[Maui, United States = DigitalToday reporter Daegeon Seok] On-device agents will need cloud-level quality and user context that links devices to be widely used, according to an assessment. Models running on devices currently have a quality gap versus cloud models, and agents must learn a user’s preferences separately on each device. Qualcomm CEO Cristiano Amon (크리스티아노 아몬) and LiquidAI CEO Ramin Hasani (라민 하사니) raised the issue in a keynote dialogue on Tuesday, the second day of the Snapdragon Summit, and proposed a memory layer that links context across devices as a solution.

Hasani first pointed to a gap between AI capabilities and real value. "Personal productivity boosted by current AI models is far ahead of corporate change," he said, presenting "production-grade AI" that connects capability to value as a task.

He cited quality as a problem, saying five things must change for agents to become the computing platform for billions of devices. "One reason the device AI and local AI revolution has not happened yet is the quality bar," Hasani said. "We need to narrow the gap between cloud frontier-level performance and devices."

Linking context is also important. Users want to move the AI experience and preferences they used on one device to another, and that requires a memory layer that ties devices into one, he said. Hasani stressed that trust and personalisation need to be handled on-device to be embedded in context.

"The reason I trust this agent is because it runs on my device," Hasani said. "Scale increases capability, and adaptation gives personalisation." Amon echoed the point, saying, "If you understood what Hasani said, you will know why it is pointless to argue about cloud versus edge." He added, "In the end everyone will have their own device, and what they do with that device is unique to each user." Hasani saw a hybrid environment that complements cloud-focused attention with local AI as the next opportunity.

LiquidAI’s proposed solution is "Liquid Context on Snapdragon." The product consists of two layers: a memory layer that reads, compresses and records accumulated experiences while a user uses a device, and an agent layer that receives this information and processes tasks. The recorded context can be moved between devices so that AI on another device can inherit it. "This innovation made the agent know when it should deliver value," Hasani said.

He also said the previous generation approach is having users tell a system to do work, and the goal is to process tasks before a request. This process is handled across multiple computing units in Snapdragon. According to Hasani, the agent runs on the Hexagon NPU, and the CPU handles agent sampling.

AUTOMOTIVE, SMARTPHONE AND WEARABLE OEMS TARGETED; ALREADY APPLIED IN MERCEDES-BENZ VEHICLES

Qualcomm explained in a pre-briefing a structure in which the Oryon CPU handles planning and orchestration and the Hexagon NPU handles AI inference, which was interpreted as an example of a partner applying it in a product. Hasani cited as the reason for choosing Snapdragon that as use cases become more sophisticated, they need to be near devices, and users must always be able to access systems that keep collecting information. LiquidAI said it designed its second-generation Liquid Foundation Model, or LFM, to fit Snapdragon.

LiquidAI supplies the product to OEMs in a B2B form. Hasani introduced it as "an example OEMs can turn into a final product" and said it would apply it across automotive, smartphones, laptops and wearables. In cars, it is already working with Mercedes-Benz to implement fully local in-vehicle intelligence inside Snapdragon-based vehicles.

"This is production-grade AI actually deployed on Snapdragon chips," Hasani said. "Automotive OEM interest is high in changing the in-vehicle experience." He added, "In mobile and wearables, we see shared context across devices and proactive agents as the biggest opportunity."