I'd be particularly interested in if this has changed your thinking on the effectiveness of compute governance as an AI governance agenda, if either (1) China waking up to AGI just means they are willing to spend disproportionately large amounts on compute or (2) compute governance is pointless unless you do the 99th percentile of things (large BIS budget, global collaboration, on-chip mechanisms, and more).
I think of compute governance as quite a broad topic, covering any attempt to monitor and control AI capabilities via compute. I guess you mean the power of export controls in particular. I didn't realise most one could impose via export controls on China is a ~3x increase in costs (by keeping them ~2 generations of chips behind), which isn't nothing, but is less than I expected.
More useful analysis on this topic from Epoch: https://epoch.ai/gradient-updates/us-export-controls-china-ai
I'd be particularly interested in if this has changed your thinking on the effectiveness of compute governance as an AI governance agenda, if either (1) China waking up to AGI just means they are willing to spend disproportionately large amounts on compute or (2) compute governance is pointless unless you do the 99th percentile of things (large BIS budget, global collaboration, on-chip mechanisms, and more).
I think of compute governance as quite a broad topic, covering any attempt to monitor and control AI capabilities via compute. I guess you mean the power of export controls in particular. I didn't realise most one could impose via export controls on China is a ~3x increase in costs (by keeping them ~2 generations of chips behind), which isn't nothing, but is less than I expected.