Kakao Reveals In-House LLM Development Journey and AI Agent Technologies at if(kakaoAI)2024

Kakao shares its current AI technology landscape and future vision through the if(kakaoAI) 2024 conference. This event features technical sessions categorized into AI Agent, AI Platform, Multimodal AI, and AI Safety.

In-House LLM Development and Serving Optimization

Kakao reveals the entire development process of its proprietary Large Language Model (LLM). The sessions address challenges and solutions from data collection to Pre-training and Post-training, while also introducing experiments and strategies to optimize inference costs. Additionally, they share experiences with implementing an LLM Serving Framework used in services like Kakao Pay, along with optimization techniques to reduce latency and costs.

AI Agents and Service Application Cases

Announcements include Context-Aware AI Mate, which understands user context, and a RAG-based bot developer for improving internal work efficiency. Through real-world application cases such as AI code reviews using CodeBuddy, the webtoon preview service Helix Shorts, and an LLM-powered conversational wine recommendation service, the event showcases an AI Native environment.

Multimodal AI and Enhanced Safety

Building on the achievement of its proprietary multimodal LLM being selected as a highlight paper at CVPR 2024, Kakao shares its training methods and insights. The event also introduces technology that improves performance by transferring the language capabilities of LLMs to speech recognizers. Furthermore, through the AI Safety Initiative, Kakao presents its efforts to provide safe AI services, including phishing detection, malicious user group detection, and spam response, as well as improvements in Explainable AI (XAI).