What:
High Performance Computing has been a key driver to the success of AI. From -> , compute for AI Training has increased by a factor of 350 million. Governments could:
- Track compute used by a company
- Limit / subsidise compute for certain AI projects
- Enforce rules / guardrails.
It’s Easy To Govern:
It’s easy for governments to monitor how advanced AI systems are becoming because they can see the 10’s of thousands of GPUs going towards these projects. How easy?
- Detectable
- Excludable: If the government wants to stop someone in particular from accessing thousands of chips, they can.
- Quantifiable: Think energy as well as chips
- Supply chain is incredibly specific / concentrated.
But Should We?
- As Algorithms improve, the advanced models will require less compute (think Llama 2→3).