KAIST said on March 17 that a research team led by Yoo Hoe-jun (유회준), a professor at its Graduate School of AI Semiconductor, developed 'Soulmate', a personalised large language model (LLM) accelerator that evolves on its own to match user characteristics.

At the core of Soulmate is on-device AI technology that processes data on the device itself without going through an external server (cloud).

The team directly implemented in the semiconductor retrieval-augmented generation (RAG) technology, which creates customised answers based on remembered conversations, and low-rank fine-tuning (LoRA) technology, which learns by immediately reflecting user feedback. This enabled a real-time personalised AI system that responds to users in 0.2 seconds while learning at the same time.

It also applied a mixed-rank architecture that optimises processing methods based on the importance of information, sharply reducing power consumption. The chip can perform complex learning and inference simultaneously with ultra-low power of 9.8 milliwatts, about 1/500 of the power consumption of a smartphone processor.

In particular, it implemented a 'security-complete AI' structure in which all personal data is processed only inside the device and is not transmitted to an external server, fundamentally blocking concerns about personal data leaks. The team said it expects the technology to combine with next-generation platforms such as smartphones, wearable devices and personal AI devices to open an era of truly personalised AI services.

Yoo said, "This study lays the technological foundation for AI to develop into a user's true companion by imitating the process of people building friendships with each other."

The study, with doctoral student Hong Seong-yeon (홍성연) participating as first author, was selected as a 'highlight paper' at the International Solid-State Circuits Conference (ISSCC) held in San Francisco in the United States. The Soulmate AI semiconductor is expected to be commercialised around next year through faculty startup OnneuroAI.