← Using AI at Work: AI in the Workplace & Generative AI for Business Leaders
Using AI at Work: AI in the Workplace & Generative AI for Business Leaders
103: Using AI in Manufacturing: Generative vs. Predictive and Autonomous AI with Bryan DeBois
00:52:07
Loading summary
Transcript49 lines
- [00:00]
Chris Staigle
Most manufacturers are chasing the wrong kind of AI for knowledge work is not the same thing as AI for operational work. So the listeners, what do they need to understand about these kind of two paths of AI?
- [00:13]
Brian Dubois
The client says we've got this mandate around AI and what they're thinking is they're thinking about the generative AI and that's what my clients are typically thinking that we're going to be talking about. And I immediately pivot and I pivot to analytical AI, predictive AI, computer vision, and of course autonomous AI.
- [00:30]
Chris Staigle
So what would be kind of like the clearest distinction between what's possible with a GPT Pro account or a Claude Pro account versus what you guys are doing at Rose?
- [00:39]
Brian Dubois
Unless you're comfortable with your surgeon stopping mid surgery and going and checking ChatGPT for what the next step in the surgery is, if you're not comfortable with that, then you shouldn't be comfortable running it on the plant floor. This is still considered a high risk environment.
- [00:52]
Chris Staigle
Brian Dubois is a leading voice in industrial AI and manufacturing automation, helping companies understand why factory floor AI is not the same as ChatGPT. With decades of experience in industrial systems, he educates manufacturers on using predictive analytical computer vision and autonomous AI to improve safety, quality and operations.
- [01:13]
Welcome to Using AI at Work.
- [01:15]
I'm your host Chris Staigle. Each week we'll be learning how today's business owners, entrepreneurs and ambitious professionals are getting more done with smart use of tomorrow's tech.
- [01:25]
Let's get started.
- [01:26]
Right now, every business leader is asking the same question. What are we going to do about AI? If this is you, chiefaiofficer.com has the answer.
- [01:34]
We give you a simple path forward
- [01:36]
where we provide executive and team training so your people know exactly how to safely use generative AI in their day to day. We also manage the deployment and implementation to make sure tools actually get adopted and deliver results. And we'll also guide company wide transformation so AI becomes part of your operating system, not just another shiny object. The companies that act now will increase productivity, cut costs and grow faster than their competitors. Those that wait will get left behind. So if you want to make AI work in your business, visit chiefai officer.com and see how we're helping companies of all sizes finally get results from AI.
- [02:13]
Hey everybody and welcome back to another episode of using AI at work. this point we're well into the hundreds of episodes that we've done, which is pretty fantastic and thank you all for the support that you've given us throughout all of this Today it's going to be a little different spin on AI. Normally when we do this podcast, we're talking about AI's application to knowledge work. You know, a lot of times, well, what about the field? What about the factory floor and those sorts of things? And that's not really an angle that we cover a lot on the podcast. So today let's fix that. So our guest today is Brian Dubois. He's a leader in industrial AI at the company Rovisys. Now this is a company that's got 2,000 engineers across 22 locations working on something that is not your chat, GPT, query and response. Their core lane is AI for manufacturing operations, analytical AI, predictive AI, computer vision, autonomous AI, and even maybe not optimus robots, but cobots. I mean, he's literally surrounded by AI enabled robots right now as we're doing this thing. The main, I guess the gist of this conversation today is that most manufacturers are chasing the wrong kind of AI. They have been to the industry conference, they've listened to a podcast, and a lot of them are thinking that AI for knowledge work is what they need and AI for knowledge work is not the same thing as AI for operational work. So, Brian, before we dig into it, what do you think that the listeners today, considering that we've got a broad spectrum of folks listening, what do they need to understand about these kind of two paths of AI?
- [03:59]
Brian Dubois
Yeah, and thanks for having me, Chris. I think that the way you framed it is exactly right. It's when I sit down with clients and we initially start this conversation around and the client typically says, look, I want AI. We've got this corporate mandate. And of course all my clients are manufacturing and industrial customers. So the client says, we've got this mandate around AI. Now I don't even know. They'll say, I know how to spell AI. And that's about it. And so where do we even start? And what they're thinking is they're thinking about the generative AI that you alluded to. So for some of your listeners, what is generative AI? So it's this umbrella term and it includes all the LLMs, which are the large language models, which are like examples would be ChatGPT and Grok, but then also all of the image and video generation tools all is under this umbrella of generative AI. And that's what my clients are typically thinking that we're going to be talking about. And I immediately pivot and I pivot to the types of AI that you outline there. So analytical AI, predictive AI Computer vision and of course, autonomous AI. And we can talk about any and all of those today. But those are the. That broadly, those are the types of AI that we're using on the plant floor. Those are the ones that we can operationalize today that I have no qualms, I've got no reservations about putting onto a plant floor because I know that any one of those can have a huge impact today on the types of operational problems that we face on the plant floor.
- [05:30]
Chris Staigle
So some of the things that you talk about generative AI can do some analytics or can do some prediction based on external stuff. So what would be kind of like the clearest distinction between what's possible with a GPT PRO account or a Claude Pro account versus what you guys are doing at Rovisys?
- [05:50]
Brian Dubois
Yeah, so there's kind of three significant limitations today to generative AI that really kind of disqualifies it for the type of work that we do on the plant floor. So we can recap those three limitations real briefly. So the first one is hallucinations. Most people who use generative AI on a regular basis are familiar with hallucinations. It's when the generative AI just makes stuff up out of whole cloth, just makes things up. And I'll give you an example. So this is back from May of last year. The Chicago Sun Times ran an article and some people may have seen this on X. They ran an article and it was kind of a fluff piece. It was the summer reading list for 2025. For the summer of 2025. Here's the 15 books that you should be reading, right? And they ran this article and it was in their Sunday circular. And here's the problem with the article. There were 15 books listed. Only five of those books actually exist. The other 10 were completely made up, didn't exist. And, you know, the Chicago Sun Times was obviously very embarrassed. And so they did an investigation and they discovered that, yes, the author of the article did admit that he had used generative AI to generate the article. He didn't bother to check that any of those books existed. His editor didn't bother to check that any of those books existed. And then they run the article. Article. And this is the problem when we just kind of trust this AI kind of blindly. And we're still in the early days. And that happened. And it was out there and it took less than a day for someone on X to post it. And they were the laughing stock in Chicago sometimes. Actually, my brother lives in Chicago. And I said, hey, get your hands on one of these articles because I want to hold it up when I speak at trade shows and stuff like that. I want to hold this up as a prop. And immediately the Chicago Sun Times had gathered, within a couple days had gathered all the copies and destroyed them all. When that happens kind of out in the real world, we can all laugh about it. And if generative AI generates some funny little marketing blurb or crazy email, we forward it to each other and we laugh about it and then we go in and fix it. The problem on the plant floor is that this is still considered a high risk environment. So if you give the wrong information to the wrong person at the wrong time, you could kill somebody. And sometimes it's not that dramatic, but you could blow something up, you could break something. There's a lot of issues. You could destroy a whole batch, which in some cases could be hundreds or millions of doll dollars, you know, if youif you ruin that batch. So yeah, we still consider that a high risk environment. So here's what I typically say to people. Unless you're comfortable with your surgeon stopping mid surgery and going and checking ChatGPT for what the next step in the surgery is, if you're not comfortable with that, then you shouldn't be comfortable running it on the plant floor, this is still considered a high risk environment. So hallucinations is the number one thing. But there's two other limitations that are worth talking about. The second one is the lack of reasoning. Apple did two landmark studies where they proved once and for all that GPTs LLMs are not able to reason. They can't reason, they don't understand causal effects, they don't understand if then types of things when we think that they're reasoning, when we are actually ascribing that to them. There's no underlying mechanism. Again, Apple proved it. There's no underlying mechanism where they actually are reasoning about things than then understanding that causal effect on the plant floor. Again, those are critical types of decisions that we need. And you have to understand that if I make this change that this is going to be the downstream effect of that. So that's the second thing. And then the third thing is in this one is I don't know that you'd ever be able to fix this, but it kind of eliminates our ability to use it. Is that it's non deterministic. Yeah. So if you ask Generative AI five times the same question, you will get five different answers. Right. And that's just kind of inherent in the probabilistic nature of how it answers questions. Unfortunately, on the plant floor, we can't have that either. Like 100 times out of 100, given the same inputs, I have to get the exact same output. I can't have that variability. So those three things in particular, at least today, disqualify generative AI for us to use to solve operational problems on the plant floor. I typically tell customers, if you want to use them to solve knowledge problems in the carpeted space within a manufacturer, by all means use that. That's what it's good at. But for the operational problems that we face on the plant floor, it's not appropriate, but. Okay, so then what is? Well, we have all these other types of AI that we talked about. Those are all great uses of AI on the plant floor to solve those problems. And honestly, before. So we started this. I've been with Rovisys for 26 years. We started this industrial AI division in 2019. So we were years before the release of ChatGPT and we were still doing AI. So all we're saying is there's lots of proven AI that's been around for decades in our industry that we can rely on, that we don't have to worry about these limitations of generative AI.
- [10:57]
Chris Staigle
So we're talking about obviously your experiences in the manufacturing environment. But the same risks are exist in construction if they're. They aren't clear on what they're doing here. Life sciences, where they're. Some of our clients are developing cancer drugs. Right. Like deterministic needs to be part of that formula logistics, obviously sending stuff to the wrong place like this is. So if you're listening to this and you're not manufacturing, this applies to you. Now, when these manufacturers are reaching out to you or they're seeing you on stage and they're saying, oh, yeah, yeah, yeah, we've been thinking about using AI. What do they actually mean?
- [11:37]
Brian Dubois
Yeah. So typically they have some ideas about what they would like to do on the plant floor, and we try to corral them into some kind of swim lanes that make the most sense. So the first one would be Advanced Process Optimization. So they would like to make more with less. They would like to increase their throughput. Throughput, decreased scrap, increase yield, increased quality. Those are the types of metrics that they're looking at doing utilize their assets better. A big part of that and what's driving that is this expertise loss. People have called it a silver tsunami, this loss of primarily baby boomers out of the working industry. And they are taking with them all of that expertise It's a little scary right now in the manufacturing world. Go into a manufacturer and I'll say, well, who knew how to do that? Well, it was Bill, but he left six months ago and all we have is the pile of papers on his desk. As those baby boomers have left the industry, they've taken that expertise with them. Many of my clients for process optimization are primarily looking to in some ways just get back to how good they were able to put out product five years ago. They're just trying to get back to that because all they have, the incoming workforce, they have high turnover. A lot of people just treat a factory job as like an entry level job, like a fast food type of thing. They're there for six months or less and then they're off to the next entry level job. A lot of folks are not, this is in the US workforce, a lot of folks are not really interested in spending the next decade of their life learning kind of the fundamentals of this process. How can I get better results out of it? They're just not interested in that. The clients are like, look, I still have to make product. Everyone expects these products to just appear on the shelves. They don't care how they get there. And so they're looking to AI. So advanced process optimization real quick, go through the others energy reduction, especially with a lot of the things happening in the geopolitical space right now like energy and both with data centers kind of sucking up all of the electrical energy. But then we have oil and gas type of shortages. So energy reduction and optimization there product quality again. I am very nervous that we're going to start to see major quality issues coming out of manufacturing because of this loss of expertise. How can we improve quality then issue detection, that's going to be things like asset health. How can I make sure that this asset, this piece of equipment, multimillion dollar piece of equipment stays running in top shape. Then also things like safety and how can I make sure that people don't venture into this space when a crane's moving? Those types of. So those four general areas are where, when customers come to me, those are typically the types of questions that they're asking about.
- [14:39]
Chris Staigle
Yeah, you know, you bring up something that is, I'm hearing pretty universal with clients that we talk to across all industry. And that is if like okay, in the knowledge work, Dario Amade from Anthropic said that, you know, soon 50% of entry level white collar jobs just won't exist. AI will be able to handle those. Right. Something, you know, that you're experiencing, but probably for a different reason in the manufacturing side of things, but the same result. The people at the top are saying, well, we used to groom these individuals throughout the course of their career so that they could take my job. And if we're not getting those people, whether it's manufacturing or the knowledge work, what are we going to do when I retire? So what are you suggesting or seeing or helping people plan for in the manufacturing environment to address that?
- [15:27]
Brian Dubois
Well, and so there's a number of things and there's no silver bullet, I'll tell you right now. Right. But there's a couple different things you can do. So if you're a manuf manufacturer and you're faced with this labor shortage, which they all are right now, they all are. In fact, it's easy for me to answer the question that I've been answering for seven years now. Are you putting people out of work? No, they can't find enough people. They're not firing people.
- [15:50]
Chris Staigle
Interesting.
- [15:50]
Brian Dubois
That's not the issue. But if I'm a manufacturer and I'm faced with this shortage, there's a couple things that I typically kind of advise them to do do. One is, is that those manufacturers that they need to change the perception of manufacturing being kind of like this dingy, dark factory kind of like. And some of the leaders in our industry are already starting to do that. Right. So they're, they're, they're improving that aspect of it. They're focused on culture. They're, you know, some of the things that maybe they had been able to forget about in the past, they're now, they're now focusing on those things around, you know, advancement and, you know, investment in your people. And those are all great things. Those are important. But then the second thing though is that in the adoption of AI, there's another aspect to it and that's that they are able to project themselves as a forward technology thinking company. Right. So a company that uses technology and one of the things that, one of the reasons why they're seeing high turnover is that these are folks who grew up with an iPad. I mean, my kids had an iPad as soon as they were born. Right. Like they've had technology in their hand from the beginning. And so when they walk into a factory and they're using green screens, they're using really old technology, you know, they're unimpressed. And so if we can, if those manufacturers then can portray that, you know, we are a forward thinking company and we leverage technology here, that's something that they can use because they are effectively competing for these folks, they're competing for these resources and, and those entry level folks actually have a lot of leverage and a lot of different places they can look. I don't think I understand this concern about AI replacing the white collar worker. I would say, I would still consider entry level factory jobs as a blue collar job. I would say. And so I would not say that this is something. In other words, let me rephrase it. If you're looking, if you're an entry level person and you're looking for a place to start your career and build your career and you're worried about the threat of AI, go into a factory and go to a manufacturer. Because I am telling you, and I've talked to manufacturers, this is what I do every day. I talk to manufacturers and they've reiterated this back to me, that if you go in there and you're hungry and you just want to learn and you want to work and you're excited about the possibilities there, you can accelerate through these organizations very fast. It will not take you long to be put into a supervisor position, to put in a management position. You could build a career very easily at these manufacturers and they are excited to bring in that talent. So I would say that if you're looking for an industry that's relatively immune to some of this, this AI threat, that manufacturing is definitely one of them.
- [18:44]
Chris Staigle
So Mike's really big on exactly what you're just saying. If you're going for the traditional office job and that sort of thing, AI is a threat. And with, with that job comes the debt from your college degree and all that kind of thing. Right. So he's really pushing on. Like a story that I saw just yesterday is that I don't know what the role is, but electricians, individuals who are working in data centers are starting out north of $200,000 a year. They're supported by AI, sure. But AI is not, AI is not doing the job. So this, this is falling in line with this kind of, this paradigm of the jobs that used to be looked down on perhaps. Right. As manual trades are actually the safer choice for a career path.
- [19:39]
Brian Dubois
And we really did a disservice.
- [19:41]
Chris Staigle
Very interesting.
- [19:42]
Brian Dubois
You know, over the last 20 plus years we did a disservice by looking down on those jobs. And there were countries that didn't. There were countries that if you went to college, that's fine, that was your path, but that wasn't like somehow portrayed as the right path. Like if you went to A trade. They were like, great, that's awesome. Like, as long as you're a productive member of society, nobody cared where you ended up. And we did a disservice to our young people, to our country, by taking this weird stance that everyone had to go to college and funneling everyone into college. That was definitely a mistake. So I think there's a chance to rebound here and fix that. But AI is kind of forcing our hand either way, so.
- [20:23]
Chris Staigle
Well, now let's talk about this. You know, we've got the issue of they can't find enough people. An industry that we do a lot of consulting in. On the AI side, on the knowledge work side, is construction. And, you know, we kind of have this vision of, of Optimus showing up and starting to move bricks. So you don't need the people anymore. You've got the robot that can actually, the bipedal or whatever, that can, you know, start moving the bricks or running the lines and that sort of thing. Why isn't that a viable option for manufacturing?
- [20:55]
Brian Dubois
So, okay, I've had to comment on this a lot lately, obviously, with all the focus on it. And I always qualify by saying I'm not someone who makes a business of betting against Elon Musk. It seems kind of crazy to bet against Elon Musk. However, I am not convinced of this bipedal humanoid robot angle. I don't think it's the right angle for this, and I don't think that that's really going to be what brings about this robotic future. And let me explain a couple thought experiments. Why. First off, why is. And I recognize that the company makes Roomba, just went out of business. But why is a Roomba or a Roomba, like robotic vacuum? Why is it shaped the way it is? Why is a robotic lawnmower shaped the way it is? It doesn't look like a human. It wouldn't be recognized in any sense as a humanoid robot. But it's made that way because that fits the purpose of what that robot is made to do. When you look inside of a construction site and you look inside of a factory, we have robotics and we have big machines, automation, and think cranes and things like that. None of them look like humans. The robots that we have in the factory, they have big safety cages around them. They move very fast, and none of them would be recognized as a human or humanoid. I don't know why we would make them as humans. It would undermine the whole purpose of why they're built to do what they do. And then the last kind of point here is if you're really, really committed to having humanoid robots working alongside humans on a factory that exists. They're called cobots. And I don't know a single customer that runs them. They've been around for 15 plus years in our. And they are exactly what you would think they are humanoid looking robots. They have all kinds of safety systems so that they can sit alongside of a human on a factory line and not hurt somebody. And that's important. And I don't know a single customer that has a cobot. They've existed for a long time and nobody uses them. So I'm not sure what the point is here. Where are we going with this? What are we trying to accomplish now? If you're trying to build something like Rosie and the Jetsons or something like that, that's going to live in houses and kind of work alongside, okay, that's fine. That's not my industry though. So I don't really know or care about that. But if you're talking about what you're talking about industrial construction and factories and stuff like that, I don't have any idea what the point of a humanoid robot is. And I've been in some advanced. I've been in the gigafactory, I've been in Amazon fulfillment centers. There are lots of robotics in there. None of them look like a person. And I don't know why you would make it look like a person. So, yeah, that's kind of my take on it. I don't know, I could be completely wrong on this and I'll have egg on my face. You sure? You know, Elon proves me wrong. But like, yeah, I don't, I don't get it.
- [23:51]
Chris Staigle
Well, I'm sitting over here visualizing Terminator as a Roomba.
- [23:58]
Brian Dubois
Well, to be honest, if you ever saw. There was a Black Mirror episode on Netflix called Metalhead and it was about those robot dogs. If we armed them and gave them autonomous ability to kill. And I'm like, yeah, that's scary. That's something that would be. But the Terminator, I mean, honestly, making the Terminator shaped like a human doesn't make a lot of sense. It's a lot scarier if it's smaller and it's more mobile and more agile. But anyway, yeah,
- [24:30]
Chris Staigle
I think that was probably the first episode of Black Mirror that I ever saw. So I want to go back to this concept of we've got. And this isn't just in manufacturing, this is in any industry, as I mentioned. And the knowledge work they're concerned about, hey, when I retire, who's going to take over what happens happens to the, the domain experience, the subject matter experience that, that I can't get from asking chat GPT a question or I can't get from sitting in a classroom. I actually have to have sat in those meetings and spent them, you know, burn the midnight oil on the floor, fix whatever those, those life experiences are that make somebody's judgment and discretion when it comes to problem solving in those environments better. So you mentioned that. Well, that guy, he retired six months ago and all we've got is a stack of papers. Are you, are you or does anybody at Rova sis or do you guys partner with anybody that helps these companies start to capture that information now so that they can perhaps use that to inform the robotics that are coming or the, the predictive AI or the. Is that a process that you're seeing happen where the generative AI is supporting what's ultimately going to be producing?
- [25:44]
Brian Dubois
Yeah, I think that there's still some things that can be done there and I think that we get involved in that. It's somewhat tangential because again we're typically focused on the plant floor, but we also have customers that are looking to capture that knowledge. And there are some forward thinking companies out there right now, vendors that are offering options and software that will start to marry the data from the plant floor to that more knowledge based information that's stuck in PDFs and SOPs and all of that kind of stuff, that type of data. There's some software vendors right now out there that are starting to build what they call an industrial knowledge graph. And so what that is is a combination, what they typically refer to IT as a combination of it, OT and ET data. So IT data would be that knowledge type of data. There's all those PDFs and all of that as well as things like maintenance records or knowledge about documents about how this machinery functions and what its operating parameters are. All of that type of data and being able to mine all that data, that's typically the IT data as well as data from more operational systems like SAP and those types of systems being able to pull that data, marry it to ET data. What's that? That's engineering technology. Think things like plms and those types of CAD systems and the systems that are either designing the equipment or designing the widget that you're trying to build, the automotive part or whatever. All of those types of systems are generally called engineering systems. So pulling all that data together and then relating it to the OT data. The OT data is the operational technology data. It's the data coming from the plant floor. And so you're building an industrial knowledge graph where you're relating all of that data together and you're saying that rather than in the past, we have a very myopic view, depending of, kind of the lens, we're looking at that myopic view of saying like, okay, this is the current operating parameters of that piece of equipment. Okay, but what about the maintenance information? What about how it's supposed to operate? What about the information from the supplier on the raw material? Material that's coming from SAP? Like there's all kinds of other sources of data. And so the industrial knowledge graph is an effort to pull all of that data together into a single place and then layer on an LLM that can then query that industrial knowledge graph. So the goal is, is, look, you may never be able, we may never be able to get rid of hallucinations. I think that they're probably inherent in the transformer architecture, which is the underlying algorithm under generative AI. And we may never be able to divorce ourselves from hallucinations. But, but at least maybe we've got the best chance of grounding it with this industrial knowledge graph. From my opinion, I think that that's the best way to ground it in. The answers that it generates would be to send it into that knowledge graph to get its answers. So that is to your question, that is the way that we're trying to solve some of these problems.
- [28:49]
Chris Staigle
So to help me better understand this knowledge graph, you're on X. Obviously Andrej Karpathy was creating some buzz about how he was using Obsidian.
- [29:00]
Brian Dubois
Obsidian.
- [29:01]
Chris Staigle
Are you familiar with what I'm. Where I'm going with? Okay, so Obsidian is a, it's, it sounds like it's, it's doing this, the, the knowledge graph. It's, it's taking in all this information, but it's actually creating an architecture that makes it easier for AI to, to access and consume. It's an architecture of the data, I guess. So, yeah, this is similar. So. And I'm also seeing more stories about, in some countries, they're having the factory floor people wear headgear helmets with cameras in it that are recording, you know, what they're doing so that they can synthesize that into something that AI can ingest. Is that something that we've seen?
- [29:42]
Brian Dubois
So I have not seen as much of that in the US if it's just a camera, I think that we're probably. Okay. One of the Things, if you remember like Microsoft Hololens and some of those efforts to make kind of a heads up display or an augmented reality display, there's been lots of efforts over the years to try to get that to land in US manufacturing. The problem is that it just becomes a safety hazard. You can't have anything blocking. You have to be. I do our training, our on site training to all of our new cadets at Rovisys. And I mean one of the first things I tell them is when you are in a plant, you need to have your head on a swivel and you need to be constantly watching, watching for yourself because nobody else is looking out for you. You've got to be watching out for yourself. The number one way that people die in a plant, it's not from some crazy chemical spill or some wacky thing, it's from forklifts. That's the number one way that people die or get injured in a plant is because of forklifts. And the modern forklift. I don't know if you've seen they have all kinds of protections now around it to try to protect pedestrians and things like that. But if it's a burden forklift, I mean the visibility is very low on that driver to be able to see. And so if you're walking around with this thing in front of your face, it's just so easy to step out in front of a forklift and that's the end of you. So yeah, most customers are not really looking at doing that. Now I have heard about that though, where companies are just more like a GoPro kind of on a headband and recording everything they do. I'm in favor of that. I think that's a really interesting idea. I don't know that we have the technology yet to really be able to work with that, but you can imagine we will. And maybe not that far off to actually be able to really process that and do some interesting things with it. I'm fine with companies trying to do that today to capture some of that data
- [31:37]
Chris Staigle
going back to. Because I know that this is going to be an issue, like you said, the Silver Tsunami, whatever it is, the individuals who built their career pre AI, where you had to be tactile, you had to still think about things, you had to still like use cognition. You weren't just able to go and hey, what's the answer, right? Like people can now with, with these models, what kinds of institutional knowledge are the hardest to capture? If a company, you know, they start to see some efficiencies and they start to cut that stack.
- [32:09]
Brian Dubois
That's a really good question. I don't know that I've ever been asked that. What kinds of expertise on the plant floor are the toughest to capture? I mean, I'm split on two. I'm going to go with. So it's either going to be maintenance issues. So how best to run and maintain this equipment is definitely up there. Having like, we're in a situation now, and I actually have a buddy that just retired from maintenance in the steel industry after 30 years. And, and he, and many like him can immediately turn around and go into consulting and make a good amount of money, just work part time. Because the expertise they've gained over all those years of how to maintain this equipment and especially a lot of these companies, yeah, maybe they bought it as kind of an OEM or off the shelf, as close to off the shelf as we get in our industry equipment. But they've customized it so much over the years that, that I don't even know how you maintain this equipment if you weren't the person or trained by that person who's been maintaining it. So I think maintenance is high up there, but then I gotta go almost equally as complex and difficult would be kind of the process itself. So many of these processes. A lot of people like to think that a factory is like, inputs come in and then machines do a bunch of stuff and then outputs go out. But the reality of it is, is that the quality of the output is in a lot of cases is directly correlated to the skill of the operator. And so kind of the black art of knowing how to get good product and good results out of this equipment is also, I would say, equally as complex and difficult to capture that expertise. That's that tribal knowledge that gets passed down from one operator to another, like, hey, if it's a hot day, you know, you, you have to know that you got to turn that knob down or else you will not get good quality. Like that's. Where is that written down? You know what I mean? Or if a. If it's rainy out, you know, you got to turn up the thing on here and you got to make sure you add more additive here or else we're gonna get gooey this out of there. Like nobody is writing that stuff down. So. Yeah, and. And then the question becomes, how do we capture that? And it's difficult. Like the way that we do it is, is we kind of have to catch those folks before they leave. And we basically go through a process called machine teaching where we say, train us like you would a new operator, and we capture all of that knowledge. And then we. But because it's not for generative AI purposes, we have to capture it in a way that can actually be consumed by the autonomous AI. So the autonomous AI actually, so we build it almost like a workflow, a decision tree, and then we hand that decision tree to the autonomous AI and it uses it then to train off of, along with a simulator. And then it trains in that simulator using that machine teaching so that it can learn to get better. That's really the only way we found how to capture that.
- [35:24]
Chris Staigle
So for the listener, you may think, well, I'm not manufacturing that process. That. That activity that was just shared with you is something you should be considering. How do I do that in all parts of my business, whether it's the floor or whether it's the office? Because that's exactly what we do. You want that person teach me, like, it's my first day on the job learning what you do. And if you can get that. That. That deep domain expertise, that person to. To document and codify what it is that they do, that will go a long way to you not finding yourself in a situation to where it's like, person retired. Oh, now we got to hire him as a consultant. 250 bucks an hour just to come to the job.
- [36:06]
Brian Dubois
That is not an easy process that I just described, because besides the hours that it takes to sit down with that person and have them teach you, then you have to observe them doing that. And I'll give you an example. So we had a situation where we were building an autonomous AI for a company that makes flat glass. Okay? So think flat glass is, like, used in windows and doors and, you know, big sheets of flat glass. And the trick about flat glass is that the glass is extruded. It's molten. And they extrude it onto a bed of molten tin. And then the glass actually will. Well, gravity will do all the work. It will flatten as much as possible just by floating on top of the molten tin. And then an operator will. They have these arms that come out, and they will actually hook into the edges of the glass. And then you can push and pull it like taffy. And that's actually how you make different thickness and widths of flat glass as you literally push and pull it like taffy while it's floating on top. It's really cool on top of this bed of molten tin. So we sat down with one of their expert operators, and we were trying to Automate. We were trying to build AI to help them automate the process of doing what are called changeovers. Now, changeovers is when they change that width or thickness of glass. And it's a very. It's a very difficult and complex process. And if you do it wrong, you will lose control of the sheet of glass and it will go crashing into the side of the bath of molten tin. And then real bad things happen. So. Can happen. So one of the things that they said, first things they said is, so we build a changeover plan. And it was. They literally write it in pen on a piece of paper and they say, okay, so to do a changeover, like, you may have to increase. Decrease the thickness for two minutes, and then you have to let the glass kind of rest and kind of reset, and then you can make another change. You can't make any quick changes. So it's a process to move to a different thickness or width. And so they said, we write it down. And so 10 minutes we're going to be at this, then we're going to change the angle and then we're going to increase the heat. Like all these different parameters. And like, we write this plan. Now, we said, once you write that plan, do you ever change, do you ever vary from the plan? No. Once it's written, it's written in stone. We follow it to a table, we write it. That's it. Right. Now, again, if all you did was the machine teaching you, sat down in a conference room and had them teach you that, that's what you would go and implement. So now we're watching over their shoulder and we watch a couple changeovers and they're going along and suddenly we see them scribble out a number and write in a new number. We go, you just changed it. You just changed it. And they go, oh, well, yeah, we had to change it because such and such happened. And so, you know, the real world happens and you got to react to. To it. We're like, okay, but we have to know that, that this process can change. When we built our system then we had to build it in such a way where it was dynamic. We had to build a system now. So instead of the old MapQuest that you would in the olden days, where you would actually get the directions and print them out, that's not going to work. We have to build it like a GPS where if the operator takes a wrong turn or the glass doesn't react in the way that they think it's going to, it's got to Reroute, like Google Maps does in real time. It's got to reroute. And so we had to know that so we can build our AI to be able to do that. Now, that is one part of one process, this manufacturing facility, right? It was. I mean, it's a whole effort to get that. Now that it's done, it's done. And you would never have to do that again, you know, at that, you know, we've captured it now for that manufacturer. But that process now has to be be repeated across every process. Everything that that company does and every company does. In order to make AI be able to do it, we have to do that process for every single one. And so to your point, that is the process. That's the effort ahead of us. That's the mountain that we have to climb to be able to get there.
- [40:01]
Chris Staigle
And for those of you that, again, aren't manufacturing, the same thing happens for that big spreadsheet that somebody maintains or running payroll, they're like, oh, I always do it this way. Except when, yes, you got to capture those things or else AI will be an additive, but not a replacement for that process. So here's a question. In your opinion, at least in the manufacturing environment, who owns AI?
- [40:28]
Brian Dubois
What's the role within manufacturers? I would say ownership needs to be shared between the plant floor and it. So they both play a role in this. Right. So when I started my career in manufacturing, it was still the Wild West. On the plant floor, if you needed to hook up two pieces of equipment, you ran to Staples and grabbed a hub and just threw it on the plant floor. Like, who cares? Like, nobody. You know, like, it was the Wild west. And that is just not the case anymore. And for good reason. Right? Like, it has brought a certain amount of governance and discipline to the plant floor, which is great. We needed that. Right. You can't have Windows machines that have any kind of Internet connectivity on the plant floor that are running Windows 98. Right. But we still see that now in a lot of cases, the machine is isolated so it can't get out to the Internet. So that's how they consider it safe. But even then, it's still like. So it has to play a part in all of this. And the other thing is that the AI models that we're building are typically being trained in the cloud. So immediately you're in the realm of it. We're training these models in the cloud, but so now we're in IT land. And so we're doing big data and we're Doing all the types of cool things that it likes to do then. But here's where we transition. Once that model's built in the cloud, we export it out of the cloud and we get it running on an edge device in the plant. Now OT has to take ownership of that. Now OT has to be the one that's responsible for the care and feeding of this AI model. And here's the other thing. And I always, I mean, we train our customers on this. You know, you're, you're adopting a puppy. Like this is an investment. Like this is something that has to be part of your standard maintenance. And every AI model drifts. It's called model drift. And it, it exists in every AI model where what that is is where the AI model starts to depart from reality. And so it can be just the machines just over time, it's going to change the way it operates just because of physics and stuff wears down and whatever. There's a lot of factors why, but AI models drift. So you're going to have to have a process. It doesn't have to be done probably more than once a year, maybe twice a year, but you have to have a process to retrain that model and redeploy on the plant floor. Because it's not just a set it and forget it kind of thing. You're adopting a puppy. So that's all part of it. So back to your answer. I think it's both. I think that it has to own it in the plant floor. What we call OT has to own it as well. It's not just an either or type of thing.
- [43:06]
Chris Staigle
Now, one of the things that we talked about that you seemed to leverage as a good example of what's possible with combining different AI approaches was Meta's Cicero project. Can you explain to the audience what that is and kind of out what the lessons I learned from that?
- [43:25]
Brian Dubois
I can. I'll try to make it as brief as possible. So when we talked about autonomous AI, and it's come up several times now, what is autonomous AI? So autonomous AI is built on a different underlying algorithm. It's not built on the transformers that generative AI uses. It's built on an underlying algorithm called Deep reinforcement learning, or DRL. And that DRL, where did that come from? That came from Google, DeepMind back in 2016. So some of your listeners remember AlphaGo and all of that, and what a big deal that was. So that didn't go away. It's moved into different industries and in particular it's moved into the plant Floor and we call it autonomous AI, but it's drl. Now, what makes autonomous AI different is all of those limitations that I talked about with generative AI. Autonomous AI doesn't have. So it's deterministic. It will give you the exact same answer 100 times out of 100. As long as the inputs are the same, it'll give you the same answer. It knows how to reason, so it can build long term strategy. It understands causal connections. And if, then, if I do this today, this is going to be the result tomorrow. It understands all of that and it doesn't hallucinate. If it doesn't know the answer, it's not going to make something up. It will say, I'm over my skis, which is what you want. Want. Like I'm out of my comfort zone, I'm returning control back to you like an autopilot in a plane. Like that's what you want, right? So that's. Those are the strengths of autonomous AI. And because of that, it's really, really well suited for the types of problems we see on the plant floor, those operational problems. Okay, so where's Cicero fit into all of this? So Meta set about to solve a problem. They wanted to build an agent that could play the game of diplomacy. Now diplomacy is, is this old board game. Now, people don't even play it in the board. They play it on the Internet now. But it's an old board game and it's a map of old Europe. And the idea behind it is, is that if you're going to attack, if you're one country, if you're France, you're going to attack this other country. There's a lot of negotiation that has to happen. Hey, I'm going to, I have to go through you to attack this other country. Don't attack me though. I'm not going to attack you. I'm attacking the country on the other side of you, right? And so then the other player, and there's a lot of backstabbing and stuff like that. There's a lot of negotiations. Negotiation. And that's why Meta picked this problem, because they were like, this is a great problem to attack. And so they used generative AI. They used an LLM to solve the problem of the person to person communication. So they use the generative AI for what it's good at, which is the natural language, which is the human and machine interfacing. Right? Like being able to win over a human being and convince them a lot of, of people who play against it don't even know they're playing against a computer. To win them over and to negotiate effectively in natural language with them, they use the LLM for that. But to win the game, they used autonomous AI. They used that to build the strategy because that's what it's good at. You can't use generative AI to win the game, but you can use autonomous AI to win the game. The Meta Cicero project was the marriage of both of those. It was a hybrid AD that used both generative AI and used autonomous AI both to their strengths. And to me, that, and that was something I had talked about for many years that I felt like that was probably the future was this hybrid AI between the combination of those two. When I saw that was very vindicating for me because I was like, okay, this is what I've been talking about. This is a perfect example of both of those playing to the strengths of those two prominent AI approaches and getting something better as the result of the combination of both those. So that's really. That to me is why I'm so excited about that and I hope to see many other examples of those two types of AI approaches being used together.
- [47:09]
Chris Staigle
No, that makes a lot of sense actually. Huh. Well, Brian, thank you so much. This has been, we've got manufacturing clients, but honestly we don't spend a lot of time on the manufacturing floor. We spend a lot of time, time, sure, with the knowledge workers. And I, I knew that there were huge opportunities for these clients, but I wasn't the person to bring that, that factory direction to the table. So this has been very informative for me. But I'm, what I've also taken away from this is you can call whatever industry you're in, it doesn't matter. There are similarities in how, how the people fit into this. The governance is necessary. I mean, even if you're, you know, we talked about probabilistic and deterministic outputs from the models and that sort of thing. I mean, best case scenario, you're always getting deterministic unless it's something like creative writing or image generation for, you know, ad creative or something. So understanding the, I think just introducing that, that's the reality for people. People, a lot of these listeners may not have understood that, oh, if I ask the same question to chat GPT, don't change anything. Copy, paste, copy, paste, copy paste five times. That's not going to give me completely different answers, but it's not going to be. I can't copy and paste and have a complete match of that output. That There is, there's going to be this, this license that AI has taken to create that answer. And as you mentioned, in construction, in life sciences, in manufacturing, we can't afford that. We need the determinism and that generative AI is not the resource. If determinism is the difference between an injury, a lawsuit, a deal, won or lost, you cannot rely on generative AI.
- [48:59]
Brian Dubois
That's exactly right. Yep. Yep.
- [49:01]
Chris Staigle
Big wins. Thank you. So, so for if we've got, and I know that we do, we've got clients who are in manufacturing that are and, and listeners who are in manufacturing that are listening to this and if they want to, if they said, hey, this makes a lot more sense than what I'm hearing from perhaps the consultant or the salesperson who's trying to sell us something and they want to like, stay on top of how you're approaching this and what you're seeing and thinking about the space. Where do they go to find more information about what you're doing?
- [49:31]
Brian Dubois
The easiest way is to follow me on LinkedIn. So just my name, Brian Duboisbier. Y n d E B O I S if you just search that on LinkedIn, I should be the first person result and follow me there. Every podcast that I've done over 35 appearances on various manufacturing and AI related podcasts and I post them every time I post them on LinkedIn. You'll see those there, you'll see where I'm appearing next to speak and then go to The Rovisys website, R-O-V-I-S-Y-S.com and find us there. And happy to have a conversation with anyone about this. And this stuff's all evolved very fast. So happy to talk to folks about all this.
- [50:10]
Chris Staigle
Thank you for that, that's very generous. And we're going to be putting those links in the show. Notes for the listener. Again, we're not like exclusively focused on manufacturing, but from what I've heard today, Brian's position on this is extremely sound and not new. As we mentioned, you know, they've been working on this well before anybody even heard of ChatGPT, so. So that's who you want helping you with these, answering these big questions when it comes to the manufacturing side of AI meets manufacturing. So thanks Brian for taking the time and dear audience, thank you so much for listening to yet another episode of Using AI at Work. If you get value out of this, forward an episode along, leave a review. Anything you can do to help us get this in front of more people would be fantastic. As you know, there's, there's no charge for this. We don't really have any advertisers or anything like that. We do this because we feel that this is probably one of the more important topics for everybody to be up to speed on. And we do our best to make sure that you leave each episode with a little more understanding of what, what this means for the world, for the future, for your future. So again, just thanks for joining us on the episode. Thanks, everybody. We'll see you on. You're welcome, Brian. Thank you so much for being here. We'll see you on the next episode, everybody.
- [51:49]
Brian Dubois
Sam.