From Models to Products + Unit 1 Review

A trained model sitting on a researcher's computer isn't a product — it's closer to a finished engine sitting in a garage with no car built around it yet. So how does that engine end up running your phone's recommendation feed or your video app's "up next" list?

Once a model is trained, using it on new, real-world input is called inference — the model isn't learning anymore, it's just applying everything it already learned to make a fast prediction. Training a model can take weeks of computing time; running inference on your one search query takes a fraction of a second. Companies often take a general-purpose trained model and fine-tune it — keep training it a little further on a narrower, more specific dataset — so a broad language model becomes specialized at, say, customer support replies for one company.

But a model on its own still can't talk to your phone's app. That connection usually happens through an API (Application Programming Interface) — a defined way for one piece of software to ask another piece of software to do something and get an answer back. When an app "calls an AI model," it's almost always sending a request through an API, getting a prediction back, and displaying it to you, often within milliseconds.

This is exactly how the recommendation engine deciding what video to show you next works: a model trained on patterns of what similar viewers watched is running inference, through an API, on data about your recent activity, dozens of times a day, every time you open the app. The same pipeline — train once, fine-tune, deploy behind an API, run inference constantly — powers everything from your voice assistant to AI image generators, which is exactly the recipe behind generative AI tools that create new text, images, or audio on demand rather than just classifying or predicting a number.

Today wraps up Unit 1: you now have the full pipeline in your head, from "what counts as intelligence" all the way to "how a trained model ends up running an app on your phone." Unit 2 turns a more critical eye on that pipeline — starting tomorrow, we ask what can go wrong.