LG’s hyperscale AI EXAONE 2.0 to be launched for drug development this year

Bae Kyung-hoon, head of LG AI Research, introduces the AI platform EXAONE 2.0 at the LG AI Talk Concert held at LG Science Park in Seoul on July 19. [Photo by Yonhap]

LG AI Research has unveiled EXAONE 2.0, a hyperscale artificial intelligence (AI) language model that can be used for expert applications in the development of new materials or medicines.

During LG’s AI Talk Concert event held on Wednesday, the research institute said the upgraded version of EXAONE services will be available for LG Group researchers, starting in the second half of this year.

Hyperscale AI refers to AI that thinks, learns, and makes decisions like human minds by learning large-scale data based on a computing infrastructure capable of large-scale computation.

This new version of EXAONE, first released in December 2021, has reportedly learned about 45 million specialized documents, including patent documents and academic papers, as well as 350 million images. The dataset provided to EXAONE 2.0 was secured through partnerships.

In particular, LG’s latest AI model focused on lightweighting large language models (LLMs) and multimodal models that process multiple information such as images and languages simultaneously to address cost issues.

Considering that much of the existing expertise data is in English, LG developed the hyperscale AI as a bilingual model that can understand and answer questions in both Korean and English. The amount of training data fed to it was also increased by more than four times compared to the previous model.

While maintaining the same performance level as the existing model, the language model of EXAONE 2.0 reduced inference processing time by 25 percent and memory usage by 70 percent, resulting in a cost savings of about 78 percent, according to the research institute.

Bae Kyung-hoon, head of LG AI Research, introduces the AI platform EXAONE 2.0 at the LG AI Talk Concert held at LG Science Park in Seoul on July 19. [Photo by Yonhap]

The multimodal model, which enables bi-directional generation between language and images, uses twice as much memory as the previous model to improve image generation quality but reduces inference processing time by 83 percent, saving about 66 percent in costs.

EXAONE 2.0 can be customized, too, in terms of model size or type to meet individual purposes and budgets, the company added.

During the event, three new EXAONE platform services were released: Universe, Discovery, and Atelier.

Universe is an interactive AI platform for professionals of certain fields. It can provide evidence based on pre-trained data as well as the latest specialized data in each field to present its answers.

Discovery is a platform specialized in the chemical and biological fields, which is expected to help develop new materials, substances, or new drugs. The multimodal AI technology applied to it enables it to database non-text data such as molecular structures, formulas, charts, tables, and images. With Discovery, the time required for research and development in the relevant fields is expected to be reduced to 5 months from 40 months, according to the research institute.

Atelier is a multimodal AI platform that can express images with language and visualize language as images based on 350 million pieces of copyrighted image and text pair data.

LG plans to launch the three EXAONE services for researchers across the LG Group, starting in the second half of this year.

“LG is the only Korean company that has succeeded in commercializing both bilingual and interactive multimodal models. We will continue to create AI models differentiated from other generative AIs in the market based on collaboration with domestic and international partners,” said Bae Kyung-hoon, head of LG AI Research.