LLMs/google and whatever articles I've read have failed me on this, so I'm hoping to find someone with some insight on this simple question. How is compute distributed in research? Sutskever's 20% for superalignment is confusing to me. 20% …