We set out to create a language that would allow analysts to create, train and execute machine learning models in a natural way.

It also turned out to be a great language for agents to use to interact with the world.

LLMs are typically good at two things:

1. Reasoning about the world at a high level, so in this case playing the role of an analyst and in that context come up with a plan for how to analyse a piece of data
2. Generating contained snippets of code for a common programming language, like SQL

I would say they are (depending on model obviously) great at 1 and okay at 2 - in the sense that they are more likely to get 2 wrong than 1.