This is a special post for quick takes by Thomas Kwa. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Having labs adopt a policy of holding out their best alignment measurements rather than training against them, hill-climbing, or making narrow patches seems extremely high value. Multiple percentage points of x-risk could be at stake, and it seems pretty tractable.
Situational awareness seems so high though. Fable 5.2 IDed its alignment tests cold and then reasoned about how it should respond to shape the model release it knew would result.
Nevertheless I think the most likely future is that we'll have the ability to measure alignment in a way that correlates with x-risk propensity, but not robustly enough to train against.