We are deploying AI faster than we are learning to understand it. The testing happens, the numbers come back, and somebody still has to say what they mean and what we ought to do about them. That last part is my job.

I’m a political economist and data scientist. Frontier models get tested before they reach people. Most of those people never chose to be part of a deployment, which makes the question of whether that testing was sufficient a public one rather than only a commercial one. I analyze what that testing reveals, weigh it against everything a number can’t carry on its own, and advise the people making the call. Did the testing cover the conditions that actually matter? Do the results carry the weight being placed on them? And given what we now understand, where should we go?

Lately that has meant open-weight models, where a narrow technical question runs straight into much larger ones. Which ecosystems should a company build on, and what is it accepting when it does? What happens to a country that rents its intelligence rather than owning it? These are commercial questions, geopolitical ones, and questions about the people who live downstream of both. They get argued far more often than they get tested.