KAIST Unveils SoulMateI, On-Device AI Chip for Ultra-Personalized, Privacy-Preserving Learning

A KAIST team in Daejeon has unveiled SoulMateI, an on-device AI semiconductor designed to learn a user’s preferences, speech patterns, and even emotional cues in real time. Led by Professor Yu Hae-jun of KAIST’s AI Semiconductor Graduate School, the project is described by the researchers as the world’s first ultra-personalized AI that operates entirely on a single device. The first author, Hong Seong-yeon, a PhD candidate, presented the results at a January 16 briefing on KAIST’s main campus.

Unlike conventional large language model accelerators that rely on cloud servers for ongoing learning, SoulMate integrates memory and learning directly into the chip. It employs retrieval-augmented generation, or RAG, to customize responses and a low-rank adaptation technique, or LoRA, to adjust the model based on user feedback, all inside the device.

The hardware centerpiece is a 28-nanometer chip. The system stores user conversations locally, enabling it to remember a person’s tastes and even their speaking style. As interactions continue, the AI progressively tunes its replies to the individual, with the design ensuring that no personal data leaves the device.

Efficiency is a standout claim. SoulMate can perform complex learning and inference at about 9.8 milliwatts, delivering responses in roughly 0.2 seconds. The researchers say the energy use is about 1/500 that of typical smartphone AI processors, highlighting a convergence of low-power design with on-device learning capabilities.

During the briefing, the team showcased the actual SoulMate chip and a demonstration board. They argued that all personalization happens within an on-device environment, addressing concerns about privacy when data is sent to the cloud and reducing latency that can occur with slow or unreliable network connections.

The work has drawn attention within the research community. At the International Solid-State Circuits Conference in San Francisco last month, the study was highlighted as a notable contribution. The eight-member team has reported strong interest from major U.S. technology firms, and Hong Seong-yeon is set to begin an internship at Nvidia in May.

KAIST plans to advance the project toward commercialization through its university-affiliated startup, Onnuro AI, with an eye toward productizing SoulMate around 2027. The initiative reflects Korea’s push to develop domestic, low-power AI hardware capable of running advanced AI features without reliance on cloud services.

For U.S. readers, the development matters because on-device AI chips could reshape how consumer devices handle privacy, latency, and data ownership, potentially reducing dependence on cloud-based inference for smartphones and wearables. If such edge AI hardware scales, it could influence supply chains for AI silicon, collaborations with American chipmakers, and the timing of deployments for platforms that prioritize security, user experience, and rapid responses. The international interest also signals a competitive acceleration in the global race to commercialize personalized AI hardware.