This is my first post here. Hi All.
TL;DR Verification of internal agreements will take a lot of people. I am worried because I have not seen any discussion on the topic of how many people would be needed.
Proposals for verification talk about what could be checked at data centers and in the supply chain so that trust is not required for an international agreement slowing or stopping AI R&D. I am worried that no one has thought about how many people this would require. I have not seen this discussed anywhere. I have read all (that I could find) of the published literature on the topic and not seen an inspector count mentioned. Based on some very back of the envelope calculations using mostly made up numbers I think it would require thousands of people to start a verification regime. Where are these people coming from? Is there any groundwork that can be done to help get a verification regime started as quickly as possible in case of an agreement on the personnel side?
Has anyone thought about how many people would be required for verification to enforce an international agreement to slow down AI progress as suggested by MIRI, plan A, ect.?
I have seen a lot of writing on technical means of verification and am aware of a number of groups doing R&D to make these mechisms possible. I have tried to think this through on my own and it seems like any of the proposed solutions would take thousands of people to implement. This calculation is a first attempt. I would welcome any feedback or if anyone would like to collaborate on a second, publishable version.
Based on the calculation I think it would be a good idea to have something like VERTIC (Verification Research, Training and Information Centre) which supports the governmental bodies that are doing international verification regimes. I think that any groundwork that could be done to stand up an inspectorate for data centers or the supply chain could speed the process of having a way to enforce a treaty. This could look like having a curriculum, running mock inspections, building an on-call roster, running a timing duty to see how long inspection tasks take in a test environment, and red-teaming verification mechanisms (both existing and new ones as they are developed).
I don’t have anyone to discuss this with in my immediate circle and would welcome any thoughts on this topic from anyone who has been thinking about verification.
