That model collapse argument assumes pre-training teams are just scraping raw web garbage without curation.