Where this ends is general intelligence though, where all more challenging tasks can simply be done by the model.

The scenario I fear is a "selectively general" model that can successfully destroy the field I'm in but keep others alive for much longer, but not long enough for me to pivot into them before actually general intelligence.