I’ve observed that people saying “LLMs are just a tool” usually compare them to language server implementations, compilers, and more. I disagree with this view because the tools they’re compared to are usually deterministic in the sense of they’re not just a blackbox that sometimes answers one way, sometimes another depending on whether the API provider changes the model weights, the temperature, etc.

Tool users.

They're a new category of thing. They aren't really "just an" anything.

When the wheel was invented cavemen probably said "it's just a stone".