Lesson 4 of 6
Foundation models as a platform
7 min read
Old-school machine learning meant a new model for every task. Today one model writes code, answers email, and reads images. What changed — and how do you build on it?
Classic ML vs the foundation model
In classic machine learning, you built one narrow model per task: a spam classifier, a churn predictor, each trained on its own labelled data. A [foundation model](glossary://foundation-model) flips that — it's a single, general model trained on vast data that already handles thousands of tasks out of the box. You don't train a new one; you direct the one that exists.
Classic ML: one narrow model per task. Foundation models: one general engine for many tasks — you steer it instead of training it.
The model as a platform
Because it's general, a foundation model behaves less like a program and more like a platform — the CPU of a new kind of computer. Around that engine you attach parts: its [context window](glossary://context-window) as working memory, files as storage, tools and the live web as peripherals. Tap each part to see what it maps to.
A foundation model is a platform: a general engine you extend with memory, files, tools, and the web — not a single-purpose program.
This is why 'AI engineering' is a real discipline. Classic ML was mostly training; building on a foundation model is mostly integration — wiring the engine to memory, data, and tools.
The shape of it
- —Classic ML trained one narrow model per task, each on its own data.
- —A foundation model is one general engine that handles many tasks out of the box.
- —You build on it like a platform: attach memory, files, tools, and the web.
- —The work shifts from training a model to integrating one.
How is building on a foundation model different from classic machine learning?
Continue in the app
Take the whole Becoming an AI Engineer course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.