The Vibe Coding Apocalypse (systemsandsociety.com)

“Vibe coding” — letting large language models write your code — is already practical and uneven: demos look impressive, but real use exposes hallucinations (confidently wrong answers), verbosity and subtle bugs. LLMs are excellent for examples, API exploration and answering complex questions faster than search, yet not yet trustworthy for full production systems. A recent METR study captures this tension: developers report being 24% more productive using coding LLMs, while observed productivity fell by 19%, highlighting mismatches between perceived help and real outcomes. For the AI/ML community that means immediate gains in prototyping and knowledge work, but new risks in correctness, maintainability and verification. The longer-term stakes are structural. It’s plausible tooling that turns formal architecture or specs into working apps could displace many engineering tasks without requiring “general” intelligence, but getting there needs solutions for model truthfulness, reliability and new ecosystems — languages, libraries and APIs designed for LLM-driven workflows. Progress will likely be incremental, over years or decades, leaving time for new roles (tool builders, spec designers, verifiers) to emerge. The practical implication: invest in robust tooling, testing and specification standards, and treat LLMs as powerful assistants rather than drop-in engineer replacements.