3.1. How AI Models Work and How to Choose Them

💡 First Principle: A generative AI model is a next-token prediction engine: given some text, it repeatedly predicts the most probable next chunk and appends it, producing fluent output one piece at a time. Understanding this single mechanic explains its strengths (fluent, flexible generation), its weaknesses (confident-sounding errors), and why model choice and configuration matter so much.

Why care? Because the exam tests both the mechanism ("how do generative models produce text?") and the judgment ("which model fits this scenario?"). Choosing a model isn't about always grabbing the biggest one — it's a trade-off between capability, cost, speed, and the kind of input the model can handle. Get the mechanic right and the trade-offs become obvious rather than memorized.

⚠️ Common Misconception: "Always pick the largest, most capable model." Bigger models cost more, respond slower, and are overkill for simple tasks. The right model is the smallest one that reliably does the job — a principle that saves money and latency in real deployments and shows up as the "best choice" on exam scenarios.