Alignment pretraining: AI discourse creates self-fulfilling (mis)alignment Von Hackernews 18. Mai 2026 security ORIGINAL QUELLE:arxiv.org Quelle: Hackernews Comments Tags: 2tainment, comments, disclose, rethinking, Security