Have you read "Why does deep and cheap learning work so well?" It's referenced in "When and Why Are Deep Networks Better than Shallow Ones?", I liked the explanation of how the hierarchical nature of physical processes mean that the subset of functions we care about will tend to have a hierarchical structure and so be well-suited for deep networks to model.
For the Skunk Works and SpaceX examples, I did find myself wondering whether some aspects like the powerful decisive managers are strictly better or merely increase variance and so appear more often when looking at the most successful projects. I haven't done much reading of the primary and secondary sources for progress studies, how easy would it be to find details of the practices of average or failed projects to compare against?
The Curse of Reversal seems to match the lack of bidirectionality of ROME edits mentioned here: https://www.alignmentforum.org/posts/QL7J9wmS6W2fWpofd/but-is-it-really-in-rome-an-investigation-of-the-rome-model
RE: claim 25 about the need for research organisations , my first thought is that government national security organisations might be suitable venues for this kind of research as they have several apparent advantages:
However, they may introduce problems of their own:
Has this option been discussed already?
Vermeer's "Evaluating Select Global Technical Options for Countering a Rogue AI" mentions that "longer wires will pick up more of the energy from the pulse than shorter
wires do, so devices attached to very long cables are more vulnerable."
This made me wonder whether hardware modifications like attaching such cables to GPUs to increase their susceptibility to HEMPs would be worth exploring as a cheaper and less technically complex short-term alternative to the proposals for hardware-based chip shutdown options.