[saxenaabhi]

I'm not sure you can make that assumption even when a human wrote that code. LLMs are competing with humans not with some abstraction.

> The issue is that if it's struggling sometimes with basic instruction following, it's likely to be making insidious mistakes in large complex tasks that you might no have the wherewithal or time to review.

Yes, that's why we review all code even when written by humans.