LLMs have a codex of lots of examples of React apps embedded within them.

They can replicate them on demand, with tweaks for any given context.

LLMs have absorbed this knowledge by being bombarded with innumerable examples.

Techniques like RLAIF also allow them to be force-fed AI-generated and auto-ground-truthed examples.

This codex is at a slow pace layer; it takes months for the LLMs to be trained and then deployed.

It's also indirect.

In a world of infinite software you'll want a tiered system, with a faster, more direct loop that allows faster adaptation^[om], to complement the lower pace layer of LLMs.