short non-fiction · tech & the future
Some AI may move closer to the device
Smaller language models create a different design space from ever-larger remote systems.
A language model need not be the largest available to be useful. Compression, quantisation and task-specific tuning can reduce memory and computation enough for some models to run on phones, laptops or local servers. That can lower network delay and keep selected data on the device, while a larger remote model handles harder requests.
Small does not automatically mean private, efficient, accurate or safe. The surrounding app may still transmit data; hardware manufacturing and repeated use consume resources; benchmark gains may not survive a particular task. A plausible future is a layered system that routes work between local and remote models. Its benefits would depend less on the label “on-device” than on disclosed data flows, measured quality and user control.
Based on the work of
Mustafa Suleyman with Michael Bhaskar
· 2023
Why this bookA practitioner’s broad argument about powerful, proliferating systems—best read critically beside technical evidence.
What to keep in view
Model size alone establishes neither privacy nor energy benefit; evaluate the complete system and task.