Common pitfalls when building generative AI applications

Chip Huyen catalogs six common pitfalls in GenAI application development: over-applying generative AI where simpler solutions suffice, misattributing poor UX to the AI, premature complexity (agentic frameworks, vector DBs, finetuning before simpler approaches are validated), over-indexing on early demo success, skipping human evaluation in favor of AI-as-a-judge, and crowdsourcing use cases without strategic direction. The piece is grounded in real case studies from LinkedIn, Intuit, and unnamed startups, and emphasizes that for most AI products the differentiator is product/UX — not the model. Particularly relevant to engineers building agentic or LLM-powered tools who may be navigating the 80%→95% quality gap.