Kyunghoon Bae
2024.08.05
Toward Expert AI : EXAONE 3.0 and ChatEXAONE
Since the announcement of EXAONE 1.0, Korea's first bilingual and multimodal generative AI model, in December 2021, LG AI Research has been on a path to develop expert AI that can be applied to real-world industries by releasing upgraded models and commercialization results every year.
Today, LG AI Research introduces EXAONE 3.0 model and ChatEXAONE. From the perspective of the expert AI that LG is committed to, I would like to explain the differentiated performance of EXAONE 3.0 model, why we are open-sourcing the 7.8B model, and the implications of ChatEXAONE, enterprise AI agent.
The differentiated performance of EXAONE 3.0 model
Models that show high performance in various benchmarks are often published, but rarely do they show high performance in the real world. This is because existing benchmarks focus on measuring the performance of a single command, and it is difficult to reflect the evaluation of users or the field.
To ensure that EXAONE 3.0 serves as the core of the 'expert AI,' we aimed to develop reliable AI for effective use in various industrial applications. It outperforms similarly sized open models in MT-bench, Arena-Hard, WildBench and AlpacaEval, which were developed to evaluate it. This means that EXAONE model is at the top of the class in important aspects of human preference, such as multi-turn conversations, instruction following, coding, and math, and can be combined with domain-specific documents and data from experts and enterprises to produce more convincing results.
It is especially meaningful that these performance evaluations are measured in English and published in a technical report. Our customers go beyond our LG group, and many are international. We will transparently publish the results of our world-class performance evaluations in line with global standards so that they can be validated and made available to more customers.
Why we are open-sourcing EXAONE 7.8B model
Today we are open source EXAONE 3.0 7.8B model for research purposes. We believe that we can make a small contribution to revitalizing the research ecosystem by enabling the use of advanced models from a research perspective.
There were concerns that transitioning EXAONE, previously used exclusively within LG, to an open model might mean forfeiting certain technical and business advantages. However, considering the rapid advancements in AI technology, it was determined that releasing the current model could have a positive impact on the AI ecosystem. Therefore, we decided to make EXAONE 3.0 7.8B model publicly available.
EXAONE 3.0 is a lineup of models of different sizes depending on the application, from ultra-lightweight models for on-device AI, to lightweight models for general purpose, to high-performance models for specialized applications and future research. The open-sourced EXAONE 3.0 7.8B model is the most versatile model in terms of performance and affordability within EXAONE 3.0 models. We hope it can be used by many academics and organizations. We also hope that the open ecosystem will stimulate the discussion of EXAONE research, which will lead to the open-sourcing of more models. We would like to remind companies that our partnership channels are always open to companies that want to use our models for commercial purposes.
The implications of ChatEXAONE, enterprise AI agent
Today, we are also unveiling ChatEXAONE, which is developed based on EXAONE 3.0. ChatEXAONE is an Enterprise AI Agent for improving the productivity of enterprise employees. It combines business documents and external information to generate reliable answers and provides a variety of convenience features and professional insights tailored to the user's job category. It will be available to LG Group users first, revolutionizing everyday life and making a case for streamlining work that requires expertise.
Another significance of ChatEXAONE is that it is a reference AI system based on multiple EXAONE 3.0 models. EXAONE 3.0 models are designed to perform well in real-world use cases, but to be applied in the real world, models cannot operate alone. Information extraction or search engines such as Retrieval Augmented Generation (RAG) must be combined, connections to databases and organizational charts within the enterprise are required, and a variety of external tools are needed to provide specialized capabilities. As a reference AI system, ChatEXAONE builds all the necessary components for an enterprise AI system to secure use cases and continue to evolve. Then, when appropriate, it will be customized and offered to external customers to meet their company-specific requirements and areas of expertise.
Toward Expert AI
LG AI Research has been aiming for expert AI that enables generalists to have expert-level capabilities and experts to have higher-level capabilities by utilizing AI. LG's expert AI is not an AI that replaces humans, but an AI that works with humans to reach a higher level. This is in line with the LG AI Research's mission of Advancing AI for a better life.
Until now, the performance and results of EXAONE models have only been available to selected customers because more people think that general AI or language-specific AI is important, and only a few prefer expert AI. But now, more people understand it is more important to use AI in real industries, and that specialized performance and affordability are required for this. Therefore, in this announcement, we are releasing evaluation results and some models so that as many developers and partners as possible can understand and use EXAONE. LG AI Research will continue to strengthen its ecosystem and partnership activities to bring the benefits of expert AI to all businesses around the world.
Dr. Kyunghoon Bae, President of LG AI Research