Open-weight AI models are policy's hardest problem

Open weights are an irreversible release mechanism, and that one design choice now sets the terms of every AI policy fight that follows. The pattern across the last eighteen months, with DeepSeek in 2025 and Kimi K3 in July, is that the frontier no longer lives inside a few sealed labs. It has been copied to tens of thousands of machines worldwide, and once a model is duplicated at that scale, regulation cannot retrieve it.

Nageswaran's column turns that fact into a framework. He names three reasons blanket bans fail. First, the diffusion is already complete: any student or startup can download and run the weights for almost nothing. Second, the well-meaning users vastly outnumber the malicious ones and are the best operational defence. Third, prohibition would simply return the frontier to the largest incumbents and their governments, concentrating the very power the policy was meant to disperse.

The defensible move, in Nageswaran's framing, is to draw a line where the harm cannot be patched. A cyberattack built on an open model can be blocked, caught, and patched; a pathogen design cannot be recalled at scale. Biology is the one place where the asymmetry between release and repair is total.

The mechanism is portable: ask whether the model can still be recalled, count the well-meaning users against the malicious ones, and only draw a hard line where remediation is impossible.