I find the canary strings interesting at this point as evidence about “civilizational competence” regarding AI training. I enter the picture pessimistic, but I think if AI labs actually did reliably filter training data, it would mean I need to make certain updates about how things work. Disappointingly, things match my expectation that is generally not the kind of effortful, “conscientious” act with no immediate pay-off thing that people reliably do, especially at org scale.
In the same vein, we’re not going to get filtering of material where AIs are depicted doing bad things, or real sustained boxing. Maybe Control will get enough attention and sustained focus to be implemented actually, but the same underlying reality feels at play to me.
I find the canary strings interesting at this point as evidence about “civilizational competence” regarding AI training. I enter the picture pessimistic, but I think if AI labs actually did reliably filter training data, it would mean I need to make certain updates about how things work. Disappointingly, things match my expectation that is generally not the kind of effortful, “conscientious” act with no immediate pay-off thing that people reliably do, especially at org scale.
In the same vein, we’re not going to get filtering of material where AIs are depicted doing bad things, or real sustained boxing. Maybe Control will get enough attention and sustained focus to be implemented actually, but the same underlying reality feels at play to me.