My experience is very very different. LLM code is often useless for Frontend stuff for me while it's okay for things like Java and Rust etc.
For frontend stuff the churn seems to be so big that I almost never get valid code because the LLM only know about some old versions or combinations of versions. The "statistical average" they learned is seldom correct for me.
I have a semi-technical friend who's built a whole SaaS with paying customers (currently $x,xxx/mo) using nothing but AI and an occasional bit of hand tuning.
He has some knowledge of the space and technical terminology from managing devs and projects, but was never a developer or engineer himself.
I watched his approach once and it's totally different from mine when I tried AI. He tells the AI what to do and if there's an error, he copy+pastes the error from the dev tools and says "fix this". And he'll keep cycling.
When this happens to me, I think too much about the error and perhaps I'm too specific to the AI on how to fix it. He just lets the AI do it over and over until it gets it right.
He doesn't care how the code looks, what libraries are used, etc.; the only thing that he cares about is "when I click this, does the right thing happen". It's actually kind of insane what he's built solo over a period of 4 months or so.
For frontend stuff the churn seems to be so big that I almost never get valid code because the LLM only know about some old versions or combinations of versions. The "statistical average" they learned is seldom correct for me.
I wonder what we do differently.