How I Think

Process notes — how I pick problems, when I think I’m wrong, what I actually read.

A short note on how I work, written for collaborators and search committees who want a sense of style rather than résumé content.

This is meant to be honest, not aspirational. It will change.

Last updated: 2026-04-25.

How I pick a problem

I look for the intersection of three properties:

  1. It bothers me — I find myself returning to it across unrelated contexts. If I’m not pulled back to a question across weeks, I don’t trust that I will sustain it across years.
  2. There is a measurable handle — at least one observable that would meaningfully change my mind. If I can’t construct that, the problem isn’t a problem yet, it’s a mood.
  3. The intersection of fields is uncrowded — most NeuroAI questions are popular and well-staffed. I prefer the edges where neuroscience methods meet AI artifacts (or vice versa) and few people sit comfortably on both sides.

A clarifying test: if I were the only person on Earth working on this, would the world notice it being missed? If the answer is no, the problem is interesting but not for me right now.

How I know I’m wrong

The thing I mistrust most is my own intuition when it is well-formed and fluent. Fluency is correlation with what I’ve already absorbed, not with truth.

Operationally:

  • Falsifiable claims first. Before believing something, I write down the observation that would refute it. If I can’t write that, I’m describing a feeling.
  • Cheapest experiment that could end the question. Many “complex” problems collapse under a 30-minute diagnostic that nobody bothered to run. I default to running it before building anything.
  • Adversarial review. I treat my own results like another author’s draft — what would a hostile reviewer say first?

I find I am most often wrong about scope (I overestimate generality) and about novelty (I underestimate what’s already been done by someone with slightly different vocabulary).

What I actually read

Mixed:

  • One classic in cognitive/computational neuroscience per month (something pre-2015 that earned its citations) — keeps the long view.
  • arXiv recent, narrowed by author and venue I trust, not by topic.
  • Working code of recent papers I want to understand. Reading code catches assumptions that papers gloss over.
  • One book a year outside science (literary fiction, philosophy) to refresh the way I think about evidence and interpretation.

I read less than I used to and re-read more.

How I work with people

  • Slow defaults, fast course-correction. I aim for clarity on the framing of a collaboration before optimizing speed.
  • Externalize disagreement. If I disagree, I would rather say it clearly and risk being wrong than hedge and create noise downstream.
  • Credit accurately. Authorship and acknowledgment are signals to the next collaborator. I take both seriously.

What I’m trying to get better at

  • Writing earlier in the project instead of at the end.
  • Saying “I don’t know” without rushing to fill the silence with a guess.
  • Choosing which papers not to read.

For the specific questions I am chasing, see the Research Questions Registry. For the technical work, see the main page or CV.