Comment by bonoboTP

Fixing existing bugs left in the codebase by humans will necessarily be harder than writing new code for new features. A bug can be really hairy to untangle, given that even the human engineer got it wrong. So it's not surprising that this proves to be tough for AI.

For refactoring and extending good, working code, AI is much more useful.

We are at a stage where AI should only be used for giving suggestions to a human in the driver's seat with a UI/UX that allows ergonomically guiding the AI, picking from offered alternatives, giving directions on a fairly micro level that is still above editing the code character by character.

They are indeed overpromising and pushing AI beyond its current limits for hype reasons, but this doesn't mean this won't be possible in the future. The progress is real, and I wouldn't bet on it taking a sharp turn and flattening.