“AI is evolving from data-centric to human-centric, and from software-centric to hardware-centric. “As technologies like neuromorphic chips become commercialized, hardware-based AI that mimics the human brain will emerge, potentially marking a new turning point for on-device AI.”

Professor Kim Gyeong-gi of Daegu University presented on 'On-Device AI Technology Trends and Future Outlook' at the '2025 e4ds Tech Day' event held on September 9th, and emphasized the importance of hardware and small models, saying that AI is changing from the cloud to devices.

As digital transformation accelerates, the evolution of artificial intelligence (AI) is entering a new phase.

In particular, 'on-device AI' that enables learning and inference on the device itself is attracting attention, moving away from cloud-centric, centralized AI.

Professor Kim Gyeong-gi reviewed the development of AI technology, mentioning the full-scale start of deep learning in 2012, the emergence of AlphaGo, and the recent large-scale language model (LLM).

Professor Kim emphasized that “now, beyond simple generative AI, revenue models will be formed with agent-based AI,” and that on-device AI will play a key role in this.

On-device AI processes data on the device itself without going through a server, providing fast response and high security.

Professor Kim said, “We need AI that can respond in real time while protecting personal data,” and explained that reducing reliance on the cloud is essential in areas such as self-driving cars.

A video presented during the presentation showed a food scientist using AI to provide personalized support based on his work style and data.

Professor Kim cited this as an example, saying, “AI understands me and only I can understand it.“The era has arrived where we operate based on data,” he said.

A representative example is a service that learns and creates content based on personal files, such as Google's 'Notebook LM'.

This trend leads to ‘hyper-personalized AI.’

AI, which learns users' habits and preferences to provide customized services, is expected to play a significant role in resolving future security and privacy issues.

Hardware advancements are essential for implementing on-device AI.

Professor Kim pointed out that domestic small and medium-sized enterprises are having difficulties developing commercial-level chips, saying, “Low-power, high-efficiency chips such as MPUs and neuromorphic chips must be supported.”

“It’s not just performance that’s important, but power efficiency and peak power management,” he added.

On-device AI requires a “small language model (SLM)” rather than a large model.

SLM is a model with fewer than 8 billion parameters, has low inference cost, and is suitable for running on devices.

Professor Kim emphasized, “On-device AI can become a reality through SLM.”

Professor Kim noted that the United States and China are leading the way in AI technology by investing massive amounts of capital, and diagnosed that “it will be difficult for Korea to win the money game.”

On the other hand, if you develop specialized agents or on-device AI in a specific field, you can be quite competitive, he predicted.

He said, “On-device AI has various advantages such as real-time processing, low latency, reduced internet dependence, and reduced operating costs.”He emphasized that “this is a technology that must be paid attention to in order to secure long-term cost efficiency and autonomy.”