Regulating AI currently looks like a race to set barriers so high that only trillion-dollar companies can climb them. To avoid this, we should shift focus from regulating the entity to regulating the compute and the output. Instead of demanding massive, expensive auditing bureaucracies for every startup, governments could implement hardware-based tracking or usage-based licensing. For instance, monitoring large-scale GPU clusters at the chip level allows for oversight based on actual training scale rather than the size of the company's legal department.
We could also adopt a tiered regulatory framework. Small researchers working on niche architectures shouldn't face the same scrutiny as a lab training a trillion-parameter model. By using scalable thresholds, we ensure that safety requirements grow proportionally with the potential risk of the model. Open-source development also serves as a vital counterbalance; promoting transparent, verifiable safety benchmarks allows the community to audit models without requiring a central authority to oversee every line of code.
Lastly, funding public compute infrastructure can prevent a monopoly on safety research. If government-funded labs have the resources to test frontier models, safety becomes a public utility rather than a luxury good reserved for the biggest players in Silicon Valley.
