Aron Semle on MCP, Agents and the Data Foundation You Can’t Skip

Speaker 2 (00:00) Hello, welcome back to the ITOT Insider podcast. This is David. Willem is not with me this time he's enjoying some well-deserved holidays, but I have a very, very interesting guest here. Actually, I have to say welcome back to Aron Semle the CTO at Highbyte. We had a podcast exactly, well, exactly, give or take a couple of weeks, exactly a year ago. um And I'm super happy to have your back because in that year, so many things have changed, have evolved in the industrial data landscape. So, Aron, welcome. Let's start with the first uh short introduction, reintroduction of your self and Highbyte. Speaker 1 (00:44) Yeah, thanks David. Happy to be back. Yes, Aron Semle from HighByte, CTO at HighByte. So my background is many years ago, joined Kepware back in 2008. And that's we'll see if we can actually have a Kepware t-shirt on the show. This is like one of the old, I have things for your internet t-shirts. You had to be back in like 2015 to get, that was before the PTC acquisition. We were doing internet of things, industrial internet of things I'm wearing today. But yeah, started. Speaker 2 (01:09) it's going very high if you go to the collector's markets. No, no, it's priceless. Speaker 1 (01:17) exactly. Uh, but yeah. And so it's kind of started my career there. Industrial manufacturing was acquired by PTC. did it a year or two at PTC, uh, briefly jumped out and got into healthcare. And that was doing fall prevention in nursing homes for a couple of years in a startup. And that was healthcare is its own, its own beast. Um, I've heard it. One person told me it's government healthcare. and then manufacturing in terms of just complexity. And I would agree, I would agree with that. And then joined HighByte in 2020 now. So it's been over five years. So back then, HighByte was formed in 2018. So I joined a year or two in. We had a very, would say, John likes to say MVP product, emphasis on the M for minimal. And we've grown it since then, but we were talking about industrial data ops way back then before AI was even part of the discussion. And obviously 2023. the release of ChatGPT changed all that. I think it was 2023 if I'm getting that right. But yeah, but happy to be back. like you said, so much has changed in a year. You're like, oh, what have you guys been doing? I had to go back and look at some of our releases and stuff because we're just moving so fast. Like I even forget some of what we've done. Speaker 2 (02:24) Yeah, it's crazy. It's also uh when you just look to the entire uh industrial data landscape, um it took a year or two years maybe after ChatGPT was released. uh But now we are in this super fast changing dynamic, um which also introduces some confusion, um I have to say. So maybe we can touch upon a couple of those. Yeah, 100%. Yep. confusion in this podcast as well. let's start with just looking at that last year. obviously for those who haven't listened to our previous podcasts, why don't you listen to that one as well? But like in that last year, from your perspective, what are the major changes in the markets for HighByte? let's start with that. Speaker 1 (03:23) Yeah, I would say oh in the market, especially when one trend we've seen is like AI is pushing everything, right? So like we were doing industrial data ops before AI. And I remember I tell this story like starting at HighByte John Harrington he's one of the co-founders was just on calls every day talking about industrial data ops. think 5 % of the people he talked to knew what he was talking about, but we were, you you got to be able to contextualize your data at the edge. It's got to be contextualized when it gets to the cloud. We don't do that anymore. And I think you can go back like maybe a year, maybe a year and a half and like, we don't sell the problem anymore. Like people understand the problem. So we have customers now, prospects coming to us saying, Hey, we know we need to clean up this data to get AI ready. And we need a solution that we're looking at industrial data. And you and Willem, David, I mean, jumped on earlier too. And like, you probably see it as well. Like industrial data ops is now a segment of the market. Sebastian from Frost and Sullivan, had him on a, we did these data ops days late last year where we brought customers in and did a few around the states and he gave a great presentation as an analyst at Frost and Sullivan where he made the argument that AI and data ops are hand in hand, coupled together and he was showing the growth of that market and he has a great visual of that. I think he's right, I think he was brilliant in that analysis and we're seeing that. um Speaker 2 (04:39) interesting to just I don't want to interrupt you but like it's interesting to see that and I have the same experiences so a couple of years ago you start with this evangelism almost on but you have you really have to build a foundation you need something right and then that was a very hard sell and then you were I think especially it's now better, but especially in the peak um of the hype around uh LLMs, everybody went like, yeah, we need AI. And then I again came in and said like, but you know foundation first? And I have that feeling that now, like uh people understand that. Like now people come to me and say like, we want AI and thus we need to build a foundation. 100%. Speaker 1 (05:34) Yeah. We're seeing this the same thing. It is. It happened relatively fast. Right. And I think it's more because of some of the hype, you know, the, sea level folks are like, my gosh, we need to be doing this. And that kind of travels down to the ranks where someone at the, at the factory level goes, okay. we need this to do it, right? And that comes back. But you're seeing that loop close. And it is interesting how the conversations have shifted. So yeah, looking back at HighByte, like a year ago, I think we had first released our MCP server, which is, we'll get into a little bit of that, but MCP is one of the protocols that agents can use to call and interact with data sources. So that was new last year. We've continued to enhance that so that we're actually seeing customers and we can get into it, but use cases around that, like interacting with factory systems more now. kind of in chat bot stuff. But we've been doing non-AI stuff too, right? Like we've enhanced pipelines to do looping logic, which, hey, LLMs are not great at loops. That's one of the things that they can get kind of tied to knots on. So stuff like that, central data, so being able to pull data after you contextualize it in remote sites up into central so that, and a lot of those use cases are, hey, we have AI and data context in the cloud, but we need to run that in the cloud, but we need an easy way to tunnel that data up on demand. So those kinds of interactions. See, I'd have to go back and look at our whole roadmap. The other place we've been putting a lot of time in lately, which is really interesting is not just, so we kind of separated data ops for AI, right? So that's everything we're talking, like trying to contextualize your factory data, get it ready for AI, be able to use it, MCP, all that stuff. The other thing is AI for data ops, which is also a big part of the market, which is, okay, can I use AI to help me contextualize and do this faster? And... my own job as an engineer easier. Yeah, I don't want to do this drag and drop click this all that. Like I know what I need to do. Just like we're using LLM's everywhere else. Can I have it help me? So in our last release that just went out a couple of weeks ago, we put in what we call a pipeline AI agent, which is really like you can hook up to an LLM and you can sit there and prompt it. And pipelines are like our ETL tool for moving data, but it can analyze and edit. And what's cool about it is when you talk about AI, there's like, there's totally manual, right? I want to just get into this and drive because I know what I'm doing. There's, hey, I kind of know what I'm doing, but I want some AI assistance, because not because I'm lazy, I'm just busy and I need some help here, which AI is really good at that, human in the loop. And then there's totally autonomy, where it's just like, just go do this, I don't have the time, I trust you in this task. So what we did with pipeline AI agents is the middle ground. like. I can prompt it. What's cool is the UI comes back and shows you what it edited in line and you can review that except for Jack continue to iterate. So it's kind of like I talked about as that AI slider, um but it's cool. I mean, we Speaker 2 (08:23) that's interesting, that's interesting because I think... uh Again, if you talk about people who are using Claude or whatever, LLM, to uh develop, um there is this... To develop. Yeah, that's the way to put it. Yep. like there is this misconception in my opinion that everybody can now create anything. Like, yes, sure, the very small, you know, household app or the proof of concept which might still expose some API keys in the Speaker 1 (09:02) source code or whatever. Yeah, it's like point. and it's like, okay. Speaker 2 (09:12) But I think this is really interesting because the people who benefit, in my opinion, the most of this evolution are the ones who understand what they are doing. But you just want to give them superpowers. Make it like very simple, very easy for them. Speaker 1 (09:29) The other side of that is when you don't know what you're doing and you're flying blind, allowing the agent to do too much and you don't have the acumen to review and edit and that's where it gets a little like, that's where you leak keys and you start doing some things that potentially are harmful. Yeah. absolutely. Speaker 2 (09:42) And unfortunately, I came across a couple of people I know oh who developed something and, for example, created an application which got hacked. And then through that hack, some data leaked. And then you talk to these people and you go like, yeah, but... you're using an API key with admin privileges and you've exposed that one. And then the response is, what is an API key, David? Anyways, very interesting. I um think the first thing we could focus on is that foundational part. So I think that it's fair to say that Having a contextualized data foundation is a must have, preferably using or having data of good quality as well. But at a minimum, being this contextualized data sets, we have our capability map, our listeners, our readers are definitely gonna know the capability map. We had our first iteration when we talked last year, now we have our second iteration, but I think the foundation remains unchanged. um Let's go back to that foundation. How would you start building such a foundation today? What are the typical customers you come across? How advanced are they? Are they still using historians? Let's go to the foundational part first. Speaker 1 (11:29) Yeah, there's the digital maturity like grab or slide, right? Like if you've got equipment and you've got nothing connected to anything and it's a totally manual process, like you are not ready to use AI on your factory. Like you need to be able to connect to your machines. Hopefully you have some HMI SCADA capabilities. Hopefully you have an historian. If you're at that level, then you're probably ready for the next level, which we would call that's industrial data ops. That's the data foundation. So the way we position industrial data ops is it's not going to overlap with any of that existing functionality you have that's running your plant. That's different across all your plants. but it's a layer that you put in to say, I can connect to all those systems. I can contextualize the data and then I can provide access to it from other systems, AI being one of those, right? And the trick there is like, how do you consistently represent or model or put context around equipment that could be the same across facilities or different, right? Because you want, that's the scalability piece. You want some harmonization such that if you go to factory to factory and the OT systems are completely different, at that data ops layer, whoever consumes that doesn't need to know the difference, right? And that's a lot of the work of industrial data ops is that, okay, how do we commonly contextualize it and do all the work to make it look the same? Because we know under the hood in the factory, it's not like, this production order is coming from an ERP system or an MES, and it's great. That's all. This one's actually coming from a CSV file that's on someone's desktop, but that's where they're at, right? But if I'm consuming that from AI or anything, start up a plan. need to know that difference, right? So I think companies that kind of already have some of that connectivity skate age and my historian are in a better position to layer and industrial data ops and then get value of AI companies that don't even have the connectivity piece. It's 2026. But I'm not, I'm not here to judge either, right? There's certain companies that their margins are so great. And you look like my dad ran a factory in Maine that did biological safety cabinets, right? And I remember early in my career walking through his factory and they've got CNC machines and cutting machines and Allen Bradley PLCs. And I was a kepper. I'm like, why don't you connect and all this stuff? And he's like, why do we need to like our margins are great. There's not a lot of competitors, you know, like why would we invest there? So it's not for everybody. oh But yeah. Speaker 2 (13:44) So in our work, started making the, it's not really two groups, but we talk about the OT light and OT heavy type of companies and certain industries, they just naturally... already have data available. They have these, I would say, widespread skater, DCS systems, have these historians, they have certain legal obligations to report data to a government or to their customers or whatever. ah But yeah, then you also have a very big set of manufacturers who like your dad's company, but also... I've been working with so many bakeries where they buy a machine from that vendor and then they buy a machine from that vendor and another machine from that vendor. And there is a local HMI over there and local HMI over there. Speaker 1 (14:40) all bundled up into the machine. Speaker 2 (14:42) And that's perfectly fine for them, but they also now want to get on that AI train. then for them, think there is no immediate need to go and buy a traditional historian and or traditional SCADA system. ah They now have this power on the edge available as well. Yeah, yep. Speaker 1 (15:06) And we see that too, like being able to get from data ops, connecting into those local HMI skaters running on the equipment that the OEM put in there, which could be SQL databases or whatever's there. Yep. Speaker 2 (15:18) It makes it easier for them. I think with the tool set they have available today, that makes it easier. Having said that, I wanna bridge into context and does also a bit unified namespace. um I think our listeners know that unified namespace is a topic I also talk about quite often. And I think one of the most important things from my perspective is that uh it's not just about having an MQTT broker. It's about a concept. about how do we... It actually starts with a uh data governance approach. And that's kind of tricky, especially for those who are starting from nothing. Like where do you start? How do you start bringing your data together? Does it need to be an MQTT Like where is your thinking right now? Speaker 1 (16:14) Yeah, UNS is evolving, right? If you look, maybe 10 years from now, we'll look at the history of the, UNS, the UNS movement. I'll call it. We'll be able to like fully describe it. But my, my take is that, you know, five years ago when we were doing UNS and it was just, you know, it was the start of data ops. Like we're going to put data ops at the edge, do all the heavy lift of connecting to this stuff, contextualizing, and then we're going to publish it. um So why was that so appealing? And you know this, right? Because let's take that bakery. Like all these machines are independent. There's no way to get access to the data unless you go up to the physical HMI and you're giving me a solution where I can connect a broker, subscribe to all, and I can just see it all. If you're coming from nothing and you can get to there, that's value, right? You look at it and go, wow, I can see the real time state of the factory. Advance, you know, five, six years to where we are now and AI is helping drive this conversation. The part that sticks is okay, that contextualization, that, that, uh, creation of the data. That's great. Publish just publishing it to MQTT, which this is not like, think, again, I think that strategy sticks around. I don't think that's going away. think MQTT is here to stay and the broker is here to stay, but that's really just the transport of real time data out. Right. And I think the market, we didn't like, we thought this five years ago, but we weren't, we're not going to go like the market wasn't ready to hear it. So we're not going to get really bullish on it. But that is, it's just a transport. Now you need to get it somewhere where AI can come in and make requests. You need to solve the problems of what about historical data, stuff that doesn't traditionally come through an MQTT broker. I think you can look at a lot of the broker providers too and the capabilities they're adding, which is also kind of this context of, okay, that was the start, but we need to do more. to provide the other data patterns that we need in the factory to get access to this data. So it's not that anything we did is nulled. It's just, was the first step of the journey towards like something bigger that is, that supports the broader use cases that we have. Speaker 2 (18:13) Yeah, and then we come to having access to the historical data. How do you approach governance across multiple sides and across multiple contextual dimensions? Speaker 1 (18:30) guarantee that that occurred and the transaction and all those kinds of things. So I think it's just industry maturity on it. And to me, comes back to like the nice part for HighByte is we're industrial data ops. So we're that layer on the factory regardless. That's the hard part. Honestly, it's not like once you've defined that providing the interactions, publishing MQTT and all that is a lot easier. You have optionality. oh But the nice part for us is like our... the problem we saw has changed, And we've kind of stayed even while the AI stuff, like we're still doing the same thing we were doing five years ago, eight years ago when we founded the company. um The problem is just more visible now. Speaker 2 (19:07) Yeah, because you touched uh upon AI, um let's step into LLMs, maybe even agents. That's the new kids on the block. um So, okay, let's start with a definition first from your perspective. um So using an LLM, are now experienced. Most people are... have now have some experience in the last in the last years what an LLM is now we have these agents ah first of all what what's what's the what's the difference where do they what do they do Speaker 1 (19:53) Yeah, so LLMs are, I that's the large language model, that's the large neural network. Most of us experience those, the most through chatbots. So it's like, put in a request, in comes the response, I put in the request again, and the entire context of the chat goes in, or some compressed version, and a response comes out. Now you have tool, tool calling has always been part of LLMs or early on. is, this is where MCP came from. So if the LLM doesn't know the response, it has a set of tools. can reach out and get answers. That's also just the loop that in that's part of LLMs. So tool calling is really the actual, the LLM can detect, I can't respond to this. I need to go call something. Agents are really the code around the LLM. Right, so an agent is in a chat bot, it's just like putting in the response. If the LLM comes back and says, I need to call a tool, agent is code that's gonna go call that tool over MCP or some other mechanism and then respond back and put data in the context window. An AI agent is really just an LLM in a loop. So take the human out and that conversation is like happening by the agent and the agent decides when it's done, right? And that really simplistically. That's what an agent is. How autonomous it is is up to the code that you write around the loop. uh But make no mistake about it. We're just building text documents and feeding them into a large language model. And then stuff comes out that's sometimes in the JSON format that we can code against. And that's what we're doing. It's not beyond that. uh Speaker 2 (21:19) But and that's interesting because you know I can ah obviously I can imagine using this for doing some research on the on the internets or summarizing texts or helping me, yeah helping me find even that helping me find some I'm but maybe that's maybe a bit old school but I still like googling for an answer. For me like Google is still my first because somehow I don't know I it's like I can see the source and I can click on it and I can then determine whether I trust it or not. I don't know but on the other hand if I ask a 16 year old right now what is the capital of I don't know what country? like chances are really big, they just open chatGPT and all that to chatGPT, right? But one of the problems I see in the industrial uh context setting, and again, you mentioned the word that we did as well, is the fact that we want something which is uh deterministic, where we know, where we go from point A to point B to point C, and it's always the same. seems to be still a problem. Speaker 1 (22:45) Yeah, so determined, I mean, that's the biggest, someone would say it's a feature like LLMs are hallucinating all the time and they just happen to be right, you know, a percentage of the time. That's part of the nature of their probabilistic, right? They're not deterministic. Early on, like a year ago, and I probably haven't used this in a while, but I would say, you know, if you're trying to do a genetic AI or whatever you're doing, you're going to get a cat meme response at some point. And like, you need to, you need to be able to handle that in whatever you're using. And that's the truth. think we were joking before, like I've got a trip coming up, a hiking trip. um in Europe, it's like one of the long seven day or four or five day backpacking trips. I'm doing some research and like you, I'm using the LLM to cheat like, hey, how could I break this up? And one of the core things is on this hike is, is it east to west or west to east? And the LLM comes back and says, oh, you want to hike this west to east. And I had done some reason I had done some Googling because I'm old and I'm old school. And I was like, that doesn't sound right. So I like prompt the LLM again and goes, no, no, it's west to east. I'm like, all right, pull up Google, look at all trails. No, it's east to west. Right. So like the LLM not only was wrong, hallucinated, it like doubled down on the hallucination and wouldn't relent without like an external source. But that's the nature of it, right? So that fundamentally limits in manufacturing and across industry, like how we can use, it doesn't make it not useful, but it limits how we can use it. So if we're talking about fully autonomous agents that we're gonna get control of our machines, very bad idea, right? Agents that that monitor the machines and make suggestions to an operator that knows the machine probably a better idea um as long as the operator is trained but kind of to speak I guess more candidly the two use cases I see that are probably going to drive the most value for manufacturing and AI are The chatbot right and we're seeing this now. I think industry is everyone's talking like a jet to gay I totally Thomas what we're seeing is folks are connecting data ops to industrial systems, exposing MCP tools that are custom to them. we, at Hannover in a week or two, a week, guess, I'm presenting in the AWS booth around a customer that's doing this, right? And they connected to a validated PI system. They exposed some tools for event frames and not the whole system, but just like tools that they, data they wanted to expose. And then the guy that installed the system 15 years ago, prompted the LLM, their internal LLM, with the tools. And he got interesting responses, which is not surprising, right? And he started to look at that go, wow, if we can start to rely on some of these responses, I don't have to go into the clean room, because this is a pharmaceutical, to go get the answers from the engineers that are in there. There's so many um higher level applications, dashboards. reports that we built that no one looks at that we could just generate on the fly if we can start to rely on this. So that's, I mean, it's fairly disruptive, but it is a real use case that adds value. The other one is the development of applications by people on the factory floor. like, and we, I just saw this last week. So we met with a large manufacturer and they're like, oh yeah, we have a hundred plants and one plant, this really smart guy or person created an app. that drove value for that plant. Now we are taking that into our ITOT organization and trying to figure out how we take ownership of that app, because it was vibe-coded, and distribute it across the rest of our facilities, because it's driving real ROI. But now we need to manage it, right? Speaker 2 (26:12) That's still a tricky one. Speaker 1 (26:15) It's tricky for the IT organizations, but I think the ability to hand over to allow the person with the problem to create a solution for it using language, that's huge, right? Because... Speaker 2 (26:29) And that also you mentioned the, we have a validated PI system for that Pharma customer. And then we have this abstraction layer. And then we have, let's say your playgrounds, like do whatever you want. that I think that's a very, very important one where... you want to make sure that your core systems, they remain core and they remain validated or at least there is a clear deterministic path like A goes to B goes to C and this is how we run operations because operations is deterministic by definition. ah But having that playground available, that's room for exploration. That's interesting and it's also an organizational question. This is definitely, well, I would say, I would even dare to say this more organizational than technical because you need an organization which fosters that. But then also like your example, at a certain point in time says like, okay, but you know what? We're gonna make this a bit more formal. We gotta take ownership. Yeah. Speaker 1 (27:43) Yep. And what's interesting is the part that doesn't change in all that is that data foundation, that industrial data ops layer, right? So what you're seeing more broadly in the market is that application layer. It's a scary place to be right now, right? Like, uh, because AI is very disruptive on that, on that side. So you're seeing a lot more focus on like, you're not going to go vibe code a SQL server database. No one's going to do that. Right? Why would you do that? You just use the database, but applications on top of that. more disruptive for AI. And I would argue that the data foundation, Industrial Data Ops, is more like the database for manufacturing than it is the application. But it's really interesting times, because like I said, that was just two weeks ago. We were having that conversation. It's like, oh, how's AI, LLM? What do you guys do? it's like, oh, know, guy, vibe coded an app that drives value. And now we're trying to take ownership of it. And I thought it was fairly progressive for. absolutely. to try to figure that out. But I'm also like, I'll check in in a few weeks. Like, Hey, how's it going? Cause I'm just curious. Like how, how are you doing it? Speaker 2 (28:42) even if you would suppose that the value would be uh clearly shown, then the organization could even decide to take the app as the proof of concepts and to start over using maybe a scalable framework or a more secure framework or whatever. So in this case, okay, the app was taken over, which might be a perfect solution, but just having the ability to show value in... Hours slash days? That's the difference. Speaker 1 (29:15) Yeah, yeah, 100%. Yeah, I think you're right. I would not want to take the Vibe Code and go put that in production and try to manage it. uh If I had an application platform that I could then easily kind of bring that into that I've already managed, I think that's a better way to do it. Speaker 2 (29:31) Now, because you mentioned MCP, I want to also take a bit of time to step into that very interesting topic. um actually the my confession, my introduction in MCP were your videos you recorded a year or so ago, I don't know. Speaker 1 (29:53) 2020, it was right before MCP was announced. had a video on function calling and then that video went out. Someone's like a week later is like, you should look at MCP. And I was like, this is exactly the problem I tried to solve in my code. So I like instantly saw the value of it. And yeah, that was probably 20, like 2024. Yeah. Speaker 2 (30:11) and those videos still exist. And you actually also have, and I'll put a link in the show notes, but you also had this, I don't know how many hours workshop you recorded on building your own MCP server or your own application, I guess. So I'll make sure to put that link in the show notes as well for those who are interested. But again, in your words, like what is MCP? Is MCP really a thing? it Speaker 1 (30:40) MCP is getting beat up lately. uh And I get it, right? Anything that has like the rising edge is very steep. Like it's gonna have, it's gonna get beat up on the backend a little bit. But I think MCP is here to stay. And the way I position it is it's remote procedure calling for LLMs. It's like how, it's not how I'm gonna, I'm probably not gonna use it locally, but if I need to reach out to a remote machine and make a call. from the given the MCP tool. So it provides discovery and remote calling and a little bit of instruction in between. Where it gets beat up is again, the rise is really fast. Technologists look at it and go, this should just be a REST API, open API, like why are we reinventing the wheel with a new protocol? The funny thing is that we actually go look at the latest version of MCP. It is essentially a REST API. they kind of got rid of a, server sent events and all that is still in there, but it is essentially an API. So when people are like, it should just be a REST API. It's like, well, it kind of is like you need to go look at it. The only thing that could challenge, and here's the take, like people are like, oh, MCP is dead. If LLMs continue to get good enough that they can just call APIs and APIs, Speaker 2 (31:52) Yeah. Speaker 1 (31:52) get high level enough that they're not the crud developer tools anymore, they're actually like interactions, more like MCP tools, then I could see a de-emphasis of MCP. For us, it doesn't matter, because you're still, like you're building the industrial data ops layer, the contextualization, you we have pipelines to expose uh custom APIs or expose as MCP tools. So you're building all the stuff you need, how you call that, whether it's MCP or REST, it doesn't matter. Yeah. like that's not the hard part. The hard part is like when this gets called, I need to go read some data from the OPC server. I need to check the state of this. I need to sit in this loop and then write it. It's like all that logic that happens as part of the interaction. um But I think MCP is here to stay. What it does, it does really well. It stays inside its boundaries and I just don't see, and it is an API. Speaker 2 (32:46) But still, if you go back to having something which is deterministic, where we can trust the result, how would you position that today? using MC, is that, production is the wrong word because we're not going to use it in production, but is it stable enough to be trustworthy or is it still something which is really for exploration use only? Speaker 1 (33:16) MCP I think is stable enough. function calling is pretty, I mean, nothing in an LLM is deterministic, but if an LLM messes up a function call. ah It's not the end of the world, right? It can typically recover. And the important part is like you define the MCP tool, its inputs, its outputs. If you get a call that is like trying to do SQL injection in your MCP tool, you detect that and you stop it with deterministic logic, right? And that's why like our implementation of MCP from the gate was not a here's our static tool set that you need to use. It's not, need to actually develop your own tools as if they were APIs and manage what those tools do because they need to be deterministic. oh So yeah, think MCP used in the right way is fine. Speaker 2 (34:00) Yeah, And to throw in a new words, um i3X um very recently announced. Before we dive into that, um we will probably, well, most likely do a podcast with uh John Dyck and Jonathan Wise in a couple of weeks, really on that topic. I know you were, I think you were very heavily involved in that, right? Speaker 1 (34:27) Yes, I'm on a group chat with Jonathan and that would be a great, yes, I won't get into too much, but with him and Matthew Paris, so we're working on version one of the spec along with a larger working group of folks. So I'm heavily involved in i3X. But the story there was, if you go back to 2020, right, when we were starting, we saw UNS and MQTT. Even in the early days, people were like, this is really cool. How do I run a transaction through the broker? How do I access historical data? So we saw all the problems, right? So I was in one, there was a podcast I was on. Like I knew something was coming. Like I was like, there needs to be a new standard to kind of manage this API literal access to everything we're doing inside of UNS. ah So even in our product that we released something called a UNS client, which is really, it's just an MQTT client that you can use in the product to go see what's in the broker. But I called it UNS client. I got some heat for this, like external people, like it was just an MQTT client. Why didn't you call it that? And I call it that because I think there's gonna be something else eventually that this will also support to do rights, to do all the things you talk about. And so a year ago, I forget, it's been a while, but when I first saw the RFP and interacted with Jonathan on i3X, I'm like, this is it. This is the thing that I've been waiting for. So as soon as I saw it, I'm like, no, I'm all in. We 100 % need this in industry. I position it as an API layer for the factory, a standardized one, because every vendor has this already. We're just doing it all our own way. But if we can standardize the access of contextual data across vendors, it just removes so much inefficiency in the system. So yeah, I'm super bullish. I'm glad you're going to talk to Jonathan about it to get a little bit more into the details. it's not out yet. um I think before the end of Q2, I think we'll have version one out, though. Speaker 2 (36:16) I would say, in short, it is this extra layer on top of different vendors, different products, ah not to replace that product, but just to standardize the way you get data in context from those different sources. Speaker 1 (36:35) Yep. So what I'm doing now is like, I'm, I'm trying as hard as I can. I'll use this, your podcast as a, as an opportunity to do it. I'm trying to go to all the cloud vendors and be like, you need to support i3X ingest because in a future state, factories are going to expose this interface. It's like this is my data. This is my namespace. This is how my data is contextualized. You can come in and subscribe to it. You can read it. can, methods won't be in the first version, but they'll be coming. Like this is the layer in which you can interact. And if cloud vendors then, then step up and say, okay, we're going to build clients. Like that is going to be the ingestion highway in and out of the factory. And I think again, Cloud needs contextualized data. Like we're doing all this today, but I think it's a different way to do it that'll remove a lot of inefficiencies from the Speaker 2 (37:25) So obviously our podcast is the most listened one to within Amazon, Google, AWS. So no worries, Aron, your messages are received. Anyways, jokes aside, uh yeah, I think it sounds interesting and I'm looking forward to the discussion we'll have with both gentlemen in a couple of weeks. So if you are not subscribed yet, make sure to do so. Speaker 1 (37:52) And you get that podcast as well. Yeah. Speaker 2 (37:56) Maybe to end this conversation I was thinking like I want to do something else. Willem is not here. So, you know, I now have all creative freedom to do otherwise it just goes like no David bad idea, bad idea. So I'm going to do bullshit bingo with you, Aron. So I'm just going to throw a couple of terms in your direction and I would like you to um to score them on an axis which goes from very much hypey hype to reality production ready like or I don't know or something somewhere in between. So let's start with a simple one industrial data ops. Speaker 1 (38:41) Oh, 100 % hype. Yeah, it's just, no, I wouldn't be at high bite if I believe that. No, I think that's foundational. I think that's here to stay. Just given the realities of we got factories of legacy systems that are a mess. We need that layer to clean it up. mean, that's it's not, it's the not fun part of AI and everything else, but it's the absolutely necessary part. Speaker 2 (39:03) absolutely. To use your words, so AI for data ops, using AI tools to develop my pipelines, to test ideas, that type of thing. Where are we there? Speaker 1 (39:19) Very real, if I went, let me go zero to 10 where 10 is hype and then zero, I don't know if that's the way to do it. But yeah, zero is reality. I would say that one's probably around two or three. We're seeing it today. People are just, our backend is JSON. So we have people just generating JSON and just dropping it in the product and it works, it's pretty functional. So we're building some more governance around that. And in four or five we're working on the whole data modeling contextualization. So how can I bring in an LLM? and similar to what we did with the pipeline stuff, say, okay, I've got connectivity to all these machines, help me go through this, and here's my data models for my enterprise, or maybe I don't even have the data models, get me started. Crawl these machines, look at the data that's there, and then let me review what you built just to accelerate the contextualization piece. And think that's, we're experimenting with that now, and I think it's very real. Yeah. yeah. Speaker 2 (40:11) Although I think people would still love their Excel sheets, right? With their... Speaker 1 (40:15) Honestly, Excel sheets are a big consumer. It's like, give me a PID diagram, give me an Excel sheet, give me access to your OPC server and help me create something. Human in the loop to create deterministic output, 100 % needed, but accelerate me as the human to be able to review this, yeah. Speaker 2 (40:32) interesting, Where would you now position MCP on that scale? Speaker 1 (40:39) Yeah, MCP it's, it's real. mean, like I said, we got customers using it now getting value out of it. So I would say probably two or three. It's very, very real. Yep. Speaker 2 (40:48) then maybe as a last one, em going for a fully autonomous AI driven lights out production plant. Speaker 1 (40:57) it, don't do it. The closest I've seen is suggestions, right? So you'll see like, I'm gonna push down suggested set points and allow you to reject and or accept. So it's still a human in a loop. That's not fully autonomous. So fully autonomous, if I had anybody doing that with HighByte right now on production data, I would be very nervous. just don't think, I don't think, the tricky part is you go on LinkedIn and like that's everything, everybody's talking about it. But I think anywhere someone's using it, it's like the most trivial task. were talking before, like you probably don't even need LLMs to do what it's doing. You could probably do it in a deterministic way really much faster. Speaker 2 (41:37) really touch upon the uh RPA, robotic process automation uh part, but let's be honest, think most um comments um robots in the comments and I'll send you over my super secret playbook, which made me, which is, think it's 99 % fancy RPA. Speaker 1 (42:02) Yeah, anyways. uh like the more more a solution relies on an LLM, the less of a moat that solution has too, right? That's the other bit. Like I see some of the posts, it's like, oh, it's totally like, like 80 % of what you're doing is the model. So where's the value of the solution? We're kind of in this really weird space, but it's hard on LinkedIn to like separate what's the LLM versus like what's actually deterministic outside of it. Yeah. Speaker 2 (42:30) it's, there are many empty box solutions, right? When one of the big LMP providers, they change their pricing model, all of a sudden their business case becomes something. Speaker 1 (42:43) on the horizon, like you are seeing the monetization piece this year is becoming much more real and you're just not sure. My hope for all this is that inference becomes cheap enough and easy enough that we start to move away from these, you know, Nvidia GPUs only can run this stuff and we need giant data centers and we can run this more locally and it becomes more democratized as a technology. Speaker 2 (43:07) Yeah, I think it'll happen. Speaker 1 (43:09) And that will be very disruptive to the current market, but ah you will see. Speaker 2 (43:16) Alright, interesting stuff. um Lots of things to talk about, to think about. um So we talked about this MCP stuff and I highly recommend our listeners to, again, I'll put a link in the show notes and in the article to take a look to your workshop. um It's a rather lengthy set of videos, but it's all available for free. So that's super interesting. m You will be at Hannover Messe. uh year. Yep. assume. So you will be at Hannover Messe at the AWS booth, at the Siemens booth and at the Microsoft booth. Speaker 1 (43:56) think so. Don't quote me on that. I will be roaming around all three of those booths because we have good partnerships with all three. I don't know if we have presence in all of them, but again, I'm just a CTO. So humble. Speaker 2 (44:08) ah So anyways, for those who are still, um yeah, we'll go to... Actually, the funny thing here is, Aron, when I was there last year, I actually came across people who um stopped me and said like, David, on my drive over here or on my flight over here, I actually listened to a couple of your podcasts to prepare myself for Hannover Messe. So maybe those people do the same thing this year. So that was really cool. ah That's a wrap for this episode. uh Thank you very much to HighByte for sponsoring this one. Thank you, Aron, for sharing your insights. And of course, to our listeners for tuning in. uh And if you enjoyed the conversation, don't forget to subscribe to uh our podcast or to our blog via itotinsider.com. So uh me and Willem, next week we will be at Hannover Messe. We will actually record a... live. We will record a podcast over there. I'm not going to stream it live, but we will do one over there where we will share some of the other things we saw and some announcements. uh So make sure to tune in next week again. And until then, um take care and bye bye. Cheers.