It isn't certain that this framing is true. As part of learning to predict the outcome of the work token by token, LLMs very well might be "doing the work" as an intermediate step via some kind of reverse engineering.
They're already have the full work available. When you're reading the source code of a program to learn how it works, your objective is not to learn what keyword are close to each other or extract the common patterns. You're extracting a model which is an abstraction about some real world concept (or some other abstractions) and rules of manipulation of that abstraction.
After internalizing that abstraction, you can replicate it with whatever you want, extends it further,... It's an internal model that you can shape as you please in your mind, then create a concrete realization once you're happy with the shape.
I.e. the theory of the program as it exist in the minds of the development team might not be fully available for reconstruction from just the final code and docs since it includes a lot of activity that does not end up in the code.