Language Is the New UI

I think our tendency to look at Large Language Models and see a new species is fundamentally flawed. Arguably, it’s a step in a completely irrelevant direction, and from that ill-chosen direction the entire recent AI hype appears skewed and distorted.

What we should see instead is a new abstraction layer. We should consider LLMs another step in the long saga of moving the interface further from the metal. You’ve probably heard it all before: we moved from punch cards to Assembly, from Assembly to C, and from C to JavaScript. Each step introduced towering abstractions. Each step was taken to phrase extremely complex mechanical behaviour in ways that were increasingly hospitable and natural for us.

In fact, I suggest we should consider Large Language Models as a new type of user interface — LUI (Language User Interface), if you will. I think, if we were to predict how AI will affect us in the medium term, we should seek inspiration from a comparison between the different modes of human-computer interaction:

• Command Line Interface: Instructions are given as precise, objectively meaningful commands.
• Graphical User Interface: Instructions are given through interaction with graphical elements on the screen.
• Language User Interface: Instructions are given in natural language, as if the user were talking to a coworker or a peer.

At the other end of the spectrum, we have CLI. While it’s clear that it’s a small minority of users that interact with their devices through a CLI, it’s a relevant minority. Namely programmers, researchers, and the various other experts who have a need to express their instructions in extreme detail.

It’s noteworthy why CLI has remained relevant. One of the reasons is that GUI tools are fundamentally susceptible to complexity: a sufficiently complex GUI tool is arguably much more difficult to use than a CLI tool of similar complexity.

There are other concerns too. In a CLI environment, automation is trivial (just run the commands from a file), instructions are easy to give (just message the user with the command they need to run), and so on. In many situations, a proficient CLI user is also much more efficient than a GUI user (think, for example, recursively deleting files of certain type).

Of the interaction modes listed above, LUI seems like the most intuitive one, and, at a glance, the most powerful one too. I think, however, that this merits a careful analysis. It seems reasonable to expect that the GUI-to-LUI transition features many of the same phenomena that appear in the CLI-to-GUI comparison.

One of these, which is highly relevant for programmers, is how language introduces “fuzziness”. Something I have written about before is that natural language does not excel in expressing precise instructions. It’s arguably not very efficient to use natural language to perform work that needs to satisfy a strict specification. While it will most likely have its place within the way software is built in the future (it’s not difficult to imagine it serving a major role), it is difficult to imagine LUI completely displacing CLI and GUI, and (hence) us as the users.

I think there is a significant amount of room for speculation here. And I mean speculation from a fresh perspective. To expand upon this, I’ll end this post with an opinion: I really hate the term AI. AI is too big a concept for its own good. It comes with connotations so heavy that it’s very difficult to discuss the subject without the discussion drifting into a couple of thoroughly trodden paths.

People should really look at these tools without any preconceived notions of what they are and how they should be used. This is impossible when they are labelled essentially “you, but better”.