Discussion about this post

User's avatar
Felix Choussat's avatar

Does the team have any thoughts on what the best proxies for the different VCAD categories would be? Suppose you were an American inspector monitoring a Chinese datacenter and unwilling to trust their individual or government definition of, say, internal or external inference. What could we check for instead?

I think verifying compute allocations could be extremely valuable, but I’ve been blocked on clear ways to demarcate compute use as “for safety/capabilities R&D/external customers/internal agents/experiments” outside of company judgement.

Jonas Vollmer's avatar

> Slowing software progress, rather than only delaying training runs, makes pacing much more effective.

Could this be a false dichotomy? E.g., if you impose the requirement "all training and experiments must be done using [specific safety measure that comes with an alignment tax, such as monitorable/legible CoT]", you might end up with most of the desirable properties of VCAD (no overhangs), but also you made your systems safer.

Or pacing via other desirable features, e.g. companies agree to offer security bug bounties in the amount of $100*[ECI-150]^3 for white hat hackers obtaining algorithmic secrets (and 100x more for exfiltrating model weights).

9 more comments...

Ready for more?