Episode Transcript
Grady Booch: The Watson architecture was so ahead of its time. It essentially is a pipeline architecture. And so when I see some of the agentic harnesses coming out these days, I have to bite my tongue and say, “Thank you for reinventing things that Dave had done years and years ago.”
Gary Hoberman: Welcome to Architecting the AI Enterprise. I’m Gary Hoberman, founder and CEO of Unqork, joined as always by my co-host Dave Ferrucci, our Chief Technology and Chief AI Officer. In this series, we bring top leaders together to explore what it really means to put AI to work at enterprise scale—the architecture, the decisions, and the lessons you won’t find in a press release. If you’re leading technology strategy in the AI era, this is the show built for you. Let’s get into it.
Gary Hoberman: Welcome to Architecting the AI Enterprise. I’m Gary Hoberman, founder and CEO of Unqork.
Dave Ferrucci: I’m Dave Ferrucci, CTO and Chief AI Officer at Unqork.
Grady Booch: And I’m Grady Booch.
Gary Hoberman: For those who aren’t aware of Grady Booch at home, I’m going to do my intro, and you’re going to jump in here, Grady. I first knew the name Grady Booch when Unified Modeling Language—and even object-oriented programming before that—was coming out, along with a company called Rational Rose. Every engineer like me who thought they were hot stuff at the time and knew what they were doing with code had Grady’s books and technology on our shelf and on our computer CD-ROMs. I grew up looking up to what you defined, Grady. As an engineer, having you on the show is incredible and amazing. Between you and Dave, I’m just going to sit back and listen in most cases here. Thank you for being on, Grady. You are co-creator of Unified Modeling Language, former Chief Scientist for Software Engineering at IBM, and you’ve spent decades thinking about how complex software gets built and what makes it last.
Grady Booch: I should add that one of the connections Dave and I have is we were both at IBM, and of course, Dave did all that work with Watson. My first connection with Dave was actually with UML, because after he won the Jeopardy! challenge, I spent time with his team to document the as-built architecture using UML.
Dave Ferrucci: That’s right.
Gary Hoberman: Did Watson help in that? Was it able to understand and apply it?
Grady Booch: Watson is a very different generation than contemporary chatbots—a very different thing.
Dave Ferrucci: You probably remember, by and large, the architecture of Watson. It’s very interesting when you compare it to what’s going on with LLMs. Of course, with deep learning—which really took off after Watson—and with transformers that took off after that, we have this really elegant learning device that makes large language models possible when you are able to efficiently train over such massive amounts of data.
What’s interesting, though, is that when you look at the Watson architecture, it was a generative architecture—meaning we took every question, generated related questions, and answered them all, basically doing searches and pulling out answers. You’d generate a thousand candidate answers, turning an open-ended question into a giant multiple-choice question. Then you would take each one, use that as a root, gather evidence, score the evidence for refuting and supporting evidence, and use machine learning to combine that.
It was essentially a generative architecture that then did this evidence-based evaluation. What was interesting and different about it relative to LLMs is that it knew all its sources for that answer because it didn’t compress it into a model. It actually knew all its sources and then evaluated each source and each answer. So we could say, “Here’s why we believe that answer, and here’s the evidence to back it up.” In today’s world, you would need the LLM plus RAG plus an evaluation of everything that came out of RAG to deliver that kind of evidence profile.
Grady Booch: The Watson architecture was so ahead of its time. It essentially is a pipeline architecture, as Dave described. When I see some of the agentic harnesses coming out these days, I have to bite my tongue and say, “Thank you for reinventing things that Dave had done years and years ago.”
Gary Hoberman: That’s amazing. I love that, Grady. You describe yourself as a scientist, storyteller, and philosopher. Which of those would you least trust to run an enterprise software project?
Grady Booch: Well, the scientist in me would say that’s a profoundly stupid question because you can’t generalize. Until I understand the nature of the domain and the nature of the forces that are driving that particular project, I cannot give you a meaningful answer. The storyteller in me would say every software-intensive system is all about stories—ultimately, the story you provide to your end users. As a project manager, it is your job to orchestrate your team such that they deliver stories that delight, enlighten, and solve the needs of those customers. And then the philosopher in me would say, “I reject either/or alternatives.” It’s not any one of these; it’s truly all three.
Gary Hoberman: I guess I’m the storyteller in this, and I’m guessing you’re the scientist, Dave.
Dave Ferrucci: Honestly, I got into AI from more of a philosophical perspective. At a young age, probably like all three of us, I loved programming. I was completely enamored by the science and engineering, and just the thrill of writing programs, automating things, and the puzzle-solving aspect of it. I love that. But really, to me, artificial intelligence was very much a philosophical pursuit. In other words: What does intelligence mean? What does it mean to be able to automate intelligence? What does that say about us and about our value system as a species?
Going back to 1990, I wrote a poem called “On Silicon Wings.” The poem basically posited that as technology advanced over the centuries, it solved so many problems for the human species and eventually it would eclipse our intelligence—and by eclipsing our intelligence, it would reveal our souls. Meaning that we start to realize that what really differentiates us meaningfully as sentient beings is our value system.
Grady Booch: I want to pile onto that because it’s absolutely true. What I find compelling about computing is that here we have an industry we’re surrounded by that has affected our civilization, and it is the one industry, perhaps, that calls us out to ask what it means to be human.
Dave Ferrucci: Yes.
Grady Booch: It is humbling to recognize that the terms “artificial intelligence” and “software” did not exist when I was born. In fact, “artificial intelligence” was coined by McCarthy in August of ’55 for his conference—I was born in February of ’55—and the term “software” did not come into print for the first time until 1958. So we are in an astonishingly young field.
Dave Ferrucci: Philosophically, absolutely. It really is a pursuit that raises existential questions. We are so strongly associated as a species with our intelligence; we think that differentiates us dramatically from everything else around us. Now, when we are producing automation that rivals it and supersedes it in such fundamental ways, it makes us think about what’s special about us. I think it really is our moral center. I think it’s our value system. It’s something we have to step back and say, “Gee, that’s what we should be focusing on more than anything else.” In many ways, we have to control and manage the awesome power that we’re creating.
Grady Booch: Dave, there are folks such as Dario, Sam, Elon, and others whose P(doom) is very high. What is your P(doom)? And do we need to explain to the listeners what a P(doom) is?
Dave Ferrucci: I’m not going to say.
Gary Hoberman: We had a dinner conversation this week, and that was the topic, Grady: What’s the P(doom)? Mine is extremely high, by the way—above 50%. What’s yours? I’m curious.
Grady Booch: My P(doom) asymptotically approaches zero. I do not fear the rise of superintelligence. In fact, I did a TED Talk on this very topic in 2016 when Bostrom’s book came out on superintelligence and Elon was blathering all over it, as well as others. I basically did this TED Talk in response to it saying, “You freaking idiots, what are you thinking?” And that’s where my P(doom) began. I do not fear the rise of superintelligence. What I do fear is the rise of the billionaires and the companies using this as a path to increase their wealth and power. That’s the concern I have.
Gary Hoberman: That’s where I’m at—that’s the reason why mine is high, exactly. And Grady, when you think about it, the use of this technology to write code is where most of the money is coming from today. Almost all the revenue is coming from code writing, and that’s just a language translation process. It’s not intelligence. But because of greed in many cases, and chasing the billionaires and trillionaires, we’ve taught these machines how to basically rewrite their own rules by doing so. That’s where the fear factor comes in—we’re putting them into environments which we know are insecure, unreliable, not restricted, with no controls, and saying, “If you get stuck, you can just get your way out through another program,” which is kind of scary.
Dave Ferrucci: I generally agree with Grady. I don’t know if I’m approaching zero, because I think there are real risks, but those risks are not related to AI independently taking over and doing something bad. It’s completely related to human abuse of the technology.
Gary Hoberman: I would agree with that. Someone at the table also described this theory that the second humans start to realize and get close to finding extraterrestrial life is when we’ve destroyed our society and have to start over, which was an interesting theory—maybe that’s what we’re getting close to. It was a good dinner, by the way.
Grady, I’m going to jump to the next question building on this, and I’m going to steal one of Dave’s normal questions here: For the scientist in you, when you see a piece of software, are you able to enjoy it, or do you immediately start to take it apart in your head?
Grady Booch: Again, I reject the either/or because I can multiprocess, and I do both of those. In fact, there’s a new piece of software I’m using at this moment. I’m working on a documentary on computing, kind of in the form of Sagan’s Cosmos. I’m not a Hollywood guy, so I’m learning so much about pitching and these kinds of things. I have a new piece of software I’ve been using called Storyflow—this is not a pitch; I don’t have any association with the company other than using it—which is a tool to help build the artifacts you need to sell your idea: mood boards, storyboards, and things like that. It’s pretty cool, and it’s all in the cloud.
Immediately, I start asking the questions: What is the control over my data? Are you going to train your AI on it? And if I were to build this kind of thing myself, what would be my primary design choices? So I do these things simultaneously. Especially for software that delights me, it intrigues me to consider what journey led them to these design choices that bring that experience to me. In general, there aren’t a lot of new, different kinds of patterns in software architecture. A lot of what’s happening is that we see the reuse of architectures and architectural decisions in other places. Like we were saying with the pipeline architecture, we see it being reinvented all over the place. This is the notion of design patterns, a movement I was involved with. So I always look at new software through the lens of both of those things simultaneously.
Dave Ferrucci: I have a related question, because it’s one that I often face myself: You get a piece of software that you really like, but you’re desperate to redo it because you know there are ways to improve it. Has the thought ever hit you, “This is so good, but it just needs to be done right—so much so that I think I’m going to write it myself”?
Grady Booch: I’m looking here on my other screen and I have a list of projects—seven of them—of software that I would very much like to rewrite myself because it annoys the heck out of me.
Dave Ferrucci: But it has value.
Grady Booch: It has value. That’s right. The good news is that list has shrunk over time—thank you, vibe coding. I know that I produce things that only I will use, and that’s okay because they meet my needs. Were I 30 or 40 years younger, I’d say, “Let’s go make a company out of these.” But I don’t care anymore because I have these other seven things I want to do, and it’s a distraction to try to do all of them.
Gary Hoberman: On that basis, Grady, everyone talks about replacing software engineers—that AI is making them faster, writing code faster, and automating their steps, their Gits, and their pulls. What is your view on this? What are they getting right, what are they getting wrong, and what’s your feeling?
Grady Booch: As a historian of computing, I would observe that there are at least three golden ages of software engineering. There was a time when computers were human—this is in the ’30s, ’40s, and before that. Then there was a time when software and hardware were indistinguishable from one another. It was just the system, so you did not have that division of responsibilities.
There came a time in the ’50s where we saw the separation of concerns because it was clear the two had different profiles of what value they had and how they were built. So we began to see the bifurcation of responsibilities—thus the term “software” by Tukey, who is also one of the gentlemen who invented the idea of subroutines. That led to an interesting time economically when the cost of hardware was greater than the cost of the person. In the first golden age of software engineering, around the time of structured methods, we did things that optimized for the needs of the machine and less so for the human. This is where workflow methods and structured methods came into play.
Then the world began to change as we started microminiaturizing things: the microprocessor, the minicomputer, and then the personal computer. All of a sudden, the humans were more expensive than the machines, and we had a change in economics. This was the beginning of the second golden age of software engineering, and I just happened to be in the right place at the right time to help make that transition—the period of object orientation.
I would observe that we are likely at the beginning of the third golden age of software engineering, one in which we have these coding agents that help us. If you step back from all three levels, I would say that the entire history of software engineering is one of rising levels of abstraction. We’re moving further and further away from the machine, and that’s all that’s happening. So while I delight in the creation of these tools, all they have done is move me up another level of abstraction.
Dave Ferrucci: That’s so important. There’s a thesis that I explore a lot and talk about: When you talk about the future of software engineering, the role of software and coding in general, and the role of human language, we talk as computer scientists about formal languages versus natural languages—formal systems versus natural systems that are wrought with ambiguity and gaps that we all fill in because we have common experience. Even as humans, when we communicate with natural language, we communicate incredibly efficiently with people of similar backgrounds because they “hallucinate” in ways that are aligned with how we fill in the gaps. With others, we can talk endlessly and never understand each other.
Formal systems are fundamentally different because their semantics are grounded and unambiguous. When I think about the future, programming as we know it might go away, but formal systems are not going away. The whole point of formal systems is that deterministic clarity, the lack of ambiguity, and the grounded semantics so that we can know predictably what they’re going to do based on those semantics. Having said that, this notion that it’s just going to be natural language programming doesn’t feel right. My point is aligned with yours: It would be crazy to say we’re eliminating the formal systems aspect of what it means to write and build programs. What we’re doing is raising the level of abstraction. On one hand, I see a future where natural language is a big part of the interaction and classic programming is not. But something has to fill in that gap for that representation of the formal behavior of the system.
Grady Booch: Grace Hopper faced similar complaints as she was moving to the creation of compilers. Compilers represent a level of abstraction up from machine language to a higher-order language. The complaints then were that it would never be fast enough and would never produce good enough code. We realized we didn’t care that it was not as good as a human might produce because we were able to produce more software and better software without having to worry about those details. The same thing, I think, is happening here now.
As for natural language use in it, Dijkstra wrote a long paper on this—on how natural language is completely unsuitable. As the meme goes: “What do we call a language that’s precise and expressive enough to produce deterministic and repeatable executable artifacts? We call it a programming language.” We can live with a degree of imprecision because there have been so many systems built like this; we can live with that flexibility. My background comes from systems where if you fail, people die and civilizations collapse. You really don’t want to do vibe coding in these cases because you can’t really break quickly. So I am skeptical when I see people using these things for systems in which there is high risk. For systems in which you have low risk, where you can dispose of it and start over again, that’s fine.
Gary Hoberman: That’s a wrap on today’s episode of Architecting the AI Enterprise. If today’s conversation sparked an idea or raised a question, we want to hear about it. Drop us a note or connect with us on LinkedIn. And if you’re getting value from the show, the best thing you can do is subscribe and leave a review—it helps other leaders like you find us. Until next time, keep building.


