An Interlude Before the Third Act of “The Assemblage of AI”

In the conclusion of my last blog, Chains of Unstable Correlations, I wrote:

When you work around experts long enough, you notice that experts sometimes (usually?) forget what isn’t obvious to non-experts. Conversely, people in the field often see things experts overlook because they live inside the operational texture of the system.

After a conversation with an old friend yesterday, I realized that after my recent barrage of 45+ minute blogs since the publication of my book, Time Molecules, in June 2025 (which probably amounts to an entire book in itself), it’s time to take a step back and regroup with the two or three readers of my blogs about the bigger picture of the intelligence of a business.

I’m a BI Architect/Developer who misunderstood the “I” in BI all those years back in 1998 when I first heard of BI. I just started my dev position with the SQL Server OLAP Services 7.0 team at Microsoft, and never heard of an OLAP, BI, or DSS. I thought the “I” was in BI was like the “I” in AI. It wasn’t until about 15 years later that I realized the “I” in BI was like the “I” in CIA—as in, gathering intel to give to the decision makers. Subconsciously, I think I didn’t want to see that since the “I” in AI is a lot more fun than the “I” in CIA. So that lead me down long and windy road where I was perpetually out of sync with the mainstream BI crowd and sort of locked out of the AI crowd.

So it’s perfectly understandable that sometimes I need to step back to ties things back up.

FYI, here is a list of most relevant of those “recent” blogs:

1. The Complex Game of Planning: Plans are about how to get from State A to State B. It’s usually not a single step, the map isn’t fully materialized, and the sequence of tasks often requires parallel threads.
2. Stories are the Transactional Unit of Human-Level Intelligence: This is the way we passed down knowledge before we could write.
3. Beyond Ontologies: OODA Loop Knowledge Graph Structures: This offers a few examples of story formats.
4. Conditional Trade-Off Graphs – Prolog in the LLM Era – AI 3rd Anniversary Special: Mapping pros and cons of configuration changes we make. Being cognizant of them, even the ones we’d rather sweep under the rug, is a key to good decisions.
5. Long Live LLMs! The Central Knowledge System of Analogy: This defines the roles of LLMs in my framework. That role is of the know-it-all friend who knows a lot about a lot of things. This friend can also organize.
6. Reptile Intelligence: An AI Summer for CEP: This is an explanation of System 1 in my framework. A web of functions, each fast, simple, highly-probable but not exact.
7. Outside of the Box AI Reasoning with SVMs: The freedom to push the boundaries with some idea of how far we can go before things get scary. This prevents us from needing to be overly-cautious because we don’t know where the “red line” (a boundary we are not supposed to cross) is at. This is really Part 1 of Chains of Unstable Correlations.
8. Analogy and Curiosity-Driven Original Thinking: The way we really create.
9. Thousands of Senses: Not just the five senses we learned about in school.

This break is especially timely since my next blog, “The Products of System 2”, is indeed intended to reign things in. Originally, this what I’m writing here was supposed to be the introduction to “The Products of System 2”, but it’s too long for that purpose and I want to be sure I clarify what I have in mind.

The upcoming post, The Products of System 2″, is like the “third act” (the closing arc in which prior forces converge into outcome) of my virtual book, The Assemblage of AI. It’s not the end of the trail, but the beginning of the end. It just up the central question:

What is the actual product of all the Assemblage of AI?

It’s the same as the product of human intelligence: stories, plans, procedures, strategies, designs, recipes, checklists—the intellectual artifacts no other creature on Earth comes close to producing. These are the solutions we’ve figured out—seeded by the solutions created through nature’s hundreds of millions of years of evolution, then observed by us as a library of examples, and adapted to our needs, and added to our corpus of culture. And our culture has replaced the seeds of invention created by evolution as the primary fountainhead of further intelligence. From thousands of sensory inputs, through layers of transformation networks, the end result is an organized body of information that forms the floor of the wisdom level of DIKUW (Data → Information → Knowledge → Understanding → Wisdom).

Let’s revisit the premise that has run through everything I’ve written, as distorted by my misunderstanding of the “I” in BI.

The core insight of my book, Enterprise Intelligence, is that enterprises have been building analytics platforms for decades that correspond—mostly unintentionally—to components of human intelligence. It makes sense: we treat corporations as entities, and human intelligence is the only real model we have. So our tools evolve in parallel, mimicking the same functional needs.

Over the decades, new “flavors of the year” arrived, each hyped as the next big thing because the previous one fell short in some way. But those flavors weren’t replacements—they were necessary components that got installed, grabbing the spotlight while the former star receded into the background. The hard part was always wiring them into a coherent whole. Evolution handled that wiring for biological intelligence over eons. For us, it’s our own creation—so it’s up to us to do the integration.

I thought about this back when the Japanese Fifth Generation Computer Systems (FGCS) project was underway. I never saw it as a true failure. They simply underestimated the difficulty: crafting rules at scale, building consensus around them, and maintaining them in a constantly changing world. Symbolic expert systems abstracted concepts in ways that lost the rich contextual details—symbols are inherently lossy, even if the intent was deterministic rather than probabilistic. It hit a roadblock, but the direction wasn’t wrong.

When FGCS was abandoned, it didn’t mean Prolog (or any logic programming) didn’t work. It meant Prolog alone wouldn’t suffice—nor will the Semantic Web alone, nor Big Data alone, nor machine learning alone, nor LLMs alone (that’s the “NoLLM” theme running through The Assemblage of Artificial Intelligence). It’s more about AND than OR (replacing what didn’t work with something else from a different angle). The trouble with AND is that means there are more moving parts, more things to worry about. With OR, we can just move on to the next thing to see what happens. Well, of course. The world is made of interacting systems, not things. We architect systems from components. Intelligence is a system, not a thing.

There’s a big difference between something that didn’t work because it’s only part of the solution and something that plain didn’t work.

Closer to my heart in the BI world are those OLAP cubes that seem so 2005. Caching when we know something will be used often—and having the power to update it (even if not in real time)—never goes out of style. BI itself has endured many pains and evolutions: Bill Inmon’s formalization of the Enterprise Data Warehouse long ago, the rise of OLAP cubes, master data management to integrate domains, the shift from ETL to ELT, and the progression of visualization tools from single-user Excel to multi-tiered Power BI.

In 2004, my work situation led me to build an AI SQL Server performance tuning engineer. (I wrote about the origin story of my AI journey in Enterprise Intelligence.) I used Prolog, but needed to write my own .NET version (I worked at Microsoft at the time) with two key changes. First, most non-software developers who code tend toward monolithic structures—hard to write and maintain. I made it distributed, snippets of Prolog fragments composed into larger ones, modular logic blocks.

Second, automatic rule generation. Machine learning was emerging, and deployed models were already autonomously making business decisions. Software code was another source. Between those, plus pulling “MetaRules” from SQL queries (reference tables as MetaFacts), rule creation became at least partially automated. But roadblocks remained: managing definitions was brutal, Semantic Web was barely known, no graph databases existed. I couldn’t develop all that as I was writing the Prolog engine and building a data warehouse supporting automatic rule generation and maintenance (what was like MLFlow today).

I detailed the origin story of Enterprise Intelligence in the book: “Intuition for this Book”, page 51.

Over the next decade and a half, those roadblocks started dissolving. More data science tools, massive data volumes and cloud infrastructure, master data management, Data Vault, Data Mesh, data virtualization—all arrived as pieces. Integration stayed tough; some even introduced new issues (Data Mesh risking silos, Data Vault slowing queries, MDM needing smarter matching when NLP wasn’t mature enough for complex entities). Semantic Web and knowledge graphs gained traction, and graph databases like Neo4j became practical.

But through my experience with BI, SCL, Map Rock, and Reps, I didn’t need to try my hand at AI again because I knew there were still roadblocks.

Then ChatGPT went viral in November 2022, and we finally had a major missing piece, LLMs good enough to provide light inference and act as glue between components. For the first time in a decade, I didn’t sense any readily visible big obstacles for my plan. Of course, more obstacles are out there, but the path seemed clear to move until the next hurdle. My book, Enterprise Intelligence, was about exactly that: integrating these disparate pieces (BI, data mesh, knowledge graphs, knowledge workers, AI) into a whole. See my blog, Enterprise Intelligence: Integrating BI, Data Mesh, Knowledge Graphs, and AI

LLMs became the mortar. Why are LLMs good as mortar? They are high-end translators that can serialize and deserialize communication between loosely-coupled or even decoupled components. When we’re having a conversation with someone, we don’t emit some encoding of the trillions of synapses in our brain. We serialize it into language and the friend we’re conversing with deserializes it and hopefully it properly links itself into his synapses. Like a high-end translator LLMs can take into account fuzziness and language nuances. It may not be fluent in the subject of the conversation, but it’s more than just a word-mapper.

To be clear, my approach to AI, at this point, doesn’t require great technological leaps to set up a baseline. It’s data-driven at the foundation and LLM-driven for the glue. We don’t need to invent massive new swaths of tech. What enterprises have already built—data warehousing, business intelligence, OLAP cubes, big data, data science, cloud SaaS platforms, event processing, master data management, data virtualization, methodologies like Data Mesh and Agile project management, ETL/ELT processes, MLflow, semantic layers—is like assembling the organs of an intelligence of a business.

I once read somewhere (I can’t recall where) that Cro magnon of 100,000 years ago were physically like us today. But they didn’t act like us. How those conclusions were made, I don’t know. But an interesting hypothesis about how we could look the same, but not act the same is that the heavy evolution was the tweaking of our neurotransmitters. I don’t know how much truth there is to that. The takeaway is that the last mile to AGI—the last mile of a 10-mile trip that is wrought with all kinds of potholes and mud— is primarily about fiddling with the interfaces.

It may not be a plug and play yet, but the components exist. The dream is like the electric self-driving car, the major parts exist in good-enough and improvable state to put something viable together. In my case, Python and Java scripts with assists from LLMs and knowledge graphs like Wikidata.org, and LLMs (with translation skills far beyond word-mapping, plus fine-tuning and prompt engineering) are effective intermediaries between the parts.

The one potential exception is something like Reps, from my 2014–2016 work on Reptile Intelligence / Automata Processor. It was a hefty component for massive parallel rule/pattern recognition, but it ran into roadblocks too (that’s a whole other story). I don’t see it as crucial now—MLflow and modern streaming engines are the core of System 1-style fast recognition well enough.

RAG, chain-of-thought, mixture of experts, and connecting protocols like MCP are the strong core for System 2 reasoning and organization. LLMs play a dual role here in the intelligence of a business—centerpiece for deeper reasoning/organizing and mortar between components.

Lastly, my target audience are enterprises, not AI vendors. As a Principal Solutions architect for Kyvos Insights, a vendor of a killer Semantic Layer, that makes sense. My job is to highlight how Kyvos’ semantic layer fits into a modern enterprise’s layer of intelligence.

I wouldn’t say that success among enterprises is only “survival of the smartest”, but it’s at least a major pillar. Enterprises know they are in competition with peers as well as non-peers that want to set a flag on other territory. Like a human, that enterprise needs superior situational awareness, a strategic mind, and the skill of persuasion (thanks, to the late Scott Adams). The enterprise should be concerned with building superior private capability, the enterprise-developed strategic asset of all strategic assets, as opposed to jumping the current bandwagon because that works for others.

Look out for “The Products of System 2” in a couple of weeks (late Feb to early March 2026).