What a Foundation Model Actually Is

Before there's a chat app, a writing assistant, or an image generator, there's a foundation model: a massive neural network trained on huge amounts of text, images, and code until it can recognize patterns and predict what comes next. That's Level 1 of the stack, and it's the layer everything else in this course sits on top of.

You don't interact with a foundation model directly very often. What you actually use, day to day, is a product built on top of one, a chat interface, an app, a plugin. The foundation model is the engine, the assistant you open on your phone is the car built around it. Same idea as our "pattern-matching, not magic" lesson from Level 2, just one layer deeper.

Foundation models are trained once, at enormous cost, by a handful of companies, and then reused by everyone building on top of them. That's why you'll see the same underlying model show up inside dozens of different apps and tools with different names and different interfaces.

Why bother knowing this if you never touch it directly? Because when a tool description says "powered by GPT-4" or "built on Claude," you now know exactly what that means: it's using someone else's foundation model as its engine, and its actual value to you comes from whatever they built on top of it.