> 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).
Yes, in our taxonomy that's the "require safeguards" option. Agree that it's nice that your example doesn't create a software overhang. It would still be very helpful to have VCAD in place to enforce it well, so you can verify that all the relevant workloads are implementing hte safeguard.
> 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).
I haven't thought about this sort of bounty, I'm not sure it would naturally qualify under pacing? So seems out of scope for this post.
The economic cost of having a high capability level combined with poor security would incentivize companies to slow down unless they greatly improve their security.
Makes sense regarding VCAD being important for verification.
> 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).
Yes, in our taxonomy that's the "require safeguards" option. Agree that it's nice that your example doesn't create a software overhang. It would still be very helpful to have VCAD in place to enforce it well, so you can verify that all the relevant workloads are implementing hte safeguard.
(edited my above comment to add another example)
> 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).
I haven't thought about this sort of bounty, I'm not sure it would naturally qualify under pacing? So seems out of scope for this post.
The economic cost of having a high capability level combined with poor security would incentivize companies to slow down unless they greatly improve their security.
Makes sense regarding VCAD being important for verification.