I think its good not to overthink this idea of a self-replicating machine. In the GPT-4 paper, the system was given some access to some cloud apis and was able

I think its good not to overthink this idea of a self-replicating machine. In the GPT-4 paper, the system was given some access to some cloud apis and was able to do things like requisition more resources from the cloud. I imagine the first stage of "self-replication" for these LLMs or their successors is going to be like ordering CPUs from Amazon rather than setting up silicon fabs. You need to get far beyond single human level intelligence and effort to start building chips and we are probably not even there yet!