…. creating checkpoints, etc. …

Unfortunately I think a bootstrapping intelligence is closer to a nuclear explosion, than a building being built. Either prior learning fully supports the next “layer” to be learned, or it does not. If it does not, then learning will quickly peter out, since it is not capable of sustaining repeated building upon what has been learned before.

this is our current situation, and it is very very roughly like the study of fissile processes before we had processes that would yield more than one new fact (er’ I mean atom) from one prior one.

Just as with a nuclear explosion there is no good way to control it. you either have a process the peters out, or it blows up. So I think the check pointing you imagine will be pragmatically impossible, unless all such work was happening within very regulated test environments.

Not sure what the answer is. Seems like we must simply accept whatever outcome comes, or we must stop building bigger computers.