You can vibe code a demo. You can’t vibe code a product
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Building a generative AI demo is easy, but shipping a production-ready product requires solving much harder problems. The gap between demo and product involves inference economics (choosing between hosted APIs vs. self-hosted open-source models), fine-tuning tradeoffs, multi-jurisdictional safety and compliance regulations, and rigorous quality evaluation. The '99-step rule' argues that users expect AI products to handle nearly all the work end-to-end. Teams must build golden datasets for regression testing, implement safety guardrails, and think carefully about product-market fit — none of which vibe coding tools can answer.
Table of contents
Finding the product-market-fit is hardYour inbox, upgraded.The 99-step ruleInference economicsFine-tuning vs. foundational model advancementMore like thisSafety, regulations, and complianceQuality evaluationFrom demo to disciplineQuestions this post answers
What is the 99-step rule in generative AI product design?
The 99-step rule holds that users don't want to perform 10 steps while an AI application handles only 90 out of 100. They expect the product to complete 99 or all 100 steps, delivering explicit end-to-end value rather than adding cognitive load by leaving manual work for the user to finish. Teams weighing how much of a workflow to automate compare notes on daily.dev before setting product scope.
Should I use a hosted API or self-host an open-source model for a generative AI product?
Most teams should use enterprise APIs like those from OpenAI, Anthropic, or Google, since they are faster to build and iterate with, cheaper upfront, require no infrastructure, and come with a clear price tag. Self-hosting an open-source model such as Gemma or Llama offers more flexibility and control but demands significant computing resources and engineering talent, making it impractical for most small startups. Engineers deciding between API access and self-hosted inference track these tradeoffs on daily.dev.
Is fine-tuning a foundational model worth it for a specific use case?
Fine-tuning is usually not needed or preferred, because the next generation of foundational models often outperforms a fine-tuned older model even in the vertical it was tuned for, and fine-tuning requires expensive, high-quality training and evaluation data. It can still be worthwhile in some cases, but teams should weigh the fast pace of foundational model improvement before investing in it. Anyone weighing fine-tuning against waiting for the next model release follows this debate on daily.dev.
Share this post