Very thoughtful take! A nuance that maybe you didn't consider, maybe didn't make it into the post, or maybe just isn't very useful: technologies can be centralizing or decentralizing in different contexts, especially at different points in time. Nowadays guns are largely a decentralizing technology (e.g. it's easier for a gang to wage war on a police force than it was for medieval European peasants to wage war on plate-armored knights), but initially they were highly centralizing (only trained armies with access to the capital and expertise requires to smith guns). I wonder if AI proliferation will turn an initially centralizing technology decentralizing? Idk
It's a good point. I think that I'm trying to grasp at something broader in some sense. Something like, "the existence of guns as a class is decentralizing even when the guns are held by a small number of people." I'm not sure I could coherently make that argument, I'd need to reread the langdon winner article i cited in the post (he does a better job of it). Winner talks about nuclear weapons in this way -- the possibility of accidental detonation is so catastrophic that they basically require / create strict hierarchy, militarized command structure, constant surveillance, and authoritarian power. So my argument may be something like, 'once you get your hands on a gun, you are immediately at level with everyone else, and so any centralization actually happens outside the gun ownership / in order to contain it. This is pretty hard to do since guns are pretty small, easy to move, and inert until used' something like this
If you haven't read Brynjolfsson's "Use of knowledge by AI in society", it basically argues that in a Hayekian world, knowledge is difficult to define, and inherently local, so distributed systems tend to work better. But in a post-Hayekian world, as more knowledge moves from tacit to codified, then it's easier to pool information, which removes barriers to centralization.
The precise point that happens is the balance between how much is gained by having "one single planner" who can take advantage of all that knowledge, versus how much benefit accrues to individuals by NOT sharing their knowledge. I can imagine that open-weight models increase local knowledge, but also that pooling info produces better results. So I dunno if I have a clear call on which it is.
Very thoughtful take! A nuance that maybe you didn't consider, maybe didn't make it into the post, or maybe just isn't very useful: technologies can be centralizing or decentralizing in different contexts, especially at different points in time. Nowadays guns are largely a decentralizing technology (e.g. it's easier for a gang to wage war on a police force than it was for medieval European peasants to wage war on plate-armored knights), but initially they were highly centralizing (only trained armies with access to the capital and expertise requires to smith guns). I wonder if AI proliferation will turn an initially centralizing technology decentralizing? Idk
It's a good point. I think that I'm trying to grasp at something broader in some sense. Something like, "the existence of guns as a class is decentralizing even when the guns are held by a small number of people." I'm not sure I could coherently make that argument, I'd need to reread the langdon winner article i cited in the post (he does a better job of it). Winner talks about nuclear weapons in this way -- the possibility of accidental detonation is so catastrophic that they basically require / create strict hierarchy, militarized command structure, constant surveillance, and authoritarian power. So my argument may be something like, 'once you get your hands on a gun, you are immediately at level with everyone else, and so any centralization actually happens outside the gun ownership / in order to contain it. This is pretty hard to do since guns are pretty small, easy to move, and inert until used' something like this
Oh, interesting! I think I didn't understand the sense of "structural" that you were using here. I'll need to chew on this more!
If you haven't read Brynjolfsson's "Use of knowledge by AI in society", it basically argues that in a Hayekian world, knowledge is difficult to define, and inherently local, so distributed systems tend to work better. But in a post-Hayekian world, as more knowledge moves from tacit to codified, then it's easier to pool information, which removes barriers to centralization.
The precise point that happens is the balance between how much is gained by having "one single planner" who can take advantage of all that knowledge, versus how much benefit accrues to individuals by NOT sharing their knowledge. I can imagine that open-weight models increase local knowledge, but also that pooling info produces better results. So I dunno if I have a clear call on which it is.
https://www.nber.org/system/files/chapters/c15303/c15303.pdf