We are no magicians. We cannot create code which does not exist. There is simply not enough Delphi code around for AI training. It is easy to have good coverage for JavaScript and similar where you literally have bazillion web pages available for scraping, where plenty of them virtually repeat the most common, required functionality. Pushing for more publicly available code without considering its quality, can also backfire.
What we need is better non-AI code completion. If you need to generate larger chunks, then you don't have to do that directly within the IDE. Also you can easily use some other editor, like VSCode to give AI access to context and generating code, and then simply reload changed files in IDE. This works fine in both ways. This is not a showstopper.
Ditching VCL would be the most stupid idea ever. It would be suicidal. There are huge amounts of code out there that use VCL, and moving all those to FMX would be impossible. And this is not just about old code, people use VCL for writing new code, too. Because they already have all the other infrastructure built around VCL. I am certainly not going to start new Windows application based on FMX, unless I really need some of its features.
What could help this transition would be restructuring VCL and FMX to use common Application layer which would enable mixing VCL and FMX frameworks in the same application. However, this is also something that is not very likely to happen as both frameworks are rather mature at this point and such restructuring could have impact on backward compatibility. Maybe having support for multiple helpers in scope and opening up private parts of VCL and FMX allowing more customizations from the ground up could help in such transition. But this would be long term and slow process.