KAIST researchers have developed the world's first AI semiconductor that learns and mimics a user's speech patterns, preferences, and emotions in real time—essentially creating a digital "soulmate."
KAIST announced on the 17th that a research team led by Professor Yoo Hoi-jun at the Graduate School of AI Semiconductor developed "SoulMate," a personalized large language model (LLM) accelerator that evolves to match individual user characteristics.
The technology is considered a core semiconductor breakthrough that will accelerate the era of "hyper-personalized AI for me" beyond the current "AI for everyone," learning and responding to users' conversation styles and preferences.
Large language models like ChatGPT that we use daily can answer countless questions proficiently, but they cannot remember users' minor habits or previous conversation contexts. This explains why AI still feels like a stranger despite being deeply integrated into our lives.
The key to "SoulMate" is on-device AI technology that processes data directly on the device without going through external cloud servers. The research team implemented Retrieval-Augmented Generation (RAG) technology, which generates customized responses based on stored conversations, and Low-Rank Adaptation (LoRA) technology, which immediately reflects user feedback for learning, directly within the semiconductor.
This enables "SoulMate" to achieve a real-time personalized AI system that responds to users in a remarkable 0.2 seconds (216.4ms) while simultaneously performing learning.
The team also applied a Mixed-Rank architecture that optimizes processing methods based on information importance, dramatically reducing power consumption. The semiconductor can perform complex learning and inference simultaneously with just 9.8 milliwatts of ultra-low power—1/500th the power consumption of smartphone processors—enabling operation on mobile devices without battery concerns.
Notably, the "security-complete AI" structure ensures all personal data is processed only within the device without being transmitted to external servers, fundamentally eliminating privacy breach concerns.
The research team expects this technology to usher in a true personalized AI service era when combined with next-generation platforms including smartphones, wearable devices, and personal AI devices.
"This research establishes the technical foundation for AI to evolve into a true companion for users by mimicking how people build friendships with each other," Professor Yoo said. "Future AI will go beyond being a simple tool to become like a 'best friend' who understands me best anytime, anywhere, while perfectly protecting personal privacy."
The research, with doctoral researcher Hong Sung-yeon as first author, was selected as a "Highlight Paper" at the International Solid-State Circuits Conference (ISSCC) held in San Francisco starting February 16th, drawing attention from academia worldwide.
At the conference, the research team demonstrated the semiconductor chip successfully changing response styles in real time based on user reactions, proving the excellence of Korean AI semiconductor technology. The "SoulMate" AI semiconductor is scheduled for commercialization around 2027 through faculty startup "OnNeuroAI."