"AI-relevant compute is a particularly effective point of intervention: It is detectable, excludable, and quantifiable, and is produced via an extremely concentrated supply chain," the researchers argue.
These factors, along with others like supply chain constraints on semiconductor manufacturing, offer policymakers the means to better understand how and where AI infrastructure is deployed, who is and isn't allowed to access it, and enforce penalties for its misuse, the paper contends.The paper highlights numerous ways policymakers might approach AI hardware regulation.
They further expanded on that mechanism, and to us it sounds as though they are suggesting the accelerators could self-disable or be remotely disabled by watchdogs: Specialized co-processors that sit on the chip could hold a cryptographically signed digital “certificate,” and updates to the use-case policy could be delivered remotely via firmware updates. The authorization for the on-chip license could be periodically renewed by the regulator, while the chip producer could administer it. An expired or illegitimate license would cause the chip to not work, or reduce its performance.
For nuclear weapons, these security locks are designed to prevent one person from going rogue and launching a first strike. For AI however, the idea is that if an individual or company wanted to train a model over a certain threshold in the cloud, they'd first need to get authorization to do so. Compared to hardware,"other inputs and outputs of AI development – data, algorithms, and trained models – are easily shareable, non-rivalrous intangible goods, making them inherently difficult to control," the paper reads.