Summer Mengelkoch gave a talk this week about remote trials — getting data from people in their actual lives, instead of pulling them into a lab where the measurement changes the thing being measured.
I've been turning it over since, because it names the oldest problem in studying people: you can have rigor or you can have proximity, and the usual infrastructure makes you choose. Lab or life. I think that's about to change — and I think most of this room is better positioned for the change than they realize. Let me make the case plainly.
Using AI today, for most people, is requesting: you type a prompt, take what comes back, and hope. It's useful, and it caps out fast — not because the model is weak, but because requesting leaves all the judgment on the table.
There's another posture, and it's the one that's about to matter. Call it directing: deciding what an intelligence should remember, and teaching it to distrust its own stale beliefs. Controlling what you put in front of it, because what it attends to shapes the answer more than the model does. Treating its instructions as something you test and refine against a result, not something you guess at. Running several of them behind a check that catches what any one got wrong. And refusing to believe the work is done until you've verified it actually did what it claims.
Read that back. None of it is a coding skill. It's methodology.
Skepticism, decomposition, verification, provenance — the exact instincts that define what a good researcher even is. The frontier has quietly moved from what the model can do to what you build around it, and what you build around it is method. The person who can point a few agents at a literature sweep, keep honest track of what's been checked, hold back the low-confidence findings, and prove the pipeline did what it claims isn't "using AI." They're working at the edge — and they got there by being a good scientist, not a good programmer.
I spent a day this week in the deep end of how this works under the hood — the plumbing that lets you hand an agent a narrow, accountable set of things it's allowed to do, and nothing else. It's genuinely technical, and I won't make you read it here. But one thing from that day is worth carrying out, because it answers the first honest objection anyone raises.
Someone next to me put it simply: when agents act on your behalf, can't someone poison what yours reads, or hijack what it does? Yes — and it's the right question, sharper still the moment you imagine a community exposing its own knowledge to anyone's agent. But the answer isn't "trust the model to behave." It's design: you can build the thing so the harmful move isn't even available to make — so an injected "delete everything" fails not because the agent nobly refused, but because deleting was never on the menu for that visitor. You shrink what can go wrong by deciding, in advance, what's reachable at all. It doesn't make misuse impossible. It makes it structurally hard — and the gap that remains is exactly the kind of problem this room is good at.
Which brings me back to Mengelkoch, and the thing I actually care about.
A community holds enormous knowledge about how it really works — who gets fed, who gets watched over, how people hold each other up under pressure. Almost none of it is written down, and the usual way to capture it is to send researchers in, extract it, and publish it somewhere the community will never see. The knowledge leaves the place it came from. Most people doing that work don't want to be extractive — the infrastructure just hasn't offered another way.
Here's another way. The community holds its own knowledge — kept current, owned by the place that holds it — and a researcher is granted access to ask, inside limits the community sets and can see. The community decides what's shared and who can ask; where every answer came from is built in; and the knowledge stays meaningful because it never had to leave the life it describes.
That's not a bridge between researcher and community. It's a way for both to hold the same knowledge at once, without either one flattening the other.
What would it look like to direct this kind of intelligence in partnership with the communities you study — rather than on them?
I think it's buildable now. And I think the people best suited to build it are the ones who spent a decade learning to direct inquiry rigorously and are about to find out they already know how. If that's a lane you want to walk, I'd like to walk it with you.