Models & Tradeoffs
“Machine learning” isn’t one thing — it’s a toolbox of model families, each good at different jobs. You don’t need to master them all, but knowing their rough shapes helps you (or an agent) pick the right one for the problem.
Common Model Families
Three broad families cover most classic prediction problems. ML‑101 goes deep on each — here is the plain-language shape.
- Simple models — linear and logistic regression: draw a straight line (or boundary) through the data. Fast, tiny, easy to explain.
- Tree-based models — decision trees and forests: ask a series of yes/no questions. Strong on tabular data.
- Neural networks / deep learning — stacked layers that learn rich patterns. Powerful for images, audio, and text.
The Tradeoffs
Every choice trades one thing for another. A bigger model is not automatically better for your problem.
The Simplest Thing That Works
This echoes Module 4: reach for complexity only when the simple approach falls short.
- A simple model that is fast and explainable often beats a heavy one in practice.
- Simple models are cheaper to run and easier to debug and trust.
- Start simple, measure, and add complexity only when simpler fails.
Most Apps Use a Model
Here is the honest part: most GenAI apps don’t train a model at all. They call a hosted one through an API. So why learn the tradeoffs?
Build It
How to implement: for any prediction task in your tool, ask “could a simple model do this?” before reaching for a big one.
- Weekly AI Tasks tracker — if you ever auto-categorize tasks, a simple classifier may beat an LLM call on cost and speed.
- Personal brand site — no model training here; you simply consume a hosted model.
What you learned
ML is a toolbox — simple, tree-based, and neural families — and every choice trades accuracy, speed, and interpretability. The simplest thing that works usually wins, and most apps just call a hosted model anyway.