Generative AI technology is becoming popular in R&D in various fields. Beyond digital transformation, it is expected that securing scenarios that use LLM and SLM in combination will become important in the AX era.

On the 16th, the KIEI Industrial Education Research Institute held the '2024 Generative AI Introduction and Practice/R&D Utilization Plan and Actual Case Seminar'. The seminar covered the following topics: △Generative AI trends and corporate introduction strategies, △Generative AI utilization plans and cases by field, and △Generative AI security strategies.

Large Language Models (LLMs), such as ChatGPT, Gemini, Copilot, and Lama2, have proven their utility in various fields such as music, video, healthcare, new drug development, construction, and energy management, and competition to expand the use of AI services across industries has begun in earnest.

In particular, it is predicted that 'multi-modal AI' will have an impact on search tools and content creation by simultaneously processing various types of data such as text, images, audio, and video. For example, 'Sora' announced by OpenAI is a text-based video creation platform. Through this, users can easily understand and interpret prompts and create scenes such as specific characters, actions, and backgrounds. Now, customers can easily access customized content based on AI and human interaction rather than a simple listing of information.

Microsoft has developed and developed the vision of the AI tool, 'Microsoft Copilot'. It is being introduced across solutions to innovate work productivity and business processes in various fields. According to MS, user satisfaction increased by 70% through AI agents that support document work such as PPT and Word and coding.

'RAG (Retrieval Augmented Generation)' stands for search-based generation model, and is a technology that connects a large amount of user data by combining search engines and generation models. Lee Geon-bok, MS Developer Leader, said, "The details of GPT5 are not yet specific, but it is gradually developing into a service form that uses various LLMs together rather than simply using one GPT function."

Recently, the lightweight AI has been attracting attention to supplement the problems of the enormous operating costs and power consumption of super-large AI. 'SLM (Small Language Model)' is an economical and high-performance language model, and its usability is expected to expand further as it plays a role in executing commands in special areas at the edge.

SLM is utilized as an 'on-device AI' that is installed on specific devices such as smartphones. It consists of billions of parameters and requires less time and resources for learning, so it can be easily executed on mobile devices. It can also be used offline when the Internet is not supported, and has high accuracy by using selected high-quality learning data.

For example, if AI functions are installed in kiosks in an offline environment, they can be applied as SLMs rather than LLMs. Leader Lee Geon-bok predicted that “even if SLMs are released, LLMs will still be needed, and the mixed use of models will become common.” Google provides open source models called Gemini Nano and MS Phi-2. Phi-2 has 2.7 billion parameters and can provide tasks such as coding and text faster than LLM.

Meanwhile, the lecture that day introduced various cases of introduction of generative AI and R&D utilization. Aim Bio CEO Heo Nam-gu predicted that although developing new drugs using AI is difficult, it will grow based on investment from big tech companies within 3-5 years. CEO Heo said, “Currently, there is a limitation of insufficient verification data for performance evaluation of AI new drug developers, but deep learning-based antibody and protein prediction are expected,” and “Based on the paper, high growth in the synthetic new drug sector is expected.”

In addition, Professor Moon Hyun-joon of Dankook University claimed that AI-based building energy and indoor environment management is now possible. Energy management methods vary depending on the characteristics of building facilities, and AI can achieve energy savings based on building data such as schools and fire stations. For example, Google used data from sensor cooling systems in DeepMind’s deep neural network to minimize energy use during the deep learning process, thereby reducing cooling costs for data centers by 30%.