ML Engineers Split on Vibe Coding

The core takeaway is that ML engineers don’t seem anti-AI; they’re anti-blind trust. In ML, where code is often tied to experiments, metrics, and theory, a wrong-looking-but-working answer can be more dangerous than a syntax error.

• –AI is useful for boilerplate, throwaway scripts, refactors, and getting from zero to a working baseline faster
• –The biggest failure mode is silent correctness drift, where generated code changes assumptions, theory, or pipeline behavior without obvious breakage
• –Several commenters frame the right workflow as supervised use: ask for help, verify outputs, and keep control of the architecture
• –ML work may be a better fit for AI coding than generic software in some cases because pipelines can often be tested end-to-end with clear metrics
• –The discussion suggests “vibe coding” works best as an accelerator, not as a substitute for understanding