There is a massive gulf between "correct code" and "correct implementation" in many real-world scenarios.
Business logic and baking in domain expertise into your data model is most of the work. Making the code work efficiently doesn't matter if your code doesn't even do what it's supposed to.
Normally this is an argument in favor of human-in-the-loop LLM-based development -- "the human just needs to curate and verify!" However it seems all too easy to me (especially having witnessed it more than a few times) that subtle discrepancies emerge between the stakeholders' desires for the function of the code and the developers' understanding of those requests. Hopefully we reach a best-case scenario where that's all developers need to focus on, but more likely we'll see some pretty egregious things slip through the cracks (the wave will likely start with security/privacy issues before the phenomenon is recognized) as this technology matures into the common workplaces.
Business logic and baking in domain expertise into your data model is most of the work. Making the code work efficiently doesn't matter if your code doesn't even do what it's supposed to.
Normally this is an argument in favor of human-in-the-loop LLM-based development -- "the human just needs to curate and verify!" However it seems all too easy to me (especially having witnessed it more than a few times) that subtle discrepancies emerge between the stakeholders' desires for the function of the code and the developers' understanding of those requests. Hopefully we reach a best-case scenario where that's all developers need to focus on, but more likely we'll see some pretty egregious things slip through the cracks (the wave will likely start with security/privacy issues before the phenomenon is recognized) as this technology matures into the common workplaces.