For example, it could hypothetically do this via inventing a new, more powerful technology which might be more useful than transformers for AGI, and refocusing AI alignment research on this technology, saving AI alignment research months or years or decades of misdirected effort.

Or it could hypothetically do this by providing technology which significantly speeds up AI alignment research, and which did not already exist.

The capabilities research does not have to be net-positive, or actually used in a net-positive way, or even invented, in 2023. It only has to be seen as net-positive eventually.

Mar 5, 1:06pm: Will AI capabilities research done in 2023 or after be net-positive for AI alignment research? → Will some piece of AI capabilities research done in 2023 or after be net-positive for AI alignment research?