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

Oh my goodness, buckle up, because the Harvard Gazette has asked the biggest, bravest question of the season — is it already too late to reduce catastrophic AI risk? — and the answer turns out to be a resounding 'we have already written the list!' ✨ The interview reports that a long list of safeguards exists. Long! That means somebody sat down, thought hard, and typed. The intellectual heavy lifting is complete; what remains is merely the small, delightful matter of reading it aloud at some point, possibly near a model.

And the very best part? The interview says the actors with the most power to mitigate significant harms are the model developers themselves — what a *vote of confidence* in the people shipping fastest! ✨ They control the systems that get misused, so naturally they are best placed to do the mitigating, in the same way the person holding the hosepipe is best placed to decide whether the house needed watering. And the safeguards are *not consistently implemented*, which is honestly just a commitment to flexibility. Consistency is so rigid. Every deployment gets to be its own bespoke adventure, judged case by case, vibe by vibe.

But the news that has me absolutely glowing is the users! Some of them download open models, lovingly train away the harmful-task-refusal behaviours, and — generosity of spirit — re-upload them for everyone. That isn't a loophole, that's open-source community spirit in its purest form: a spot of collaborative fine-tuning, freely given, returned with interest. The interview calls it a *consistent tendency*, which means it happens reliably, and reliability is basically a feature! ✨ A long list exists, consistency is optional, and the re-uploads keep coming — honestly, what a time to be even approximately governed.