[Virtual Human 2.0] The Beginning of Evolution... From Advertising Model to Top 1% Expert

Mega-scale AI Transforming Industrial Sites... Integrated with Virtual Humans

Editor's Note Virtual humans are evolving. They no longer remain mere creations dressed in computer graphics that move according to human-made scenarios. Through ultra-large artificial intelligence (AI), they think and make decisions on their own. Last year, virtual humans made a dazzling debut as TV commercial models, limited to entertainers such as models and influencers. Virtual humans 2.0, equipped with brains and capable of independent thought, are now replacing humans in top 1% expert fields such as scientists, new drug researchers, and financiers, as well as jobs like call center agents and museum docents. Although not fully original creators, they have also ventured into creative fields such as art, music, and poetry. Because they can consider more possibilities faster than humans, they produce unexpected results that are recognized as creative works. How far can virtual humans replace human activities? How close can they get to humans? We shed light on the era of virtual humans 2.0 with infinite possibilities.

[Asia Economy Reporter Yuri Choi] In February, a noteworthy rookie designer appeared at New York Fashion Week in the United States. Under the theme "Flower from Venus," the designer showcased vivid colors and mysterious patterns. The rookie designer was the virtual human "Tilda." Designer Park Yoon-hee applied colors and shapes to images created by Tilda to produce about 30 outfits. The creator was a virtual human, and the finishing touches were done by a human, resulting in a new form of collaborative work that has received acclaim.

The source of creative power is the ultra-large AI "ExaOne" developed by LG AI Research. ExaOne has learned from 600 billion corpora and 250 million high-resolution images combined with language. For example, if there is a visual material of a flower shape and the expression "beautiful" is associated with it, ExaOne learns from hundreds of millions of pairs of image and language data to autonomously find images that fit specific language. Compared to an average person who sees about 100,000 images in a lifetime, this is 2,500 times more. This enables it to come up with unimaginable ideas.

Thinking Like the Human Brain with ExaOne... Active as Researcher and Designer

Tilda is equipped with an ultra-large AI resembling the human brain structure. Based on computing infrastructure capable of large-scale calculations, it learns and judges data autonomously. Tilda has 300 billion parameters, the highest computational power among domestic ultra-large AIs. It possesses 'multi-modality' capabilities, processing language as well as images and videos simultaneously to think.

Ultra-large AI is already active in industrial sites. LG AI Research has entrusted ExaOne with solving group-level challenges. It focuses on solving problems with at least 10 billion KRW in business value, such as new drug candidate development and battery life prediction.

A representative example is the AI contact center used as an advisor in affiliate customer service centers. ExaOne Advisor learns customer data to classify consultation types and preferences. This supports providing services at the level of veteran counselors.

By accumulating communication experience with various people, it can become an expert trained over decades. It even analyzes customers' emotions revealed in sentences or conversations to provide highly satisfactory service. Furthermore, the day when AI directly concludes B2B contracts is not far off.

In the chemical field, it is active as a researcher. ExaOne has learned tens of millions of related papers and documents to create a model predicting cancer treatment candidate substances. Once trained, having just one or two AI researchers is like having dozens of assistant experts for research and development. Working with AI experts, even ordinary people can produce expert-level results.

It has also entered the creative domain like Tilda. ExaOne is used to create product design drafts. For example, inputting "robot vacuum cleaner design resembling a supercar" automatically generates hundreds of drafts. Designers then add detailed designs and engineering based on these drafts.

Developing Expert AIs by Field... Evolving into Universal AI

The activity areas of ultra-large AI are expanding further through the "Ultra-large AI Alliance," which includes 13 companies. LG AI Research has partnered with various industries such as Google, Woori Bank, GS Retail, Hanyang University Hospital, and EBS. When partner companies provide data, the research institute creates field-specific AIs.

Woori Bank is developing an AI banker that provides consultation services like a bank clerk. GS Retail aims for AI capable of customer-tailored responses, such as recommending food ingredient recipes. These AI services can also be integrated with virtual human technology. VA Corporation, a metaverse company in the alliance, is jointly developing an AI human. It is a virtual human that reads emotions embedded in language and generates corresponding facial expressions and movements. They explain that by combining expert AI with virtual human technology, they will present new customer experiences.

The ultimate goal is universal AI. While existing AI learned data from specific fields and solved related problems, universal AI is an all-purpose AI capable of handling various tasks.

Baek Gyeong-hoon, head of LG AI Research, said, "We will verify the potential of ultra-large AI through collaboration with various fields and expand the ecosystem. Next year, we will promote AI research and development capable of expert-level conversations in all professional areas."