Jimmy Carroll: [00:00:03] Hi everybody and welcome to the Manufacturing Matters podcast, where we aim to get to the heart of the technologies and trends shaping the global manufacturing industry. My name is Jimmy Carroll. I’m the vice president of operations at Tech B2B Marketing, and I have the pleasure of being joined by our guest Frederick Reinink, who is a senior business development specialist at Musashi AI North America. And of course, my good friend Winn Hardin, managing director of Tech B2B. Today we are going to provide an insightful — Frederick, I should say, will provide an insightful look at some of the real-life applications and challenges in AI. So we’re really looking forward to that. Before we jump right in, though, Frederick, I just want to thank you for taking the time today. Really appreciate it.

Frederick Reinink: [00:00:39] Thank you for having me, Jimmy. Happy to speak about some of these issues around AI and how to deploy AI.

Jimmy Carroll: [00:00:45] Well we’re looking forward to it. I’m sure this could go quite a long time given the amount of questions I have about it. So we’ll be mindful of your time though and limit it to 30, 40 minutes or so at the most. So anyway, Frederick, correct me if I’m wrong, but Musashi AI is a unique company in that you’re part of Musashi, which is a Japanese global automotive supplier. So that in and of itself is unique. Tell me a little bit about the company’s background and how Musashi AI fits in.

Frederick Reinink: [00:01:13] Yeah, great question. Well, we see that uniqueness also as an advantage. So to your point, Musashi Seimitsu has a long history of being a tier one automotive manufacturing organization. Obviously throughout that history, well aware of manufacturing processes and quality processes. And that’s really where my group got its foundation was through observation that within manufacturing facilities, about 20% of resources and costs still were going towards final part inspection or quality-related processes. So the organization, looking at some of the new advancements that were happening with AI and capabilities, put resources across a few teams in the globe. My team is working out of Waterloo, Ontario, Canada, and then from there, with the results that were shown, it officially spun off as a separate business, a technology-focused business, and that’s where my group, leveraging a lot of the experience that Musashi has as an automotive manufacturer but also recognizing that the same issues and problems that Musashi was facing, other OEMs and tier one and tier two were also having those same challenges,  kind of recognizing the market for the technology. That’s really what led to the commercialization of the business unit.

Winn Hardin: [00:02:43] That’s great. Again, thanks for joining us, Frederick. So I think it gives folks a lot of confidence when they know that you’re eating your own cooking. You’re putting your own systems and softwares through the trials and really high-paced operational environments. So I look forward to talking more about the tips and tricks and the advice that you can give folks. For sure, when we’re talking about getting ready for these applications, setting them up, and what to do during operational performance. But we’re coming right off of Automate. We’re recording this show not too long after the Automate show up in Chicago this year. So you want to give us a quick rundown on what was new and cool for you at the show?

Frederick Reinink: [00:03:17] Yeah, sure. So Automate, this was our third year attending and exhibiting. Musashi had a larger booth presence this year compared to past shows but also a new location for us. The two previous shows were in Detroit; this year in Chicago. Our main booth, so we designated three areas where we were demoing different product. Our largest footprint was around what we call our automated inspection system. So that’s a cell that can accommodate automated inspection. And the part that we were showing was an automotive engine part. We received a lot of compliments on what that system demonstrated. So it did a really good job explaining that hardware side that Musashi also supports during vision applications but also all of the software that we would deploy in a production setting was running on that system as well. So was able to also handle the conversation on what does software look like and what’s the capability on that software. So cell that was really beneficial to have there. And I thank the team that put a lot of hours into getting that ready for us. Some of the other stations were more on some of the software products. So post-inspection, we like to have data be available through an application that’s browser-based.

Frederick Reinink: [00:04:42] So it runs in the AWS cloud, quality staff, operations, floor staff have the ability to see what’s actually happening through that system, all of the output, any of the trends. So we had a second station where we were demoing capabilities of that software platform, just a large screen. Again, a lot of good feedback because quality, traceability, especially in automotive is such a huge forward-looking requirement. And with that platform, we can  give an indication of how we can deliver traceability from a quality perspective. And that got a lot of positive feedback and questions. And then following that, the third station was around a mobile-based product that we’ve created to assist more with quality sorts. So unfortunately, sometimes things get sent to an end customer or you catch something and you quickly got to assess the status of parts. So it’s a mobile-based application that you can use on your device where you can scan the part ID or the laser engraving, and it will give you images and the record of that part. So it’s a way that can accelerate, like I mentioned, quality, sort, or containment use cases.

Winn Hardin: [00:05:58] And helping to protect those customers who may not want to have to receive an entire shipment sent back to them because someone saying that the defect happened at your end when, well, no, we’ve got good image proof and archives that that’s not the case. It shipped out fine. It met all specifications, so that’s important. If I could ask a follow-up question, do those demos and Musashi’s approach, do they include both traditional machine vision capabilities for inspection as well as AI or is this pure AI?

Frederick Reinink: [00:06:23] No, of course, we leverage traditional machine vision where it’s appropriate, things like presence checks sometimes or some gauging checks that we’re doing. You’re going to still utilize some of the traditional machine vision algorithms and code, but a lot of what we do is AI-based or specifically deep learning–based.

Winn Hardin: [00:06:50] So one more before I let Jimmy jump in again. I’m sorry, I always have too many questions, Frederick. So we talked about ease of use for sure being a hallmark of the Musashi application or software platform environment. SPC, reporting back. Are there any other primary value differentiations that you were talking about at Automate either number of checks, autonomous or unguided self-learning to be able to follow changes in a process? SPC and ease of use are obviously the top two. And I’m loving hearing people talking more about SPC, because so often that data in a traditional machine vision environment has been maybe captured but never really used. So I think it’s fantastic when people talk about how does AI need to get into manufacturing? It’s that attention to detail and that ability to get that siloed data into an environment where you can extract intelligence that we haven’t seen, and I’m glad that we’re seeing it now. But again, any other key value propositions from the Musashi platform?

Frederick Reinink: [00:07:51] I would say from an inspection standpoint, really our value proposition is around automating full part inspection of complex parts. So this is where there’s a lot of vision out there. And it can be the right application. Less complex parts can be solved with off-the-shelf systems and those types of products, where our value proposition is, for example, like a really complex automotive transmission part, where there’s going to be some customization, either from the code itself to manage how the inspection happens. But also the AI models themselves are probably going to need a lot of application-specific data, let’s just say that. And that’s where our engineering group, we apply a turnkey approach to delivering that. So really from the end user or the customer, we see them as subject matter experts. So they know their parts, they know their processes, but from a quality standpoint, they give us direction as to this is what the the quality check is going to be on this part. They give us the part drawings, part samples themselves, inspection documentation and workflows and requirements. And then it’s our engineering group that works with them, collaborates with them. And then we create the software to do that. So that’s a differentiator. We deploy or hand over turnkey software.

Winn Hardin: [00:09:22] I suspect we’ll talk more about that too. Thanks for bearing with me on that, Frederick.

Frederick Reinink: [00:09:25] Good. And I agree with what you said. Data is such a wonderful byproduct of AI. So not only do you get  consistent, accurate, reliable inspection with AI, but AI can process what’s happening in an image in an instant and it can automate some of that difficult, or time-consuming data capture. And that’s really where, to your point, once you have that data now you can add additional value and insights on top of that. And that’s really where organizations can really leverage AI.

Jimmy Carroll: [00:10:01] Frederick, some of these responses are really refreshing to hear, because if you go back and anyone who’s listened to multiple podcasts has heard me say this, but it’s hard not to repeat it because it’s just true, right? If you go back seven or eight years ago, deep learning was sort of being not reintroduced, but deep learning emerged in the industrial space. And a lot of people were hyping it as this technology that was going to take things over and then, yeah exactly, it was just shiny. It was flashy, it was exciting.

Winn Hardin: [00:10:32] Those are my Monty Python hands, by the way. There was a little hum in the background.

Jimmy Carroll: [00:10:36] Yeah. So in recent years, though, and again, like AI and deep learning, these aren’t new technologies, but in the industrial space, it seems like in recent years that the technology has sort of settled into a well-established group of applications where it adds real value. I think Musashi AI is a really good example of that. Like you’re targeting the inspection of complex metal parts and pairing it with traditional discrete rules-based machine vision algorithms and technologies. And now we’re starting to see AI as more of a larger tool for the toolbox than a magic bullet that’s going to come in and change everything. And it’s just really interesting. And over the last year or two, I think it’s really come about and there’s not really much of a question behind that. I just wanted to tell you that and I guess follow up and say, you’ve touched on it a little bit, but what’s your fundamental approach to AI? What’s your general approach I guess?

Frederick Reinink: [00:11:41] Our general approach, as we talked in the intro, customers need things that perform. They need an AI system or software that doesn’t require a lot of intervention. So at the end of the day, performance is going to be really, really important to them. Maybe one of the unforeseen hurdles or challenges is around data. Data to get an AI that can meet performance expectations. So our group, a lot of the investment in R&D, a lot of our intellectual property, our patents are around that software piece. And we’re really looking to bring down the amount of data required but improve the performance of the system.

Winn Hardin: [00:12:32] We’re really specifically talking about customizing some pretrained models, right? Basically optimizing it for the customer application itself?

Frederick Reinink: [00:12:39] Yeah and we’ve got what we call our active intelligence algorithm at Musashi, which is a fusion of two types of algorithms: anomaly detection and segmentation-based models. And there’s advantages to both algorithms. And we’ve done some enhancements. We’ve created this architecture, which we call our active intelligence, which is not self-learning. So it’s still human-in-the loop with engineers, just due to concern about things like model drift or performance. But we’ve seen that have really good impact on the ability to create a production-ready model in an accelerated time. So the time-to-value of the solution is really what we want to achieve with the customer and deliver to the customer.

Winn Hardin: [00:13:22] Would you say that developing those initial optimized datasets to be able to train the AI models, that still remains the largest challenge for customers when they’re trying to implement AI?

Frederick Reinink: [00:13:32] It’s certainly one of the largest challenges. I think when we engage with the customer, it’s really important that they assign someone that’s going to be  their subject matter expert and remain consistent so that they can define what the system requirements are, what the quality inspection requirements are. Because without that, you’re going to spend a lot of time going in different directions, trying to accommodate different people’s opinion. You really do need, we call it ground truth. You need them to agree on what ground truth, data, and requirements are. And then that’s where our group can begin to work towards delivering that. Once you have that established, then I would agree. Yes. Having the data becomes the largest hurdle to clear to get a production-ready model.

Winn Hardin: [00:14:22] Absolutely.

Jimmy Carroll: [00:14:24] Frederick, given your role as an automotive parts supplier or your parent company’s role as an automotive parts supplier, I do want to ask about different applications and where deep learning and AI have become really popular. I guess EV battery manufacturing would be one of those. So how has the development of AI and deep learning technologies over the past few years impacted battery manufacturing?

Frederick Reinink: [00:14:51] Yeah, sure. So we’re looking at these applications, and the feedback that we’ve had through meetings is that the same benefits that deep learning has delivered to other industries, including traditional automotive or medical or heavy industry, it’s around that performance piece, and performance can mean better detections of true defects. But it’s another maybe point that’s not really thought about is the false positive rate. This is where you get a lot of issues with systems historically where the number of rejects or kick-outs produced by the system, it’s just not something that can actually ever get deployed without consistent or constant interventions by production groups. So going back to EV, yeah, there’s going to be new use cases, new types of products being made, new processes being used, laser welding. And we’ve had some interaction with laser welding companies, and there are quality issues that are presented visually. So this is where a deep learning-based 2D or 3D system is going to add a lot of value. And it’s my understanding or it’s been communicated to me that the challenges of a lot of the existing technology providers is around that false positive rate. So it’s just something that creates issues. And if AI can bring the false positive rate down, in theory it’s going to be a better technology to manage inspection.

Winn Hardin: [00:16:36] And related, it’s a challenge that’s been with us for a really long time. Being able to set that perfect threshold. How much waste am I willing to put up with versus what’s the downside or liability with shipping bad-quality parts. And every application’s got a different sweet spot in that area. But I love what I’m hearing from you, that combining traditional inspection approaches with AI helps to reduce those false positives, which means overall you’re going to get better-quality goods going into the marketplace and especially in the case of EVs, where it’s really a safety-critical system, this is super important, right? And the only other option is to just throw a ton of manual laborers. Hope you catch ’em on a good day and or manual inspectors and that kind of method and hope and pray, as we’ve done for literally hundreds of years for those types of applications.

Frederick Reinink: [00:17:29] If those manual inspectors are available, given the persisting labor shortage.

Winn Hardin: [00:17:34] Very true. And didn’t have a late night the night before. It’s not the sixth hour of their eight-hour shift. It’s very difficult to keep that level of awareness for a whole shift, which is traditional. Frederick, have you done any looks in terms of or can you give us a generic case study? How much have you been able to cut false positives for some of your customers? What can people expect to get if they go through this process?

Frederick Reinink: [00:18:00] Yeah, sure, I can talk about projects. I can even say, like, I know when you speak to someone that’s interested in deploying an AI system, of course they’re going to say 100% detections, zero false positives. That’s always like, yeah, that’s the perfect system.

Winn Hardin: [00:18:17] It goes back to the Holy Grail. That’s where I bring Monty Python back. But anyway, sorry.

Frederick Reinink: [00:18:21] But if you quickly talk to them, you’ll be able to explain to them why that’s just not achievable. Unless it’s a really simplistic application, it’s not achievable. But it’s been our experience that, and we’ve done like proof of concept projects with organizations where they’re very new to AI, very complex parts. They didn’t know necessarily how well this software was going to perform with their specific part or for this application. Typically what we see, let’s call it an eight-week, project, proof of concept project. Initially that first week, there’s going to be a lot of image acquisition, and we’re building datasets at that point. And we’re making optimizations. By the second week, you’re going to have a much better-performing model than you did as you started the project, week one, and then typically by the third week and the fourth week, assuming that you’ve had steady throughput and you’ve got new data, by week four, you’ve got something that could pretty much go into a production setting.

Frederick Reinink: [00:19:24] So this has opened a lot of eyes. We try to educate the customer prior to the project, like, this is going to be the life cycle of the project. You can’t expect something week one, just due to how AI is dependent on data. And that’s really it. The more data, the sooner that we can acquire more data. And it’s not just any data. It has to be “representative of defect” data or “good part” data. But a lot of times it’s more the defect data that’s of interest because you’re training segmentation-based models on no-good data or defect data. So we communicate that and they understand it, but when they actually go through it and they can see, oh, a part that was run week one was missed. But week three now there’s an algorithm. And every time we run it, now it’s caught. And they understand it. They get it.

Winn Hardin: [00:20:23] Absolutely, and a complex engine part may have dozens of inspection points. So the fact that you’re even talking about building a system within four weeks, when a rules-based programming approach, assuming it was fixtured even, which is a big assumption, right, would at least take that, if not more, for bugging and troubleshooting. So that’s super impressive. And I think we’ll probably touch back to this as we talk more tips and tricks for the customers out there when we’re considering these types of systems. But the fact that you have a real-world approach to this. It’s still HIL, human-in-the-loop during the initial training phase and that you’re providing that engineering support. Your systems are easy-to-use web-based browsers for sure. You don’t need data scientists to be able to evaluate, but these people may not be inspection experts. They’re experts in building powertrain parts. So the fact that you’re willing to support them from the beginning to the end of deployment I think is very realistic, and the only way that you’re going to achieve that high success rate in a four-week time frame, which seems to me to be very compressed. Being able to to produce a system for complex inspections within a month is killer. That’s a big deal.

Frederick Reinink: [00:21:34] Yeah. Don’t get me wrong. There was a bunch of work done to produce the hardware element of that solution. So that four weeks is really when that system was at the customer site and we’re taking in their data. So don’t try to make it sound like we can deliver a project in four weeks. The software life cycle, certainly once you have data, it’s really impressive because, like I said, I’ve seen that “Wow” with the end user or the end customer where they said, “Okay, this has gone from, wow, we weren’t quite sure how it was going to perform to it’s exceeded our expectations in days or weeks.” So that’s really the benefit of AI. And to your point, the amount of time and resources to create these algorithms, and it’s not a single what we call a region of interest, or ROI. We’re looking at 10, 15 ROIs for 20 to 30 defect classes, not a single defect class. So you need something that’s very robust, and that’s maybe the benefit of AI. You can create these really, really robust, adaptable solutions with the customer.

Winn Hardin: [00:22:38] And that hardware development would be done whether it was through a traditional method or an AI combination hybrid method. So you were never going to get away from image acquisition, lighting, and trying to figure all that out.

Frederick Reinink: [00:22:51] You nailed it Winn. Yeah, exactly. The only change is that AI is the software that’s actually processing images or inferencing, as we say. You’re exactly right. Everything else would be standard. Whether this was an AI-based or non-AI-based solution.

Winn Hardin: [00:23:05] And many of those cases, the traditional rules-based model would never have been able to complete these kinds of inspections anyway. Not with anywhere near that level of confidence.

Frederick Reinink: [00:23:14] A hundred percent right, Winn. And that goes back to why my group was created. So some of the applications that we’re currently developing for were previously attempted to have automated using rule-based non-AI-based systems. And all of those projects failed, unfortunately. It would have been great if they could have solved it, but it wasn’t feasible and for the reasons that we talked about.

Winn Hardin: [00:23:39] And now here we are five or six years later, and I’m sorry, Jimmy. Five years ago, six years ago, a lot of integrators were very concerned about trying to adopt these AI solutions. They didn’t really understand them. The level of integration process and capability and the software itself wasn’t as user friendly. They might have spent an awful lot of time on NRE trying to solve an application that could have never been solved previously. And here you are solving those applications within a month, two months with total optimization, programming, debugging, done. So we’ve come a long way in a relatively short period of time, it seems like from an old guy anyway.

Jimmy Carroll: [00:24:20] That’s actually a perfect segue to one of the questions I wanted to ask Frederick, which is about — it’s almost like maybe some of the initial challenges were, AI is not a good fit for the job, and people didn’t really quite understand it yet. It’s almost as if it’s industrial AI 2.0. What are some of the challenges you’re seeing now, now that deep learning and AI have been sort of established tools? They are established tools, obviously. What are some of these new challenges that you’re seeing? And do these challenges change if you shift from small, medium, large enterprise?

Frederick Reinink: [00:24:53] Okay, yeah, good question. So me being in the business development role, customer-facing, a lot of interactions with customers, I would say educating on the delivery of AI, keeping expectations where they need to be realistically. I think things like ChatGPT just put people’s expectations for what AI can do out of the box at a level that’s not achievable. They’ve got a fantastic product there, and it’s a specific type of AI for a specific use case. So for me, it’s about managing and making sure that the end user or the customer understands how the process is going to unfold with their application, and that means data, of course. We have that conversation around data, getting a sense of how soon we could put together something that will be deployable. But even long term. So this goes into any kind of production deployment. What happens if there’s a change? Something changes. You might have a different supplier for a particular part, and that could have a minor impact on the surface appearance. That’s really where our data product, so over time, any changes, end users have the ability to create what we call notifications or they give us insight into, “Okay, the AI said this. Can you look at this? Can you evaluate with us if we need to make changes to the system?” So that keeps that ability to track how the AI is doing long term across the deployment. So, yeah, I think the biggest thing around now is making sure that people don’t expect too much. But at the same time wanting to provide them a good indication of where it can get to. So not over-promising, being very realistic in where an AI-based solution can get them

Winn Hardin: [00:26:54] We’re notifying them when yields are starting to drop down because of variations from various suppliers, whatever the issues, and then giving them the tools to be able to say, do we want to modify our basic model or is this a temporary blip and ergo no. I like this because it’s more specific. So many times when we talk with folks, it’s, “Our AI platform is really super easy to use.” Okay, well that sounds great, but how, and you’re talking about consensus building or ground truth as I think you called it earlier in the day, which is very common thing that most folks don’t talk about when they’ve got inspectors, not just multiple inspectors in a single facility but now you’ve got 20 facilities around the world, it gets incrementally, exponentially more problematic. And other things you said, like having multiple 10-, 20-class defect classifications within 12 or 15 different ROIs. Each one can have that same list subset.

Jimmy Carroll: [00:27:53] Bad data will produce bad results.

Winn Hardin: [00:27:56] Every time.

Frederick Reinink: [00:27:57] You’re right, Winn. Of course, our solutions can’t create new problems for the manufacturing group. So that’s where we feel we have to understand with them if there is a change, how can we react in the way that they want us to, that they want the system to. Some of these new things, let’s call it an anomaly. Something comes up. This could be something that they look at. They evaluate. Their quality group says, “Yep, we’ve looked into it and it meets specification. So can we have your model either detect it or pass it or just have your model not even detect it anymore?” That’s really the conversation that we have post-deployment of a production algorithm.

Winn Hardin: [00:28:41] Absolutely. So you’ve talked about some of the challenges: data acquisition, optimization, growth over time. If someone’s considering an AI model and we talked also about setting expectations properly. This is not a large language model. This is a much more complex element that we’re talking about here versus a ChatGPT. But do you have three or four suggestions that when people are considering this application you tell them, “Okay, before you come talk to us, these are the things you really want to think about.”

Frederick Reinink: [00:29:11] Great question. I’m going to think on it a little bit.

Winn Hardin: [00:29:17] Sure, we’ll give you, like, three whole seconds, frederick. It’s always defining the specification, right?

Frederick Reinink: [00:29:24] Well, that’s what I was going to say. But the specification, really, there should be a good business case behind it, right? Because if it’s really your organization’s first interaction with an AI system, you’re going to want to, like you mentioned, build consensus for this technology at that site. So you want it to have impact. Choose maybe an application where the AI group says, “Yes, this is a good problem for us to solve. It’s not outside of the scope of what we could deliver.” But then also, the customer site is going to say, yeah, this is going to have a major impact either through improved quality detections, less claims or warranty issues, or whatever that could be. Maybe it’s a cost element. Maybe automating this inspection is going to allow that organization to put human capital into other areas and that could have a lot of benefit. So that would be one of the conversations I would have with anyone considering AI is just make sure you’re choosing a use case that’s going to meet what you want out of the system from an AI perspective but is going to deliver maximum value.

Winn Hardin: [00:30:34] From a TCO, total cost of ownership, including soft returns about the ability to move inspectors to production line monitoring and other places where you’ve got labor gaps and you’ve got challenges going for a while.

Frederick Reinink: [00:30:45] I’ve talked to organizations where they’ve got a particular defect that they’re having issues with and they’re evaluating different technologies. And when it comes to AI, a lot of the time they’re unsure how an AI model would do with that particular application or that defect. And then this is where proof of concept. So maybe that’s another suggestion is, if you’ve got something in mind, what is priority for your organization to have validated or confirmed by an AI system? And if you can identify something, you could do this in a very low-cost way. AI gets trained, you’ve got some software, you can run it either in the technology organization site or you could potentially do something at the customer site or user site. The buy-in is so strong when they can see that software operating in their facility with their parts. So proof-of-concepts or proof-of-principle projects can be a really, really great way, low-cost way to build, to get that buy-in, de-risk some of the concerns that they might have about adopting an AI solution. And if you choose wisely, you don’t really have to say, well, we have to do everything that a production solution would do. It’d be like, no, if you can prove it out on these two defects and on these two regions or surface areas, we’ve got enough trust in the system to go forward with a production deployment.

Winn Hardin: [00:32:20] And luckily the customers have a partner in Musashi who’s willing to help them through at least some little part of that POC. And most of the customers they don’t really think about it. Everyone brings pie-in-the-sky ideas to so many of our mainstream vision automation integrators and looking for that silver bullet scenario. But it can be hard on your end, right? How do we manage which projects we are really going to investigate and try to help solve where we believe we have a good confidence level of success? Do that TCO, do that due diligence beforehand and come up with a subset and a reasonable POC. And then that can make a real difference.

Jimmy Carroll: [00:33:04] Frederick, we generally like to close with asking if you have any trends or predictions for the industrial automation space. But before I do, you brought up ChatGPT, and I just have to ask, not specifically about ChatGPT or large language models, generative AI, but if you’re having a conversation, whether it’s a person you meet on the street or a family friend, or if you have kids, the kid’s parent or something, and you tell them what you do for work, and you explain to them that your company works in AI, how do you talk them down off the ledge? So many people fear AI, and seven or so years ago, there was this conversation about robots are going to take our jobs. Well, that’s been proven wrong. And now there’s this new conversation of, like, AI is going to take over the world. Well, no. How do you how do you explain.

Winn Hardin: [00:33:50] It might be better if it did, but . . .

Frederick Reinink: [00:33:55] Great question. So, when people ask and I try to give them an overview of what I do or what my company does, I immediately start to talk about quality, quality applications. So we’re an image-based AI processing company. Our algorithms really are for processing of images. Everyone’s got an opinion about AI, myself included. I think obviously it’s going to deliver a lot of value, to manufacturing, just to the general public. Certainly it has to be done in a way and delivered in a way that doesn’t compromise anything around safety. We’re still under obligation through our projects to meet all of the requirements that they have regarding a safe workplace, making sure that our systems are reliable. They do reliability tests with our AI systems in the same way that they would do with a non-AI-based system. So it’s not like, oh, if it’s AI-based, it’s a shortcut to having something put into a production deployment. Not the case at all. We still have to meet those requirements. But overall I am very positive on what AI is going to do, especially with impact to manufacturing. I think there’s a number of ways, quality applications, of course, but even outside of quality, there’s so much to do around just general optimizations. I think most of us are using something like large language models on a daily basis anyways, just in our regular jobs. And we see the benefit, we see the impact of it, and hopefully it is allowing us to do things quicker or spend less time on those activities where we could use that AI.

Winn Hardin: [00:35:46] Even if you’re not a huge ChatGPT fan, if you’re asking Alexa for information or Siri or anybody, you’re using a large language model, whether you realize it or not.

Frederick Reinink: [00:35:57] Yeah, like Microsoft has AI baked into it now. Even Excel, and I like it. It works. It delivers things in a way that it almost seems intuitive. When people have that experience with AI, where it’s more, “Oh, this is better. This is done faster or this is more consistent.” That’s where people say, “Okay, I get it. This is a good application for AI.” And then to the point about people worried about it taking away their jobs, that’s where there’s concern, and I don’t think AI is ever going to replace someone’s job. It’s really that you learn to to use AI in your job.

Winn Hardin: [00:36:43] Yeah. Over and over we’ve seen automation increases head count, companies are more profitable, and they’re able to hire more people. So there are no outliers where that’s not the case.

Frederick Reinink: [00:36:53] For sure.

Jimmy Carroll: [00:36:55] So, Frederick, the closing question I like to ask, and I’d like to frame it in the context of Huang’s Law, Jensen Huang from NVIDIA, who said years ago that the rate that GPUs are advancing is significantly faster than traditional CPUs. And I think it’s not necessarily a coincidence that we’ve seen a lot of AI deployments increase as these other technologies increase. So just with that in mind, are there any exciting developments you see in the next two or three years based on the advancing capabilities of not only industrial computing but deep learning models and cameras and beyond?

Frederick Reinink: [00:37:37] Yeah, great question. So my group, we work with NVIDIA a lot. We attended a session where Jensen was speaking last February. And even now as far as the industrial compute market has progressed, they’re still anticipating further enhancements that are going to allow AI-based companies to deploy the latest and greatest, new types of models, new types of architectures that, again, really are going to have a more positive impact on outcomes of projects, either from cost perspective or performance perspective, so we’re excited by that. And that’s really where our software group, they spend time and effort making sure that they’re keeping up to date with what’s available, what shows promise, and what we could utilize within our technology stack. Regarding cameras, even things like 3D-based image acquisition has certainly come a long way in the last five, if not longer, 10 years. But again, quite exciting to see where that’s going from a cost perspective, from a capability perspective. And I do feel like that’s really going to lead to a lot more value in vision applications. And then some other technologies as well, probably too many to name, but certainly quite excited. Certainly, I think there’s just going to be more ability to integrate AI across more applications is where I see that going.

Jimmy Carroll: [00:39:15] Yeah, cheaper, easier for sure.

Winn Hardin: [00:39:18] Yeah,  I can’t wait to see what AI does when it’s regularly using three-dimensional data, and not just in the industrial space. Because we always break the trails for the most part. Although academics are out there doing amazing things and they’re bringing new ideas and we’re leveraging that, so I don’t mean to steal their thunder. Without them, none of this would exist, but our deployments in industrial seem to be more refined. We don’t do everything, but what we do, we develop really, really well, and I would see more of the flow going from an industrial application leveraging 3D for AI than, say, autonomous vehicles, which are using low-resolution, multilayered sensor networks. I guess it’ll be interesting to see, but that’ll be fun.

Frederick Reinink: [00:40:05] Nobody knows what that new great technology could be a year or two years from now. But yeah, excited where the general direction of AI is going and then some of these other modalities as well. So quite exciting.

Jimmy Carroll: [00:40:22] Yeah. Absolutely. Like in the same way that we’ll follow along with developments in computing and cameras and so on, it’ll be interesting to follow the progression of Musashi AI. I’ve had the pleasure of meeting up with your team in the last two Automates, and next year I look forward to seeing what you’ve got. So, Frederick, again, really appreciate the time. if anybody would like to learn more about Musashi AI, it’s musashiai.com. Or you can reach out on LinkedIn or reach out to us at manufacturing-matters.com. We’d be happy to pass along any questions to Frederick and his team. And, once more, Frederick, really appreciate it.

Frederick Reinink: [00:40:55] Thanks for having me, Jimmy and Winn. I appreciate it.

Winn Hardin: [00:40:59] It was a pleasure, sir. Thank you.