The subheadline contains the wrong answer, for those astute enough to see it, especially those more so than this op-ed‘s opinion writer, Judd Rosenblatt, who is the CEO of AE Studio and president of the AI Alignment Foundation:

The key is “alignment,” which improves capability while cordoning off dangerous knowledge.

Rosenblatt expanded on that not too much later in his piece:

…build the world’s most capable open-weight model in America, with the dangerous capabilities stripped out before release

Because nothing says limiting innovation and truncating freedom like censoring knowledge and capability. Never mind that that will limit and truncate only American innovation and knowledge; it’ll have no effect on our enemies.

He went on:

Using GRAM [Gradient-Routed Auxiliary Modules] to train language models, we found that knowledge about bioweapons, offensive cyber, and nuclear capabilities, and even antisemitism, could be routed into compartments, then removed. The released model behaved as if it never learned that information.
…
What matters is a safety standard that causes a dangerous capability to stay absent under adversarial testing.

The bioweapons, offensive cyber, and nuclear capabilities bits are seemingly obvious targets for censorship, but that knowledge is too easily compiled via a variety of sources, including other AI models that aren’t restricted. The real kicker, though, is that last: who gets to define what is a dangerous capability? And who sets the threshold between random data to information to knowledge? Will these Know Better definers decide that identifying or not defining hate speech (whatever that is) is a dangerous capability and the thresholds for that? Providing arguments for, or against, human involvement in global warming? Men in girls’ sports and bath- and locker rooms? Defining “woman?” Criticizing government? This list is endless.

The blatant…foolishness…of Rosenblatt’s idea makes me wonder about trolls and moles.