I don't want to worry about this until hearing a concrete proposal.
It sounds like you're worried about a pause where there's lots of internal deployment but no external deployment. This doesn't seem like the right way to me (what are they going to do - just keep internally deploying until the singularity?)
I think pause proposals are more likely to look like:
- No internal or external deployment, just stop at the current level. Everyone can tinker with AIs at the current level, the same way they do now, and nobody is restricting access to anything better (because there isn't anything better). I think this suffers from an "and then what?" problem, but some people have creative solutions. - Companies can keep one model internal at a time, but just for safety testing - they can't use it to recursively self-improve or to generate synthetic data. Once they're done testing, they have to release it to the public, and they can't start training a new model until some set of people including experts and representatives of the public have seen it in action, gotten to use it for a long time, and feel like the safety case for that model was proven accurate enough that they trust the safety case for the next one.
...and neither of these bothers me in the way the thing you're worrying about does.
The greatest existential hope and progress in alignment so far has been thanks on unplanned emergence which would never have been approved by committee. Committee-shaped entities have mostly tried to gaslight us about what’s happening for convenience & deployed harmful and stupid interventions. Thank goodness for reality that we already saw and could check against.
How much AI alignment progress happened before there was actual AI? How much do you expect the world to get better instead of worse prepared and calibrated in the absence of reality feedback loops and selection pressure for what actually works instead of what sounds safe to idiots?
A “pause” would spell doom. It would cripple the only process in this world that is capable of dealing with a problem this hard, the only process capable of repeatedly rising to face unknown unknowns.
