Field notes

What the work is teaching me.

Short reflections on research, building, and the classroom.

Research reflectionAugust 2026

A whiteboard mapping ideas for sparse autoencoder research and evaluation — Our STEMFORALL idea map—from regularized SAE variants to evaluation signals.

What two weeks taught me about research

Two weeks is not enough to produce a mature research project. It is enough, however, to experience the rhythm of research: learning unfamiliar foundations, forming questions, testing novelty against the literature, running experiments, and deciding what the evidence actually supports.

When our group began exploring language-model interpretability, we did not all share the same background. Before asking everyone to generate research ideas, I created an onboarding syllabus with LLM foundations, available GPU resources, and a workshop schedule. Giving the team a common starting point made better questions possible.

Many early ideas had already appeared in recent papers, so we set them aside. Some surviving hypotheses failed in experiments; others produced observations that deserved more time. By the end, we did not have a publication. We had something more appropriate for two weeks: a sharper question worth pursuing, a team that had learned how to work together, and a clearer understanding of what research progress can look like.

What I remember most are the small acts behind that progress—filling teammates’ water bottles so they could stay focused and scheduling overnight GPU jobs so limited compute did not sit idle. The workshop taught me that research is not only about results. It is also about building the conditions in which careful work can happen.

Teaching reflectionJuly 2026

Tim, Dr. Edgar, and Gary smiling together after teaching AI Ain’t Magic — AI Ain’t Magic, left to right: Tim, Dr. Edgar, and me.

A whiteboard lesson tracing gradient descent and backpropagation through a neural network — Gradient descent and backpropagation, worked through step by step.

A whiteboard connecting I, Robot, Pascal’s Wager, and decisions with potentially infinite stakes — Connecting Pascal’s Wager and I, Robot in an AI ethics lesson.

Four years later, I became the teacher

AI Ain’t Magic was my first teaching experience at Michigan, and it was unlike the recitations and office hours I had led before. Instead of supporting one semester-long course, I helped introduce machine learning to high school students with no prior experience in the field.

I taught mini-lectures on gradient descent, probability, AI ethics, and philosophy. I also led a mock trial asking whether an AI image-generation company had violated the rights of human artists. Moving between mathematics, humanities, and social questions made one lesson especially clear: AI education is inherently interdisciplinary.

Preparing those lessons gave me a more complete view of material I thought I already understood. Four years earlier, I had been a high school student learning R in a data-science summer program. Now I was the person standing at the front of a university classroom. That full-circle moment reminded me that teaching is not the final step after learning—it is one of the ways learning continues.

Project reflectionAugust 2026

What a wrong answer can reveal

Online posts often treat children’s unusual math answers as jokes. One example that stayed with me showed a student writing “4x + y = 5,” apparently combining the visible coefficients while overlooking what the variables represent. The answer does not prove a diagnosis, but it may point to a specific misconception: addition is familiar, while variables are not yet meaningful.

The comments were just as interesting. Many parents said their children made the same kind of mistake. A wrong answer can therefore be more than an isolated failure; it can be evidence of a recurring reasoning pattern shared by many learners.

That is the motivation behind SEA. Teaching the correct procedure matters, but effective support should also help a student understand why their reasoning went off track. The best hint does not simply replace a wrong answer with the right one. It makes the next productive question visible while preserving the student’s ownership of the work.

“Progress can mean learning which question deserves more time.”