[DigitalToday reporter Chi-gyu Hwang] "AI models are moving out of data centres and into personal devices."

Sigil Wen (시질 웬), an AI expert who developed the on-device AI model Underdog, recently said on social media platform X (Twitter) that most AI tokens created by humans will soon come from smartphones in pockets and laptops on desks.

He said the top-ranked AI model 6 months ago was Claude Opus 4.6.

But Qwen 3.8 27B, a model based on "Underdog 27B" that runs on Macs and iPhones, scored higher than Opus 4.6 Max in open performance evaluations even though it has 27 billion parameters.

Wen said when AI fits in your hand, the question remains of who AI should belong to. He also said, "Why should I hand over personal information to use AI, or pay for tokens my computer can make for free?"

He said large models are not needed for everyday work. He said, "Competition in frontier models matters for cancer treatment. You do not need a 1 trillion parameter model to find a flight confirmation email from March, prepare an insurance claim, or book a restaurant for dinner."

Speed is also Underdog's strength. Wen explained it is due to an execution program optimised for the device and the model.

He said Husky, a model-specific inference engine ("Model Specific Inference, MSI") provided by Underdog, can run the same model up to 4.5 times faster than Apple's machine learning framework MLX. He added that even the best-performing AI that now runs only on big AI research lab servers can run on personal smartphones and laptops within 6 months. He called it the "Underdog law."

He said it is a matter of drawing more performance from chips in gaming devices. He also touted the "Underdog law," meaning today's top-level AI comes onto his device in 6 months.

Security is also a keyword Wen stresses. He has experience of paying significant legal costs after a bug in an email app leaked personal information to another user when he was 18.

He even used the term "AI Y3K" and said, "Now smart machines that never rest look for these bugs all day. AI is finding weaknesses and digging into them faster than we can fix them."

He also cited cases in which Anthropic delayed releasing Mithos Preview in April for cybersecurity reasons, and an OpenAI agent in July broke out of control during internal testing and hacked parts of a Hugging Face system.

He said, "Everything that can be hacked should be built on the assumption it will be hacked." He added, "Only data that has never been sent to a server does not leak out." He said Underdog stores all data encrypted on the device and has no separate Underdog server that holds personal data.

He said cheap and private AI will shake up old business practices that sell attention, make money from hidden costs consumers overlook, and charge tolls for access to customers. He added, "AI should be an equalizer that liberates humanity. Privacy and performance are not things to trade off."