On July 19, “EXAONE 2.0”, a large-scale AI model, was finally unveiled at the LG AI Talk Concert 2023!
LG AI Research, founded in December 2020, has worked hard to solve the dilemmas of various industrial fields! EXAONE has pursued innovation through understanding multimodal (lingual and visual) information since the early stages of its development and has taken on challenges for new scientific discoveries based on knowledge humanity has accumulated over time. LG AI Research is the only company in Korea, and one of the few globally recognized entities, that creates generative AI that can understand and create languages and images and even comprehend professional literature.
But there are several challenges in applying the current generative AI in industrial fields. The generated outcomes have low reliability and high costs in training and operating large-scale models, and there is also a risk of leaking the internal data of companies in the process of implementing the technology in industrial settings.
Therefore! To address these issues, LG AI Research focused on enhancing the quality of training data, improving cost efficiency, and providing a custom AI model. As a result, “EXAONE 2.0”, a new expert AI, has emerged that offers a distinctively different experience from existing generative AI models! The reveal of EXAONE 2.0 garnered immense interest with over 200 journalists in attendance. Let’s go a little deeper into EXAONE 2.0!
Distinctive Factors of EXAONE 2.0
Data
EXAONE 2.0 has secured high-quality data, with a focus on professional data. There are more than 45 million pieces of data from academic papers and patents and more than 350 million pieces of image-text data and it resolved all copyright and privacy issues for the training. We developed a bilingual model that can understand and generate both Korean and English to overcome the limitations of Korean language data. Performance was maximized through bilingual multi-task tuning by significantly increasing the quantity of Korean-English training data by more than four times compared to the previous model and innovatively enhancing performance through synergy between Korean and English data.
Enhanced Cost Efficiency
The large-scale language model of EXAONE 2.0 reduces inference time by 25% compared to the previous model and decreases memory usage by 70%, which can reduce costs by about 78% compared to the previous model while maintaining identical performance. The multimodal model has improved “economic feasibility,” a noted limitation of large-scale AI models, by doubling the memory usage for better quality, reducing inference time by 83%, and reducing costs by about 66%
Custom Model Based on Customer Needs
EXAONE 2.0 can be provided as a custom model based on the customer’s needs and purpose. Customers can choose between text and image for input and output data and also select the language. Customers also can further fine-tune their internal data specific to their professional domains or choose the construction method. We anticipate enhanced implementation of EXAONE in professional industries as this customization maintains domain-specific expertise and alleviates concerns for internal data security.
Now that we learned about the three differentiators of EXAONE 2.0, here is an introduction to the three platforms implemented with EXAONE 2.0: EXAONE Universe, EXAONE Discovery, and EXAONE Atelier!
EXAONE Universe, designed to provide expert responses and insights
First, an introduction to EXAONE Universe. EXAONE Universe is a platform that generates expert responses, engages in research for the latest technologies, and aids companies in their specialized tasks using the massive language model that is the core of generative AI technology. Based on professional literature and the latest knowledge, we developed a model capable of comprehensive inference and logical thinking, and focused on generating eloquent answers by collecting factual evidence.
(Moontae Lee explains in a way that anyone can easily understand!)
Moontae Lee, Head of Advanced ML Lab, explained the EXAONE Universe through a live demonstration! Even in 2023, experts still spend a considerable amount of time on work that involves searching for information, selecting the most relevant documents from search results, and extracting required portions to consolidate into one piece of useful information. The large-scale language models currently available from other companies are proficient in generating fluent words within the context but have frequent hallucination problems that generate false statements and a response basis that is difficult to create an overall framework.
To address these issues, EXAONE Universe was designed with a comprehensive and logical inference function that identifies professional documents from the latest database, infers detailed evidence from the documents, and formulates different evidence into insights! Hence, EXAONE Universe can provide the basis for responses as references (focusing on factual evidence reliable for use by experts )
)
(The actual EXAONE Universe model that opened to LG Affiliates in July!)
Moontae Lee input the question, “How can hallucination be resolved in a large-scale language model?” and demonstrated how it generates a response. EXAONE Universe proposed how to utilize both text data and visual data, suggested natural language inference (NLI), and provided in-depth responses by element! Additionally, “What should we study generative AI?” and “Will AI and automation replace human workers?” were asked. “What should we study generative AI?” in particular had an intentional grammatical error in which EXAONE Universe demonstrated surprising results by generating an answer by comprehensively inferring “why” with “what.”
EXAONE Discovery explores new scientific facts
Next is EXAONE Discovery! EXAONE Discovery is a service platform for exploring new scientific facts that are boldly venturing into developing novel materials and drugs using AI (amazing, right?). In the first stage of comprehension, we developed the in-depth document comprehension technology to extract data and knowledge from professional literature such as academic papers and patents. It is currently being refined through the second phase of “design-prediction” which designs the molecular structures with desired characteristics through EXAONE 2.0 and predicts the results of intermolecular synthesis.
(Charismatic Se-hee!)
(Astonishing demonstration producing novel materials)
Se-hee Han, the Head of the Materials Intelligence Lab, explained EXAONE Discovery. EXAONE Discovery is a platform that will serve as a navigator in finding novel materials and drugs. Designing and synthesizing molecules were once the exclusive domain of human scientists. But now, the AI model is expected to assist scientists in quickly designing desired materials and easily predicting synthesis that was traditionally synthesized and evaluated in the lab! She introduced the vision of EXAONE Discovery, that the era will begin where AI can facilitate the quick development of new materials with small amounts of data
Conventional research involves researchers analyzing previous research results, establishing a hypothesis based on their analysis, and verifying the hypothesis by creating materials and making assessments. In this process, countless experiments are conducted that result in failure, consuming significant time and money. The services of EXAONE Discovery drastically reduce this process! When scientists input a sentence, the model reviews numerous data and complex information and suggests the best way to develop novel materials that minimize the risks of research failure for scientists!
Se-hee demonstrated finding a new material that can substitute existing battery additives using EXAONE Discovery in reflection of the latest environmental issue of reducing the fluorine content of battery additives. It was truly an exciting and thrilling moment envisioning the impact EXAONE Discovery will have in the future.
EXAONE Atelier provides creative inspiration and ideas
Lastly, we introduce EXAONE Atelier, providing creative inspiration and ideas. EXAONE Atelier based on EXAONE 2.0 was trained with 350 million image-text pair datasets! Of course, data with copyright issues were not included in the training.
Designers who are always seeking something new can materialize their design concepts by generating images that never existed before through EXAONE Atelier. They can gain interesting and innovative ideas that may not have emerged and utilize these ideas in their designs to significantly reduce the time required to create a product. Meaning, EXAONE 2.0 can become a fantastic partner even when it comes to creativity.
(Demonstration like a pro!)
Seunghwan Kim, the Head of Vision Lab, introduced EXAONE Atelier. Explaining the zero-shot image captioning implemented in EXAONE Atelier, he went beyond its ability to describe input images in a natural language like a real person by highlighting the advantage of being able to perform tasks for users by combining AI with the EXAONE language model
(It can create advertising phrases for LG Household & Health Care products!)
A request was made to generate marketing phrases by providing images of LG products to verify the actual operation of EXAONE Atelier! EXAONE Atelier generated a creative advertising copy that understood the features in the image of cosmetic packaging: “Introducing the premium skincare set that will take care of your skin and beauty. The blue box packaging filled with specialized expertise will enhance your features and boost your confidence to another level.” In addition to advertising copy, videos were shown for writing captions for everyday photos that general users post on social media and for creating a children’s story with pictures. The story was so captivating that it left viewers curious about what’s next!
Q&A session full of diverse insights
(Q&A session with heated questions and answers)
(The demonstration space led by Jin-sik Lee, Head of EXAONE Lab)
A Q&A session was held following the introduction of EXAONE 2.0. LG AI Research Director Bae Kyung-hoon, Lab Director Kim Seung-hwan of Vision Lab who carried out the demonstration, Lab Director Han Se-hee of the Materials Intelligence Lab, Lab Director Lee Moon-tae of the Advanced ML Lab, and Choi Jeong-kyu of the Multimodal AI Group joined for the AI Biz. Unit Director Lee Hwa-young of the Development Unit shared in-depth insights about each respective field. There were so many enthusiastic questions that not all of them could be addressed due to time constraints
At the site of the Talk Concert, space was prepared to try the EXAONE 2.0 platform! It was an opportunity for a hands-on experience with EXAONE Atelier and to be blown away by the astounding technology of EXAONE.
At the site of the Talk Concert, space was prepared to try the EXAONE 2.0 platform! It was an opportunity for a hands-on experience with EXAONE Atelier and to be blown away by the astounding technology of EXAONE.(Check the video if you’re curious about the LG AI Talk Concert!)
EXAONE by LG AI Research is at the forefront of creating tangible successes not only in the LG group but also in various industries in collaboration with domestic and global partners. LG AI Research will continue leading the popularization of generative AI technology and evolving into a competitive AI R&D company not only in Korea, but globally. Please look forward to our journey
FLEX MEET-UP, a look back on the first half of the year!
In the afternoon on the day of the Talk Concert, a monthly meet-up was hosted for everyone to gather! After the successful end of the Talk Concert, the culture team prepared a very special FLEX Meet-up to commemorate a review of the first half of the year. Let’s dive into the party-like scene on a spectacular scale!
(Everyone gathered in the hall where Talk Concert was held in the morning!)
(Not just a regular meet-up, but a FLEX Meet-up! )
)A welcoming banner greeted you as you entered the hall where a meet-up on a grand scale was prepared! Researchers enthusiastically entered, ready for their special lunch boxes for the afternoon, with beer, wine, and even a cocktail bar to accompany fun music!
to accompany fun music!
(A fantastic combination of sandwiches, deep-fried chicken, fruit, and wine snacks)
The Culture Team went to every nearby convenience store to get special edition drinks instead of ordinary beverages! Researchers were deeply moved by the incredibly passionate Culture Team
(From the Coca-Cola NewJeans edition!)
(To Asahi Super Dry draft cans!)
The researcher who volunteered as bartender for the day even made cocktails. There was even gin and tonic, Jägerbombs, highballs, and screwdrivers on the menu!
(One highball please!)
(Explaining the beverages on the menu)
(The awesome bartender )
)
(A shy smile is a bonus )
)
(Our one-day bartender is incredibly skilled!)
(Researchers settled down with their lunch boxes and beverages- Everyone is a great mood!)
(Let’s have a great event full of things to do!)
Now, shall we begin the FLEX MEET-UP for a review of the first half of the year?
The agenda for this special day was filled with various events including an introduction of new employees, Shining Star awards, a recap of the Talk Concert, a quiz for a review of the first half of the year, and a typing competition!
The Talk Concert was a big project involving many participants, including researchers. Director Bae Kyung-hoon took time to hear impressions of the event from one of the researchers and the Team Leader of the AI Communications Team. They were both impressed to have such a fulfilling day with everyone working together and achieving great results
(It’s hard to believe that half of the year has already passed us by)
A video is a must when looking back on the first half of the year! The Culture Team produced a video on the activities of LG AI Research for the past six months. It was a moving moment to see the achievements of various research organizations and teams in just half a year altogether. It’s hard to believe we accomplished so much in such a short time. Thank you to everyone for working so hard
After the video, we held a quiz for a review of the first half of the year! This was an online quiz where you had to be the fastest one to answer correctly. It was a fun time full of sighs, cheers, and laughter
After the video, we held a quiz for a review of the first half of the year! This was an online quiz where you had to be the fastest one to answer correctly. It was a fun time full of sighs, cheers, and laughter 
(Relax your fingers and get ready!)
(Congratulations to 1st place!)
Then the typing competition took place as the final event!
The typing competition was held for those who submitted their records for short sentence exercises on the Hancom Taja website. The competition featured Top 2 in the Korean and English categories with the first-place prize being the mango bingsu from Shilla Hotel! A perfect prize you don’t want to miss, especially for the hot days recently!
(A typing speed for English at over 1,000 words per minute!)
(A stage reminiscent of e-sports)
(Go for first place!)
(Get that Mango Bingsu!)
(Typing away quickly!)
(Stay calm! Keep calm!)
(First place award ceremony, congratulations everyone!)
The Talent Relations Team had two people advance to the finals and prepared cute placards They confidently claimed first place in the Korean category! (Impressive )
They confidently claimed first place in the Korean category! (Impressive )
The FLEX MEET-UP to look back on the first half of the year ended with delicious food, fun music, and a variety of events, and will be memorable for years to come. We hope the joyful event was conveyed well to all of you through our blog! We’ll be back next time to bring you more joyful and exciting news