Personal Hardware AI Models Lag Datacenter Versions by Years
Local models require optimization for intelligence per token per second.
848174.7K
“While the open/closed model gap is on the order of months, the runnable-on-datacenter-hardware vs. runnable-on-personal-hardware gap is on the order of years... For many purposes this doesn't matter, but for some it does (e.g., emergency situations, strong privacy requirements)”
Miles Brundage@Miles_BrundageTECH#57Story Brief
In a post on X, AI policy researcher Miles Brundage argued that the gap between open and closed AI models is measured in months, but the gap between models that run on datacenter hardware and models that run on personal devices is measured in years. He said that difference matters most in emergencies or under strict privacy requirements, when cloud access can vanish or become a nonstarter. In follow-up posts on X, he pushed the point further, arguing that losing internet access can set your AI help back by years.
“To make the point re: emergency situations bluntly, losing cell reception/wired Internet sets your (AI-augmented) intelligence back by years of AI progress, which is a lot + could be catastrophic in a world with atrophying skills + increased dependence on AI across sectors”
Miles Brundage@Miles_Brundage