Orchestrate all the Things - Mapping the Jagged Frontier of AI: Libraries, Ladders, Loops and Graphs. Feat. Carlos E. Perez, Intuition Machine Co-Founder

Episode Date: September 15, 2026

What does a three decade-long path of working with AI and asking what separates human cognition from a machine's look like? The answer says as much about the machines as it does about us. At eleven years old, Carlos Perez was already choosing between two possible futures. One was first contact with an alien civilization. The other one was building AI. He picked Inner space over outer space. The route there wasn't direct. After deep learning broke through around 2012, Perez eventually concluded that These are systems that you actually grow - more similar to biology than to physics. Perez co-founded The Intuition Machine in 2015, and he has published a number of books and essays on AI. He also works with Fannie Mae as a a Consultant, providing AI strategy and architecture guidance for the Enterprise Architecture organisation.  We explore his path of discovery and tinkering: from AI's uneven capabilities to what's actually changing between model generations, enterprise architecture, AI harnesses, loop engineering and graph engineering for AI agents. Article published on Orchestrate all the Things: https://linkeddataorchestration.com/2026/09/15/mapping-the-jagged-frontier-of-ai-libraries-ladders-loops-and-graphs/

Welcome to orchestrate all the things. I'm George Anadiotis and we'll be connecting the dots together. Stories about technology, data, AI and media and how they flow into each other, saving our lives. What has a three-decade-long path of working with AI and asking what separates human cognition for machines look like? The answer says as much about the machines as it does about us. At 11 years old, Carlos Perez was already choosing between two possible future. One was first contact with an alien civilization. The other one was building AI.

He picked inner space over outer space. The route there wasn't direct. After deep learning broke through around 2012, Ferres eventually concluded that these are systems that you actually grow, more similar to biology than to physics. Ferres co-founded the intuition machine in 2015 and has published a number of books and essays on AI. on AI. He also works with Fannie Mae as a consultant providing AI strategy and architecture guidance

for the enterprise architecture organization. We explore his path of discovery and tinkering, from AI's uneven capabilities to what's actually changing between model generations, enterprise architecture, AI harnesses, loop engineering, and graph engineering for AI agents. So if you start run GPT3 back in 2000, It was really good at writing out very coherent text sentences, right? But there was no causally behind it. It didn't really understand what it's actually going on. So over time, they introduced instruction following, so it could actually follow instructions

because prior to that it wasn't following instructions. You just gave it examples and it would complete an example, kind of like an analogy test, right? then you trained it on programming code. So what happened was it learned causality, so it's big, right? It learned cause and effect that it did not have prior to that and so forth. So the thought process behind this is a recognition that we already know that in software engineering because of AI, the concerns of software engineering migrates to something different from just actually writing code.

So in the same way, right, we've always been doing enterprise architecture, right? And the question is, when we now introduce AI technology, right, how does enterprise architecture actually change? So that's what I call native AI enterprise architecture, right? which is a kind of like a different thing or from enterprise architecture, but it has the same goals, right? And because a lot of times when you're doing enterprise architecture, you find yourself, aren't we just documenting things, right, after the fact, right?

I hope you will enjoy this. If you like my work and orchestrate all the things, you can subscribe to my podcast available on all major platforms. my self-published newsletter also syndicated on Substack, Hackern, Medium, and DZone, or follow, gesturate all the things on your social media outsource. When I was 11 years old, right, I was thinking, what is the most important thing that will change humanity? What kind of an event would change humanity?

And I came up with two options, two separate paths. One path was we would have a first encounter where to, aliens, number one. And at that time, I was very interested in astronomy, right, in space. And the second one was, we would invent artificial intelligence. Those are two paths. And I sort of realized the first path I don't really have a control over, right? It's up to the aliens to choose whether they want to make themselves visible or not.

So I went with the other path, the inner space versus the outer space, the mind, so to speak, right? So that was my intention at when I was 11, right? And eventually went into physics for my undergraduate degree, but discovered, yeah, I don't think I was smart enough for physics. So I went to something easier, computer science. And I did my master's in computer science, and I eventually ended up in IBM's research division early in my career. doing all sorts of stuff mostly around the internet and business to business protocol, that sort of thing. And then after that I moved into different startups, mostly tangentially involved in AI. So we had different kinds of AI during the day, right?

Sub-Sumption architectures, constraint engines, and that sort of thing. at least the classical, good old-fashioned AI, right? But by the time it hit 2012, there was this thing called Deep Learning that came along, right? And I said, what is this, right? I remember 10 years prior or when I was doing my master's, I was interested in what they called neural networks at that time, right? But there was not a lot of interest back then, right? So when Deep Learning came along, I said, yeah, this,

was probably something that is in my domain of expertise because I was coming from physics, right? And I also had computer science. So it's just like the nice overlap that deep learning, most of its equations were something that is quite simple from the point of view of someone trained in physics, right? They're just gradient descent and that sort of thing, right? Continuous real mathematics, right? So I got myself involved in deep learning. And I was thinking back then was, Yeah, with my background of physics, maybe I could basically take what I learned there and basically contribute to the field, right, or understand the field better because I'm coming from a different space than the people who are coming from machine learning, right? So I had this perspective that

physics, knowledge of physics would actually give me a leg up in understanding deep learning, which eventually proved false, right? because I just realized, no, these kind of systems are a kind of systems that you actually grow, right? So these things are more similar to biology than they are in terms of physics, right? So that eventually changed my understanding of deep learning and I moved into this notion of using what they call, it's called semiotics, which is the study of science, right? So it is a study of meaning. Right.

And then so I wrote several books in 2017. I wrote a book, Artificial Intuition, which sort of like made the metaphor, okay, that deep learning systems are actually intuition-based systems rather than logical-based systems. So intuition is a kind of different kind of inference system. Anyway, long story short, right? Since then I've been writing several books. And basically the reason why I wrote these books was to be able to get a handle of the progress of AI.

Thanks that was actually very illuminating. I was wondering how come you developed, because my understanding was that your background was mainly software engineering, so I was wondering how So I was wondering how come you developed this interest, this early in fact, interest in AI. You were ahead of time from most people. Yes.

It is. It was your first attraction in a way. Yeah, Mary. Well, well, my, yeah, my, yeah, I have a physics background, I mean a physics education So people who have gone through the physics undergraduate and graduate when I understand, we go through the same kind of development of our thinking process. So that's my background. But on the other hand, I was always good at computers.

It was like a secondhand skill. So my master's was in computer science, but it was in neural networks. So, right. I found the most of computer science. Well, as a profession, I was a software. I am a software architect. So it's not like I have ever practiced physics. It's just that's my background.

Okay. Which is usually the case. Most people who finish physics can't get a job anyway. True. Indeed. And actually, we'll get to the enterprise architecture part. But before we do, let's. Let's start with something a bit more accessible maybe for a wider audience.

And I know that you have been writing under your own brand, which is called the Intuition Machine, for a long, long time and I have been following your writing. And I picked one of your latest pieces, one that's called the Jack Frontier, our understanding AI is an even revolution. And I picked that because it describes It shares an experience that I think most people can relate to. It talks about how progress in AI capabilities is uneven, creating what you call islands of superhuman performance surrounded by surprising gap. And you make a series of arguments there.

You claim that AI has uneven capabilities and you give them an example, such as illegal AI moods, which may be able to handle routine contract analysis with superhuman efficiency, while completely failing contextual reasoning that any human is capable of, basically. And then you also say that this darkness is not random, but rather it follows the contours or the patterns of how different cognitive processes can be digitized. And people also judge AI capabilities unevenly. So our capability to judge AI is itself constrained by the very same patterns that AI exploits and fails or fails to exploit. And as a result, we have a feedback loop where jagged AI capabilities interact with jagged human assessment to create even more unpredictable

outcomes. And based on these premises, then you talk about the paradox of what you call the paradox of constraints and freedom and the temporal dynamics and co-evolution of how people using AI systems are both influencer are influenced by that. I would like first to ask you to summarize this argument, because then in the second part, I want to ask you something more specific with how to do with your specific craft. Yeah, so the, yeah, I think we've all experienced that,

the Jagannism AI, and I guess over time, we actually see that there's less, it does less stupid things. So today's models are pretty capable, right? But the jagged is a consequence of the curriculum that it was presented, right? In other words, whatever was the curriculum was the kind of like the low-hanging fruit that was available for the training. So these, I consider these deep learning systems, these large-active models, as what I call intuition machines. And what I mean by intuition here is that these are machines that learn through experience, right, through their training set. They're not learned from rules, but rather from being exposed enough times to experiences such that they eventually discover patterns and they use those patterns.

So most of the progress of AI is the consequence of what was available at that time. So in the beginning, when we're talking about GBT3, then we're talking about the corpus of the air. internet, message boards and Reddit and social media, et cetera, was your and books and et cetera were used as the original foundation. And over time, that was refined over time. So as a consequence, there are certain information, certain the jagadus that we see as humans that we can recognize in the AI is a consequence of the difference in how humans themselves develop.

We developed as people, agents with bodies. So we kind of understand what it means to fall, what it means to fall from a certain height, what we understand that jumping out of a cliff could mean that we would be hurt really badly because of the height and so forth. So we have these kind of intuitions that are developed as a consequence of us being biological creatures, right? And those are because they're intuitions, right, they're obvious to us, right? But for an agent, a AI agent, it did not have a body. It doesn't come with a body. It has a reverse kind of development where it starts off with text, right?

And therefore, there are gaps in its understanding that we kind of see. It's obvious to us that that's completely stupid, right? Because they have a different development path, right? So that's where the jaggedness comes from, right? And the feedback loop is also, right, as we work with these kind of systems, right, they also compensate for our inabilities themselves, right? So we have a limit to our attention. We have a seven plus two conceptual limitation while these kind of same systems don't, right? So when they actually render things, they would render things that would actually take a human a while.

So, like, if you ask you to do 10 top things of some subject, right, we would actually have a hard time actually coming up with the 10 top things, and they would actually instantly render it. So that has an effect in our cognition, right, in how we actually work with these kind of things. And it is like any technology, right? One of the paradigms I use very often to understand humans and their interaction with technology is, a kind of framework that Marshall McLuhan, he wrote this book years ago, is probably in the 70s, the medium is the message, but he's talking about the effects of media, right, that media actually amplified certain things and make other things obsolete. Right, so that the way you interact with media, right, exercises a certain cognitive skill at the same time atrophies a certain cognitive skill

so where just like an example like GPS people who use GPS eventually lose their ability to navigate right so so that kind of thing so AI is kind of the same thing is a media any media right and the more we use it certain things atrophy certain things become more enhanced right so that's the interplay the co-evolution so to speak of us and AI yeah actually I like the GPS examples very much for a number of reasons first because it's kind of near and dear to me I deliberately refrain from from using GPS at all precisely for the reason that that you just described and I've seen this skill atrophine people around me who do regularly yes and there's also like

you know published studies about how exactly this is how exactly this happens. This is a well-known fact. It's documented. So based on that, I would like in your own experience and also the fact that the craft of software engineering is probably the one domain currently in which we see the most rapid uptake of AI.

I would like to ask you, how do you see this fact shaping the skills of software engineering, software engineers working in the field today and how you may extrapolate or how you can predict based on that on what's going to happen to the crafts going forward because there's already been you know some studies that show that first the people who are able to benefit the most from using AI for coding are the most proficient software engineers and told that the number of so basically that hiring juniors is in decline, both in software but also another profession. So how do you see this trend?

Yeah, you're right. So with that, yeah, that's a very important topic that's coming back to light with the latest discovery by OpenAI with the solution of the Navier-Stokes equation for the Clay Institute, million million price so the mathematician Terence Tao is probably one of the best mathematicians today that we have today was saying precisely that right as we have AI that's able to solve the real real difficult problems what that happens is that the people who need to learn how to do mathematics right probably are less motivated to do it or lose certain

Yeah, they lose the incentive to do it, right? Because a machine kind of does it. It's sort of like I remember when I first started using computers and I rendered a 3D plot, right? And I said, wow, right? Back in those days, it was an impressive thing. But decades later, it was like unimpressive you actually generate 3D, right? So the question is, where's your incentive to actually learn? So that is problematic in our engagement with AI.

But going to software engineering, yeah. So what's happened now is that because generative AI is so prolific in generating code, right? Then the responsibility moves on to another side, right? With abundance, we got a scarcity, and that scarcity is, is in very, verifying results, right? So basically validating if it's actually correct.

So now there is a lot of effort now, like there's, or last of example, there's been a lot of complaints now that because AI generates all the code and fixes all the problem, the poll requests in most open source repos are becoming too big to actual handle. So there's a lot of work now navigates towards actually. just validating, verifying the correctness of the actual work. So that's where software engineering has to also changes. Now, with regards to the observation that more advanced practitioners benefit, right, is that because there's still gas in how it generates code, right?

And it's always beneficial, right, that you can steer the models to maintain a particular architecture, right? But understanding architecture is an advanced skill, right? That's actually developed over time for software engineers, right? When you start off with a software engineer, you're probably given the usual education, but usually it's only at the level of optimizing a particular algorithm, but it's not optimizing the entire system, right? is only through work that we actually learn how to do software engineering and architecture.

And that's a gap, right? So if everyone's just byte coding, then they don't like to learn architecture. Or maybe they might advance quickly to architecture. I'm not sure, right? But it's not emphasized enough, right? We know it. We just know it as people who have practiced software for a long time, and we developed a sense of architecture.

We develop a sense of design patterns and that sort of thing, right? So that when we use by coding, right, we can, there is a, yeah, there's an overarching framework or shape that we can manage, even though we don't see the individual modules, right? But we can nudge the AI to keep that shape. And that's a developed skill or acquired taste. And that could be missing for the new branch, the new, the new. people going into software engineering. However, on the other hand, they can they can probably graduate to that level quickly because you could use AI to basically exercise that kind of skill sets. I guess we'll have to see because we're kind of early in the in the life cycle of getting like a new generation basically of software engineering, our software engineers and observing how their development through time is going to come up.

Yeah. I think in general, in terms of the bigger picture, I think we just have to get a better understanding of what makes us human, right? And basically move the kind of work that we do to the kind of work that expresses our humanity, right? Because, I mean, I mean, I guess there was this notion, I mean, culturally, that because of the industrial evolution, that we should become more like machines, right? And then, but right now, right, whatever is machine like is being surpassed with AI, right?

So we have to go back and say, okay, but what are the things that actually makes life mean? What are the what do we do that? We quite generate meaning, right? In a lot of cases, it's like the problem is that the stuff that is meaningful for humans are stuff that are inconvenient, are difficult. That we, like if we climb a mountain, right, we climb on Everest. Why is it meaningful? Because it was difficult. if we took a helicopter to the top of the mountain ride it's not as meaningful right yes we have the experience but it's the actual process the actual effort right that makes anything meaningful

and that's just kind of like the that's that's intrinsically human right all right so i i like that because it's the perfect uh well i like that for a number of reasons including the fact that it's a perfect segue to uh to move on to the next question that i want to ask you in the sense that it's kind of philosophical. So another interesting claim that you make that has to do with what you call the LLM capability ladder and how LLMs climb the ladder in the way that they do. You say that there's surprisingly little about AI models,

the raw ability of AI models that changes from one generation to the next. What changes is the division of labor. Basically what we trust them with. You elaborate on a metaphor of how people onboard new hires and you say that there's a parallel there to how we entrust AI models with more responsibility basically. And the model goes from producing texts to taking instructions to conversation and feedback and then to reasoning, retrieval, action, persistence and eventually goal setting. And so when I was reading that, part of me was nodding and then another part was kind of questioning the premise. Because while I do see the parallel and I do acknowledge the engineering steps that you outline,

I also wonder, don't you think that there is also a progress in terms of model capabilities which in turn enables this gradual shift in responsibility? because don't you think that's kind of prerequisite, and that's also what all of these benchmarks show that we have like a gradual increase in actual model capability? Yeah, so, okay, let's see. Yeah, if we look at the evolution of large language models, and we can start with GPT3 as your baseline.

I say GPT3 was in the year 2000. So I actually considered a GP3 AGI. But basically, that model essentially has remained the same until GPT5 version. In fact, it was just recently rumored that all the progress of OVN AI starts off by using GPD-40, that model. and basically they have been permanently improved the curriculum to do certain things from yeah so if you start run GPT3 back in 2000 right it was really good at writing out very coherent text sentences right but there was no causally behind it it didn't really understand what it's actually going on so over time they introduced instruction following so it could actually follow instructions because prior to that it wasn't following instructions. You just gave it examples and it would complete the example, kind of like an analogy test, right?

And then you trained it on programming code. So what happened was it learned causality, so it speak, right? It learned cause and effect that it did not have prior to that and so forth, right? So as it could, we continue to develop when you added reasoning models, it improved its reasoning and so forth, right? it got to the the eventual state that we had G5, but GPL5, but it was the same essential underlying model. And what that kind of means is that it just needed more experience over time. Just like a baby, right, is the same brain that we, that exists when they're born, right?

And it just developed over time and you get to a five-year-old that's more capable than baby, right? They can walk in itself. Same thing, right? AGI, GPD3, just keep improving the curriculum. And then we got like a six-year-old version of that GPT, right? So, yeah, so that is how LLMs develop, right? In a way of we're giving it instructions first. In the beginning, it didn't know what the instructions meant.

And over time, that notion of self-awareness. why these instructions were important was actually developed over time. So until we have today, right, that they're able to be performed very well in agentic goals in terms of fulfilling goals and actions and that sort of thing, right? Which is actually different from the way humans develop, right? Because we humans as were born, right, we have billions of years of evolution, right? that give us the mammalian baseline cognitive capability. So it's not the same developmental path, right?

We aren't given instructions in the beginning, right? We're actually, we interact with the world today. So the gap that we call jaggedness is because our development paths are different, right? Human development is in our 3D space, and LLM development is in a word space, a vocabulary space, so to speak, right? So all the stuff that it's achieved has been in things that can be verbally or symbolically expressed explicitly. So when it does mathematics, it does it very well.

So there's tacit things, right, of interacting with the world. That's still TBD, right? our robots are still very slow right and because the training set is not there right it's not developed under those circumstances so i think maybe maybe the gist of what you're saying would probably come down to something like the core architecture of models like gpt has not really evolved that much what has evolved these techniques around these architecture such as reinforcement learning from human feedback plus the training data set not so much in terms of the static part let's say but mostly about the part where these models interact with users and

this is what offers the kind of developmental evolution that you seem to be yeah yeah i would argue that is true until gpt um five because we now have gpt six right, which is a new pre-trained, right? So I believe GPT6 is a step change from the previous models, right? So it's doing something different, right? We're still yet to discover how different it is, right? But it's definitely has a different, from my experience with it, it has a different feel, right?

But previous to that, right, it was all restricted from a baseline experiential view of just recognizing symbols and words. But this newer kind of AI, Astra, or GPD6, has a sense of how objects, complex objects, like buildings and machinery are organized together to make into a useful machine. So it's quite different.

So I'm not going to make the same argument for the newer model. I'm just making the argument for the existing models, right? Okay, I have no opinion whatsoever on GPT6 because I haven't had the chance to use it yet. So I'll take your word for it. Right. Yeah, my framing here is true up to a certain point. And then right now, after that point, we'll still have to discover it. One of the things that you've already mentioned that you actually have a name for in your writing is the library metaphor.

So you argue that today's AI models resemble a person who knows a lot about the world, but whose knowledge is secondhand. It's like a person who has been locked into a library, their entire lives. They've read everything there is in that library, but they have no real experience of the world whatsoever. Yes. And based on that premise, you offer a duality in terms of how these models may evolve in the future. So the first possibility is that models may grow their own felt sense from everything that we have shown them. But that sense may not necessarily be identical to ours. And we don't really have a way of knowing what exactly that's going to be.

The second possibility is that models will basically keep doing what they're currently doing. So manipulating language, but no actual grounding. So my question was, which one do you think is more likely and why? Yeah, so the question is, well, will they ever find grounding? and and and and um and um well it seems to me that they would eventually need um if they want to have the kind of grounding that humans have then they would actually need to have human bodies or bodies essentially right because the whole idea is that I think the problem with the training sets that we give to the LLMs to learn experience is that they're already curated.

They're always the kind of, they're already cleaned up, so they don't show the failures that we encounter when we're trying to understand something, right? That we understand something because we do something in the world and it fails, right? And then we learn from that failure. Now, of course, I think the newer models learn from failures when it does a genetic program, So when it's programming and it writes certain things and it fails, right? It knows how to debug it because it's all in the virtual world, right? But what's missing, of course, is the access to the physical world.

So that's really the next step, right? Being able to access the physical world, right, gives it feedback on what occurs in the physical world and what are the actual compensations for it. but there's not enough data for that kind of thing. Right. So yeah, so that's, yeah. So I think eventually it will learn those kind of things because I mean, previously it did not

because we didn't have the kind of training sets that had the that presented the failure modes. But we kind of do have that kind of, We have that now because we've trained through people who have been vibe coding and basically discovering all the failures, right, and asking it to figure it. So in the domain of programming, that's available. Probably in the domain of mathematics, I don't know if that's available. Maybe it discovered it on its own. I don't know, right?

But in the domain of physical AI, there are many domains where we don't have enough data, right? But I guess over time, as we automate the, we have like robotics that automates the experimental process, then I think over time that would be actually compensated for. But the question is, the bigger question is, will it have the actual values that we have as humans? Well, I think that's an open question. Indeed. Okay, so going back again to something a bit more specific and you already mentioned Vipe coding. So again that makes it easier to kind of switch into that. So in one of your latest pieces you elaborated on something relevant to enterprise architecture.

As a former enterprise architect myself, this caught my attention. So basically to make the claim that in AI-native organizations, the enterprise architecture and the AI harness are the same artifacts. And, well, first, before we go into the details of the argument, I will have to ask you to very, very briefly explain, or for people who may not be familiar with terms, what do you mean by enterprise architecture and also by AI harness? And then we can zoom in on their similarities. Yeah.

Yeah. So the thought process behind this is a recognition that we already know that in software engineering, because of AI, right, the concerns of software engineering migrates to something different from just actually writing code, right? So in the same way, right, we've always been doing enterprise architecture, right? And the question is, when we now introduce AI technology, right, how does enterprise architecture actually change? So that's what I call native AI enterprise architecture, right? which is a kind of like a different thing or from enterprise architecture, but it has the same goals, right? And because a lot of times when you're doing enterprise architecture, you find yourself, aren't we just documenting things, right?

After the fact, right? We're just documenting how the enterprise works with our IT technology. So what really what we're trying to do is trying to make legible, right, trying to make understandable, the big mess that we have under an enterprise, all the software components and so forth, right? And we convince ourselves that we have a handle of it because we actually documented all of it, right? But in reality, the documentation gets stale, right?

And it's not reflective of the actual reality, right? But when the context of native enterprise architecture, what you actually do is everything, right, about the enterprise, is actually made explicit within the context of the large language model. So you can actually ask it questions and resolve tensions, right? And basically you make legible the enterprise as a consequence of the power of the technology. That would be right.

But the thing is, the very odd thing about enterprise architecture is that we're, we're, we're, really managing boundaries within organizations. The organization has all these different boundaries. We're building software artifacts that respect those boundaries, so to speak. And the way enterprises have always worked, right, is the reason we give organizations enrolled and so forth, right, is that we want to, it is kind of like the modules of responsibility, so to speak.

Right. Right. But when we talk about another world, we're talking about a native AI organization, right? Those boundaries, those roles, right, are actually artificial. So basically, there's no reason why you have like an accountant, right, right, because the, a separate accountant role, because the AI can do all of that, right? Right, you can do accounting, it can be coding,

you can do the project management and surf, right? So the boundaries are technically artificial. So if we, if you wanted, if you went towards a native AI organization, there seems to be like there's no roles going on, right? They're just doing all the work and out pops the actual outcome, right? But what, but the question. is why is that an actual problem, right? The problem is because humans, we need legibility. We need to understand what in the world is going on. So we go back to having some sort of organization,

artificial or human base, right, that gives us a sense of what's going on, right? So there's this always this tension between legibility and fluidity. So I make this argument in another essay, right? This essay about a British admiral, right, who when they used to fight awards in these naval battles, right, they had to basically preserve the line, right? And that's how it would fight. If you veered off the line, right, then that would be – that would lead to punishment. So basically the admiral was shot because he left the line, right? The reason why they kept the line was so that they understood that everyone was following the rules.

Everyone was there was some sort of, it was a signal that everyone was following the rules rather than actually good battle tactics. And then it eventually evolved such that Nelson, right, basically when he fought, I think, the French, French or Spanish, right? I forget, right? But basically told his commanders that it's okay. The only thing that I want you to do is when you, when you, when you confront the energy, you would basically confront the enemy, you would basically line up beside it and fire all your guns. That's all right. But it doesn't matter if you kept the line or not, right?

And he eventually died, but they, they were successful because he gave agency for all the different commanders of the ship. right but the the the point is that um um we're always trading off legibility with effectiveness right within enterprises right and the the eventual evolution of uh enterprises which would be native AI right is that we'll lose that legibility right in favor of a lot of fluidity but how do we actually keep it so that we can maintain accountability and oversight over the the enterprise I would argue actually that, well, yes, obviously legibility is benefit, let's say, of modularity. But I would argue it's not just because of that that we design systems the way we do in a modular way. Even if you have, let's say, a fully AI native organization, I think that yes, in theory, you could have like one monolithic AI.

do everything, but what we are observing either today in, let's say, glimpses of such systems is that most of the time we actually have these systems being sub-agents that divide the labor per tasks. And I think the reason for that is not just necessarily for a human to be able to inspect what the system does, but also for division of labor and modularity and scalability and also for the system itself to be able to regulate and control itself. Wouldn't you say that this is also an important requirement? Yes. Yeah, we would like to, yeah, we divide the work and we divide it to specialists, right?

Because why do we divide to a specialist? Because we want the, yeah, like, like, we might divide, yeah, so one of the good agentic pattern is like the, the, the, the AI that's generating the code, so I speak, right, shouldn't be the same AI that's actually doing the auditing, right? Because there is, yeah, you want the kind of specialized, so it has one, one kind of goal, right? One is the audit and one is to generate the code, right? And if they overlap, they can actually interfere with the performance.

It's kind of like when you're writing, you don't want to be writing and editing the same time, right? You just want to be writing and coming up with all the ideas, and then a different phase, you actually become the editor, right? So you can do, it's much more effective if the goals are more narrow, right? So if you're a tester, right, versus the planner versus the architect and so forth, right? The roles are specialized, and that's how we, that's how we would architect the agentic platform, right? They would have special skills for the particular role that they're in, right? And that works out. I think that works out well within bigger organizations, right?

Not so much in more creative endeavors, right? Because in creative endeavors, you really don't know what the roles, the roles that are themselves might be a little bit more undefined, right? Or multi-discipline, that sort of. But yeah, I agree the important specialization and specialization sort of like implies some sort of accountability, some sort of capability bar, so to speak. So one of the other things that you mentioned

about enterprise architecture. and how it works today and how it has historically worked as well, is the fact that it works a little bit after the fact, let's say, you have a nice documentation of everything, but gradually your documentation sort of loses touch with reality, what's actually happening on the ground. And interestingly, one of the things that I've been keeping an eye on lately in the last few months,

and is an effort made by some people to port very probably the most well-known standard in enterprise architecture something called the Archie-made to port that to do RDF for those who are familiar with that is standard for knowledge representation in the good old symbolic AI world and the reason that they're doing that is precisely that to be able to have like this live feedback look between actual systems and their documentation in enterprise architecture. And I came to this from the angle of context graph because another role that people have figured out that enterprise architecture may have to play is precisely that. So for the last year or so people have been talking about context graphs which essentially comes down to

a graph of all the decisions that have been made in a certain organization as well as the decision traces that lead to those. Some people have been making the case that this can be built on the foundation of enterprise architecture because enterprise architecture is what connects all the systems and therefore can give you access to all those traces. And the argument goes fine, okay, we can use enterprise architecture for this and in turn we can use this porting, this live, this bi-directional feed to have enterprise architecture that's actually in touch with the reality on the ground. So how does that sound to you?

Yeah, that is, yeah. The effectiveness of a large-laction model is highly dependent on its context, right? The more context you give, the less is going to hallucinate, right? Hallucinate, meaning fill in the gaps. Right. So if you can bring in all the context of how your enterprise operates, right, including the architecture decision records, the ADRs, right? Then the LLM has a better chance at reasoning of why it takes one approach over another approach, right?

I mean, right. So it makes sense to bring everything into the LLM's context when, which is getting pretty big nowadays right just two million right to basically able to reason about your next steps within your architecture right what's the next steps for if you're going to develop a platform right give me all the scenarios so so I've been building one for Fannie Mae right where we have all the different divisions in the different groups right within the anime and then and basically a whole bunch of architectural principles and you're basically able to run the LAM to give you

different scenarios and different reasonings behind those scenarios so that you can make better judgments right because if you think about it right how does it actually work within an enterprise architecture group like you go on and people have different subsets of the entire enterprise within their head right and they're gonna go in a meeting and there's sort of like let's reason about this But is that reasoning as good as a reasoner where it has all the context within the LLM? And you can actually spend time exploring more alternative scenarios rather than coming up with your reasoning in a 30-minute meeting. Let's go this way.

And yeah, that's one aspect, right, is the efficiency. The other aspect is, yeah, architecture is like governance, right? It's government, right? And you're basically balancing these rules that you give across all the different subgroups, right? And the reason why you have this group is so that you can scale, so to speak, right? It makes things less costly, right? At the same time, every group should have its own autonomy, right? its own agency, right, too, because the architectural group does not have

intimate access of the actual detailed requirements of a particular subgroup, right? So there's always this balance between governance, higher level governance, and the agency of individual groups, right? And by balancing that, right, you can sort of like, because if every group was, for example, totally independent, right, they'll be making their own stuff and reinventing the wheel, right, hundreds of times. And that's probably the natural thing. So when you look at the architecture of big enterprises, you see so much redundancy.

You see kind of like the same kind of functionality, duplicated, and we're paying five times for the same functionality, right? Because it was decided by different groups, right? But how do you actually get a handle of that, right? so that there is uniformity at the same time you preserve serve agency. And I think there's a lot of potential there in terms of framing enterprise architecture in the context of using today's AI. There's a lot of potential right there that we can see, right?

Which is different from, this is also one of the, which is different. from software engineering, right? We can use AI for software engineering, it's a whole bunch of techniques, right? But we can also use AI for enterprise architecture, a whole bunch of other techniques. All right, so as we're getting close to wrapping up, we have something about five minutes left. So I would like to wrap up by asking me to share like a very, very quick preview of what was actually the reason that made us connect in the first place. So your writing on graph engineering. This was what caught my attention because I was attracted to all things graph.

I run a conference around knowledge graphs, I have a newsletter on graphs and therefore this was very, very interesting to me. So I would like to quickly ask you to share your thinking of how you have been using looks specifically in the context of ages, where looks work, where looks may go wrong, and how graph engineering may alleviate that. Yes. So that essay was triggered by a post by Pete Steinberger. He's the one person who developed OpenClawe.

And he's at the forefront and actually running fleets of agents to do his work. He's been famous for basically approving 5,000 PRs in a month or something. So he's doing some crazy, scalable work. And a few months ago, he said we should all be doing loops. We should all be doing loop engineering. And what he meant by that was like it is these loops. that are self-improving. So basically you set up the AI so that it runs in a loop and it's constantly automatically improving itself. And he was saying, everyone should be doing loop as engineering,

right? That was a few months ago. Then one day, right, maybe a few weeks ago, right? He says, are you still doing loops? Are you still doing loops? You should be doing graph engineering, right? And maybe he was just joking or whatever, right? But I basically parsed out what he meant, right? And what he meant, what basically I think was missing there. You probably should have said loop graphs or something like that, right? But basically these are loops, different loops that are all interacting with each other, right, to basically improve the system.

So it's a graph of loops, so to speak, right? So, so it, it, and the question is, I guess what's being addressed is, what happens when a loop breaks, right? Right. And because you have a graph of loops, right, and you basically have basically other loops that are, that actually audit your loop, right, and actually resolve any conflicts, right? And as you build an organization, right, you'll have all these different loops, right, that all interact in different ways, right? And how do you actually structure those kind of loops? So the presentation that I'll be giving at CDL, right, is basically introducing loops and then this notion of loop graphs, right, and showing how you would build it within a harness engineering framework.

So that would be it, right? So there, but basically, yeah, it is this idea that organizations have multiple loops going on, and these multiple loops are actually interacting, right? And there's a question of who's observing who, who constrains who, who can override who, where is that evidence coming from, and that's our thing. So yeah, in a lot of ways, it's also related to enterprise architecture, right? Because that's really what you're asking about, right? How do the different entities within the organization actually interact with each other, right?

And keep everyone on track. I'm really looking forward to learning more about your experience and your actual thinking and how this could be developing because like I said, this is actually a very, very new thing. athletes in terms of naming it. Perhaps there are practitioners who have been experimenting for it, with it longer than we have actually known the term, but it's still very, very fresh. So very much cutting it.

So thanks a lot for the very interesting conversation. And like I said, looking forward to actually meeting you in person in London. Okay, yeah, thank you so much for having me. and I'll see you soon in London. Thanks for sticking around. For more stories like this, check the link in bio and follow link data orchestration.