LG Uplus is set to enter the AI business-to-business (B2B) market next year, leveraging its artificial intelligence (AI) token optimization strategy. The company aims to dominate the market by enhancing AI agent performance while reducing token costs through proven technologies.
During an AI Token Optimization Study held on November 12, LG Uplus Chief Technology Officer (CTO) Lee Sang-yeop stated, "We have validated numerous practical services, moving beyond theoretical efficiency." The company plans to launch its token business next year through partnerships with collaborators and the establishment of its own infrastructure.
Lee emphasized the ongoing efforts to optimize token costs during the process of user inquiries and responses.
Kwon Gi-deok, head of the AI Transformation (AX) Engineering Lab, presented the results of the token optimization research, explaining, "By patterning question types during the harness engineering process, we can achieve over 85-90% efficiency by reducing exploration and evaluation repetitions."
In the model routing phase, the company reported a 25% reduction in general question processing costs and over 30% savings for coding agents. They are also developing technology to prevent unnecessary searches and validations during AI agent operations.
LG Uplus is working to streamline its AI models and increase response generation speed. The size of its on-device AI model has been reduced from 2.4-3 gigabytes (GB) to approximately 800 megabytes (MB), about one-third of its original size. Kwon noted, "We have improved performance to 95% compared to other big tech clouds, and our token production is significantly higher than competitors."
Additionally, the company has advanced technology that allows smaller models to generate responses quickly while larger models handle validation. Kwon also mentioned that two papers from their research were accepted at the AI conference NeurIPS.
At the infrastructure level, LG Uplus is pursuing a Chip-Mix strategy, collaborating with NVIDIA and AMD for graphics processing units (GPUs) and working with domestic AI semiconductor company Furiosa AI for over two years to validate neural processing units (NPUs).
CTO Lee stated, "We plan to distribute tasks that require large memory and computation, as well as those that depend on memory bandwidth, across cost-effective semiconductors to enhance efficiency."
The company attributes its success in token optimization to its experience in managing large-scale traffic, exemplified by its AI call agent, ixi-O.
Initially, LG Uplus utilized both its on-device model and Google Gemini during the launch of ixi-O, but after practical tuning, it combined its own AI model with EXAONE to improve service quality and cost efficiency.
The AX team is also applying token optimization technology within the company. Developers are using their own AI model in coding tools to maintain existing work environments while reducing AI usage costs.
In collaboration with Kakao on the 'Everyone's AI' project, LG Uplus is responsible for operating the AI models. CTO Lee remarked, "While Kakao handles the front-end, we manage the model area, and various recipes and optimization methods are being researched to develop diverse token factory service technologies."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.