What am I interested in and what is my research vision?

I am interested in philosophical problems. You could also approximately call them “open-ended” or “semi-structured” problems. I take philosophical problems to involve systematic inquiry into foundational and/or abstract questions dealing with values and interpretations, that we make progress on with human judgement. These include questions in the discipline of Philosophy (e.g., are moral statements truth-apt? what is the basis for knowledge?), law (e.g., how should we interpret this case?), politics (e.g., what should my views on this political issue be? what political action should I take?), therapy & personal life (e.g., how should we understand death? should I have children?), and AI Alignment (e.g. how do humans supervise AI that may become vastly more intelligent than us, possibly through other AIs?). Such problems also arise in science (e.g., how might this drug which fixes one issue with a patient create other issues that adversely affect their wellness? how do we make sense of the reproducability crisis in science?) and math (there are an infinite number of true mathematical statements, but which ones are the interesting ones?).

I think a key feature of philosophical problems is that human judgement is needed to make meaningful progress. As a thought experiment, imagine Dario’s “country of geniuses in a datacenter” running for a thousand years to prove (or disprove) a bunch of mathematical conjectures. It is conceivable that they will have made genuine progress at the end of these 1000 years (although they may also prove conjectures that they propose but we don’t think are that interesting or useful). Now imagine a “country of geniuses in a datacenter” running for a thousand years to discover “philosophical truths”. They come back to us humans and give us the list of truths they have discovered: 1. Humans actually don’t have free will, 2. ought implies can, 3. … I think this seems a little bit ridiculous, not because these AI geniuses may not be right, but because without my human judgement — my capacity to reflect upon and endorse statements that I take myself to be responsible for — these 1000 years of “multi-agent debate” don’t mean anything (to me, which is the point). (My language and thinking here is taken heavily from the philosophers Robert Brandom and Wilfred Sellars.)

Okay, so human judgement is essential to making progress on these open-ended questions. So let’s throw AI out of the window because it, in principle, lacks the capacity for this kind of judgement. No: judgement has a strong relationship with what philosopher Brian Cantwell Smith calls “reckoning” — a calculating prowess for uncovering patterns and reasoning out. Good judgement emerges from reflecting on the fruits of reckoning, or even doing reckoning ourselves. Great friends don’t tell us what to think, but rather help us see other ways of reckoning that we had not reckoned with before, and which we may judge to be ways of reckoning that we want to endorse. Now ideally everyone would have access to these kinds of friends, but we don’t, and even if we do, wouldn’t we want more?

To realize this vision, I think we need to give AI more conceptual structure. That is, we need to allow people and AI to collaborate on “shapes of thinking” — structures or scaffolds to set the terms and goals of inquiry into an open-ended question — and, with the iteratively developed endorsement of human judgement, realize what kinds of ideas those shapes of thinking will lead to. As AI models get better and better, they will better be able to realize the shapes of thinking that we judge to be worth pursuing.

Many of the open-ended questions we bring to AI involve a cascade of decisions that spills into a landscape of possibilities — how to frame the question, what to value, what assumptions to make — but we typically receive a single result, presented as if it were the only one. We have no way to know how much hinges on decisions we never saw. In data analysis, “multiverse analysis” addresses this by systematically varying these choices and reporting how the result changes. We ask: what would multiverse analysis look like for the kinds of value-laden, open-ended problems people are bringing to AI?

We introduce the conceptual multiverse: an interactive system that surfaces the conceptual decisions an AI’s answer depends on, and the alternatives each decision opens up. Instead of taking a single answer on faith, a person can navigate a tree of branch points — transparently inspecting, intervenably changing, and checking each step against the reasoning norms of the domain. We develop a verification framework that enforces properties of good decision structures (like unambiguity and completeness) calibrated by expert-level reasoning, and we represent the multiverse concretely as a Python program that agents can build, verify, and regenerate.

We calibrated and built the system in three domains — philosophy (an open-ended question), AI alignment (a sensitive user request), and poetry (a generation prompt) — and ran in-depth studies with 15 participants. Across all three, the conceptual multiverse helped participants develop a working map of the problem, surfacing thinking that uncontextualized answers would have left buried: philosophy students rewrote essays with sharper framings and reversed theses; alignment annotators moved from surface preferences to reasoning about user intent and harm; poets identified compositional patterns that clarified their taste. arXiv and website coming soon!

Loading papers…

Loading papers…

I was also mentored by Jared Moore, Quan Ze (Jim) Chen, Sebastin Santy, Matt Wallingford, Carina Fourie, and others.

I wrote two books on deep learning for Apress/Springer Nature: Modern Deep Learning Design and Applications (solo) and Modern Deep Learning for Tabular Data (with Andy Wang).

I was a TA for 7 quarters across intro Python programming (CSE 160), intermediate Python (CSE 163), and discrete math for CS (CSE 311). I also TA’d English Composition and Literary Analysis at the Robinson Center, and Introduction to Machine Learning at The Coding School.

I entered university at 14 through the Robinson Center Early Entrance Program. My worldview was especially transformed during Transition School by Amanda Zink teaching English and Michael Reagan teaching History.

One of my close friends who also went through the Early Entrance Program with me is Vivek Prakriya, who is the youngest city councilmember to win a contested race in a city with over 75k people (Redmond, Washington). In Summer 2025 I managed social media for his campaign — running content for his run for Redmond City Council seat no. 2.

I have reviewed for CSCW, the NeurIPS Moral Psychology & Moral Philosophy Workshop, and the ICML AI + HCI Workshop, and have served as a research panelist and student host at events for the Allen School and the Robinson Center.

I spent summer of 2022 at Deepgram working on speech transcription models, advised by Andrew Seagraves.

An unorganized and growing archive of quotes, ideas, and works that influence or represent my thought. Order is randomized.

Academic papers and articles that influence my research and thinking. By the way, clicking on one of the squares in the background takes you to a random article.