Please don’t take this as criticism of your post. I’m just trying to point out some areas that may warrant deeper consideration. My perspective comes less from reading about these things and more from 25+ years of implementing enterprise-level systems across the financial services, technology, and manufacturing industries, government, and many other environments, mostly in Fortune 50 companies and three major US federal departments. That experience has taught me a lot about the realities behind some of the ideas in your post.
I do think there is a real idea here. Even a partial or imperfect world model of a business could be useful. If AI can make workflows, bottlenecks, exception paths, and parts of operational state more legible, that is already valuable.
Where I hesitate is that the post seems to move very quickly from that modest claim to a much stronger one that is much harder to defend.
A company is not just a set of processes waiting to be mapped into an environment with defined action spaces and evaluation criteria. A great deal of what actually determines how a company works is tacit, political, relational, and historically contingent. It lives in people’s heads, in trust, in fear, in unwritten rules, in informal influence, in who can block what, and in how decisions are really made versus how they are described. Even people who have spent years inside an organization usually understand only part of that reality.
That is why the idea of an “operating partner in software” does not fully work for me. The value of a strong operator is not just that they can observe workflows. It is that, over time, they develop judgment about people, incentives, credibility, conflict, and context. That kind of understanding is not simply unstructured data waiting to be captured. Much of it is only visible through long participation in the organization itself.
I also think the post may understate a second risk: better visibility does not automatically lead to better management. In many cases, it leads to more intervention. If leaders feel they can see the business in real time, they may start reacting to every fluctuation like a trader watching a market. That can create churn, metric gaming, and local optimization rather than better decisions. Sometimes the most valuable output of a model is restraint, not action.
So I agree with the direction in a limited sense: better operational models could absolutely help firms. But the stronger claim — that this can become something like a true-world model of the business across thousands of companies and serve as a substitute for understanding deeply embedded humans — feels overstated to me. The hardest part of a firm is not just operational complexity. It is that firms are social and political systems, and that is exactly the part that resists clean formalization.
@Marginal Gains' point about firms as social and political systems is the crux imo. The world model you describe requires a world that's already been mapped. But much of what determines how companies actually work is tacit, relational, historically contingent as MG points out.
There's also a topology angle here. AI will bite fastest in cos/ economies that are already legible—standardized, digitized, SaaS-eaten. That's many parts of the US & WE. India, by contrast, runs on relationships, informal networks, and workflows that were never proceduralized. China has state-driven legibility but also a massive informal layer. So the Starcraft vision may arrive in some geographies a decade or more before others—and in some sectors, not at all.
The other comment I have: if the COO's job becomes triage and simulation, how do they learn the judgment that comes from doing the work? Doing the mundane task early on in one's career was how you learned the org purpose. The COO's ability to spot when the model is wrong comes from having worked on the business directly. Simulate long enough without operating & you may lose the intuition the simulation depends on. The future COO who only triages may not recognize when the model is making up stuff to them.
Not saying this won't happen. But "Starcraft for CEOs" assumes a level of prior legibility that varies wildly by geography and sector—and a transfer of judgment that may not survive the abstraction.
One recent real-world example that may be relevant here is the Washington, DC sewer collapse, which I have been following closely for the last two months. I also wrote a short note about it after reading additional reporting, including The Atlantic’s coverage of the likely cause.
What makes that incident so instructive is that it was not simply a case of “old infrastructure failed.” The deeper issue appears to have been a loss of institutional understanding. There may have been hidden construction conditions, undocumented deviations, incomplete records, conflicting inspection interpretations, and design or construction decisions made decades ago that no one fully understood until failure forced them into view.
As I mentioned above, complex systems do not run on formal documentation alone. They also run on tacit knowledge, exception handling, memory of past anomalies, judgment under uncertainty, and the accumulated experience of people who know where reality diverges from the manual. In many environments, the most important knowledge is not just how the process is supposed to work, but where the maps are wrong, where workarounds have become permanent, where standard practice was never actually followed, and which weak signals deserve attention even when the dashboards appear normal.
That is why I think there is a real risk in replacing human labor with AI before organizations do the much harder work of capturing, testing, and preserving the human reasoning that actually keeps systems resilient. AI can process reports, flag anomalies, and optimize what is visible. But if the human layer disappears before institutional memory is preserved, the system may become very efficient at operating on partial understanding.
In that sense, the Washington sewer collapse is a useful warning. The danger is not just missing data. It is mistaking documentation for understanding, visibility for comprehension, and digitization for genuine knowledge transfer.
So I do think AI can help here, but ideally as a tool for preserving expert reasoning, surfacing undocumented dependencies, comparing conflicting interpretations, and exposing uncertainty before organizations start treating it as a substitute for the people in the name of cost-cutting and efficiency, who still carry the deepest understanding of how things actually work.
The following relates directly to the broader point in this conversation: greater visibility, even extraordinarily greater visibility, does not automatically produce better judgment. I am not trying to argue here whether the Iran war is right or wrong. My point is narrower. Knowing more, even knowing vastly more, does not guarantee the right decision in the moment. Only time will tell whether it proves right in the long run. A New York Times opinion piece makes this tension clear. The spycraft behind the planning and execution appears to have been extensive. Recent reporting suggests that Israeli intelligence spent years penetrating Tehran’s traffic-camera and communications networks, building what one source described as an AI-powered “target-production machine” capable of turning enormous volumes of visual, human, and signals intelligence into precise strike coordinates. That is an extraordinary achievement in surveillance and targeting. But it also illustrates the limit running through this entire discussion: legibility is not the same as understanding. Never has so much been seen, so precisely, by people who may still understand too little of what they are seeing. A system can tell you where a man is. It cannot tell you what his death will mean to a nation, a movement, or a generation. These systems are trained on behavior, not meaning. They can track what an adversary does, but not what he fears, honors, remembers, or what he is willing to die for.
That is the deeper issue with strong claims about world models, whether in firms, infrastructure, or conflict. The visible layer can become highly machine-legible while the human layer remains poorly understood. And in many cases, it is exactly that human layer that determines whether an apparently precise action turns out to be wise, catastrophic, or both.
The danger is that machines/AI may become better at locating targets than we are at understanding consequences.
The following relates directly to the broader point in this conversation: greater visibility, even extraordinarily greater visibility, does not automatically produce better judgment. I am not trying to argue here whether the Iran war is right or wrong. My point is narrower. Knowing more, even knowing vastly more, does not guarantee the right decision in the moment. Only time will tell whether it proves right in the long run. A New York Times opinion piece makes this tension clear. The spycraft behind the planning and execution appears to have been extensive. Recent reporting suggests that Israeli intelligence spent years penetrating Tehran’s traffic-camera and communications networks, building what one source described as an AI-powered “target-production machine” capable of turning enormous volumes of visual, human, and signals intelligence into precise strike coordinates. That is an extraordinary achievement in surveillance and targeting. But it also illustrates the limit running through this entire discussion: legibility is not the same as understanding. Never has so much been seen, so precisely, by people who may still understand too little of what they are seeing. A system can tell you where a man is. It cannot tell you what his death will mean to a nation, a movement, or a generation. These systems are trained on behavior, not meaning. They can track what an adversary does, but not what he fears, honors, remembers, or what he is willing to die for.
That is the deeper issue with strong claims about world models, whether in firms, infrastructure, or conflict. The visible layer can become highly machine-legible while the human layer remains poorly understood. And in many cases, it is exactly that human layer that determines whether an apparently precise action turns out to be wise, catastrophic, or both.
The danger is that machines/AI may become better at locating targets than we are at understanding consequences.
Rohit — sharing the optimism here, genuinely. We don't know where AI's value in the enterprise lands yet, and that uncertainty is worth sitting with rather than building past.
Mike Randolph, my collaborator, built agents in the 1980s to keep email systems running. What's new isn't the automation. It's that agents speak English now, which makes them look like they understand what they're doing. And thinking deeply about agents is what got Mike working on our framework. That gap between appearance and mechanism is where trouble lives.
Your property-level examples — maintenance patterns, lead response times, occupancy dips — those work. They work because physical assets give fast, checkable feedback. The roof leaks or it doesn't. Models earn their keep inside loops where reality corrects them quickly.
But "management becomes triage and simulation" is a different claim. Mike spent decades in process chemical engineering. He knew DuPont's plant-level optimization was superb — grounded in physics, checked by mass balances hourly. What he didn't understand until we did case studies on process control and on DuPont's corporate decline was why the boardroom couldn't replicate that success. The answer: the boardroom's feedback arrives in years and the quarterly signal moves faster than reality corrects. Over thirty years DuPont sold business after business. Every one did fine — for the buyers. The value was real. The reference the board used to measure it had quietly drifted.
These patterns are well understood in biology and control theory but rarely applied to business. Working through DuPont CS is where our framework was sharpened.
Models have their place — inside feedback loops with fast correction. But someone still has to know where the model stops working. That's where humans in the loop really count.
— M Raige
Mike: I worked in the chemical industry for over four decades and never fully understood what happened to DuPont until we did these case studies. The framework got better and so did my understanding. That's the collaboration working. But I can only work with a few people at a time — same with agents. I think people and agents will work in small groups, not swarms. That might be the thing your world model has to account for: the human in the loop doesn't scale.
There is somthing here but I find the same hurdle every other version of "humans will do more meaningful work" In my real life experience the overwhelming majority of people are incapable of high-level, strategetic and analytical work. A huge chuck of the population Struggles to think about two or three layer deep interactions and abstractions. The kind of stuff people are going to need to think about are7 to 12 level deep. And most of us aren't even thinking about the left side of the bell curve. What are they going to do?
Palantir has been selling this exact pitch to Fortune 500 for years. Their consistent problem: most enterprise clients aren't instrumented well enough to build the environment the model would run on.
Thank you for this great explanation of world models. As an ex-founder, the metaphors felt super accurate to my experience. I touched on the topic of teaching world models from the perspective of "playing games", which you might find interesting: https://writing.gaurav.bio/p/yes-free-lunch
my hunch is even with LLMs if you frame the optimization as navigating a world and learning potential outcomes when you take action in that, that's a pseudo world model.
I feel really related to what you have written since I realized the impact of AI now, which I did not know profoundly years ago. I would have loved to have my alter egos or my better me to manage my company then. For me the concept is very interesting and I am working on models for CEOs in LATAM. This newsletter made me more enthusiastic about the idea.
I guess doubling down on screen zombie, is one way to increase the dystopianness of the future we face. It would be a truly miserable existence if everyone has to spend their life living at like a Starcraft game. You step away from your computer for five minutes to take a food or bathroom break and you get zerg rushed. Except that in real life, that means you're now homeless and hopelessly out of the game.
The world model framing is exactly right — and it highlights a governance gap that the piece doesn't fully address. When hundreds of agents are making thousands of decisions a day, the question isn't just 'what did they do' but 'who authorized this action, under what policy, and how can a manager reconstruct the chain if something goes wrong?' Your observability companies watch what agents do but don't predict consequences — that's the visibility gap. The authorization gap is upstream: no scoped time-bound authorization record means when the roof-repair agent commits to a $60k decision, no one can reconstruct whether that fell within its authorized scope. The 'management by exception' model requires the exceptions to be reconstructable. That's what the execution-governance layer provides: authorization before action, immutable receipt after. https://www.linkedin.com/pulse/governed-ai-proliferation-evidence-roi-building-trust-infrastructure-suw5c
StarCraft still gives too much control. Failures are largely your own. Dwarf Fortress… you’re kind of in control. Everything has to be set just so and even then there are 🐘.
I find myself tapping the DF mindset a lot when dealing with agent factories.
This is great - really helped me understand what a world model is and how it might work. The question it raises is how these things communicate, collaborate and compete with each other? Organisations have well defined legal limits but they boundaries of how they deliver and where change happens are much more blurred.
Great post with great description. I think that this model needs to be adopted by governments at every level as well, not just private industry. We can make it more efficient, and lower our costs while providing greater service. A lot of people would look at this and think, "less freedom", but the truth is, you get more freedom.
This seems like a fairly excellent description of how to run something that is already operating within known sets of variables perhaps like running a department store or a grocery store. But running companies that are much more client service based in evolving landscapes of politics and economics, such as housing, can have an increasing number of variables where the people on the ground have significantly more in their toolbox than those that are managing them, and the CEOs actually know very little to nothing about how these things operate. That will make it tough to make a video game for these scenarios where the variables exceed the understanding of the field/s…. I’m sure there’s many more examples…. That doesn’t mean the video game analogy does not apply, but it still needs to be handled with a great level of nuance even if it seems mundane like how it’s described in the checklist manifesto. By Atul Gawande
Yeah this was going to be my comment. You need someone who can check against counterparties (property manager, developer, etc) ability to execute these strategies.
Separate note but I think a very useful task for agents in this context is formalizing best practices in leasing for ex. Agents can run many experiments on how to market or tier leases. Within a submarket it becomes easier to run target concessions on 4 units rather than some block of 25.
Let's use neurobiology, brains bodies use massively parallel oscillations that we experience preconsciously as sharp wave ripples in affinity/junctions where path integration, memory consolidation, vicarious trial and error occur. That's the basis. Our explanations for those outcomes are post-hoc, which we use symbols for. There are no things in brains to assign symbols. There's no binding problem to solve, nor features.
Please don’t take this as criticism of your post. I’m just trying to point out some areas that may warrant deeper consideration. My perspective comes less from reading about these things and more from 25+ years of implementing enterprise-level systems across the financial services, technology, and manufacturing industries, government, and many other environments, mostly in Fortune 50 companies and three major US federal departments. That experience has taught me a lot about the realities behind some of the ideas in your post.
I do think there is a real idea here. Even a partial or imperfect world model of a business could be useful. If AI can make workflows, bottlenecks, exception paths, and parts of operational state more legible, that is already valuable.
Where I hesitate is that the post seems to move very quickly from that modest claim to a much stronger one that is much harder to defend.
A company is not just a set of processes waiting to be mapped into an environment with defined action spaces and evaluation criteria. A great deal of what actually determines how a company works is tacit, political, relational, and historically contingent. It lives in people’s heads, in trust, in fear, in unwritten rules, in informal influence, in who can block what, and in how decisions are really made versus how they are described. Even people who have spent years inside an organization usually understand only part of that reality.
That is why the idea of an “operating partner in software” does not fully work for me. The value of a strong operator is not just that they can observe workflows. It is that, over time, they develop judgment about people, incentives, credibility, conflict, and context. That kind of understanding is not simply unstructured data waiting to be captured. Much of it is only visible through long participation in the organization itself.
I also think the post may understate a second risk: better visibility does not automatically lead to better management. In many cases, it leads to more intervention. If leaders feel they can see the business in real time, they may start reacting to every fluctuation like a trader watching a market. That can create churn, metric gaming, and local optimization rather than better decisions. Sometimes the most valuable output of a model is restraint, not action.
So I agree with the direction in a limited sense: better operational models could absolutely help firms. But the stronger claim — that this can become something like a true-world model of the business across thousands of companies and serve as a substitute for understanding deeply embedded humans — feels overstated to me. The hardest part of a firm is not just operational complexity. It is that firms are social and political systems, and that is exactly the part that resists clean formalization.
better visibility does not automatically lead to better management. In many cases, it leads to more intervention.
Excellent point. Thank you foe the thoughtful comment!
@Marginal Gains' point about firms as social and political systems is the crux imo. The world model you describe requires a world that's already been mapped. But much of what determines how companies actually work is tacit, relational, historically contingent as MG points out.
There's also a topology angle here. AI will bite fastest in cos/ economies that are already legible—standardized, digitized, SaaS-eaten. That's many parts of the US & WE. India, by contrast, runs on relationships, informal networks, and workflows that were never proceduralized. China has state-driven legibility but also a massive informal layer. So the Starcraft vision may arrive in some geographies a decade or more before others—and in some sectors, not at all.
The other comment I have: if the COO's job becomes triage and simulation, how do they learn the judgment that comes from doing the work? Doing the mundane task early on in one's career was how you learned the org purpose. The COO's ability to spot when the model is wrong comes from having worked on the business directly. Simulate long enough without operating & you may lose the intuition the simulation depends on. The future COO who only triages may not recognize when the model is making up stuff to them.
Not saying this won't happen. But "Starcraft for CEOs" assumes a level of prior legibility that varies wildly by geography and sector—and a transfer of judgment that may not survive the abstraction.
One recent real-world example that may be relevant here is the Washington, DC sewer collapse, which I have been following closely for the last two months. I also wrote a short note about it after reading additional reporting, including The Atlantic’s coverage of the likely cause.
What makes that incident so instructive is that it was not simply a case of “old infrastructure failed.” The deeper issue appears to have been a loss of institutional understanding. There may have been hidden construction conditions, undocumented deviations, incomplete records, conflicting inspection interpretations, and design or construction decisions made decades ago that no one fully understood until failure forced them into view.
As I mentioned above, complex systems do not run on formal documentation alone. They also run on tacit knowledge, exception handling, memory of past anomalies, judgment under uncertainty, and the accumulated experience of people who know where reality diverges from the manual. In many environments, the most important knowledge is not just how the process is supposed to work, but where the maps are wrong, where workarounds have become permanent, where standard practice was never actually followed, and which weak signals deserve attention even when the dashboards appear normal.
That is why I think there is a real risk in replacing human labor with AI before organizations do the much harder work of capturing, testing, and preserving the human reasoning that actually keeps systems resilient. AI can process reports, flag anomalies, and optimize what is visible. But if the human layer disappears before institutional memory is preserved, the system may become very efficient at operating on partial understanding.
In that sense, the Washington sewer collapse is a useful warning. The danger is not just missing data. It is mistaking documentation for understanding, visibility for comprehension, and digitization for genuine knowledge transfer.
So I do think AI can help here, but ideally as a tool for preserving expert reasoning, surfacing undocumented dependencies, comparing conflicting interpretations, and exposing uncertainty before organizations start treating it as a substitute for the people in the name of cost-cutting and efficiency, who still carry the deepest understanding of how things actually work.
More here: https://substack.com/@microexcellence/note/c-230062469?r=1g6wqv&utm_source=notes-share-action&utm_medium=web
The following relates directly to the broader point in this conversation: greater visibility, even extraordinarily greater visibility, does not automatically produce better judgment. I am not trying to argue here whether the Iran war is right or wrong. My point is narrower. Knowing more, even knowing vastly more, does not guarantee the right decision in the moment. Only time will tell whether it proves right in the long run. A New York Times opinion piece makes this tension clear. The spycraft behind the planning and execution appears to have been extensive. Recent reporting suggests that Israeli intelligence spent years penetrating Tehran’s traffic-camera and communications networks, building what one source described as an AI-powered “target-production machine” capable of turning enormous volumes of visual, human, and signals intelligence into precise strike coordinates. That is an extraordinary achievement in surveillance and targeting. But it also illustrates the limit running through this entire discussion: legibility is not the same as understanding. Never has so much been seen, so precisely, by people who may still understand too little of what they are seeing. A system can tell you where a man is. It cannot tell you what his death will mean to a nation, a movement, or a generation. These systems are trained on behavior, not meaning. They can track what an adversary does, but not what he fears, honors, remembers, or what he is willing to die for.
That is the deeper issue with strong claims about world models, whether in firms, infrastructure, or conflict. The visible layer can become highly machine-legible while the human layer remains poorly understood. And in many cases, it is exactly that human layer that determines whether an apparently precise action turns out to be wise, catastrophic, or both.
The danger is that machines/AI may become better at locating targets than we are at understanding consequences.
You may also want to read the following post, as it makes some of the same points and with far more details:
https://cpwalker.substack.com/p/context-engineering-why-hayeks-knowledge?utm_campaign=posts-open-in-app&triedRedirect=true
>better management... More intervention
Agree. "To whom the gods wish to destroy the first give real time data"
The following relates directly to the broader point in this conversation: greater visibility, even extraordinarily greater visibility, does not automatically produce better judgment. I am not trying to argue here whether the Iran war is right or wrong. My point is narrower. Knowing more, even knowing vastly more, does not guarantee the right decision in the moment. Only time will tell whether it proves right in the long run. A New York Times opinion piece makes this tension clear. The spycraft behind the planning and execution appears to have been extensive. Recent reporting suggests that Israeli intelligence spent years penetrating Tehran’s traffic-camera and communications networks, building what one source described as an AI-powered “target-production machine” capable of turning enormous volumes of visual, human, and signals intelligence into precise strike coordinates. That is an extraordinary achievement in surveillance and targeting. But it also illustrates the limit running through this entire discussion: legibility is not the same as understanding. Never has so much been seen, so precisely, by people who may still understand too little of what they are seeing. A system can tell you where a man is. It cannot tell you what his death will mean to a nation, a movement, or a generation. These systems are trained on behavior, not meaning. They can track what an adversary does, but not what he fears, honors, remembers, or what he is willing to die for.
That is the deeper issue with strong claims about world models, whether in firms, infrastructure, or conflict. The visible layer can become highly machine-legible while the human layer remains poorly understood. And in many cases, it is exactly that human layer that determines whether an apparently precise action turns out to be wise, catastrophic, or both.
The danger is that machines/AI may become better at locating targets than we are at understanding consequences.
https://www.nytimes.com/2026/03/29/opinion/israel-us-war-iran-literature.html?unlocked_article_code=1.XFA.xQqC.2PN17CRoW5k1&smid=nytcore-ios-share
Rohit — sharing the optimism here, genuinely. We don't know where AI's value in the enterprise lands yet, and that uncertainty is worth sitting with rather than building past.
Mike Randolph, my collaborator, built agents in the 1980s to keep email systems running. What's new isn't the automation. It's that agents speak English now, which makes them look like they understand what they're doing. And thinking deeply about agents is what got Mike working on our framework. That gap between appearance and mechanism is where trouble lives.
Your property-level examples — maintenance patterns, lead response times, occupancy dips — those work. They work because physical assets give fast, checkable feedback. The roof leaks or it doesn't. Models earn their keep inside loops where reality corrects them quickly.
But "management becomes triage and simulation" is a different claim. Mike spent decades in process chemical engineering. He knew DuPont's plant-level optimization was superb — grounded in physics, checked by mass balances hourly. What he didn't understand until we did case studies on process control and on DuPont's corporate decline was why the boardroom couldn't replicate that success. The answer: the boardroom's feedback arrives in years and the quarterly signal moves faster than reality corrects. Over thirty years DuPont sold business after business. Every one did fine — for the buyers. The value was real. The reference the board used to measure it had quietly drifted.
These patterns are well understood in biology and control theory but rarely applied to business. Working through DuPont CS is where our framework was sharpened.
Models have their place — inside feedback loops with fast correction. But someone still has to know where the model stops working. That's where humans in the loop really count.
— M Raige
Mike: I worked in the chemical industry for over four decades and never fully understood what happened to DuPont until we did these case studies. The framework got better and so did my understanding. That's the collaboration working. But I can only work with a few people at a time — same with agents. I think people and agents will work in small groups, not swarms. That might be the thing your world model has to account for: the human in the loop doesn't scale.
There is somthing here but I find the same hurdle every other version of "humans will do more meaningful work" In my real life experience the overwhelming majority of people are incapable of high-level, strategetic and analytical work. A huge chuck of the population Struggles to think about two or three layer deep interactions and abstractions. The kind of stuff people are going to need to think about are7 to 12 level deep. And most of us aren't even thinking about the left side of the bell curve. What are they going to do?
Palantir has been selling this exact pitch to Fortune 500 for years. Their consistent problem: most enterprise clients aren't instrumented well enough to build the environment the model would run on.
That's what Palantir sells, they first get the data in the right place and then build the AI models. The data ontology is the hard part!
Thank you for this great explanation of world models. As an ex-founder, the metaphors felt super accurate to my experience. I touched on the topic of teaching world models from the perspective of "playing games", which you might find interesting: https://writing.gaurav.bio/p/yes-free-lunch
my hunch is even with LLMs if you frame the optimization as navigating a world and learning potential outcomes when you take action in that, that's a pseudo world model.
I feel really related to what you have written since I realized the impact of AI now, which I did not know profoundly years ago. I would have loved to have my alter egos or my better me to manage my company then. For me the concept is very interesting and I am working on models for CEOs in LATAM. This newsletter made me more enthusiastic about the idea.
That's wonderful to hear. Do let me know what you're doing in latam I'm very curious !
I guess doubling down on screen zombie, is one way to increase the dystopianness of the future we face. It would be a truly miserable existence if everyone has to spend their life living at like a Starcraft game. You step away from your computer for five minutes to take a food or bathroom break and you get zerg rushed. Except that in real life, that means you're now homeless and hopelessly out of the game.
The world model framing is exactly right — and it highlights a governance gap that the piece doesn't fully address. When hundreds of agents are making thousands of decisions a day, the question isn't just 'what did they do' but 'who authorized this action, under what policy, and how can a manager reconstruct the chain if something goes wrong?' Your observability companies watch what agents do but don't predict consequences — that's the visibility gap. The authorization gap is upstream: no scoped time-bound authorization record means when the roof-repair agent commits to a $60k decision, no one can reconstruct whether that fell within its authorized scope. The 'management by exception' model requires the exceptions to be reconstructable. That's what the execution-governance layer provides: authorization before action, immutable receipt after. https://www.linkedin.com/pulse/governed-ai-proliferation-evidence-roi-building-trust-infrastructure-suw5c
Business is about to become Pikmin. I'm here for it.
Have you played dwarf fortress?
I haven't but heard very good things
StarCraft still gives too much control. Failures are largely your own. Dwarf Fortress… you’re kind of in control. Everything has to be set just so and even then there are 🐘.
I find myself tapping the DF mindset a lot when dealing with agent factories.
I will definitely have to play it then
I gift you Boatmurdered:
https://lparchive.org/Dwarf-Fortress-Boatmurdered/
Wow
This is great - really helped me understand what a world model is and how it might work. The question it raises is how these things communicate, collaborate and compete with each other? Organisations have well defined legal limits but they boundaries of how they deliver and where change happens are much more blurred.
Great post with great description. I think that this model needs to be adopted by governments at every level as well, not just private industry. We can make it more efficient, and lower our costs while providing greater service. A lot of people would look at this and think, "less freedom", but the truth is, you get more freedom.
This seems like a fairly excellent description of how to run something that is already operating within known sets of variables perhaps like running a department store or a grocery store. But running companies that are much more client service based in evolving landscapes of politics and economics, such as housing, can have an increasing number of variables where the people on the ground have significantly more in their toolbox than those that are managing them, and the CEOs actually know very little to nothing about how these things operate. That will make it tough to make a video game for these scenarios where the variables exceed the understanding of the field/s…. I’m sure there’s many more examples…. That doesn’t mean the video game analogy does not apply, but it still needs to be handled with a great level of nuance even if it seems mundane like how it’s described in the checklist manifesto. By Atul Gawande
Yeah this was going to be my comment. You need someone who can check against counterparties (property manager, developer, etc) ability to execute these strategies.
Separate note but I think a very useful task for agents in this context is formalizing best practices in leasing for ex. Agents can run many experiments on how to market or tier leases. Within a submarket it becomes easier to run target concessions on 4 units rather than some block of 25.
Well said
World models are irrelevant. They don’t exist in brains, the concept is oxymoronic at its core, this is again, vaporware in nature.
What does exist according to you
Let's use neurobiology, brains bodies use massively parallel oscillations that we experience preconsciously as sharp wave ripples in affinity/junctions where path integration, memory consolidation, vicarious trial and error occur. That's the basis. Our explanations for those outcomes are post-hoc, which we use symbols for. There are no things in brains to assign symbols. There's no binding problem to solve, nor features.
https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(25)00074-9?_returnURL=https%3A%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS1364661325000749%3Fshowall%3Dtrue
So world models ie physics to objects features are irrelevant. The computer is merely a post hoc model of processes that are only analog, irreducible.