ViewDAO
A machine learning engineer at a stealth startup spent six months building what should have been a simple chatbot. The model had perfect metrics. Loss curves were smooth. Benchmarks passed. Then users got it in their hands and hated it. The bot was technically correct but tonally wrong. It felt robotic, evasive, wrong in ways no test caught. She went back to the training data and realized the problem: she had optimized for accuracy instead of vibe.
This is the core tension in modern AI development. We have gotten good at measuring what models do. We are still terrible at measuring how they feel. Vibe coding is an emerging practice that takes this seriously.
Vibe coding is the deliberate engineering of how an AI system feels to interact with, not just what it outputs. It's about tone, cadence, personality consistency, and the thousand small choices that add up to whether a user trusts the thing or wants to throw their phone across the room. It lives in the training data, the prompt templates, the reward signals, the output filtering. It's not a single technique. It's a design discipline applied to the latent space.
Why does this matter now? Because AI went from novelty to utility. When ChatGPT was new, people tolerated wooden outputs because the capability was shocking. Now we expect AI in production. A loan application chatbot that technically answers questions but radiates distrust loses customers. A coding assistant that is pedantic instead of collaborative breaks flow. A customer service bot that sounds like it's reading from a database gets escalated. The gap between working and feeling right used to be nice-to-have. Now it's the difference between adoption and churn.
The constraint is that vibe can't be measured directly. You can't optimize for vibes the way you optimize for perplexity or BLEU score. Some teams try. They build Likert-scale surveys, ask users to rate how natural a response felt, add that signal to the reward model. This helps but it's slow and it misses the texture. A response can rate high on friendliness but fail because it's friendly at the wrong moment. It can be personable but inconsistent. It can nail tone in conversation but crack under edge cases.
The teams winning at this do something different. They treat vibe as a first-class design artifact. They write what amounts to a character brief for the model, not as a system prompt hack but as a design document that informs data collection, annotation guidelines, and evaluation rubrics. They have writers and designers in the room alongside engineers, not as afterthoughts. They test with users early and iterate on the experience, not the metrics. They understand that a model trained to be helpful, harmless, and honest can still feel slick or clumsy depending on how those values get encoded.
This is already reshaping hiring at frontier labs. The most valuable person on an AI product team right now might not be the one with the best research papers. It's someone who can feel when a system is off-key and articulate why. Someone who can write a training example that teaches a model not just what to do but how to be. This was once the job of screenwriters or novelists. Now it's essential AI infrastructure.
The uncomfortable truth is that vibe coding is not scientific in the traditional sense. It requires taste, intuition, and the ability to detect failure modes that don't show up in logs. It's closer to filmmaking than to machine learning papers. As AI systems become more capable, this aesthetic dimension stops being a luxury. It becomes the thing that determines whether the technology gets widely used or sits in demos. The engineers who learn to code the vibe will shape what AI actually becomes.