Research · Open questions
The questions I keep coming back to.
Not a publication list — the handful of problems I find genuinely worth chewing on. They're what most of my writing and building circles back to.
Judging an AI honestly, without fooling ourselves
It's surprisingly easy to convince yourself a model is brilliant. I care about ways of testing that survive contact with the real world — measuring what it can truly do, not what a benchmark happens to reward.
When AI that “works” meets messy, human input
A system can look perfect in a demo and fall apart the moment a real person uses it. I spend a lot of time in that gap — the part the tutorials quietly skip.
Where extra complexity actually earns its keep
Sometimes a fancy, autonomous setup is the right call. Far more often, one clear instruction does the job. I like figuring out which is which before building the expensive thing.
Keeping humans accountable for what AI produces
Automating a task doesn't move the responsibility for it. I'm interested in where a person has to stay firmly in the loop — and how to design that seam so it helps rather than rubber-stamps.
Staying current in a field that changes every week
Most of the noise is forgettable; a little of it genuinely matters. I think a lot about how to tell the difference quickly, so the firehose never crowds out the actual work.