When reading this, I was reminded about hardware-enabled governance mechanisms (HEMs). AI hardware that remains secure independent of who gains access to it would enable further proliferation.
It seems like a natural solution path, that more aware middle powers could invest in
Hardware-enabled mechanisms to govern compute exports feel like a bit of a double-edged sword for middle powers though.
The pro-HEM case is that you don't strictly need export controls for security reasons anymore, because the HEMs address all the concerning aspects of exporting. But I think the main reason to do export controls is not security/safety, but economic/foreign policy leverage. (I think security/safety is the *better* reason, but I don't think decisionmakers agree). So they keep export controls / leverage up even though you have HEMs ready. Interestingly enough, this might even be an incentive against developing HEMs - success would take away your 'security' reason for leveraging compute export controls.
But if all that's true, HEMs make that leverage even more disadvantageous to middle powers: If I get my H100 equivalents up and running as a middle power today, I'm somewhat independent of whomever I imported that from. But if they have some remote oversight over what I can and can't do with these GPUs, it feels like I'm in an even worse position, because even how I use my GPUs is up to their external control?
Good points, although to me it seems difficult to solve the diffusion problems for really powerful AIs (e.g. with significant misuse / loss-of-control risks) without HEMs. Maybe a bad bargaining position is better than no bargaining position?
What other options are there?
I mean, proposals like the Narrow Path (https://www.narrowpath.co/introduction) suggesting that transformative AIs should be built under a single secure international project could potentially alleviate diffusion problems through making powerful AIs available to everyone through API, reducing dependence on US/Chinese compute. But I don't see why the US or China would be sufficiently motivated to go for this.
When reading this, I was reminded about hardware-enabled governance mechanisms (HEMs). AI hardware that remains secure independent of who gains access to it would enable further proliferation.
It seems like a natural solution path, that more aware middle powers could invest in
Hardware-enabled mechanisms to govern compute exports feel like a bit of a double-edged sword for middle powers though.
The pro-HEM case is that you don't strictly need export controls for security reasons anymore, because the HEMs address all the concerning aspects of exporting. But I think the main reason to do export controls is not security/safety, but economic/foreign policy leverage. (I think security/safety is the *better* reason, but I don't think decisionmakers agree). So they keep export controls / leverage up even though you have HEMs ready. Interestingly enough, this might even be an incentive against developing HEMs - success would take away your 'security' reason for leveraging compute export controls.
But if all that's true, HEMs make that leverage even more disadvantageous to middle powers: If I get my H100 equivalents up and running as a middle power today, I'm somewhat independent of whomever I imported that from. But if they have some remote oversight over what I can and can't do with these GPUs, it feels like I'm in an even worse position, because even how I use my GPUs is up to their external control?
Good points, although to me it seems difficult to solve the diffusion problems for really powerful AIs (e.g. with significant misuse / loss-of-control risks) without HEMs. Maybe a bad bargaining position is better than no bargaining position?
What other options are there?
I mean, proposals like the Narrow Path (https://www.narrowpath.co/introduction) suggesting that transformative AIs should be built under a single secure international project could potentially alleviate diffusion problems through making powerful AIs available to everyone through API, reducing dependence on US/Chinese compute. But I don't see why the US or China would be sufficiently motivated to go for this.