Something like ability to influence and coordinate other agents around a desired aim

I agree that directly optimizing for this objective would be very hard

But political skill is clearly instrumental to many objectives that are easy to measure

are you objecting to treating "politics" as a one-dimensional skill, or that AI could perform well at this skill (or both?)

obviously politics involves many sub-skills, but its still intelligible (and descriptively useful) to talk about humans as "good" or "bad" at politics

Excited to see gradient routing applied to more realistic scalable oversight benchmarks, e.g. @eleutherai.bsky.social's recent work on "scalable elicitation" under different label budgets arxiv.org/abs/2410.13215

Also the toy scalable oversight task is super cool, isolating the idea that oversight failures will be correlated with distributional shifts, and the most important distributional shift might be "is oversight being provided"

Part of "their stuff" is economic populism, and ~everybody (including you) still thinks that can work

But nah overall vox is great, if for the EA content alone

Based

When does Ezra not get it