Language is incredibly complex. I remember a TikTok from a bilingual English-Korean speaker comparing the English subtitles from a Squid Game scene to what was actually being said by the characters. The nuance and info density lost in translation made the subtitles feel completely remedial. Americans were basically watching a different show altogether.
Agree. I find this route reliably produces better results. Not sure I understand why though. Intuitively I'd think the models would be able to do approximate designs with higher fidelity using code as the primary reference.