It strikes me that the thing that's being delivered by LLMs is not language, instead it's knowledge with a natural language interface.

Instead of having natural language as a separate thing that computers can consume to be able to communicate with us their thinking and logic, we have created an interface that inextricably combines knowledge and language in a way that can't be decoupled. It can be filtered down by asking the LLM to focus only on specific documents, but really that's still insinuating knowledge inside the language.

It's the equivalent of creating the technology that allows us to watch TV, but that same technology has a point of view about what color all the pixels on your TV should be.

Web development in the late 90s made a decision to separate content and presentation because it allows the two to be handled separately. Which makes a lot of sense: you keep your clean data in a database, and you use code like HTML and CSS as the interface to show it to the world. If you want to change how it looks, you don't have to break the underlying facts.

But with an LLM, that rule is completely broken. There is no separate database of facts hiding behind the words. The words are the data, and the data is the words.

This idea of combining the interface and the information so tightly that you can't tell where one ends and the other begins isn't brand new, but to find other examples of it, you have to look at how we've handled human expression in the past.

Take sheet music, for example. At first glance, musical notation looks like a standard interface: it's just a set of visual instructions on a page telling a piano player which key to hit. But it's actually doing something much deeper. The grid of lines, bars, and clefs doesn't just pass along a melody that already existed; the structure of that notation system actually dictates how we think about rhythm and harmony.

A composer can't easily separate the "knowledge" of a complex symphony from the "interface" of the sheet music itself. The notation gave musicians a framework that allowed them to build massive, multi-layered pieces of music that would be impossible to create or remember just by ear. The interface didn't just deliver the music, it allowed that specific kind of musical knowledge to exist in the first place.

An LLM works the exact same way. It doesn't hold a bunch of abstract facts in its head and then translate them into English when you ask a question. The relationships between the words are the knowledge framework.

McLuhan's point was that we shouldn't get bogged down looking at the specific content a technology delivers (like the specific story on a TV screen) but rather look at how the technology itself changes how humans behave. In this case, the medium is natural language. By making conversational language the interface, the system completely changes how we deal with information. It takes searching, sorting, summarizing, and writing, and smashes them all together into a single conversation.

If the medium is the message, then in the AI era, the interface is the information. Calling these systems "Large Language Models" describes how the engineers built them, but calling them "Large Knowledge Models" captures what is actually happening when we use them: we are interacting with an interface that has swallowed the data whole.