How ant colony physics fixes enterprise AI
Transcript
Made automatically from the audio, so it may contain mistakes. Click a line to play from there.
Trying to fix enterprise artificial intelligence right now is well it's honestly like trying to understand how a complex high performance combustion engine works by you know only looking at the glossy marketing brochure.
Right, like the brochure they hand you at the dealership.
Exactly. I mean you can memorize that brochure perfectly. You can recite the horsepower, the fuel efficiency, the sleek aerodynamic design. But if that engine actually breaks down on a remote highway in the pouring rain and you pop the hood, you have absolutely no idea what to do with a wrench.
Because you don't actually know how the alternator connects to the battery or how the fuel injection system regulates the mix.
Yeah. So today we are tossing that brochure in the trash. Welcome to this deep dive. We are opening the hood, rolling up our sleeves, and looking directly at the actual mathematical physics of the engine.
Which is frankly long overdue in this industry.
It really is. Our mission today is to figure out exactly why the enterprise AI industry is currently failing at an astronomical rate. And we're going to explore a radically proposed solution. A solution that involves mapping the underlying physics of human corporations, biological immune systems, and get this ant colonies.
Yeah, it's wild. We are dissecting a foundational text today by Marco Van Hearn from Eigenvector Research. It's called Coordination Foundation Models, from enterprise science to the mathematical theory of coordination.
And it is a dense read, but man, it is fascinating.
It is a genuinely paradigm-shifting piece of scientific literature. Because what we are exploring today isn't just, you know, a new software update or some clever management framework.
Right, it's bigger than that.
Much bigger. It is a radical scientific claim that fundamentally challenges roughly 30 years of established business, organizational, and computational theory. We are going to investigate this single unifying premise proposed by Van Hearn.
Which is the idea that human corporations, ant colonies foraging for food, biological immune systems, fighting off viruses, and like digital blockchain networks processing transactions, they all share the exact same underlying mathematical structure.
Which sounds completely like science fiction at first glance.
Oh, totally. When I first read that, I thought, okay, this is a stretch. But the text backs this up with pure, rigorous mathematics.
It absolutely does.
And our goal is to unravel how proving this mathematical link could actually solve the massive systemic failure rate we're seeing right now when companies trying to deploy AI.
Exactly.
So whether you are a CEO trying to understand why your expensive new AI agent keeps, you know, hallucinating company policies, or you're a software engineer tasked with building these incredibly complex systems, or honestly, just someone who's endlessly curious about how the hidden structures of the world operate. I promise you this.
It truly alters your entire lens on a reality. But to understand the cure Van Hearn is proposing, we first have to deeply understand the pathology of the disease.
Right, the failures.
Exactly. Why are AI systems failing so spectacularly in the real world? Because if you look at the highly controlled sanitized product demos that tech companies release, I mean these AI agents look like absolute magic.
They do. They effortlessly read complex email threads, they update support tickets, they generate these beautiful reports. They seem to understand everything about the business.
But then a company actually buys the software. They deploy it into their own real messy, chaotic enterprise environment, and the agents just simply fall apart. They fall apart completely. The source text points to a very specific and frankly alarming statistic to anchor this problem. It cites a 2023 Gartner survey, which found that fewer than 20% of enterprise AI projects actually succeed in delivering their projected value once they hit production.
Just 20%. That is I mean, that's wild.
It's abysmal. That means that 80% of the time, this highly advanced, incredibly expensive, supposedly revolutionary technology is just failing to do the job. And the author refers to this massive chasm between expectation and reality as the deployment gap.
The deployment gap. Okay, but why? I mean, why is the gap so massive?
Well, we have to be clear about the nature of this gap. It isn't happening because the AI models themselves lack like raw computational intelligence. It isn't just a matter of the training data being a little dirty or the prompt engineering being slightly off.
So it's not a user error thing.
No, not at all. The failure is entirely structural. The AI lacks a fundamental understanding of the environment it has been dropped into. So to explain this, the book breaks this epidemic down into five distinct categories of enterprise AI failure.
Okay, let's go through these. I I think the best way to really understand these failures is to walk through a hypothetical scenario.
I agree. That usually helps ground it.
Let's imagine a listener's company. We'll call it Global Supply Co. They manufacture and ship specialized industrial equipment. They just spent, you know, millions deploying a brand new AI agent to automate their customer service and supply chain logistics.
A very common scenario right now.
Right. So let's look at the first category of failure through the lens of Global Supply co. The text calls this a process boundary failure. And it notes this accounts for about 35% of all deployment failures.
Yeah, the biggest chunk. So in your scenario, a process boundary failure happens when the AI agent knows exactly what logical action it needs to take, but it fundamentally cannot navigate the physical or organizational handoff between two different systems.
Okay, give me an example.
Imagine a customer emails Global Supply Co. because a piece of equipment arrived broken. The AI agent perfectly reads the email in the customer relationship management system, the CRM. It analyzes the photos, verifies the warranty, and correctly decides I need to trigger a full refund and ship a replacement.
Which is exactly what you want it to do.
Exactly. The refund process, however, lives entirely in a different piece of software, the Enterprise Resource Planning System, or ERP.
Okay, so the AI just needs to like reach over and push the button in the ERP.
But between the CRM and the ERP, there is a boundary. Now this could be a hard technical boundary. Perhaps the ERP API requires a specific rotating authentication token that the AI agent's service account just hasn't been granted access to. Or even more commonly, it is an organizational boundary. There might be a hard-coded company rule stating that any refund over $5,000 requires a human floor manager to digitally sign off.
And the AI doesn't know that.
Right. The AI agent doesn't possess a structural model of that organizational boundary. It just tries to execute the refund, hits the invisible wall, throws a generic error code, and fails completely.
Man, it's like it's like a high-speed train running perfectly smoothly on its tracks, but suddenly it hits a national border where the tracks change from standard gauge to a wider gauge.
That's a great way to picture it.
The train itself is in perfect working order. The destination is clear, but because it can't adapt to the structural boundary of the new tracks, it just derails entirely.
Exactly. And that brings us to the second category of failure, which is governance constraint violations.
Okay, so if the first failure is a train derailing, this one sounds like the train quietly switching tracks and driving straight into a restricted military base.
That is an apt comparison because governance constraint violations are arguably the most dangerous type of failure. I mean, they are insidious. They are often completely invisible until the damage is already done.
So what triggers this?
This occurs when the AI agent successfully takes an action that is technically allowed by the software architecture, but that action violently violates a company policy, a legal framework, or a regulatory compliance standard.
Oh wow. Okay, going back to our global supply coast scenario, what does that look like in practice?
Let's say the AI is processing that same broken equipment return. To expedite the shipping of the replacement part, the AI analyzes the logistics network and finds a faster third-party courier service.
Seems smart.
Very smart. And the software API allows the AI to automatically forward the customer's name, home address, phone number, and purchase history to this new courier to set up the delivery. Technically, the AI executed a brilliant optimization.
Wait, but the AI just handed over personally identifiable information to an unvetted third party.
Precisely. It completely bypassed the company's strict GDPR compliance policies and vendor data privacy agreements. The AI was entirely unaware that those governance constraints even existed because they aren't written into the API code. They exist in the legal department's PDF manual.
Oh, that is a nightmare.
It is. The software system registers a successful transaction, the customer gets their part, everyone is happy. And you don't find out there was a massive data breach until a regulatory auditor knocks on your company's door six months later with a multi-million dollar fine.
Because the AI doesn't know the law. If a button isn't physically locked, the AI just assumes it is allowed to press it.
Right. It lacks that governance context.
Which flows perfectly into the third category, which Van Hearn calls context collapse.
Yeah.
This is when the AI does the exact right thing in a vacuum, but the absolute wrong thing given the specific human context of the moment.
The AI basically operates in a state of perpetual tunnel vision. Let's stick with global supply co. The AI agent monitors inventory levels, right? It notices that the warehouse is running low on a specific type of specialized microchip used in their machinery.
Okay.
In isolation, the algorithm is simple and flawless. If inventory drops below threshold X, generate a purchase order for 10,000 new units from the supplier. The purchase order is formatted perfectly, the math is correct, the vendor is in the system, so the AI executes the multi-million dollar order.
But what the AI doesn't know is the context.
Exactly. What the AI missed is that just yesterday, the chief financial officer held an emergency all-hands meeting to announce that the company missed its quarterly earnings target and instituted an immediate total spending freeze across all departments. No new inventory purchases until further notice.
Right. A human employee would instantly know this. I mean, the human employee absorbs the context just by being in the environment. They see the stressed emails, they hear the chatter in the break room, they actually attended the town hall meeting.
They have situational awareness.
The human would look at the low inventory alert and say, normally I'd reorder this, but given the freeze, I better escalate this to the DP first. The AI completely lacks that situational awareness. It is just deaf to the environment.
It acts purely on localized data. It has no broader context. Then we arrive at category four: temporal reasoning failures.
Temporal meaning time, obviously.
Right, referring to the flow and sequence of time.
Yeah.
Organizational processes are not just, you know, lists of tasks. They are highly choreographed sequences. They happen in a specific order for very important legal, physical, or logical reasons. So an AI that lacks a robust temporal model might attempt to take an action too early, too late, or entirely out of sequence.
So at Global Supply Co., an AI might try to automatically generate and send a finalized invoice to a client for that replacement machinery before the machinery has actually been manufactured, let alone shipped or signed for by the client at the loading dock.
That is a perfect example. The invoice generation module might have been triggered by a simple database update indicating the order was confirmed. But the AI doesn't understand the chronological dependencies of physical reality.
It doesn't know how the real world works.
Exactly. It doesn't understand the physical law that event A manufacturing the widget and event B delivering the widget must be completed and verified by a human signature before event C demanding payment for the widget can legally occur. It just fires off the invoice because the module was active. It possesses no sense of time or sequence.
It's like trying to bake a cake by putting the raw flour and eggs into the oven and then trying to stir them while they bake. The ingredients are right, but the temporal sequence is just a disaster. Okay, finally we hit category five, which I found the most fascinating: institutional memory failures. This is the failure that feels the most distinctly human to me. It happens when the AI crashes because it lacks the tacit, unwritten knowledge that human employees just naturally absorb over time.
Oh, anyone who has ever worked in a complex organization knows exactly what this is. It is the vast ocean of knowledge that is intentionally or accidentally left out of the standard operating procedures. Right. It's the knowledge that Bob in the accounting department prefers you to ping him informally on Slack to give him a heads up before you submit a specific complicated tax form, otherwise he gets annoyed and pushes it to the bottom of his queue.
Yes. Or it's the deeply ingrained knowledge that a specific 30-year-old legacy database will absolutely crash the entire network if you try to run a heavy query on it between 2.0 p.m. and 3.0 p.m. when the daily backup is running.
Right. There is no official manual titled How to Deal with Bob's Moods or Why the Server Gets Cranky at 2 p.m.
But humans navigate those hidden landmines intuitively.
They do. But when an AI encounters a situation that requires this undocumented institutional memory to successfully navigate a bottleneck, it has no reference point whatsoever. It simply follows the official documentation, triggers the legacy system at 2.15 pm, and brings the entire company network down.
Wow. Okay, so we have these five massive structural points of failure process boundaries, governance constraints, context collapse, temporal reasoning, and institutional memory.
Correct.
And here is where Van Hearn makes his most aggressive and controversial diagnosis in the text. He looks at this entire landscape of failure and blames it all on one singular architectural flaw in how we build AI today. He calls it the language model trap.
This is the absolute crux of the current technological paradigm. In the mid 2020s, the dominant, almost exclusive approach to building enterprise AI is to take a massive large language model, an LLM like GPT-4 or its successors, and just fine-tune it on a company's internal corpus of documents.
Yeah, that's what everyone is doing.
The prevailing logic in the tech industry is if we just feed this massive neural network every single PDF, every company manual, every policy document, and ten years of past email archives, the model will inherently understand how our company works. Right. But Van Hearn systematically argues that this is a profound, dangerous illusion.
Okay, I really want to push back on this a bit because it feels incredibly counterintuitive to how most people understand modern AI.
Yeah.
If I literally feed an LLM every single standard operating procedure, every rule, every memo that Global Supply Co. has ever produced, how can it possibly not know how the company works? I mean, all the information is right there.
Because we are fundamentally confusing the ability to mimic language with the possession of a structural world model.
Okay, it sounds like an actor who spent a year memorizing complex medical textbooks and surgical jargon to play a brilliant neurosurgeon on a television show. That actor can stand in front of a camera and recite the exact step-by-step procedure for a highly complicated triple bypass surgery. They sound incredibly convincing. They use all the right acronyms. But if you were to suddenly drop that actor into a real operating room, put a real scalpel in their hands, and ask them to perform the surgery on a real patient, it would be an absolute fatal disaster.
It'd be horrific, yeah.
They possess the vocabulary, but they do not possess the actual physical skill or the underlying structural understanding of human anatomy.
That is a phenomenal analogy. And from a mathematical standpoint, it perfectly describes exactly what is happening inside the neural network of an LLM. We have to look under the hood at what an LLM actually computes. It is an engine meticulously designed to model language. Right. Words. Specifically, it models the statistical probability of token sequences. It calculates, based on trillions of examples in its training data, which word is most statistically likely to follow the previous word.
It's an autocomplete.
A very advanced autocomplete. If you prompt it and ask it to describe your company's procurement process, it generates a sequence of words that is statistically highly probable to sound exactly like a standard corporate procurement process.
So it's essentially generating a highly educated, statistically plausible sentence. Something like the chief financial officer must approve the purchase order before processing.
Yes. But here is the critical mathematical distinction Van Hearn is drawing. Generating that plausible sentence does absolutely not mean the AI actually possesses what mathematicians call a topological world model, where the CFO and the purchase order are mathematically bound, interactive entities.
Hold on, let me stop you right there. You just used the phrase topological world model. What exactly does copological mean in this context? So we're talking about like maps and mountains.
Not physical geography, no, but structural relationships. Topology is a branch of mathematics concerned with the properties of a space that are preserved under continuous deformation, meaning stretching, twisting, and bending, but not tearing.
Okay, I'm trying to picture this.
Think of a rubber sheet with dots drawn on it connected by lines. You can stretch that rubber sheet in any direction, changing the physical distances and locations of the dots, but the fundamental connections which dot is connected to which other dot remain completely unchanged. That underlying unbreakable network of connections is the topology.
Okay. So applied to a company.
Applied to a company, a topological world model means the AI understands the deep underlying network of relationships and rules. In a true topological model, the concept of CFO isn't just a statistical cluster of letters. It is an active mathematical node that possesses specific authorities and capabilities.
The node actually does something.
Exactly. And the purchase order isn't just a phrase, it is a structural object that is constrained by active computational policies. The LLM does not have this rubber sheet of connections. It is just stringing together words that often appear near each other. This lack of topological structure is the absolute root cause of why AI so frequently hallucinates organizational facts.
So when our hypothetical AI at Global Supply Co. confidently hallucinated a step in the supply chain process that didn't exist, it wasn't a temporary glitch or a bug in the code. It was just predicting the next most likely word based on average supply chain documents it read on the internet.
Yes, exactly.
It sounded incredibly plausible to a human reader, but it was factually structurally false in reality.
It is not a bug. It is a core feature of the LLM's architecture. To return to your television doctor analogy, the language model can write an Emmy award winning script about performing surgery, but it does not possess a three-dimensional topological model of the human circulatory system. Organizations, Van Hearn argues, are not simply sequences of words. They are vastly complex, multidimensional networks of commitments, actors, and policies. You cannot model a network of commitments using a tool designed to model sequences of tokens.
Wow. Okay. Now, to be fair to the industry, the text does acknowledge that there has been a major attempt to solve this specific problem. It brings up a field called process mining, which has been pioneered over the last decade by massive tech companies like Salonis and Signavio. But Van Hearn argues that even process mining ultimately falls drastically short.
Process mining was undeniably a massive step forward in enterprise technology.
Right. People don't follow the rules.
So instead of looking at documents, process mining software hooks directly into a company's actual event logs.
The digital footprints.
Exactly. The raw digital footprints left behind in the ERP and CRM systems every time a user clicks a button or updates a field. It takes millions of these logs and mathematically maps out exactly what actually happened in the company. It can show you with absolute certainty that despite what the employee handbook says, in reality, 50% of all invoices in the London office completely bypass the mandatory approval step.
It gives you the absolute undeniable ground truth of the company.
It gives you the empirical ground truth. It tells you exactly what happened. But as Van Hearn critically points out, process mining fundamentally lacks the theoretical framework to explain why it happened. It lacks what he terms coordination theoretic content.
Coordination theoretic content. What does that mean?
Think of it this way process mining is like watching a high definition slow motion security video of a horrific car crash at an intersection.
Okay, you can see exactly what happened. The red car ran the light and hit the blue car.
Precisely. The video shows you the exact sequence of events. But watching that video a thousand times will never teach you the underlying laws of physics. It won't teach you the mathematical equations for momentum, velocity, mass, or the coefficient of friction of the tires on the asphalt that actually govern the crash.
I get it.
Process mining shows you the organizational crash, but it doesn't give the AI the physics engine required to prevent the next one.
Okay, this is where Van Hearn's thesis zooms out in a way that is just staggering. He argues that the fundamental reason our organizational science is failing us, the reason we keep building AI that crashes, is because enterprise science itself is in what philosophers of science call a pre-paradigmatic state.
Yes.
The systems we use, HR software, financial software, supply chain software, they all possess completely incompatible, fragmented versions of reality. If language tokens are the wrong fundamental object to model and empirical event logs are too shallow, what is the actual fundamental object we should be looking at?
This is the dramatic pivot in the text, moving from the narrow field of enterprise science to the universal field of coordination science. The radical unifying thesis of this book is that human organizations, our corporations, our banks, our governments, are not the fundamental phenomenon of study. They are merely a specific local shadow of a much deeper universal phenomenon. And that fundamental phenomenon is coordination.
Coordination, the chaotic vital process of multiple independent agents, all possessing limited information, desperately trying to work together to achieve an outcome in a shared environment.
Exactly.
And to prove that this phenomenon is universal, the text refuses to just look at businesses. It examines five completely different, seemingly unrelated domains of coordination to prove they are all constructed from the exact same mathematical material. Let's walk through these five domains because honestly, this is where the book really blew my mind.
It's an incredible section. Let's explore the scope of his argument. The first domain is the baseline we've been discussing: enterprise organizations. In this domain, the actors are human employees and automated software bots. Right. The shared environment is constructed of legal contracts, corporate bank accounts, and complex digital ERP systems. We understand this. It's messy, data is heavily siloed behind corporate firewalls, and it's full of human irrationality.
Very human.
So Van Hearn asks: where can we find a system that does the exact same thing? Coordinating thousands of agents to achieve a goal, but in a pure observable environment without the messiness of human language and office politics. He looks to the natural world, he looks to domain two, ant colonies.
We are looking to the sidewalk to fix artificial intelligence.
I mean, that is just It sounds crazy, but yes, because an ant colony is a masterpiece of coordination. The actors here are the individual ants, the foragers searching for food, the soldiers defending the nest, the queen. They do not possess language, they do not have a centralized SQL database tracking inventory.
They definitely don't have a CEO telling them what to do.
Exactly. Yet they successfully allocate resources, they wage complex territorial defense, and they build intricate architecture. How? Through chemical signals called pheromones. Right, the scent trails. Yes. A forger ant finds a discarded piece of sugar. On its way back to the nest, it secretes a specific chemical pheromone trail on the ground. That chemical trail is not just a scent. In the mathematics of coordination, it acts as a binding commitment.
A chemical commitment.
It is a biological obligation for other ants to follow that path and harvest the resource. It is a highly complex, incredibly robust coordination system functioning entirely without a central brain.
That's amazing. So we move from the digital boardroom to the biological anhill, and then the text dives even deeper into biology with domain three, the human immune system.
This domain is incredibly abstract, but mathematically precise. The actors in the system are the individual cells inside your body. The T cells hunting invaders, the B cells producing antibodies, the macrophages consuming debris. The signals they use to coordinate are molecular, utopixing complex proteins and cytokines. But what is truly fascinating about the immune system as a coordination model is its governance structure.
Right, because you have millions of killer cells swarming through your body armed and ready to destroy. How do they know to attack the influenza virus, but absolutely leave your kidney cells alone? You mentioned governance, a white blood cell has a policy.
Exactly. In biology, this is called immune tolerance. But mathematically, it is biological governance. The system has evolved strict molecular rules policies that dictate exactly what molecular signatures the killer cells are permitted to attack and which ones they must ignore. And if they ignore the rule. Blockchains represent a fascinating evolution. They are essentially pure coordination systems designed specifically for environments where the actors fundamentally do not trust each other.
The trustless environment.
Right. In a blockchain domain, the actors are the decentralized computer nodes and the individual cryptographic wallets. The commitments binding them together aren't chemical pheromones or molecular proteins, they are unbreakable cryptographic signatures. And the policies governing the entire system are the hard coded consensus rules of the network protocol itself. It is a system of absolute trustless coordination enforced entirely by cold mathematics.
And finally, domain five, which feels like the absolute polar opposite of a decentralized, trustless blockchain. The text examines military command structures.
The military represents the ultimate extreme implementation of a rigid hierarchy. In this domain, the actors are the individual soldiers, the platoons, the divisions.
And the commitments are just orders, right?
Exactly. The commitments binding them are absolute life or death orders from commanding officers. And the policies are strict rules of engagement, international laws of armed conflict, and deeply ingrained tactical doctrine. The military possesses perhaps the most explicit inflexible governance constraints of any human system, primarily because the cost of a coordination failure isn't just a lost package or a delayed invoice, it is catastrophic loss of life.
So let's look at the board here. We have Fortune 500 corporations, subterranean and colonies, microscopic immune systems, decentralized digital blockchains, and rigid military hierarchies. Five completely distinct, incredibly diverse worlds. This is where Van Hearn introduces the heavy mathematics, specifically the concept of an isomorphism.
The isomorphism theorem is the beating heart of his entire generalization. In advanced mathematics, an isomorphism is a structure preserving mapping between two seemingly different complex structures.
Structure preserving mapping.
Right. It means that if you peel away all the superficial domain specific substrate, if you strip away the fact that an ant uses chemical thermos and an accountant uses an email, or that a T cell uses a protein and a blockchain node uses cryptography, the underlying mathematical scaffolding, the exact shape of how things connect and influence each other is absolutely identical.
Hold on. I really need to pause here because my brain is struggling to accept this. Are you genuinely saying that mathematically speaking, a computer simulation mapping the chemical interactions of an anthill and a simulation mapping the flow of capital through a Fortune 500 company are the exact same mathematical object?
That is precisely what the theorem proves. They are pairwise isomorphic. Every single mechanism, every feedback loop, every resource allocation strategy in that anthill has a direct one-to-one mathematical equivalent in the corporate structure. They only differ in what they are physically made of and the time scale on which they operate.
That is just wow. Which leads directing to the flagship prediction of the book. And if this works, it sounds like literal magic. The text calls it the coordination transfer hypothesis.
If the isomorphism theorem holds true, it exposes a massive flaw in how we are currently training AI, but also provides the solution. Right now, we are starving our AI models of coordination data. Enterprise data is inherently proprietary, highly sensitive, legally protected, and frankly, quite scarce.
Yeah, you can't just hack into their servers.
Exactly. You cannot just legally scrape the internal emails and ERP logs of Apple, Google, and Amazon to train your model. The data is locked away. But if coordination is universal, you don't need corporate data.
Because you can get it from the other domains.
Yes. You can generate functionally infinite amounts of synthetic data by running massive ant colony simulations. You can download the entire publicly available history of the Ethereum blockchain, billions of perfectly recorded, cryptographically secured coordination transactions. You can utilize vast biological data sets mapping immune system cellular interactions.
So, if I am interpreting the coordination transfer hypothesis correctly, observing a computer simulation of how millions of white blood cells swarm and coordinate to neutralize a virus can literally teach an AI model how to more efficiently process an invoice and navigate a bottleneck in the accounting department of a mid-sized paper company.
Yes, absolutely. Because the AI is not learning biology. The AI is learning the universal mathematics of decentralized resource allocation, signal propagation, and policy compliance. Once the neural network deeply internalizes the fundamental shape of coordination from observing the immune system or the blockchain ledger, it can instantly transfer that structural knowledge and apply it directly to your accounting department's workflow.
That blows my mind. Okay, I am fully strapped in for this matrix-level vision of reality. But it begs a massive question. If an immune system and a corporate finance department are actually made of the exact same mathematical stuff, what exactly is that stuff? What are the universal Lego blocks they're both built from? The text outlines what it calls the five universal primitives of coordination.
This is where we distill reality down to its irreducible elements. Van Hearn argues that if coordination truly is universal, it must be constructed from foundational components that simply cannot be broken down any further. He identifies these five primitives as actor, object commitment, event, and policy.
Let's not just list these, let's really dig into them. I want to see how these Lego blocks snap together in our different domains. So if we strip an organization down to its absolute bare metal, what is the first piece we find? The actor.
An actor is fundamentally an agent operating within the system that possesses limited information. The actor is the entity actually doing the work. In our enterprise scenario, an actor could be a human employee, or it could be an automated software script pulling data from an API. In the natural domains, the actor is the individual ant, or the microscopic T cell. Regardless of the domain, every single actor is defined by three specific attributes.
Okay, what are they?
First, a role, what they are designated to do. Second, a capability set, what physical or digital actions they are actually able to perform. And third, an information state what they know about the world at any given microsecond.
Right, because the actor isn't omniscient. The forager ant only knows it smells of pheromone right in front of its antenna. The junior accountant only knows that a specific invoice just landed in their digital inbox. Yeah. They don't see the whole system, they just react to their local reality.
Precisely.
But these actors aren't just floating in a void, what are they actually interacting with? That brings us to the second primitive, the object.
Objects are the passive elements populating the shared environment that the actors are actively coordinating around, modifying, or acting upon. In the business world, objects are the tangible and digital artifacts. Purchase orders, PDF invoices, a pallet of microchips in a warehouse, a line of code.
So stuff.
Basically, yes. In an ant hill, an object is a physical food source, a twig used for building, or the chemical pheromone trail itself. In the immune system, the object is the invading virus or the foreign antigen. Objects are the stuff that possesses value or meaning to the actors.
But actors and objects alone don't make an organization. You could have 50 people standing in a room holding 50 invoices, and that's not a company. That's just a crowd.
Exactly.
The text introduces the third primitive, which it beautifully describes as the organizational chemical bond, the commitment.
The commitment is arguably the most critical primitive in the entire framework. A commitment is a binding obligation between actors. It is the structural glue. Van Hearn defines a commitment as a dynamic entity that possesses a strict state machine. A commitment progresses through life cycle states. It can be pending, it can be fulfilled, it can be violated, or it can be canceled.
Okay, I think I see.
Think about signing a massive vendor contract. The moment the ink dries, a commitment is created and its state is pending. It hangs in the air, creating a tension in the organization.
And when the vendor delivers the equipment, the commitment state transitions to fulfilled. If they miss the delivery date, it transitions to violated.
Yes.
The book explicitly uses this amazing imagery, stating that an organization without commitments is a gas, not a solid. I love that. If no one is actually obligated to do anything for anyone else, the organization has no structural integrity. It just dissipates.
The density and the topology of those commitments are exactly what give an organization its rigid shape. When an ant secretes a pheromone trail, it is literally laying down a chemical commitment. When a node signs a transaction and broadcasts it to the blockchain, it is generating a cryptographic commitment. Commitments are the invisible architecture of coordination.
So we have actors bound by commitments, interacting with objects. But how does anything actually happen? The world isn't frozen in time. That requires the fourth primitive, the event.
Events are the temporal signals that trigger state changes in reality. An event is the catalyst. It is a system notification popping up on a screen. It is the sudden release of a hormone into the bloodstream. It is the mathematical confirmation of a new block on a digital ledger.
It's the spark.
Exactly. Events are the specific occurrences that force commitments to move from one state to another. When the bank processes a transfer, that payment event is what changes the invoice commitment from a state of pending to fulfilled.
And finally, we arrive at the fifth primitive, the one that keeps the entire system from descending into absolute anarchy. Policy.
Policies are the overarching governance constraints. They sit above the actors and dictate what actions are actually permissible within the system. What is highly practical about Van Hearn's approach is how he categorizes the enforcement of these policies into four distinct mechanisms.
Okay, walk me through those.
First, there are technical controls. This is when the physical environment or the software architecture literally prevents the action. A blockchain protocol mathematically rejecting an invalid digital signature is a technical control.
The door is physically locked, you cannot enter.
Second, there are procedural controls. This requires a specific sequence of actions, like requiring a senior manager's digital signature before a large purchase order can be released. Third, detective controls. The system allows you to take the action, the door is unlocked, but a security camera is watching. The system logs your action and a financial audit will catch you and punish you later if you broke a rule.
And the fourth enforcement mechanism brings me to my biggest question about this whole theory. The text mentions social controls. Now I can clearly see how a signed, legally binding vendor contract maps mathematically to a commitment. I can easily see how a digital API trigger maps to an event. But what about the messy, irrational, deeply human stuff? What about office politics? What about tacit coordination?
You mean the unwritten rule?
Yeah, for example, the unwritten universal rule that you absolutely do not microwave fish in the corporate break room. Nobody ever signed a legal contract about the fish. There's no ERP software logging microwave events. How on earth do rigid mathematical primitives capture the nuances of human culture?
It is a brilliant question, and it is precisely the kind of edge case Van Hearn uses to stress test his theory. In this mathematical framework, human culture acts as a dense, invisible network of implicit policies and implicit commitments. That unwritten rule about not microwaving fish, it is a policy. But instead of being enforced via a technical software block, it is enforced entirely via social controls.
So if I heat up salmon for lunch.
People might be less likely to help you with a project later because you violated the cultural policy.
Okay, that actually makes perfect sense. The mathematical primitives do not care if the policy is written in Python code, drafted in a complex legal contract by a lawyer, or just deeply encoded into the passive aggressive social behavior of the human actors in the break room. Mathematically, it functions the exact same way. It constrains the set of permissible behaviors in the environment.
Exactly. To prove the absolute necessity of these five primitives, Van Hearn walks the reader through a minimality proof. It's a thought experiment. What happens if we try to build an organization by pulling just one of these five Lego blocks out of the foundation?
The structure collapses.
Entirely. If you remove the policy primitive, your AI system instantly loses any ability to reason about compliance, safety, or rules. If you remove the actor primitive, you completely lose the concept of accountability actions mysteriously happened in the system, but no entity is actually responsible for doing them. If you remove commitment, the AI cannot model obligations, deadlines, or dependencies. Every single one of the five primitives is mathematically indispensable.
He also runs them through what he calls adversarial universality testing. He actively tries to break his own model by throwing incredibly weird counterintuitive edge cases at it. For instance, he looks at negative coordination, which is when actors are actively expending energy trying to avoid each other rather than work together, and he proves the five primitives still map perfectly to that behavior.
It's remarkably resilient.
But I want to pivot here, because this raises a monumental engineering challenge. It is one thing for us to sit here and philosophically agree that yes, everything can be viewed as actors' events and commitments. It is another thing entirely to translate that abstract philosophy into a hard computable language that a silicon microchip can actually process. To actually build this coordination foundation model, we need a mathematical engine.
Which brings us to the most computationally heavy part of the book. The search for the fundamental mathematical object of coordination science. How do we represent these primitives in code? The research team evaluated 13 vastly different mathematical candidates. They looked at Petri nets, which are excellent for modeling basic linear manufacturing process flows, but they completely fail when you try to introduce complex nested data types. They evaluated graph theory, game theory, process algebras, Bayesian networks.
And every single one of them failed to capture the complete picture of coordination.
They failed because the perfect mathematical engine needs to perfectly balance two seemingly opposing requirements. First, it requires immense expressiveness, the mathematical ability to represent massive, multi-level abstract structures. Imagine trying to map the entire global supply chain of an iPhone from the rare earth mines to the retail store in a single model.
That's a massive map.
And second, it requires compositionality, the ability to easily snap incredibly small, simple structures together to build larger, infinitely complex ones, exactly like snapping our Lego blocks together. Graph theory, for instance, is great for mapping a flat network, but if you try to nest an entire subgraph inside a single node of a larger graph, the math becomes incredibly brittle and breaks down.
And the ultimate winner, the mathematical framework that took the crown and possessed both expressiveness and compositionality was category theory.
Category theory is frequently described by mathematicians as the mathematics of mathematics. It operates at an incredibly high level of abstraction. Instead of focusing heavily on the objects themselves, category theory focuses intensely on the arrows, the morphisms, the relationships, and the structural transformations between the objects. It is uniquely and perfectly suited for mapping complex nested relationships without ever losing the essential shape of the network. Van Hearn utilizes category theory to construct the ultimate engine for his AI, which he names the CTEC.
CTEC, an acronym for the constrained typed evolving category. That is an intimidating mouthful of mathematics. Let's break down the four specific components of the CTEC engine, because this is literally the structural brain of the revolutionary AI we are discussing. The formula uses four symbols C, T, K, and Phi.
Let's unpack the equation. We start with the letter C. It's a C with a CDA. This represents the category itself. Mathematically, it encompasses the entire universe of all possible coordination states and every valid transition between those states. It is the architectural blueprint of the board.
The playing field itself. Then we add T.
T represents the type system. This is where our five universal primitives, actor, object, commitment, event, policy live inside the math. The type system ensures that every abstract dot on the playing field is typed correctly. It explicitly defines to the AI. This specific node is an actor. That specific connection is a commitment. This node is an event. It gives the pure abstract mathematics, concrete, real-world coordination meaning.
And then we inject K into the engine.
K represents the governance constraint set. K acts as a filter over the playing field. It dictates the boundaries of reality by defining which organizational states are actually valid and legally permissible according to the policies. If the AI hallucinates a state where an unapproved junior employee authorizes a billion-dollar transfer, the K constraint set instantly flags that topological state as mathematically impossible and filters it out.
And finally, the engine needs to move. It gets the Greek letter phi.
Phi is the evolution function. It acts as the system's clock. It mathematically maps exactly how the entire massive structure, all the actors, the pending commitments, the active constraints, moves, shifts, and changes state as time continuously flows forward. Phi is what gives the AI its temporal reasoning capability, completely solving the temporal failures we discussed earlier.
So we have designed this pristine, beautiful, incredibly powerful mathematical engine, the CTEC. But here is the massive real world problem, which the text calls the measurement bridge problem. Category theory is flawless, pristine mathematics. Real-world corporate data, like a messy, error-filled SAP event log from a chaotic HR department is definitely not flawless mathematics. How do you powerbly map a messy, flat, two-dimensional spreadsheet of raw data into this beautiful multidimensional category theory engine?
That is the core technical hurdle of the entire endeavor, and overcoming it is Van Hearn's greatest technical achievement in the book. The solution he outlines is a highly complex algorithm called the CTEC tokenization strategy. Think about a raw event log. It is just a flat, endless, two-dimensional list, timestamp, user ID, action taken. That's it.
A spreadsheet.
Yes. The tokenization process takes this flat 2D data and transforms it into a dynamic topological map. It analyzes the sequences and mathematically infers the hidden connections that aren't explicitly written in the log.
Wait, how does it infer a connection that isn't there?
It looks at patterns. The tokenization engine sees a specific user and actor completing a specific task and event. By analyzing the chronological sequence of that event in relation to historical patterns and known policies, the engine mathematically deduces the existence of the invisible commitment that must have bound the actor to the task and the invisible policy that must have constrained how they did it. It extrapolates the hidden dimensions.
I want to make sure I'm visualizing this correctly. Category theory and tokenization sound incredibly dense. But is it fair to say that if looking at a standard flat data log is like looking at a flat, two-dimensional photograph of a towering skyscraper? The CTEC tokenization process is like utilizing that photograph to computationally generate a full three-dimensional architectural blueprint. And not just the shape of the building, but a blueprint that actually highlights exactly which internal walls are structurally load-bearing.
That is a phenomenal, perfect analogy, yes. And to put your analogy further, the load-bearing walls in that 3D blueprint are the commitments. A standard AI looking at the 2D photo just sees a wall. It doesn't understand the wall's purpose. The CTEC engine, possessing the 3D topological blueprint, understands the deep structural integrity of the organization. It knows mathematically that if you casually remove that specific wall, if an employee fails to fulfill that specific commitment, the entire third floor of the organization will instantly collapse. The AI finally understands gravity.
And once you possess a mathematical blueprint of reality that includes gravity, you cross a profound threshold. You move from simply describing things that happened in the past to discovering the actual fundamental physics of how things operate. Which brings us to the absolute climax of Van Hearn's work and the culmination of this deep dive, enterprise physics.
This is where we transcend software engineering and enter the realm of pure science. We ask the ultimate question. Do human organizations have absolute unbreakable mathematical laws, exactly like the laws of physics.
Right. If I stand on Earth and draw off an apple, gravity dictates it falls. It's a law. Does a similar law exist for a company? If I pull lever X in a corporation, does outcome Y universally, undeniably happen every single time?
To rigorously answer that, Van Hearn first sets incredibly strict academic criteria for what constitutes a genuine scientific law. He is highly critical of the business world's tendency to label catchy management buzzwords as laws.
Yeah, like Murphy's law or something.
Exactly. For a principle to be a true law in coordination science, it must meet four criteria. First, universality. It must mathematically apply to every single coordination system, from the ant hill to the blockchain. Second, it must be falsifiable. We must be able to theoretically prove it wrong with empirical data. Third, it must possess deep explanatory power. It must explain why the phenomenon happens. And fourth, predictive power. It must accurately predict future states. He explicitly compares the required rigor to Isaac Newton's second law of motion, AFNA force equals mass times acceleration.
The book rigorously tests six candidate laws against these criteria. Let's dig into the mechanics of the most fascinating ones, starting with what he calls the law of governance compression.
The law of governance compression deals with the sheer weight of communication in a system. It mathematically proves that the communication overhead required to maintain coordination in an organization scales linearly if the organization possesses formal, codified policies. However, if the organization lacks formal policies, that exact same communication overhead scales exponentially as new actors are added.
Can you break down the mechanism of that? Why exponentially?
Think about a startup company with no formal rules. Every single decision requires consensus. If there are five employees, they all talk to each other. Mathematically, the number of communication channels in a fully connected network is n times n minus one divided by two. So five employees means ten channels.
Okay, manageable.
But if you grow to 50 employees with no rules, you now have 12,225 distinct communication channels. The chaos multiplies exponentially. Every single purchase, every vacation request requires endless meetings, slack threads, negotiations, and back and forth arguments because there is no policy dictating the answer. But if you implement a policy, and if you implement a rigid policy, for example, any purchase under $500 is automatically approved by the department head. You instantly sever hundreds of redundant communication channels. An employee just submits the receipt to one person. The overhead scales linearly with head count, saving massive amounts of organizational energy.
It's the difference between driving cars on a paved, multi-lane highway with painted lines and traffic lights versus forcing a thousand cars to hack their own unique paths through a dense jungle with machetes. The policy paves the road.
It is the direct mathematical organizational analog of Newton's an object at rest stays at rest. The law utilizes a concept from complex systems theory called an attractor basin.
I need to interrupt to clarify that term. What exactly is an attractor basin?
Imagine rolling a marble across a landscape filled with hills and deep valleys. Gravity pulls the marble down into the lowest point of a valley where it eventually comes to a complete rest. That valley is the attractor basin. In an organization, an attractor basin represents a highly stable, deeply ingrained, routine or structural state. The law of organizational inertia states that an organization settled into a stable attractor state will actively resist any change, and the resistance is mathematically proportional to how deep that basin is.
So if global Soplaco has processed invoices the exact same convoluted way for 20 years, they have carved an incredibly deep valley.
Exactly. The amount of force required to change that process, the sheer amount of human energy, financial capital, and leadership willpower needed to push that massive marble up and out of a 20-year deep valley is mathematically proportional to how deeply embedded those historical commitments are in the CTEC topological map. It proves why digital transformation efforts fail so often. Leaders severely underestimate the depth of their organization's attractor basins.
And then we arrive at the apex. The law that Van Hearn himself designates as the most profound and important mathematical result in the entire book, the coordination entropy law.
This is the organizational equivalent of the second law of thermodynamics. The second law of thermodynamics famously states that in any closed system, entropy, which is the measure of disorder, randomness, or chaos, will always inevitably increase over time. A hot cup of coffee will always eventually cool down and disperse its heat into the room. Things naturally fall apart. But governed organizations are not closed, random thermodynamic systems. The coordination entropy law mathematically proves the equation H of O given K is less than or equal to H of O.
Let's translate that equation into English.
It translates to this universal truth. Governance constraints universally and actively reduce coordination entropy. Implementing policies literally destroys organizational chaos.
The text compares this mechanism to Maxwell's Demon, which is one of my absolute favorite thought experiments in physics. Can you explain the demon and how it maps to a corporate policy?
James Clerk Maxwell proposed a thought experiment in 1867 to challenge the second law of thermodynamics. He imagined a microscopic demon standing guard at a tiny door between two chambers of gas. The demon observes the gas molecules flying around. Precisely. And Van Hearn proves that a well-designed governance system acts exactly like Maxwell's demon. The policy observes the chaotic, high-entropy interactions of actors in the organization. It identifies highly disordered, risky coordination states like our AI trying to send customer data to an unvetted vendor. And it simply eliminates those chaotic states from reality by making them mathematically against the rules. The policy acts as the demon's door, actively reducing the thermodynamic chaos of the entire organization.
But wait, I have to challenge the logical extreme of this law. If the mathematics explicitly prove that implementing governance always reduces entropy and eliminates chaos, shouldn't a company just create infinite governance? Shouldn't we strive to have a rigidly enforced policy for literally every single microinteraction in the building? Why not write a 10-page policy dictating exactly how many staples an employee is allowed to use per quarter?
That instinct, the desire to crush all chaos through infinite rules, is the exact fatal trap that bureaucratic managers and totalitarian systems inevitably fall into. But Van Hearn immediately introduces the critical counterbalance, the law of governance optimality. Yes, the entropy law proves that every single new rule you write reduces chaos. However, the optimality law proves that every single new rule simultaneously increases the system's governance overhead.
Meaning the actual cost of the rule?
Exactly. The cost in time, financial resources, human friction, and lost agility required to monitor and enforce that specific rule. You absolutely could write a strict policy rationing staples. You could build a software system to track staple usage, you could hire an auditor to review the staple logs.
Why would you?
Right, because the governance overhead, the cost of enforcing the rule, vastly exceeds the financial cost of just letting employees waste a few staples. The mathematical goal of an organization is not zero entropy. The goal is finding the exact razor's edge optimality point, achieving the maximum possible reduction in chaos for the minimum possible overhead cost.
It's the quest for the perfect mathematical balance. And before we close the book on enterprise physics, we absolutely must talk about the mathematician Emmy Noether. Noether's theorem is a legendary concept in theoretical physics that connects the concept of symmetry to conserved quantities. How on earth does a theorem about the fundamental fabric of the universe apply to an office building full of middle managers?
This might truly be the most beautiful, haunting insight in the entire text. In theoretical physics, Emmy Noether proved that if a physical system possesses a continuous symmetry, there must be a corresponding conserved quantity. For example, if the laws of physics don't change as time moves forward, which is time symmetry, then the total energy of the universe must be conserved. If the laws of physics are the same here as they are on Mars, which is translational symmetry, then momentum must be conserved. Van Hearn takes this mind-bending concept and proposes its equivalent in coordination science. Actor substitution symmetry.
Meaning the symmetry of replacing employee A with employee B.
Yes. If you remove an actor from the organization, let's say a senior logistics manager resigns, and you immediately replace them with a new actor who possesses the exact same role definition, the exact same capability set, and the exact same information state. Van Hearn proves that the deep topological structure of the organization remains perfectly mathematically conserved. The organization does not break. The overall structure doesn't even notice the human changed.
That is well, it's slightly terrifying from a human perspective, but mathematically profound. It implies that the true enduring structure of the company is completely independent of the specific human beings operating inside it. The roles, the policies, and the binding commitments are the actual solid reality of the system. The human beings are just the temporary fluid flowing through the pipes.
It is a humbling realization. The geometry of the coordination remains long after the individuals have moved on.
Okay, let's take a deep breath and summarize the immense staggering journey we have just taken together. We started by looking at a broken engine on the side of the road, analyzing exactly why enterprise AI is failing 80% of the time. We realized it is because these AI models are entirely trapped in the language model trap.
Right. They are generating statistical sequences of tokens blindly hallucinating because they lack a topological understanding of reality.
They lack the 3D blueprint. So to be that blueprint, we had to fundamentally redefine reality itself. We stepped back and discovered that Fortune 500 companies, microscopic immune systems battling viruses, and ant colonies horaging on the sidewalk are all just different physical expressions of the exact same underlying mathematical phenomenon: coordination.
And we took that complex coordination and broke it down into its five universal irreducible Lego blocks actors, objects, commitments, events, and policies.
We mapped out how those blocks nap together using the incredibly high-level mathematics of category theory to build the C Tech engine. And finally, we discovered that if you build that engine correctly, you find that human organizations aren't just random chaos. They are governed by absolute unbreakable laws of physics, like the coordination entropy law.
It is a breathtakingly complete end-to-end scientific framework. It provides the exact theoretical foundation needed to finally build the next generation of artificial intelligence, true coordination foundation models that actually understand the physical and structural laws of the world they are operating in.
So, to you, the listener, here is your takeaway for today. Tomorrow, when you log into your morning team meeting, or when you are stuck sitting at a red traffic light watch the cars, or when you were reading an article about the latest cryptocurrency trend, I want you to actively try to look past the surface layer. Try to spot the underlying mathematical matrix.
It's all there if you look.
Identify who the actors are, find the objects they are fighting over. Look for the invisible commitments binding them together, watch for the temporal events triggering their actions, and try to define the silent policies constraining everything they do. You are not just looking at a meeting or a traffic jam. You are watching the universal laws of coordination executing in real time.
I guarantee that once you learn to see the primitives in the world around you, you truly cannot unsee them.
You really can. It changes everything. And I want to leave you with one final deeply provocative thought to mull over on your own time, something that builds on everything we've unpacked today. We spent this entire deep dive discussing how a new coordination foundation model could learn from billions of years of biology and blockchain ledgers just to fix our current broken corporate processes. But what happens when we inevitably ask that AI to do something more? What happens when we ask it to design a completely new type of organization from scratch?
That is the ultimate unknown frontier of this science.
If an AI truly understands the absolute mathematical laws of coordination training on the evolutionary perfection of the immune system and the flawless logic of cryptography, could it invent a structure of human collaboration that is so mathematically optimal, so hyper-efficient, and yet so utterly alien to our current traditional corporate hierarchies that we haven't even conceived of it yet? And more importantly, if an AI built the perfect organization, would human beings even know how to work inside it? Something to think about. Until next time, we finally close the hood on the engine. We know exactly how it works. Now it's time to see where it can drive us.