Semiotics interlude: Why semiotics?

So, anyone even quickly skimming through my background will notice that my academical and professional expertise is very much in the non-humanistic side of things. My fundamental shaping as a scientific thinker arose in the forges of pure math - for better or worse. After that I pivoted into the technical side of AI and computer science. So why then am I reading and writing about semiotics, i.e. about the way people make meaning of things in the world? The short answer is that I want to (or need to) understand both AI and human-AI relations better. For a longer answer we have this post.

As a small caveat to the reader: I am not very happy with this post at the moment for reasons I can't quite put to words right now. Better out than in, so I wanted to get this out here, but be aware that this might get updated later on.

Language models were not expected

I would hazard to claim that the supermajority of people were taken in by pretty much complete surprise by the versatility and power of the modern language models. (With the possible, or even probable, exception of some linguists.)

Language is powerful, and we all know that it is a wise thing (or a naively obvious platitude) to say that language is powerful. And yet when most people first saw chatGPT give coherent answers to prompts they thought "cute" instead of "industry transforming". So in this regard I don't we really appreciated what was starting to happen.

"Getting surprised" happens when there is a mismatch between your model of the world and how the world actually is. In particular being surprised is an important signal that tells you that your mental map might need updating. Thus a massive surprise about the power of language models was a sign for me to try to understand the world better, especially in various things related to language models and language. But how?

It is truth universally acknowledged that Large Language Models are too big to study from first principles - to understand them you should approach them more like natural phenomenon that you can probe at and analyze the results. So how do we start probing language models? There are various mechanical approaches, especially if you have access to the model itself.¹ But with most LLMs the models are not open source. And they are very, very large which makes analyzing them or even running them a nontrivial task. Luckily there is another approach! We can look at the end result; language produced in interactions with humans. And in analyzing natural language we have centuries of research tradition already waiting for us. Namely, linguistics.

Humanists to the rescue!

To start off, I wish to emphasize here that I do not wish to reduce the field of linguistics to a "nice little probe/tool to analyze language". Instead, based on my very limited understanding of the topic, I have the mental model that linguistics is the very core of the humanistic arts which in turn is in the core of everything we do as humans.²

Furthermore, based on which linguists and linguistic theories you subscribe to, language might even be considered as the defining quantity of humans. Or as the tool that doesn't only mirror our intelligence but enables both our thinking and our consciousness. (cf. Pinker - The Stuff of Thought.) At minimum language is what we use to store and convey a considerable portion of our daily information flow.

Humanists live in this interesting world where they study the world based on language while very well understanding and even embrasing the insufficiency of language. And this fuzzy tool called language is used not only to study the world via language, but to study language via language. The tool is constantly living, wriggling and incomplete, and with that tool you study not only the world but the tool itself. This fuzzy wriggling incompleteness is a very good framework to train your mind in, and I believe the skill of getting used to handling constant uncertainty is beneficial also for LLM research.

To conclude, my two cents for now are that any "power" there might be in a language must appear in the domain of humans. To understand (the power of) language models, we need to be able to analyze and understand language. And what better place to turn than linguistics.