I used to think slow progress meant something was wrong. Now I think it usually means something is real. The papers that took me the longest to understand are the ones I actually use. The skills I had to fight for are the ones that feel like mine. There’s something uncomfortable about that — it doesn’t make for a clean productivity framework — but I keep coming back to it. ...
The first time I worked at a startup I was surprised by how much I missed the whiteboard — the slow, speculative kind of thinking that doesn’t need to ship by Friday. But I also noticed how much faster I got at deciding what mattered. Academic research can afford to be thorough in ways that companies can’t, but that thoroughness sometimes becomes a way of avoiding commitment. I carry both habits now. I’m not sure they’re fully reconciled, but I think that tension is useful.
A lot of fMRI analysis is deciding what to ignore. The signal you care about is small. The noise is everywhere. The pipeline is a series of choices about what counts as real. I think about that outside the lab too. Most of what happens in a day is noise. The hard part isn’t collecting more information — it’s figuring out what to stop attending to. I’m still not very good at this. But at least I have a framework for feeling bad about it.
Year one me thought research was mostly about having good ideas. Year five me knows it’s mostly about finishing things. Ideas are cheap. The hard part is sitting with an idea long enough to know whether it’s worth the next six months of your life. And then doing the work even after it stops being exciting. I don’t think I’d have believed that if someone told me at the start. I probably needed to find out the slow way.
Most of the AI safety conversation happens inside CS departments and a handful of dedicated research labs. I think that’s too narrow. If AI systems are going to be deployed in healthcare, mental health support, clinical decision-making — and they already are — then the people who understand those domains need to be in the room. Not just as end users. As researchers with opinions about what safe actually means in context. ...