**Summary** Kobaidze argues that "vibe coding"—shipping AI-generated code without reading it—isn't engineering, and conflating the two obscures real risks. He defines three tiers: pure vibe coding (no review), AI-generated + human-reviewed (legitimate engineering), and AI-assisted (human-driven with AI help, also legitimate). The core claim: engineering isn't about who types—it's about who owns the reasoning, debug paths, and failure modes. The source doesn't say when this tipping point should occur technically, only that systems handling money or personal data need human accountability that review alone can't provide. The thread extends the argument usefully: real failure modes emerge at deploy and 2am oncall, not code diff; review catches logic bugs but not ops gaps; and the dangerous case is when someone retains a single accountability that spans code they can't reconstruct. One sharp tension: several commenters note that review still can't verify a system *behaves correctly under real conditions*—only that the code looks sound—yet the article treats review as the engineering bar. That gap matters for financial or health systems.