The AI assistant have some bad bugs which you must to fix:
Broken HTML Tags: The AI does not correctly place < and > between tag names sometimes, resulting into broken results like <p>Textp>. Unexcepted Halt: If the request needs a long result, sometimes (not always) the AI will halt generation anywhere - mostly after 200-300 lines generated.
Rarely, when I edited a working code without broken tags, it turned all tags into broken tags.
And it’s better to improve the coder performance, and even maybe increase priority/speed of generation for long-standing users.
And you said reducing image generator quality/speed for text generation?
An old issue, started before the broken tags/unexcepted halt issue: If the code is too long (e.g., over 1,000 lines), it errors with doctype error (Sorry, there was an error with the AI helper: Unexpected token ‘<’, "<!DOCTYPE "… is not valid JSON)
And even older issue: Sometimes AI will add html in top of code, thinking it’s in a chat, formatting code as HTML. Even sometimes descript changes/code in top of ‘```html’!
I can’t help but bite: you know this fundamentally how all “AI” works, right? It’s going to spit out some output that may or may not bear any resemblance to what you actually want, because it does not think or have any intelligence. It’s merely parroting statistically-likely sequences of characters/pixels/whatever based on its training data, and those sequences can easily be incorrect or nonsensical.
Bogus output is par for the course, and that will never, ever change without a complete paradigm shift.