South Korean researchers have developed the world's first artificial intelligence semiconductor that learns and reflects users' subtle habits, preferences, and speech patterns in real time.
KAIST announced on the 17th that a research team led by Professor Yoo Hoi-jun at the AI Semiconductor Graduate School has developed "SoulMate," a personalized large language model accelerator that evolves autonomously to match individual user characteristics.
The core of SoulMate lies in on-device AI technology that processes data directly on the device without routing through external servers or cloud infrastructure.
The research team implemented two key technologies directly into the semiconductor: Retrieval-Augmented Generation (RAG) technology, which generates customized responses based on stored conversation history, and Low-Rank Adaptation (LoRA) technology, which instantly incorporates user feedback for learning.
Through these innovations, SoulMate achieves response times of 0.2 seconds (216.4ms) while simultaneously performing learning—creating a real-time personalized AI system.
The chip also applies a Mixed-Rank architecture that optimizes processing methods based on information importance, dramatically reducing power consumption. Operating at just 9.8 milliwatts—1/500th the power consumption of smartphone processors—the semiconductor can perform complex learning and inference simultaneously, enabling battery-worry-free operation on mobile devices including smartphones.
The on-device approach fundamentally eliminates privacy concerns, as all personal data is processed internally without transmission to external servers, creating a "security-complete AI" structure.
