If I want to create a web app with a back-end, database, and some services, and I tell codex to do that with a specific stack and using specific paradigms to keep the code performant and maintainable, it's still a win over coding it by hand, as models can emit ~200char/sec compared with maybe ~10 for a really fast human. There's up front planning cost, and you will have to go back and massage some of the outputs a little bit if you're particular, but for sizeable tasks it still comes out to be a big win.

If you're just working on a single react component or an algorithm to do stuff with data, there's less chance to amortize the up front planning and verification so it comes out more of a wash.

Even when you have a clear idea of what you want, there are still hundreds of decisions you need to make while building it, both big and small. Everything from what to name your database tables and columns to what data structures are optimal and what the API payloads should look like and what the tech stack should be. Anyone with a sufficient level of experience in this field has made these types of decisions dozens of times and at some point it becomes more practical to have an AI do it for you and for you to quickly skim it.

For example I want to make it so that users receive an email when their password is changed. I can either do it myself, which requires reviewing and remembering code I’ve written five plus years ago and then wiring everything up and obsessing over the wording of the email. Or I can give a two sentence instruction to the AI, work on something more meaningful while it is doing its thing, and then test it in under 60 seconds when it is done.