If your pseudo code, sorry, I mean fine detail "architecture", is that comprehe dive you might just as well write actual code and let it fix that. IMO however AI code often doesnt even compile, has logical errors and is bloated. However YMMV.
Wrong approach. If you write code immediately, you will eventually get stuck in a rut, because an architectural concept shot from the hip typically doesn't hold water. I don't talk about javascript slop here barely holding a shitty web page together, I mean software that people actually depend on, and which you will have to maintain for years with updates and additions as the project scope moves and expands.
Once you sit down and actually design something, AI will help you by looking deeply into the system you try to build, trying to find use cases and edge cases. That's when you start refining your architecture. That's when software modules come and go, components split up and parts fuse with other parts. On each step a capable AI model will guide you through the process, highlight omissions, corner cases, inaccuracies and contradictions. Any code written at this point would be a complete waste of tokens. The refined product is so much better, that you will just shake your head looking at your initial concept. It's like a whole very capable team designed this product, when you actually did it all by yourself and just threw it at 3 or 4 different agents multiple times. That's an amazing asset for submarine projects BTW!
Second massive boost from AI is, that AI happily and quickly does all the boring stuff: advanced command line parsing, proper XML parsing, help texts. Your software suddenly does stuff, which it should and could have done years ago, but you never actually got around to do it.
I honestly believe, that we are going to see lots and lots of really bad AI slop software (or are already seeing it), but there is a strong potential to write really good software with a lot less effort than ever before.