Pseudo-coding is very very important and useful during the ideation phase, when you want to try and explore a large possiblilty space of designs without committing to any specifics. You can do that because you don't have any integrations or SLA's or customers to satisfy - it's "greenfield". As soon as the rubber hits the road and you've got business partners who need commitments from you towards reliabiltiy so they can author a business plan they can sell to the board to get funding.... It's exceptionally risky to offload all your development to vibe coding.
As someone who lives and makes a living in the operational space where consequences exist for shitty and unexplainable code, I don't see that going super well for anyone. It's effectively the same thing as offshoring all your coding development to contractors, and then wondering how you ended up with such a brittle tech stack. I've been the sonofabitch unwinding that on a production outage call before, and when your only expertise for the code your business relies on is a friggin language model, not even humans a dozen time zones away.... That's not gonna result in a very resilient business model.
I'm very interested to be proved wrong, so that I can stop being responsible for that sort of thing, but there's such a titanic degree of trade secrets and organizational information classification that I just don't see a training dataset getting pulled together to train such a model. It's not that it's technically impossible, I just don't see how you get there as someone who is a technical specialist in LLMs and machine learning generally, and who actually has to realize tangible and deliverable outcomes