The era of new AI emerging every morning has arrived. Following ChatGPT, the generative AI market has been highlighted as the main keywords of this year, with ecosystem expansion and monetization, and generative AI is proving to be a tool for improving corporate productivity.

At the seminar titled 'Gen AI Spring and Autumn Period and Warring States Period, LLM (Large Language Model) Market Outlook and Corporate Adoption Strategies' held by the Korea Information Industry Association and the CRM·BI Council at the Korea Technology Center on the 21st, Labelup DA Jongmin Kim and Organize Team Leader Taeha Yoo presented on LLM market trends and corporate adoption plans.

Generative AI is a model that generates specific results rather than performing classification such as segmentation and interpretation, and the LLM (Large Language Model) has grown in size, reaching a level where the number of parameters is currently about 16,000 times larger than five years ago. Along with this performance improvement, the number of models has also increased. In the summer of 2023 alone, there were over 10,000 new models.

DA Kim Jong-min of Ravelup recently explained the major AI trends. This year's topics include LLM and multimodal models that are applied to various applications by mixing multiple images, AI-trained AI models, model lightweighting, on-device AI, and responsible AI.

■ MS, Copilot Update… GPT 4o Installed on Azure AI

All AI fields, including AI models, GPU hardware, and NPU, have emerged as AI battlegrounds. Competition between global big tech companies such as Google, Apple, and Meta and numerous AI companies such as LG, Naver, and Kakao is intensifying. Recently, Google announced that it has begun applying its AI 'Gemini' to all of its businesses.

In response, Microsoft (MS) fired back. On the 21st (local time), MS held its annual developer conference Build. OpenAI has released the update details for its AI service 'Copilot', which was developed based on its AI model GPT, and announced that it will be installing the latest AI model 'GPT-4o' (FoO) on its cloud service Azure AI.

The latest AI model GPT-4o, released by OpenAI on the 13th, can be used with MS's 'Azure AI Studio' and API. GPT-4o is a multimodal AI model that integrates voice recognition, speech-to-text, and image recognition functions to enable natural real-time interaction in the form of a conversational interface. It is characterized by faster speed and improved understanding performance.

MS introduced AI secretary function as a core service. Through this, Copilot, which is installed in all MS products, monitors the user's email without separate commands and provides quick access to data or contacts. 'Team Copilot' collects data in real time based on MS 365 documents shared by the team, judges the progress of the project, and presents unresolved agenda items in the team's group chat room and video conference. This function will be provided in advance to customers with Copilot licenses starting in the second half of the year.

In addition, OpenAI's 'Pi 3' small AI model (SLM) was released. The three Pi 3s, Small, Medium, and Vision, are applied with on-device AI technology so that they can be easily run on mobile without an internet connection.

Meanwhile, on the 20th (local time), MS also unveiled a new PC equipped with Copilot, the 'Copilot+ PC'. Copilot+ PC is a PC that combines Copilot with the Windows OS (operating system), and is equipped with Qualcomm's Snapdragon X chip. In particular, it emphasized improved battery life by prioritizing low power consumption based on the Arm architecture.

■ Allganize, Presenting Corporate Generative AI Project Strategy

On this day, Team Leader Yoo Tae-ha of Allganize explained, based on case studies, ways to successfully settle a generative AI project. Companies are adopting generative AI and utilizing it for △Generative Answer △Generative BI △Summary △Snippet Generation (keyword-based advertisement generation), etc.

'Generative Answer' is a function in which AI searches for data from internal Q&A, Google, Bing, and other sites based on internal and external unstructured documents and DBs as if asking a question to a senior within the company. 'Generative BI' converts natural language queries into SQL and searches for and displays data from DB or BI. 'Snippet Generation' generates text desired by the user for various purposes such as advertising texts, notices, and emails. These are collectively called 'LLM Apps'. Unlike general mobile apps, this refers to apps with a chatbot UI/UX added to the end.

Team Leader Yoo said, “There are proven cases where companies that have introduced AI instead of people have reduced customer support costs by up to 95%, reduced response times from up to 45 minutes to 1 minute, and ultimately increased productivity.” He added, “AI is not a replacement for people in companies. “People who are good at using AI will replace people who are bad at using AI,” he added.

The key concept here is 'RAG (Retrieval-Augmented Generation)'. RAG is a technique that enables knowledge-based text generation, and is a strategy to complement LLM, which is weak in industry/company-specific questions.

Simply put, this means that LLM does not simply generate text with training data alone, but rather uses relevant information from an external knowledge base to reduce hallucinations and generate accurate and rich text.

For example, if a person asks AI about past financial events, if the model does not know the information because it has not learned it before, it can answer by providing context. Although LLM is still in its early stages and often shows hallucinations while using RAG, it is considered a fact-based search method that can minimize hallucinations by periodically executing the latest information updates.

'RAG' is based on three core elements: 'Retriever', 'Document Ingestion', and 'LLM'. In the process of searching (Retriever) from a specified source, documents must be processed so that AI can understand them. After searching, the prompt is augmented with the context searched from the source (Document Ingestion), and the model and augmented prompt are used to generate (LLM).

Team Leader Yoo explained, “The reason why RAG is needed in companies is to maintain the latest information when there are frequent updates in general fields. On the other hand, if the field that the company deals with is special, has little information, and is private, fine-tuning will be necessary.”