Minh-Ha Nguyen

Those two interests sound unrelated, but to me they're the same problem in different clothes: how do you draw a conclusion you can actually trust from messy, observational data? With a medication, the data are millions of patient records, and the question is whether a drug is truly harmful or whether the signal is an artifact of how the study was built. With a language model, the data are its answers to test cases, and the question is whether a benchmark score means what people say it means. In both, the hard part is knowing whether to believe the number.

My doctoral work is that problem at its hardest. I study Navy submariners — men who spent years under chronic low-dose radiation in a sealed hull, about the closest thing on Earth to a long mission in space. Does that exposure raise the risk of Parkinson's? The catch is that everything accumulates together: the men who served longest absorbed the most dose, and were the healthiest to begin with — a self-selection that can make a real hazard look protective. The craft is in designing an analysis where the radiation signal survives it.

The other half of my work brings that same craft to AI, with one liberating difference: here I can run the experiment the radiation study never allows. I build and evaluate systems that reason from a patient's symptoms toward a rare-disease diagnosis (a project I call liteOdyssey), and what I care about most is evaluating them honestly — treating a model like a population you sample from, a prompt like a treatment you assign, and a benchmark like an experiment that can be designed well or badly. Most of the field reports a score; I'd rather ask whether the score is powered, representative, and measured against a standard whose own reliability we've checked.

Underneath both, I think research should be legible to the people it's about — open access, methods you can check, and as much honesty about uncertainty as about findings.