If I’m investing in frontier AI I would underwrite it mostly on new discovery and its applications like medicine because coding and most of knowledge work is quickly being saturated by sub-frontier models.

I asked Astra to look through my published science, identify new directions, and draft five NIH R01 proposals -- the multiyear, $1M-$2M grants that support much of academic biomedical science. In about one hour, using roughly $20 in compute, it produced entirely reasonable proposals. They were well grounded in my work and structured to appeal to study section reviewers. They looked very similar to what I would write if my goal were to get another grant. Astra reproduced that logic well. It turned existing work into questions, hypotheses, and experiments that would look familiar and defensible to a study section. The output was immediately recognizable as the work of an experienced academic scientist. If AGI means being able to produce the work of highly trained professionals, this is a clear concrete example. It produced something academic scientists spend a great deal of time doing. Academic medical centers build careers, laboratories, and their finances around this process. Scientists spend years writing these proposals as the next fundable extension of their work.