Control vs Selection: Civilisation is best at control, but navigating AGI requires selection

post by VojtaKovarik · 2024-01-30T19:06:29.913Z · LW · GW · 1 comments

Epistemic status: Hypothesis that seemed interesting, but I am not confident about it.

In Control vs Selection [LW · GW], Abram Demski distinguishes between two types of optimisation. As an example of control, he mentions a heat-seeking missile which iteratively steers towards "more heat", in order to hopefully reach the point in space that contains its target. In contrast, (my) example of selection would be having a device that can teleport you to any point in space and back, and using that device to find your target.
For me personally, selection has the additional connotation that a lot of the work can be happening in the optimiser's mind, rather than in the real world: For example, when building a house via selection, I imagine considering a few different blueprints, and then only building the one I like the most. This is in contrast to building a house via control, where I just start the construction, and then add to it and modify it as I go. However, note that I am not sure whether other people would endorse this view of the control vs selection distinction.

Using this framing, it seems to me that, on the high level, the shape of our civilisation mostly changes via control, and we don't have good mechanisms for changing it via selection.

This brings up some questions:

  1. ^

    I guess that rather than saying that a problem "requires" control/selection, it would be more accurate to talk about problems being more or less amenable to control/selection.

1 comments

Comments sorted by top scores.

comment by Shankar Sivarajan (shankar-sivarajan) · 2024-01-30T19:52:02.183Z · LW(p) · GW(p)

Your distinction might be close what what is described in Seeing like a State (according to this review): the organic development of local communities vs. the High Modernism of the centralized state.