According to the Stanford video the only cases (statistically speaking) where that happened was high-complexity tasks for legacy / low popularity languages, no

According to the Stanford video the only cases (statistically speaking) where that happened was high-complexity tasks for legacy / low popularity languages, no? I would imagine that is a small minority of projects. Indeed, the video cites the overall productivity boost at 15 - 20% IIRC.