LG's Exaone 2.0 AI aims to be among the top 1 percent

“Will artificial intelligence [AI] replace the human workforce?”

LG AI Research head Bae Kyung-hoo speaks to the press during a media conference in western Seoul on Wednesday. [LG CORP.]

“Will artificial intelligence [AI] replace the human workforce?”

In answering this question, AI provides its own insight with a list of theses on which it based its answers alongside references to the excerpts, all of which takes less than five seconds. With solid references, users do not have to worry about the so-called AI hallucination whereby AI provides false information.

This service is provided by the generative AI chatbot Universe, based on Exaone, LG’s hyperscale large-language model (LLM).

LG AI Research, LG Corp.'s research arm, unveiled the second generation of Exaone, dubbed Exaone 2.0, and three associated platforms - Universe, Discovery and Atelier - each serving a different purpose, to the press on Wednesday.

The first version was launched in 2021.

The newest version of Exaone, which focuses on preventing hallucination, pored over some 45 million theses, patents and literature as well as 350 million images.

It has also become bilingual, able to understand and provide answers in both Korean and English.

“The AI industry has undergone rapid changes over the past six months,” Bae Kyung-hoon, who heads the research center, said at the press conference.

“Although AI can now answer questions fluently, it does not meet the standard on-site. And the main reason for that is that the answers are not credible, expensive and data leakage is a concern.

The Universe is specifically designed to research professional information from experts. Using a chatbot format, when a scientific question is typed in, the Universe provides its own insight based on multiple professional sources. It provides a list of theses and excerpts from which it derived its answer so that users can understand the process.

LG AI Research plans to develop the platform in the areas of chemistry, biology, medicine, medical, finance and patents.

Discovery is targeted at developing new materials, substances and medicines. It can process uploaded material, a thesis for example, and understand not only text-based information but also images such as molecule structure, charts, tables and formulas.

The platform will help researchers come up with new materials and substances without needing to conduct experiments, reducing research time from an average of 40 months to 5 months, according to the lab.

The platform will be deployed to LG's chemical and bio affiliates in the fourth quarter of this year.

Atelier is a multi-modal AI platform that can process not only text-based material but also images, videos and sounds.

Based on 350 million inputted images, Atelier can create a caption for a photo, write an advertisement slogan for a product and come up with artwork inspired by certain concepts.

"What differentiates Exaone from other large language models is its expertise and credibility," Bae said. "It is the most competitive model in this area. We are developing a top 1 percent professional AI that can unearth and understand the world's knowledge."

The Exaone will mostly be deployed in business-to-business models initially to develop some success cases, Bae said.

BY JIN EUN-SOO [jin.eunsoo@joongang.co.kr]