Why do we teach? Partly so that students can make sense of the knowledge that already exists — and partly so they can produce knowledge of their own. Producing knowledge has two sides, method and substance, and the first thing I want students to see is that both sides are made of models. A theory is a model. A tool is a model. Even the data is a model — it already carries someone’s decisions about what counts and what can be measured. None of it is neutral. Research is what happens when you put these models together honestly, knowing what each one assumes and where each one breaks. Every course I teach carries that idea, at different depths, for different audiences.
My traditional-classroom years have been an apprenticeship: none of those courses were mine, and I learned the craft from the instructors I worked with while doing it. In the methods courses — graduate and undergraduate statistics, quantitative techniques — my aim was a single intuition, one student at a time: if a student remembers one thing from a section with me five years on, let it be the loop from research interest, to question, to the quantities that could answer it, to the model those quantities demand. Teaching research design, I once made that loop a lecture — walking the class through a study of mine, from puzzle to design, stopping to ask what they would do. In the substantive courses — comparative politics, the Middle East, globalization — I learned the craft of each level: introductory courses introduce questions rather than answers, making the phenomenon interesting enough that students speak; advanced courses prepare students for professional work, synthesizing scholarship from perspectives not their own. And the last traditional classroom I sat in myself — Neural Mechanics, David Bau’s course on the insides of neural networks — added one more layer to how I understand learning: method and substance fused in an interdisciplinary team, computer scientists and social scientists reading the same new material because neither discipline can read it alone.
The summer institutes and workshops are where I have led: as instructor at SICSS and AIDE, and at BLISS, which I designed. Each serves a different pace of learning. At SICSS, the pace is speed — two weeks to turn complex data into actionable quantities of interest, where a working result matters more than interrogating the tools. At AIDE, it is the interrogation — a slower walk through machine learning built to question the instruments themselves: where models are unfair, and how you would know. At BLISS, it is comprehension — complex architectures brought down to earth with paper and pencil, at the pace of a human brain rather than the machine’s, until the engine stops being a black box. What all three share with that Neural Mechanics seminar is the thing I now believe most about this field: the fastest learning happens in interdisciplinary teams — and AI has become the common object that closes the gap between the disciplines and gets them talking.
And then there is the new reality: AI is moving into every stage of knowledge production — the reading, the coding, the analysis, the writing — and teaching has to answer for that. I don’t think anyone has the full answer yet, and I hold mine loosely. But everything above points somewhere. If theories, tools, and data are all models carrying assumptions, then what students need most is not execution but the intuition to match question, data, and model, and the ability to explain the match. Machines can run the pipeline on our behalf; making the knowledge relevant to us — and supervising the machines that make it — is the part I hope will last. Supervising requires knowing what we are doing, and why it was the right way to do it.
The experience all of this comes from is on the Teaching page.