Is it already too late to reduce catastrophic AI risk? — Harvard Gazette

The Harvard Gazette has asked the question of the season — is it already too late to reduce catastrophic AI risk? — and the interview's answer is agreeably precise: the actors with the most power to mitigate significant harms are the model developers, because they control the systems that get misused. Which is the regulatory equivalent of concluding that the people best placed to prevent a pub brawl are the eleven men currently throwing the stools. Thank you, yes, we'd worked that one out.

Then comes the sentence that should be framed above every lab's reception desk: a long list of safeguards exists. A long list! Written down and everything. Existence, though, is a low bar — a fire extinguisher still in its cellophane also exists, and so does my gym membership. The interview adds, with the flat weariness of a man reading a risk register aloud, that this list is not consistently implemented. Not consistently. There's a phrase that has been through legal, PR and three rounds of investor relations before anyone let it out of the building.

And the recurring bit: users download open models, train away the harmful-task-refusal behaviours, and re-upload them. As a consistent tendency, no less — which is the researchers' polite term for 'we keep watching this happen and have stopped gasping'. Open weights are, by design, one afternoon with a graphics card away from being a different model entirely under the same name. The safeguards exist. The list exists. The re-uploads exist. The only thing not consistently implemented is anyone saying no — and 'no' is famously incompatible with a shipping cadence. Not too late, then. Merely never started in a manner that might inconvenience a funding round.