I use Codex, etc. for basic scaffolding, documentation, updating tests, double checking my approach/ideas, code review, rubber duck-ing. Its like how I used to use Google and Stack Overflow pre-AI but on steroids. But I can't get to a point where I can delegate real coding work to it. The problem is three-fold: 1. My code reading speed might very well be slower than my code writing speed. When I read the AI generated code I have to adopt/internalize the AI's mental model of the code flow and then reason through the correctness of the code, which (for a language I'm familiar with) is about the same as just... writing it myself. 2. The AI doesn't "get better" the same way as humans do. When I delegated to junior engineers at Mega Corp I felt like I was getting something out of it, even if it would have been faster to code the thing myself. The person would grow and eventually they would get to the point where I could actually hand them a spec and not have to manually review every line of their code myself. Not true with LLMs. I still need to manually trudge through GPT5.5's code, just as I had to verify all of GPT o3's code. 3. I view my verification harness as a layer on top of my own understanding, and not a replacement of it. Are there people out there working on non-trivial code who have integration/end-to-end/functionality tests that truly cover every possible combination of user state/thread state/load/behavior patterns/etc.?
And yet, I'm in the minority, because when I ask my former coworker at Mega Corp how they're using AI in their coding flows, they say the vast majority of code is AI-written and that the productivity gains are huge. So where are the gains coming from? The only way I can imagine getting _massive_ productivity gains in my development flows is if along with offloading coding tasks, I also offloaded understanding/comprehension to AI, and that's a hard leap for me to make.