The failure mode you're describing is one I keep seeing in a very different context: individual decision-making. People often come to me having optimized everything locally, their habits, their knowledge, their relationships, but without any central principle that holds the system together. Every part is working. The whole is drifting. Decentralization fails for individuals for the same reason it fails for organizations: when there's no authority that can override local preferences in service of a larger coherence, the system doesn't collapse suddenly. It just slowly becomes incoherent. The question I keep returning to is not "how much structure is enough?" but "what is the structure in service of?" Without that answer, more coordination just produces a more organized kind of lost.
I have a number of Proudhonist impulses: it fails very gracefully, and I think there is much more virtue in trying to make self-governing communes work at the level of apartment buildings and retail stores than in trying to impose it or even demand subsidies for it. I wish more people redirected their impulses from grand political schemes and towards practical implementations that outperform, if they (and on some days I) am right, existing alternatives. I have the same attitudes towards most religious communities.
But I think it's worth drawing out a distinction between the communities Ostrom examines and AI users, implementers, and related communities of practice. Namely, Ostrom is examining communities that both benefit from consumption and want to preserve value. That can be very different from, say, conductors and taxi drivers who are eager to have the chance to say "as the relevant experts this technology is completely unsafe and can't be used" when what they mean is "you're going to take away my job". So long as that distinction exists, I am very wary of giving too much deference to communities of practice.
Pretty good. The key move is replacing “community governance” with domain-specific institutional governance.
AI is too complex for one regulator, but too infrastructural for pure participation.
Governance has to move to the layer where operational knowledge is actually produced: medicine, law, education, insurance, procurement, open source, security.
That is the Ostrom/Tocqueville answer: not decentralization as romance, but overlapping institutions with partial jurisdiction.
The harder AI-specific problem is that deployment governance does not fully solve production concentration.
Compute, model weights, cloud access, APIs, benchmarks, and security thresholds sit upstream of most intermediary institutions.
So the future is not “centralized AI vs democratic AI.”
It is whether legitimacy can be rebuilt downstream while execution consolidates upstream.
Freedom in the AI age will not come from making power small.
It will come from making power answerable to more than one institutional authority at once.
The polycentric move is right, but it skips the substrate.
Ostrom’s commons worked because people had relational thickness, shared risk, and time horizons long enough to absorb conflict. Without those conditions, “nested enterprises” becomes another architecture for capture.
Intermediary institutions do not push back because they occupy the right box on an org chart. They push back when the people inside them can survive refusal.
If refusal means delisting, lost referrals, or retaliation, medical boards, insurers, and procurement officers will not countervail power. They will translate compliance.
This is not only a governance problem. It is a capacity problem. Build the conditions for dissent, or distributed governance becomes distributed fawning.
I was surprised to see Ruha Benjamin, AJL, and DAIR mentioned here by Cosmos and Chalmers. And yet, are all of these folks being bucketed as stochastic parrot type-whiners and a wonderful implication of "most anarchist thinkers?"
Help me out here, has Cosmos published much other discussion around such groups? I'm not the the most avid reader recently, but I'm curious what else has been said.
This might be worth a broader response. But at first pass, it's perhaps unfortunate for some of the valid points made about what layers need to be built "supported by" claims made in such a fashion. I guess Cosmos will inevitably draw such stoking phrasing given what it's dealing with, but, still. 'To speak "bluntly"', even.
To be clear, are you saying that it is unfair to describe DAIR as being associated with the stochastic parrots perspective? DAIR which has as a founder and Executive Director Dr. Timnit Gebru, coauthor on the Stochastic Parrots paper? Whose Director of Research, Dr. Alex Hanna, coauthored The AI Con with Dr. Emily Bender? If it is unfair to describe DAIR that way, who on earth *could* be bucketed as having the stochastic parrots?
"To be clear, are you saying that it is unfair to describe DAIR as being associated with the stochastic parrots perspective?"
No.
And well met.
"If it is unfair to describe DAIR that way, who on earth *could* be bucketed as having the stochastic parrots?"
I agree that grouping an entire organization (or multiple organizations and scholars, for that matter) into one bucket to serve a narrative would be not particularly moving. If we want the future to have untrivialized and undiluted substance we must aim for it ourselves in the present, no?
'Who on earth could be bucketed as having the stochastic parrots' is definitely what I'm going to be thinking about for part of this evening, though - thank you for raising that question.
The failure mode you're describing is one I keep seeing in a very different context: individual decision-making. People often come to me having optimized everything locally, their habits, their knowledge, their relationships, but without any central principle that holds the system together. Every part is working. The whole is drifting. Decentralization fails for individuals for the same reason it fails for organizations: when there's no authority that can override local preferences in service of a larger coherence, the system doesn't collapse suddenly. It just slowly becomes incoherent. The question I keep returning to is not "how much structure is enough?" but "what is the structure in service of?" Without that answer, more coordination just produces a more organized kind of lost.
I have a number of Proudhonist impulses: it fails very gracefully, and I think there is much more virtue in trying to make self-governing communes work at the level of apartment buildings and retail stores than in trying to impose it or even demand subsidies for it. I wish more people redirected their impulses from grand political schemes and towards practical implementations that outperform, if they (and on some days I) am right, existing alternatives. I have the same attitudes towards most religious communities.
But I think it's worth drawing out a distinction between the communities Ostrom examines and AI users, implementers, and related communities of practice. Namely, Ostrom is examining communities that both benefit from consumption and want to preserve value. That can be very different from, say, conductors and taxi drivers who are eager to have the chance to say "as the relevant experts this technology is completely unsafe and can't be used" when what they mean is "you're going to take away my job". So long as that distinction exists, I am very wary of giving too much deference to communities of practice.
Pretty good. The key move is replacing “community governance” with domain-specific institutional governance.
AI is too complex for one regulator, but too infrastructural for pure participation.
Governance has to move to the layer where operational knowledge is actually produced: medicine, law, education, insurance, procurement, open source, security.
That is the Ostrom/Tocqueville answer: not decentralization as romance, but overlapping institutions with partial jurisdiction.
The harder AI-specific problem is that deployment governance does not fully solve production concentration.
Compute, model weights, cloud access, APIs, benchmarks, and security thresholds sit upstream of most intermediary institutions.
So the future is not “centralized AI vs democratic AI.”
It is whether legitimacy can be rebuilt downstream while execution consolidates upstream.
Freedom in the AI age will not come from making power small.
It will come from making power answerable to more than one institutional authority at once.
The polycentric move is right, but it skips the substrate.
Ostrom’s commons worked because people had relational thickness, shared risk, and time horizons long enough to absorb conflict. Without those conditions, “nested enterprises” becomes another architecture for capture.
Intermediary institutions do not push back because they occupy the right box on an org chart. They push back when the people inside them can survive refusal.
If refusal means delisting, lost referrals, or retaliation, medical boards, insurers, and procurement officers will not countervail power. They will translate compliance.
This is not only a governance problem. It is a capacity problem. Build the conditions for dissent, or distributed governance becomes distributed fawning.
i'd be wary of confusing diagnosis for treatment
I was surprised to see Ruha Benjamin, AJL, and DAIR mentioned here by Cosmos and Chalmers. And yet, are all of these folks being bucketed as stochastic parrot type-whiners and a wonderful implication of "most anarchist thinkers?"
Help me out here, has Cosmos published much other discussion around such groups? I'm not the the most avid reader recently, but I'm curious what else has been said.
This might be worth a broader response. But at first pass, it's perhaps unfortunate for some of the valid points made about what layers need to be built "supported by" claims made in such a fashion. I guess Cosmos will inevitably draw such stoking phrasing given what it's dealing with, but, still. 'To speak "bluntly"', even.
To be clear, are you saying that it is unfair to describe DAIR as being associated with the stochastic parrots perspective? DAIR which has as a founder and Executive Director Dr. Timnit Gebru, coauthor on the Stochastic Parrots paper? Whose Director of Research, Dr. Alex Hanna, coauthored The AI Con with Dr. Emily Bender? If it is unfair to describe DAIR that way, who on earth *could* be bucketed as having the stochastic parrots?
"To be clear, are you saying that it is unfair to describe DAIR as being associated with the stochastic parrots perspective?"
No.
And well met.
"If it is unfair to describe DAIR that way, who on earth *could* be bucketed as having the stochastic parrots?"
I agree that grouping an entire organization (or multiple organizations and scholars, for that matter) into one bucket to serve a narrative would be not particularly moving. If we want the future to have untrivialized and undiluted substance we must aim for it ourselves in the present, no?
'Who on earth could be bucketed as having the stochastic parrots' is definitely what I'm going to be thinking about for part of this evening, though - thank you for raising that question.