Podcast #231 Industrial Strength AI

TRANSCRIPT

Ken: This is Ken Forster, Executive Director at Momenta. Welcome to our Digital Thread podcast produced by, for, and about digital industry leaders. In this series of conversations, we capture insights from the best and brightest minds in the digital industry - executives, entrepreneurs, advisors, and other thought leaders. What they have in common with our team at Momenta is their deep industry expertise. We hope you find these podcasts informative, and as always, we welcome your comments and suggestions.

Good day, and welcome to Episode 231 of our Momenta Industrial Impact Podcast. Today, I'm pleased to host Kence Anderson, CEO and co-founder of Composabl, the platform for building industrial strength and intelligent autonomous agents. Momenta recently led the seed round for Composabl, our 53rd portfolio company. Kence is an entrepreneur and innovator in the field of autonomous systems. He has designed over 200 intelligent autonomous agents for companies such as PepsiCo, British Petroleum, Shell, AB InBev, Bayer, and others.

Prior to co-founding Composabl, Kence was Director of Autonomous AI Adoption for Autonomous Systems at Microsoft, having joined them as part of the acquisition of Bonsai. He also recently authored "Designing Autonomous AI: A Guide for Machine Teaching," published by O'Reilly. Kence is the son of a master teacher and is trained in mechanical engineering. He utilizes both aspects by researching how the principles of teaching combined with human expertise can be used to design and build useful AI.

Kence, welcome to our Industrial Impact podcast.

[00:01:50]

Kence: Thanks, Ken, for having me.

[00:01:51]

Ken: Thank you for taking the time to do this. It's great because it's right at the forefront of our announcement for investing in Composabl, which we're very proud of. You likely know we used to call this the Digital Thread podcast and recently shifted our focus to the industrial impact side. However, I still like to ask the signature question, and that is, what would you consider to be your digital thread? In other words, the defining interests and experiences that brought you to this crucial point in life.

[00:02:21]

Kence: That's a fascinating question; I resonate with the thread concept. My background is in mechanical engineering, so from a young age, I've had a passion for things that move, like jets, planes, and trains. Ironically, I began my career at IBM working with databases, and I've since dedicated my entire career to software development. Composabl marks the eighth startup I've been involved with and my second venture as a founder. The first five startups were in the ad tech and marketing tech sector, which provided valuable insights into Silicon Valley's operations and the acquisition process, experiencing four acquisitions firsthand.

Ad tech was my gateway to machine learning, as it was one of the earliest applications of this technology in an enterprise setting. My first venture as a founder was with Apptimize, where we focused on financial simulations and real-time trading in the programmatic advertising market. However, my focus has shifted towards industrial autonomy for the past seven years.

Joining Bonsai in 2017 felt like returning to my roots. Bonsai's focus on applying AI to industrial systems resonated with my mechanical engineering background, particularly in machines, manufacturing processes, and logistics. Additionally, the concept of machine teaching struck a chord with me due to my father's background as a teacher. I firmly believe that if machines can learn through machine learning, we should actively teach them rather than merely observing their learning process.

I've traveled extensively in recent years, visiting factories, oil refineries, mines, warehouses, and logistics centers worldwide. These experiences have allowed me to engage with executives and operators, gaining insights into how AI can be applied to industrial decision-making processes. My forthcoming book serves as a compilation of methodologies developed along this journey to guide the design of such systems.

[00:04:58]

Ken: We'll talk about the book in a minute because I thought it was very well done and suitable for the novice in this space, which is not truly what you think about with O'Reilly. I joke with people that the typical O'Reilly book is hitting assembly code on page 3, right? This one truly is, as you say, a novel and, one might even say, a manifesto for the use of autonomous AI in industry, so I have readily recommended this at every point. I'm glad you mentioned the background in ad tech because I did notice that, and indeed, the first startup you did in terms of Apptimize. You mentioned just a moment ago what attracted you to Bonsai. I'm curious because you joined Microsoft when they acquired Bonsai in 2018 and worked your way across their AI and research teams before leading autonomous AI adoption. I'm curious: what did you learn working at Microsoft, and truly, what are you most proud of?

[00:05:53]

Kence: That is a great question as well. Microsoft is a platform company. I learned two things about platforms that I think about in partnerships. First, the platforms. In all eight startups that I've been a part of, they were all B2B platforms, so I've never worked for a B2C company, never done a vertical solution, and so, I have lots of affinity for horizontal platforms that can tackle a market effectively, and that's Microsoft's bread and butter. That's what they do. I mean developer evangelism, building an ecosystem around platforms, building huge businesses out of platforms - I learned so much from executives, individual contributors, and scientists at Microsoft regarding what it means to develop and market a successful platform. The second thing is Microsoft is about partnerships. Even things that 'seem competitive,' like building Linux into Windows - there are so many things they do that show that they're really about open AI. It's a perfect, perfect, perfect example. But long before open AI, they demonstrated how partnerships can help you attack markets. It was a fun place to be for five years.

[00:07:06]

Ken: I had to laugh slightly at your comment about developer relations. We have another company, specifically in AI, where the founder came from a developer relations background. The company is Edge Impulse. The founder is Zach Shelby, the head of Developer Relations at Arm. We've had such good luck with founders who have that background in building up the ecosystem that we've now begun to look for founders with a developer relations background, so I'm glad that you mentioned that and your background as such. You co-founded Composabl in April of 2023, so as we were recording this, you could say, roughly a year ago. Together with Xavier Geerinck, your CTO, what inspired both of you to start Composabl?

[00:07:54]

Kence: It is this idea of a methodology for machine teaching. I hypothesize that any nascent complex high-value technology is likely best adopted through a methodology. There are many examples; you can even go back to your background, Ken, in industrial control systems. Ladder logic is the methodology for programming and building PLC systems for industrial automation. Excel formulas. Back in the '70s, in my first job, I certainly wasn't around there in my career, but my first job out of college was working at IBM Db2, the first relational database. Relational algebra was the methodology that helped people understand how to ask questions with databases. Before that, you had to write complex computer code just to ask a database a question. Now, with things like Excel formulas and a platform, methodology plus platform abstraction equals mass adoption or democratization.

One of the latest examples is prompt engineering. Prompt engineering is the methodology for working with or getting results, whether it's images or text out of a large language model. You could think of machine teaching as prompt engineering for intelligent autonomous agents. If you're trying to build these intelligent autonomous agents that can control machines and factories, control entire manufacturing lines and processes, and make logistics decisions, machine teaching is that methodology. Composabl is why we called it Composabl because the method teaches you how to break down complex tasks into skills. I don't think that's machine teaching; I think that's 'teaching' teaching. What teaching does is take complex things, whether chess or football or go or basketball, jazz music - all of those things get broken down into smaller component pieces by expert teachers and then given to students or learners to practice, and usually, the practice is sequenced. I mean, that's what a basketball practice is, or a chess coaching session is. Then, you master each of the skills by practicing and then you combine the skills. Composabl provides building blocks for engineers to build these agents from composite skills, which might be algorithms or heuristics, AI models, computer vision models, or even things like large language models. There wasn't anything out there, including Bonsai, which was a fantastic platform - but there wasn't anything out there that reflected what I had experienced in the effectiveness of this methodology. So, Composabl is the best representation that we could come up with of a reflection of this methodology in building software that helps people teach machines.

[00:10:48]

Ken: I think it's one of the areas that appealed to us in some of the early conversations we were considering an investment. For anybody who's been in a manufacturing environment, factory floor, control systems, engineers, etc., some of the newer - I'll call it advanced optimization technologies, including a lot of AI, traditionally require some form of data scientist to operate or data engineering; you hear a lot more relative to factory floors. I think what appealed to us is the fact that what you're attempting to do is go right to the people who know the domain rules better than anybody else, i.e., the people on the factory floor, and being able to train them to train these autonomous AI agents and thus skipping the - I'll call it the IT-ish data science aspect of it in terms of capturing these rules and everything else. Your book does an excellent job of describing how one would traditionally train these systems and how you can do it with your methodology now. You just described it as a kind of manifestation of the method, but you guys also described Composabl as a platform for building industrial string intelligent and autonomous agents. I will ask you to put on your master teacher hat and treat us all as laypersons because that's what we are. In layperson's terms, what does that mean?

[00:12:09]

Kence: Imagine you're flying a drone, right? I'll give you a couple of examples of drone flying that represent two potential ways that you could design an agent. First of all, there are entirely different skills in flying a drone. Takeoff and landing are two very, very, very different things. Avoid obstacles as you're flying around or how to get from point to point when your battery is running low - those are all very, very, very separate skills. Any learner, whether a machine or a person, gets confused if you try to learn all the skills simultaneously. If you try to learn takeoff and landing and avoid obstacles simultaneously, learning will take a long time, and you'll develop bad habits. A teacher breaks those skills down, practices them separately, and constructs scenarios where you want to practice them separately. For example, in a windy scenario, you need to practice landing. But in a super windy scenario, you might not have to practice takeoff for a plane or a drone because you won't be allowed to or shouldn't; it's too windy. But you're stuck up there if it becomes windy while you're flying. You have to learn how to land in windy scenarios. You have to learn how to avoid different types of obstacles. In practice, you're going to be subject to various scenarios. In Composabl, you can easily create each of these skills. Okay, I want to make a landing skill, a takeoff skill, and an avoiding obstacle skill. Then, you can use a no-code interface to drag and drop these things to arrange them because sometimes skills are orchestrated in sequences. I mean, takeoff happens before landing in all cases. But sometimes, the skills are organized in more of a hierarchy.

Most of the drones that folks use are - you're not directly controlling the rotors; that's usually cheaper drones. My son used to have one that was very hard to control because it's tough to control all four of those rotors directly, so usually, in the same way as there is an effect or a control system, there's an abstraction layer. There's a PID controller that's executing. You're telling the drone which direction you want to go and what should the pitch, yaw, and roll be, or like, this is where I want to go. Then, the PID controller underneath moves the four rotors to get you there. That's a pattern I call the 'plan execute pattern.' I've already given a couple of different design patterns that show you how you could teach skills. You could teach skills as plan execute, tell us what to do and where to go, and then figure out how to control the drone and how to get there, or you could do it with strategies or skill sequences - take off first, then fly and avoid obstacles, and then land.

[00:14:50]

Ken: What have been some of your early wins at Composabl?

[00:14:54]

Kence: My first test for Composabl involved handing out our Python SDK for creating intelligent agents to a few system integrators, such as RoviSys' industrial automation system integrator, Wood Group, Decision Lab, and others. I knew they could quickly determine whether we could build the kind of agents they knew were performing and whether we were truly ready for prime time. As a result, we signed five system integrator partners who are now reselling the platform and actively using it.

I was at RoviSys last week, where we conducted a hackathon and hosted a luncheon on this topic. It's delightful to work with engineers. While a few computer scientists are among them, the majority are mechanical, industrial, and systems engineers. Teaching them how to use our platform is very intuitive for them. Engineers are accustomed to working with building blocks and excel when provided with the right tools.

We're particularly proud that two Fortune 500 companies are using our platform to build agents to control high-value processes. For one of them, it marks the first time they make a process autonomous—typically taking over ten years to master. Seeing what our customers achieve with our platform is incredibly rewarding.

[00:16:06]

Ken: That's great. Wow, you guys accomplished quite a bit in a year, another reason we were excited to invest in you. Perhaps this question focuses on process and control system engineers who want to learn the space. How do new clients engage Composabl to solve industrial AI challenges?

[00:16:27]

Kence: I am genuinely proud of and delighted by witnessing people engaging with our methodology. When we launched our beta last August, I began reaching out and connecting with individuals. Among them were representatives from a major chemical company and one of the largest consumer packaged goods (CPG) companies globally. In both instances, I inquired, 'Do we know each other? How did you join the beta?' They responded, 'Oh, you don't know me, but I know you. I read your book.'

Similarly, several professionals from System Integrators and other organizations mentioned, 'Oh, I took your course.' I offer two courses on Coursera and teach two for A3, the Association for Advancing Automation. Seeing individuals becoming interested in the methodology through these educational resources is gratifying.

A mechanical engineer named Andy Lomizak manages an industrial engineering System Integrator practice. After completing one of my courses, he shared, 'I thought I would need to pursue a PhD or a master's degree in AI to become involved, and I felt excluded from the process. However, after taking your course on AI for industrial automation, I realized I can engage now.' I find such feedback incredibly rewarding and significantly contributes to building our pipeline.

Another avenue for building our pipeline is through our partnerships with system integrators. They collaborate with top-tier chemical companies, oil and gas companies, mining companies, and CPG companies. We believe that by empowering them, they, in turn, will empower us in the marketplace.

[00:17:53]

Ken: It's very much akin to your earlier concepts of developer relations. In this case, you'd call it engineer relations and the ecosystem built around it. What I appreciated about your book was, given my background as a former control system engineer, Process and Instrumentation Diagrams (PNIDs) or, I'd say, when looking at PID loops back then, programming those required quite a bit of skill, at least a good understanding of the math behind it. Yet, I appreciated that you drew analogies early on as a basis for moving toward more advanced optimization. After reading the book, it felt like it was written for me, or at least someone with my training, so you're certainly making it relevant in that regard.

I have a contextual question for you: Momenta has invested in industrial optimization and AI for over a decade. However, it's interesting to note that few of these systems, at least until recently, have been used in autonomous or what we might call closing-the-loop applications. What are the key motivations for many companies now finally considering this?

[00:19:02]

Kence: That is a beautiful question. First of all, there's a narrative with some truth to it. Still, I will call it a false narrative that says to industrial companies that the first thing you've got to do to achieve digital transformation is implement these new IoT systems to store all your data. Then, you can start analyzing the data and creating some insights. Then, you may begin to build predictive models, make predictions, and perceive things. Of course, you'll need many data scientists who are not part of your organization, and then maybe you can optimize something or make a real-time decision. That is a false narrative for industrials in a couple of different ways. First of all, historians have been collecting and tracking industrial data in factories for decades, so it is not true that without some new IoT system, you won't have any data to analyze. Now, with things like cameras, microphones, new kinds of sensors, IoT platforms, and data platforms that consolidate data in one place, there are certainly lots of legacy systems where data is siloed - all that's true. But the narrative that you need to start fresh because you don't track anything is false. The worst part of the narrative, though, is that you'll be able to control things sometime later after you get very advanced. The PID controller was invented in 1912 by the US Navy, so every factory has control systems in it, period. There's not a single factory without a closed-loop control system in it. What I found when I first started back in 2017, making the rounds and talking to folks, was engineers don't care about the acronyms: ML, AI, DRL, whatever; they see each new technology as a potential new tool in the toolbox, and so unlike AI research, which tends to have a one-track mind, and say, like in the '80s, expert systems are all of AI.

In the '80s, the public opinion about expert systems was the same as we currently have about generative AI. It will replace doctors; an expert system may become conscious someday. It's in the press recordings; you can see it. Then, of course, the same thing was true for neural networks, supervised learning, and deep reinforcement learning. This AI is learned by practicing and mastering the game of go and beating the best go player in the world and now, generative AI. Engineers don't have a one-track mind like that. Engineers have hammers, wrenches, and screwdrivers; you need them all. It's not like when the wrench was invented, you didn't need a hammer anymore. I mean, there are still nails, and there are still screws, so it's about different decision-making characteristics. If you take an expert system, which we were talking about earlier, an expert system is fantastic. It's fantastic at capturing existing expertise. It can be a pain in the neck to maintain all the rules and keep the rules updated, and it can be very rigid. It certainly doesn't learn independently, but, for example, safety-critical procedures might be the best way to decide. A control system that uses math to make its decisions is deterministic and predictable - it's the absolute best thing at low-level control, which is what PIDs typically do in factories. But if you don't have the math to describe the complex and nuanced situations, which you often don't, especially in chemical processes, then it's not going to make the best decisions, and so on. It's not about the best algorithm and the new algorithm to rule algorithms, it's about how you combine the right tools in the right ways.

[00:22:35]

Ken: Man, you had so many great points; I love the one about PID loops being closed-loop control, which is true. We've taken to calling this whole space just optimization, industrial optimization. Of course, given all of the hype around generative AI and the reawakening of AI in some sense, we've certainly tried to tag onto that term. I was at the Hannover Messe last week, and Edge Impulse had a booth there. We did a tour bringing a bunch of BofA analysts through, and they described themselves - Edge Impulse - they were presenting as real AI versus generative. To set the stage right off the bat, we won't talk about LLMs here, right?

[00:23:18]

Kence: I often used the term "useful AI" for that.

[00:23:21]

Ken: I like that. I like that. If there was a common theme last week at the show, it was - I'll call it, certainly, generative AI, but optimization. People were talking about various forms of AI, and one of the drivers for it is just your labor. Now, especially knowledgeable people who understand the domain rules are leaving many times with institutional knowledge as that whole workforce begins to retire, and the younger generation doesn't necessarily come in. In some sense, you have to capture that knowledge, and you have to be able to capture the skills that go with that as well, and I think that's what makes what you do so beautiful because you're helping to train a whole new generation. You're helping to train the older generation going out, and you're codifying that into intelligent agents, right?

[00:24:10]

Kence: This is my favorite aspect of intelligent autonomous agents. I mean, the word autonomy tends to suggest that it's doing something on its own. People think of people being replaced, but out of those 200 agents I've designed, 65% or 70% of them were intended specifically to help less experienced operators do a better job or give the expert operator the time back to go do something else that they need to be doing. Digital transformation, which I think we are 10 or 15 years in, you'd be better to tell me this, and then I would tell you how long digital transformation has been happening, it's foundational, indeed. But it's mostly about storing data and documents. You don't transform or revolutionize an industrial business by storing data and documents. That may be a controversial statement, but if you think of it in the context of industry runs on these crucial high-value skills. Max Petrie, who used to run the R&D Snack Food Division for PepsiCo, PepsiCo makes Cheetos and Doritos and Fritos and a bunch of yummy stuff. I asked him, "What's the most important part of your business?" He said, "The most important aspect of our business is the skills of our operators." Even before the Great Resignation, those folks were retiring. I've worked with processes where folks said, "When the expert gets sick or is on vacation, the throughput on a particular line or system goes down by 40%." Folks are getting called out of retirement just to come in for a few days and solve rare scenarios that haven't come up in a while. Intelligent Autonomous Agents are the first opportunity I ever know about to store these high-value skills, and it's not to replace folks, it's to preserve these skills. Imagine you've got a database of your 10 or 50 most high-value skills that you can use as a colleague, a coworker, a co-pilot, or whatever you want to say. However, you want to say it - to your experts and novices, and you've got a training program for your less experienced operators.

[00:26:07]

Ken: We often talk about automation to autonomy, leveraging a term that I think the CTO at Rockwell coined. In that spectrum, going from automation to truly autonomous operations, we have seen that augmentation ends up being a middle state. Exactly what you're saying, where you're helping operators, you're assisting maintenance people, etc., to better operate the process as you're moving to close more and more loops, if you will, control over time. I'm glad you brought that point up. I promise everybody that this is my last pitch for Kence's book, "Designing Autonomous AI: A Guide for Machine Teaching." Again, it's an O'Reilly book. I would recommend this as a primer on industrial AI, but you mentioned some other resources a moment ago, including your Coursera work. What would you recommend to potential clients, users, and people who want to learn about adopting optimization or useful AI in the industry?

[00:27:06]

Kence: Yeah, I think the book is certainly a commitment, but hopefully, I've made it an easier and enjoyable read. The courses can be a great place to start, especially the Coursera courses, because you can take them in bite-sized chunks, and they're video-based. Each course has 20 videos; you can take things at your own pace. In the future, I will be writing more things that are even smaller and more bite-sized, whether it's white papers or articles and things like that. But there's a lot of work to do regarding machine teaching. That's the fun but challenging part of evangelizing a methodology is it's a teaching exercise. Not to make a pun with machine teaching, it is a teaching exercise where we have to teach the methodology. Now, the good news is, and actually, this is another thing I learned at Microsoft where I was talking to someone about this framework that's called the Cloud Adoption Framework, which was invented, as I understand it, by AWS, which would make sense since they invented infrastructure as-a-service Cloud computing as we know it. This person at Microsoft explained to me that if your methodology is successful enough, it disappears because it kind of becomes the fabric of how things are done. I asked him, "Well, why do I never hear about the Cloud Adoption Framework?" Because it's just how people use the Cloud. If we're successful- which will take a lot more work and passionate evangelism- then machine teaching will be how folks design industrial control systems in the future.

[00:28:35]

Ken: Well said. In closing, I always like to ask I'm curious about how you maintain your edge as a leader. Do you have any recommendations you would like to highlight for the listeners?

[00:28:47]

Kence: I'll just highlight one. Because of my background, which often involves being in the minority, not just as a racial minority, but certainly, it was in my childhood - being a very technical person among salespeople and being a very sales-y person among technical people, there are all sorts of situations where I've been in the minority. It makes you approach things from different perspectives. A Nobel Prize winner in chemistry was asked - a Japanese professor asked, "What's your secret to innovation?" He said, "Coming at a problem from multiple different disciplines." He said, "You have to understand multiple disciplines, and where they intersect and come together is where innovation happens." That's one of my secrets: you come at it from a teaching perspective, you go at it from a programming perspective, you come at it from an engineering and industrial controls perspective, and you can come up with some innovative perspectives.

[00:29:35]

Ken: Absolutely. I've heard people describe it as pattern matching, right? Look across these different disciplines to match patterns between all of those, and that describes, I would say, your average target controller process engineer because you have to master a lot of different domains and disciplines to run and optimize a manufacturing process ultimately. Again, I promised, but I have to do it one more time. The book does an excellent job of capturing that kind of multifaceted aspect of that. Look, Kence, thank you for taking the time and sharing these insights with us today.

[00:30:11]
Kence: Of course. Thanks for having me, Ken. It's great to be part of your portfolio.

[00:30:15]

Ken: It's great to have you there, and this is only the beginning of how we will begin to promote you. We're going to be very excited as we make the announcements coming up. Again, this has been Kence Anderson, CEO and co-founder of Composabl. Thank you for listening, and please join us for the next episode of our Industrial Impact podcast. We wish you an impactful day. You've been listening to the Momenta Digital Thread podcast series. We hope you've enjoyed the discussion, and as always, we welcome your comments and suggestions. Please check our website at momenta.one for archived versions of podcasts, as well as resources to help with your digital industry journey. Thank you for listening.

[The End]