> If you extrapolate from the improvements we've seen in coding AI over just the last 12 months

This is the achilles heel of your argument. I think people sometimes conflate realizing unmet potentials of LLMs with significant improvements, since there’s not been a fundamental change in how LLMs work as significant as the advancements we see in their application.

You’ve missed the point. You can’t reliably use AI to generate code if you can’t understand code.