... "We Will Create Expert AI"

LG AI Research Institute plans to make EXAONE, an ultra-large AI, function as a top 1% expert AI across virtually all fields such as manufacturing, research, education, and finance. This will be done in three stages to establish a global ultra-large AI ecosystem.

First, it opened the open API, a gateway to use EXAONE, to LG affiliates, enabling the application of ultra-large AI across LG's businesses including electronics, chemicals, and telecommunications. LG affiliates are already applying EXAONE to enhance chatbots and discover new materials and substances. An LG AI Research Institute official said, "Unlike existing AI, EXAONE reads and analyzes literature by itself, builds databases, and allows highly skilled personnel to focus on more valuable tasks rather than simple work."

Subsequently, LG AI Research Institute plans to form alliances with various global partners in finance, fashion, distribution, education, and more to broaden the utilization of ultra-large AI. To address the critical issue of data security in this process, the institute has independently developed an algorithm called 'EXAONE Tuning.' Ultimately, the AI Research Institute is also considering ways to build a win-win environment through popularization by making ultra-large AI accessible not only to some companies but also to the general public.

Baek Kyung-hoon, head of LG AI Research Institute, said, "We want to be a research institute that constantly challenges difficult problems and creates expert AI essential to customers based on excellent research achievements." He added, "We will further strengthen research and development cooperation systems with major domestic and international universities and scholars such as the University of Toronto in Canada, the University of Michigan in the U.S., Seoul National University, and KAIST, and contribute to building a global ultra-large AI ecosystem through collective intelligence via API disclosure and external partnerships."

LG AI Research Institute, regarded as LG's representative future new business, was launched in December last year with active support from Chairman Koo. It plans to invest 200 billion KRW over the next three years to develop technology and aims to increase the number of AI experts within the group to around 1,000 by 2023.

In the one year since its launch, the institute has achieved significant results, with 18 papers accepted this year alone at the world's top AI conferences such as AAAI, CVPR, ICLR, and NeurIPS, covering topics like optimal path reinforcement learning and AI composing music. It has also introduced AI to solve industrial challenges with high business contribution and difficulty in cooperation with AI organizations of LG affiliates. This year, it solved 18 challenges and aims to solve more than 25 challenges next year.

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