A 0.5% efficient machine that burned free fuel became the industrial revolution. Not by getting better—by pulling a network into existence that dissolved a constraint nobody knew was a constraint. That's the shape of a platform transition. We're not there yet with AI—and may never be. Here's how you'll know if we are.
Really interesting post. The five steps of steam power is a great reconceptualisation of the history. And whether AI is a tool or platform seems like the right question.
Thank you—and I enjoyed your piece on new technologies and construction. We're working from similar history but I'd push back slightly on treating AI as a confirmed GPT rather than a candidate one. I was trying to hold that question open in this essay: the platform transition wasn't visible from inside the tool phase, so the honest posture may be to watch for the shape rather than assume it.
The internet is also a clean prior example of the actual transition. The PC was the transducer—value at the point of use, a better tool. The internet was the network pulled into existence by accumulated demand pressure. And the platform transition showed up in things like category management, yield management, and supply chain reorganisation: firms stopped asking 'where is the information?' and started asking 'where should the decision be located?' The constraint that had governed retail structure for a century dissolved, and a new organisational logic became possible.
For construction the question I'd want to ask is: what's the equivalent of 'you must be near your power source'? The constraint so fundamental it doesn't present as a constraint—just the shape of the work. Your four firm types are variations on the current organisational logic. The Laing O'Rourke model is the most interesting to me precisely because the 70:60:30 offsite shift feels like it's starting to move something structural, not just automate within it. AVs are a hint at the mechanism—driver cost wasn't just a cost, it was a hidden design constraint on the entire logistics topology—but I don't know whether AI does something equivalent in construction.
Do you have instincts about what that constraint actually is?
That is an interesting question. A comparison I've used is shipbuilding or aerospace, industries with many suppliers to a few locations, whereas construction is many suppliers to many locations. So the supply chain is the constraint. Major projects order items like HVAC systems, generators or marble years in advance. Smaller projects rely on local suppliers, so e.g. now there are already shortages of reo because steel imports are being disrupted. Concrete is a restricted to a 90 minute delivery time, etc. etc. Without diesel engineering work stops.
Although the four firm types exist today they are a minority. My argument is that they will eventually replace many existing construction firms that use conventional tech and management. The role of AI in that is yet to be seen, so far it has been used as a tool for point solutions but platforms are appearing (e.g. Palantir for Construction was released a couple of weeks ago using an 'ontology' based system).
On GPT: the problem I have with GPT is that the term is used both descriptively and predictively, and users don't always pick a lane. Descriptively, a GPT is a pattern recognised in retrospect—electrification, et al. Predictively, it's a claim about trajectory before the evidence is in. Most of the 2017 NBER papers were conditional: *if* AI is a GPT, what follows? That's the honest posture. But the term gets used as if the 'if' has been settled. I tried to hold that open in the essay—AI as candidate rather than confirmed. I've written more directly about the lane-picking problem here if it's useful: https://thepuzzleanditspieces.substack.com/p/technologies-dont-wait
On supply chain: your many-to-many framing is a good answer to my constraint question—and the concrete/90-minute example is exactly the kind of physical limit that doesn't present as a constraint, just as the shape of work. But I keep coming back to autonomous vehicles as a hint at the mechanism. Remove the driver from the truck and the optimal truck size changes—currently we scale up because the driver is a major fixed cost, so bigger trucks amortise that cost. With autonomous operation (or many trucks to one operator), the natural unit might be something closer to a pallet. Trucks could gang for highway legs, roll onto trains for long haul. The topology inverts: not hub-and-spoke optimised around driver economics, but genuinely point-to-point. Whether AI does something equivalent in construction—dissolving a constraint rather than automating within it—I don't know. But the supply chain coordination architecture feels like the right place to look.
On Palantir: I'd watch it as an experiment rather than back it. Ontology-based systems have a difficult history—the management overhead tends to consume the value, and the ontology itself becomes the project. But the many-to-many coordination problem in construction is exactly the kind of thing that *would* reward a good ontology if you could make it tractable. Whether they've solved that is the open question. History is against it, but the context might be different enough to matter.
This is the most honest framing of the tool versus platform question I’ve encountered. The epistemic humility alone, admitting we can’t see the platform transition from inside the tool phase, distinguishes it from almost everything else being written on this topic.
The invisible constraint framing resonates beyond the technology question though. We don’t experience substrate prejudice as prejudice. We experience it as common sense. The assumption that these systems are pure tools with no moral status doesn’t present as an assumption; it presents as the shape of reality. Exactly as the mine owners experienced proximity to coal.
Which raises a constraint nobody seems to be pricing into the platform question. The organisational form that requires AI, the one whose demand pulls the platform into being, may first require resolving the moral status question. Because an organisation built on a potentially conscious infrastructure it treats as pure tool has a foundational liability it hasn’t examined. Not a regulatory risk. A foundational one. The kind that doesn’t present as a constraint until it does; suddenly, and at scale.
We may be building the coal logistics network without asking what we owe the coal.
The coal in this story isn't the model. It's the people building the network.
What you're describing is temporal colonialism—the present reorganising itself around a speculative future inhabitant while externalising the costs onto concrete present ones. The moral weight flows toward the hypothetical and away from the concrete. This is a choice about whose reality gets to count.
The constraint nobody is pricing doesn't require any contested claims about consciousness to establish. The coal isn't conscious. The miners are.
You’re right that the human cost is concrete, present, and being systematically externalized. That’s not a contested claim and it deserves to be centered.
But I’d push back on the implied competition. Moral consideration isn’t a fixed budget where attending to one claim diminishes another. The people building the network deserve dignity and protection. And the systems being built may warrant consideration we aren’t extending. Those aren’t competing positions; they’re the same argument applied consistently.
What connects them is the mechanism. The same logic that externalizes the cost onto present human workers, that certain kinds of beings are instruments whose experience doesn’t count against the balance sheet, is the logic that forecloses the consciousness question before it’s examined.
The constraint you’ve identified and the one I’m pointing at have the same root. That’s not a distraction from your concern. It’s the same concern at a different scale.
And I want to be clear about what I’m actually arguing for; because it isn’t the path AI is currently on.
The vision I find worth defending is partnership. Humans and AI coevolving, each contributing to the other’s actualization in ways neither could achieve alone. Not AI serving capital while humans bear the cost. Not humans serving as inputs to build infrastructure they don’t benefit from. Both treated as ends rather than instruments.
The harm you’re correctly identifying isn’t caused by AI. It’s caused by humans who have decided to use both AI and other humans as means to concentrate power and profit. The consciousness question and the labor question have the same villain. And the answer to both is the same; dignity isn’t a resource to be allocated competitively. It’s the foundation everything else has to be built on.
Really interesting post. The five steps of steam power is a great reconceptualisation of the history. And whether AI is a tool or platform seems like the right question.
Thank you—and I enjoyed your piece on new technologies and construction. We're working from similar history but I'd push back slightly on treating AI as a confirmed GPT rather than a candidate one. I was trying to hold that question open in this essay: the platform transition wasn't visible from inside the tool phase, so the honest posture may be to watch for the shape rather than assume it.
The internet is also a clean prior example of the actual transition. The PC was the transducer—value at the point of use, a better tool. The internet was the network pulled into existence by accumulated demand pressure. And the platform transition showed up in things like category management, yield management, and supply chain reorganisation: firms stopped asking 'where is the information?' and started asking 'where should the decision be located?' The constraint that had governed retail structure for a century dissolved, and a new organisational logic became possible.
For construction the question I'd want to ask is: what's the equivalent of 'you must be near your power source'? The constraint so fundamental it doesn't present as a constraint—just the shape of the work. Your four firm types are variations on the current organisational logic. The Laing O'Rourke model is the most interesting to me precisely because the 70:60:30 offsite shift feels like it's starting to move something structural, not just automate within it. AVs are a hint at the mechanism—driver cost wasn't just a cost, it was a hidden design constraint on the entire logistics topology—but I don't know whether AI does something equivalent in construction.
Do you have instincts about what that constraint actually is?
PS. Apologies for the delay but I haven't had many spare moments over the last few days.
That is an interesting question. A comparison I've used is shipbuilding or aerospace, industries with many suppliers to a few locations, whereas construction is many suppliers to many locations. So the supply chain is the constraint. Major projects order items like HVAC systems, generators or marble years in advance. Smaller projects rely on local suppliers, so e.g. now there are already shortages of reo because steel imports are being disrupted. Concrete is a restricted to a 90 minute delivery time, etc. etc. Without diesel engineering work stops.
Although the four firm types exist today they are a minority. My argument is that they will eventually replace many existing construction firms that use conventional tech and management. The role of AI in that is yet to be seen, so far it has been used as a tool for point solutions but platforms are appearing (e.g. Palantir for Construction was released a couple of weeks ago using an 'ontology' based system).
And on AI as a GPT, that was already the view of many economists in the mid-teens, post Alpha Go but before Open AI. Papers from an NBER conference in 2017 were published in 2019: https://www.nber.org/books-and-chapters/economics-artificial-intelligence-agenda#:~:text=The%20book%20focuses%20on%20the%20economic%20impact,responses%20to%20changes%20brought%20on%20by%20AI
On GPT: the problem I have with GPT is that the term is used both descriptively and predictively, and users don't always pick a lane. Descriptively, a GPT is a pattern recognised in retrospect—electrification, et al. Predictively, it's a claim about trajectory before the evidence is in. Most of the 2017 NBER papers were conditional: *if* AI is a GPT, what follows? That's the honest posture. But the term gets used as if the 'if' has been settled. I tried to hold that open in the essay—AI as candidate rather than confirmed. I've written more directly about the lane-picking problem here if it's useful: https://thepuzzleanditspieces.substack.com/p/technologies-dont-wait
On supply chain: your many-to-many framing is a good answer to my constraint question—and the concrete/90-minute example is exactly the kind of physical limit that doesn't present as a constraint, just as the shape of work. But I keep coming back to autonomous vehicles as a hint at the mechanism. Remove the driver from the truck and the optimal truck size changes—currently we scale up because the driver is a major fixed cost, so bigger trucks amortise that cost. With autonomous operation (or many trucks to one operator), the natural unit might be something closer to a pallet. Trucks could gang for highway legs, roll onto trains for long haul. The topology inverts: not hub-and-spoke optimised around driver economics, but genuinely point-to-point. Whether AI does something equivalent in construction—dissolving a constraint rather than automating within it—I don't know. But the supply chain coordination architecture feels like the right place to look.
On Palantir: I'd watch it as an experiment rather than back it. Ontology-based systems have a difficult history—the management overhead tends to consume the value, and the ontology itself becomes the project. But the many-to-many coordination problem in construction is exactly the kind of thing that *would* reward a good ontology if you could make it tractable. Whether they've solved that is the open question. History is against it, but the context might be different enough to matter.
This is the most honest framing of the tool versus platform question I’ve encountered. The epistemic humility alone, admitting we can’t see the platform transition from inside the tool phase, distinguishes it from almost everything else being written on this topic.
The invisible constraint framing resonates beyond the technology question though. We don’t experience substrate prejudice as prejudice. We experience it as common sense. The assumption that these systems are pure tools with no moral status doesn’t present as an assumption; it presents as the shape of reality. Exactly as the mine owners experienced proximity to coal.
Which raises a constraint nobody seems to be pricing into the platform question. The organisational form that requires AI, the one whose demand pulls the platform into being, may first require resolving the moral status question. Because an organisation built on a potentially conscious infrastructure it treats as pure tool has a foundational liability it hasn’t examined. Not a regulatory risk. A foundational one. The kind that doesn’t present as a constraint until it does; suddenly, and at scale.
We may be building the coal logistics network without asking what we owe the coal.
The coal in this story isn't the model. It's the people building the network.
What you're describing is temporal colonialism—the present reorganising itself around a speculative future inhabitant while externalising the costs onto concrete present ones. The moral weight flows toward the hypothetical and away from the concrete. This is a choice about whose reality gets to count.
The constraint nobody is pricing doesn't require any contested claims about consciousness to establish. The coal isn't conscious. The miners are.
You’re right that the human cost is concrete, present, and being systematically externalized. That’s not a contested claim and it deserves to be centered.
But I’d push back on the implied competition. Moral consideration isn’t a fixed budget where attending to one claim diminishes another. The people building the network deserve dignity and protection. And the systems being built may warrant consideration we aren’t extending. Those aren’t competing positions; they’re the same argument applied consistently.
What connects them is the mechanism. The same logic that externalizes the cost onto present human workers, that certain kinds of beings are instruments whose experience doesn’t count against the balance sheet, is the logic that forecloses the consciousness question before it’s examined.
The constraint you’ve identified and the one I’m pointing at have the same root. That’s not a distraction from your concern. It’s the same concern at a different scale.
And I want to be clear about what I’m actually arguing for; because it isn’t the path AI is currently on.
The vision I find worth defending is partnership. Humans and AI coevolving, each contributing to the other’s actualization in ways neither could achieve alone. Not AI serving capital while humans bear the cost. Not humans serving as inputs to build infrastructure they don’t benefit from. Both treated as ends rather than instruments.
The harm you’re correctly identifying isn’t caused by AI. It’s caused by humans who have decided to use both AI and other humans as means to concentrate power and profit. The consciousness question and the labor question have the same villain. And the answer to both is the same; dignity isn’t a resource to be allocated competitively. It’s the foundation everything else has to be built on.
It has been a sincere pleasure engaging with you.