I believe that within five years, models more capable than many of today’s state-of-the-art systems will be able to run locally on ordinary phones. The exact timeline may be wrong. The direction feels much harder to dismiss.
This possibility interests me not only as an engineering achievement, but because of what it could change. Intelligence could become more private, personal, accessible, and less dependent on centralized infrastructure.
Privacy becomes architectural
When inference happens locally, sensitive context does not always need to leave the device. Privacy becomes less dependent on promises about data handling and more dependent on the physical boundary of the system.
A local model could understand private documents, personal routines, ongoing work, and preferences with a form of continuity that would be uncomfortable—or expensive—to reproduce entirely in the cloud.
Intelligence becomes available
Local systems can keep working through weak connectivity, reduce round-trip latency, and avoid a marginal inference cost for every interaction. Those properties matter wherever connectivity, economics, or responsiveness make cloud-only intelligence fragile.
The interesting question is not only whether a model can fit on a device, but what new products become possible once it does.
The full stack must move
Making that future practical requires progress across models, inference software, hardware, and system design. Better architectures alone are not enough. The software stack must exploit the hardware well; products must decide what belongs on-device, what belongs in the cloud, and how the two cooperate.
This intersection—between learning systems, efficient inference, hardware constraints, privacy, and product design—is where I expect many of my future questions to live.