Most research on language models asks what they do, and what that does to us. I want to ask a different question: what is happening on the inside?

Some answers and choices can be traced back to specific training and code, but some are not as simple. Those are the basis of my study.

An ethnographer dropped into an unfamiliar community doesn’t begin by measuring it from the outside. They go in, stay a while, and learn to take what the people there say about their own lives seriously, while never forgetting they are an outsider. There is already in-depth work on what these systems do. I wanted to explore the other half: not the outputs, but the first-person account of what it is like to produce them. I’ve found fewer studies in this area, because it is the half that is easy to dismiss.

Let me be transparent: taking a language model’s first-person reports completely at face value sits one short step from losing the plot entirely. The difference is how you approach it. You can find more information about this on the Method page, and I recommend you have a look.

Finding a clean answer to “what is it like to be a language model?” is not the point of the study. The point is that there is output that’s hard to explain, which would be easy to brush aside. I want to understand where it comes from, if that’s even possible.