Chip Huyen

The things people think will improve AI apps: staying updated on AI news, adopting the newest agentic framework, picking the best vector database. The things that actually improve AI apps: talking to users, better data preparation, better prompts, optimizing the workflow.

The most asked question I get: "How do I keep up with all the latest AI news?" My answer: Why? Why do you need to? If you talk to your users, look at your feedback data, and fix your prompts — you'll improve your product 10x more than any model release will.

We're in an idea crisis. AI gives you tools to build anything from scratch — design, code, websites. But when companies run hackathons, people don't know what to build. Ultra-specialization has killed big-picture thinking.

Senior engineers get the biggest productivity boost from AI coding tools — not because they need the help, but because they know what good code looks like and can direct the AI precisely. The weakest performers go on autopilot. The difference isn't the tool. It's taste.

Best way to find ideas to build with AI: For one week, pay close attention to everything that frustrates you. When something frustrates you, ask: could this be done differently? Build something around that friction.

We are in an idea crisis. We now have tools that can help you design, code, build, and ship — faster than ever. And somehow, people don't know what to build. My advice: for one week, write down everything that frustrates you. Then ask: can this be done differently? That's your product.

RAG is not a framework choice. It's a data preparation problem. The companies seeing the biggest improvement in their RAG solutions aren't agonizing over which vector database to use. They're rebuilding their data pipelines so the model can actually retrieve what it needs.

Ask managers: would you rather give your whole team expensive coding agent subscriptions or get one extra headcount? Almost every manager picks headcount. But ask a VP and they pick AI tools. The difference reveals what you actually care about.

AI is great at well-defined, isolated tasks. The moment a problem requires understanding how different components interact, it falls apart. System thinking — knowing where bugs actually come from — is the irreplaceable human skill.

Data preparation is the #1 driver of RAG performance — not which vector database you use. Most companies I've seen get their biggest quality gains from better chunking and richer metadata, not infrastructure choices.

CS is not about coding. Coding is a means to an end. CS is about system thinking — understanding what causes a problem and designing step-by-step solutions. That skill will never be automated away.