I think the flaw in this argument is that these AIs are plateauing. Researchers are finding new ways to get these AIs to do more things, to generalize their ap

I think the flaw in this argument is that these AIs are plateauing. Researchers are finding new ways to get these AIs to do more things, to generalize their application, but I do not think the actual usable, general intelligence they can use to do those things is growing substantially. They are amazing tools, but AIs need more-general intelligence to fully automate software development, and that will require different AIs, not more Terawatts.