SEOUL, March 17 (AJP) - Researchers in South Korea have developed a specialized artificial intelligence chip capable of learning a user’s specific speaking style, preferences, and emotions in real time. The Korea Advanced Institute of Science and Technology (KAIST) announced on March 17 that a team led by Professor Yoo Hoi-jun at the Graduate School of AI Semiconductors has created SoulMate, an ultra-personalized Large Language Model (LLM) accelerator.
While popular AI services like ChatGPT can answer general questions, they often lack awareness of a specific user’s daily habits or past conversation contexts. SoulMate overcomes these limitations by functioning as a digital companion that evolves through direct interaction. The system utilizes on-device AI technology, meaning all data is processed locally on the hardware rather than being sent to external cloud servers.
To achieve this, the research team integrated Retrieval-Augmented Generation (RAG) to recall past conversations and Low-Rank Adaptation (LoRA) to reflect user feedback instantly. These technologies allow the chip to respond and learn simultaneously in just 0.2 seconds. This speed ensures that the interaction feels natural and fluid for the user, as delays longer than 400 milliseconds typically disrupt the flow of human-like conversation.
The semiconductor is designed for extreme energy efficiency, consuming only 9.8 milliwatts of power. This is approximately 1/500th of the power typically consumed by a standard smartphone processor. This low power requirement enables mobile devices to run complex AI learning and inference tasks for 24 hours a day without significantly draining the battery.
Privacy is a central feature of the new design. Because all personal data remains on the device, the risk of sensitive information leaking from a server is eliminated. The researchers expect this technology to be integrated into future smartphones, wearables, and personal AI devices to provide secure, tailored services.
"This research establishes the technical foundation for AI to become a true companion by mimicking the way humans build friendships," said Professor Yoo Hoi-jun. "Future AI will move beyond being a mere tool to become a "best friend" that understands the user best while perfectly protecting their privacy."
The study, featuring lead author and doctoral researcher Hong Seong-yeon, was recognized as a Highlight Paper at the International Solid-State Circuits Conference (ISSCC) held in San Francisco last month. The research team successfully demonstrated the chip at the conference, showing the AI’s speaking style changing in real time based on user reactions. SoulMate is expected to be commercialized around 2027 through the faculty-led startup OnNeuro AI.
(Reference Information)
Journal/Source: IEEE Xplore / 2024 IEEE International Solid-State Circuits Conference (ISSCC)
Title: SoulMate: A 9.8mW Mobile Intelligence System-on-Chip with Mixed-Rank Architecture for On-Device LLM Personalization
Link/DOI: https://ieeexplore.ieee.org/document/11409048
Copyright ⓒ Aju Press All rights reserved.