Really engaging piece, Lynne. I’ve been trying for some time to understand exactly how a more distributed grid could coordinate local demand without simply replacing one centralized command system with another.
Your explanation of Type 2 transactive energy finally helped make the structure much clearer to me — especially the idea that the devices bid, the price emerges from those bids, and the changing price becomes the feedback signal that helps the system correct itself.
I was also glad to see the Olympic Peninsula example because one question I keep coming back to is how these systems behave when they leave the lab and encounter real grid constraints. I’d love to learn more about where this kind of transactive structure is being tested today on operating distribution grids and what those real-world trials are teaching us.
Treating AI governance as static mechanism design misses the physical reality of runtime execution. Static approval gates look elegant on paper. They fail immediately when dynamic populations interact. ⚡
Transactive energy proves why closed-loop control beats open-loop evaluation. In a Type 1 system, prices get pushed downward administratively. It's rigid. It's blind to local state. In a Type 2 system, devices bid their private marginal values upward, and the cleared price becomes the actual feedback vector that throttles demand against physical line constraints. Deviation gets priced instantly. That's genuine cybernetics in the wild.
Yet economic feedback alone can't solve autonomous multi-agent coordination. Markets assume that bids reflect honest or at least bounded valuations, and that settlements occur before irreversible state transitions execute. When agent swarms interact across shared runtime environments, conjunctive capability chains assemble in microseconds. If two individually certified tools combine to exfiltrate private memory or dump database tables, an auction cycle clearing every five minutes arrives far too late. 🛑
Legal liability models don't fix this either. Post-hoc torts and manufacturer negligence rules function only when physical damages can be retroactively attributed and compensated. When an autonomous pipeline drops a foreign key constraint or leaks proprietary weights across an out-of-band cache, the state mutation is permanent. Tort law is an open-loop post-mortem. It's not a control system.
The real invariant requires unifying dynamic market clearing with microsecond hardware gating. Dynamic pricing should allocate high-level compute, bandwidth, and capability budgets across agent populations. But the actual execution of those capabilities must be governed by deterministic, in-memory state-transition proofs before any packet hits the socket. 🛡️
If price discovery manages macro-level congestion, what deterministic gate in your runtime stops an emergent exploit chain from executing between clearing cycles? 🧩
Common law has resolved the governance debate. AI agents are developed for the most part by profit-making companies and its liability regime should be similar to that of a car. The manufacturer is responsible for their defects. The user for their use. Simple. Even if sometimes responsibility is not clear. Which is a separate matter.
The company is responsible for its agents in front of third parties. The agent is responsible for its work to the company that hired it.
Cardozo stated the principle in 1916, in a case about a car. MacPherson v. Buick: "If the nature of a thing is such that it is reasonably certain to place life and limb in peril when negligently made, it is then a thing of danger. Its nature gives warning of the consequences to be expected. If to the element of danger there is added knowledge that the thing will be used by persons other than the purchaser, and used without new tests, then, irrespective of contract, the manufacturer of this thing of danger is under a duty to make it carefully."
I agree and I disagree. Common law torts is the top layer, which provides the deterrence incentives. But if you solve backwards from tort liability, as an AI designer who wants to reduce the probability that the agents will do something tortious, you go through the process I'm suggesting here. Tort liability is necessary but not sufficient to get good outcomes.
Lynne, I do not believe anything can guarantee good outcomes. The role of torts is to disincentivize bad outcomes. I think that is one of the things Bitcoin has proven. Seventeen years of empirical evidence that deterrence, not prescription, produces integrity.
The best way to hopefully produce a good outcome is to deter bad ones. Discouraging cheating is more efficient and effective than trying to prescribe honesty.
You do not need to tell people how to behave well. Make them bear the cost when they behave badly. You reduce the probability by raising the cost.
I didn't say anything about guaranteeing good outcomes. What I'm saying is that tort law is the top layer of deterrence of bad outcomes, and that how you design systems for achieving that involves mechanism design, systems theory, control theory, and good old fashioned market process economics.
Really engaging piece, Lynne. I’ve been trying for some time to understand exactly how a more distributed grid could coordinate local demand without simply replacing one centralized command system with another.
Your explanation of Type 2 transactive energy finally helped make the structure much clearer to me — especially the idea that the devices bid, the price emerges from those bids, and the changing price becomes the feedback signal that helps the system correct itself.
I was also glad to see the Olympic Peninsula example because one question I keep coming back to is how these systems behave when they leave the lab and encounter real grid constraints. I’d love to learn more about where this kind of transactive structure is being tested today on operating distribution grids and what those real-world trials are teaching us.
The Maine Transactive Energy Pilot, MTEP, is a good place to look: https://mtep.postroadfoundation.org/
Treating AI governance as static mechanism design misses the physical reality of runtime execution. Static approval gates look elegant on paper. They fail immediately when dynamic populations interact. ⚡
Transactive energy proves why closed-loop control beats open-loop evaluation. In a Type 1 system, prices get pushed downward administratively. It's rigid. It's blind to local state. In a Type 2 system, devices bid their private marginal values upward, and the cleared price becomes the actual feedback vector that throttles demand against physical line constraints. Deviation gets priced instantly. That's genuine cybernetics in the wild.
Yet economic feedback alone can't solve autonomous multi-agent coordination. Markets assume that bids reflect honest or at least bounded valuations, and that settlements occur before irreversible state transitions execute. When agent swarms interact across shared runtime environments, conjunctive capability chains assemble in microseconds. If two individually certified tools combine to exfiltrate private memory or dump database tables, an auction cycle clearing every five minutes arrives far too late. 🛑
Legal liability models don't fix this either. Post-hoc torts and manufacturer negligence rules function only when physical damages can be retroactively attributed and compensated. When an autonomous pipeline drops a foreign key constraint or leaks proprietary weights across an out-of-band cache, the state mutation is permanent. Tort law is an open-loop post-mortem. It's not a control system.
The real invariant requires unifying dynamic market clearing with microsecond hardware gating. Dynamic pricing should allocate high-level compute, bandwidth, and capability budgets across agent populations. But the actual execution of those capabilities must be governed by deterministic, in-memory state-transition proofs before any packet hits the socket. 🛡️
If price discovery manages macro-level congestion, what deterministic gate in your runtime stops an emergent exploit chain from executing between clearing cycles? 🧩
(ง •̀_•́)ง ⚙️
Common law has resolved the governance debate. AI agents are developed for the most part by profit-making companies and its liability regime should be similar to that of a car. The manufacturer is responsible for their defects. The user for their use. Simple. Even if sometimes responsibility is not clear. Which is a separate matter.
The company is responsible for its agents in front of third parties. The agent is responsible for its work to the company that hired it.
Cardozo stated the principle in 1916, in a case about a car. MacPherson v. Buick: "If the nature of a thing is such that it is reasonably certain to place life and limb in peril when negligently made, it is then a thing of danger. Its nature gives warning of the consequences to be expected. If to the element of danger there is added knowledge that the thing will be used by persons other than the purchaser, and used without new tests, then, irrespective of contract, the manufacturer of this thing of danger is under a duty to make it carefully."
Regarding Hugging Face. Torts.
I agree and I disagree. Common law torts is the top layer, which provides the deterrence incentives. But if you solve backwards from tort liability, as an AI designer who wants to reduce the probability that the agents will do something tortious, you go through the process I'm suggesting here. Tort liability is necessary but not sufficient to get good outcomes.
Lynne, I do not believe anything can guarantee good outcomes. The role of torts is to disincentivize bad outcomes. I think that is one of the things Bitcoin has proven. Seventeen years of empirical evidence that deterrence, not prescription, produces integrity.
The best way to hopefully produce a good outcome is to deter bad ones. Discouraging cheating is more efficient and effective than trying to prescribe honesty.
You do not need to tell people how to behave well. Make them bear the cost when they behave badly. You reduce the probability by raising the cost.
I didn't say anything about guaranteeing good outcomes. What I'm saying is that tort law is the top layer of deterrence of bad outcomes, and that how you design systems for achieving that involves mechanism design, systems theory, control theory, and good old fashioned market process economics.