PI and AI

This podcast is presented by NXAI, your partner for time series. foundation models and physical AI. Hi there welcome to a new episode of the Industrial AI Podcast. my name is Peter Sieberg And I'm your host today. i am going to be talking to David Rodriguez. he's the founding partner at The Industrial Intelligence Group And David and I are going to be talking about, around large language models in the Pi system.

Hi, David! Hey Peter. Please introduce yourself to our listeners please. Yeah you got it. so David Rodriguez. as Peter mentioned Foundry Partner Industrial Intelligence Group definitely come from. First job out of college was at OSI Soft, company that originally made the Pi system. So I did my time doing technical support and field service for it and all that ended up going to a customer of the Pi System And That Was A Lot More Fun Because You Get To Actually Sit Behind The Wheel Of The System And Drive It.

Make Do The Great Things That Its Meant To Do. So I spent about six years at a customer to building out solutions and then figured it would be more fun, you know, as systems integrator an independent consultant. And partnering up with some really talented people in the process. Sounds great! How was the Aviva Enable conference?

You went there last week... Oh yeah right that's always great because At the main conference, sometimes it's hard to find really technical sessions and here that was just the table stakes. You know all of the side-sessions there were hands on labs. It is always great to get past a slide deck And actually get your hand onto the newest software That has been worked on.

Get real intuitive feel for what being work done. Very good Tell us two more lines about your company industrial intelligence group please. Yeah, sure. So I mean we're just a small shop. it's four of us and We're just bunch of. you know time series data experts. You know our expertise centers around the pie system But you know It's so.

we were familiar with a lot related technologies as well And me. The general idea is everyone says to make best use Of operational data available To accompany actually turn into form that's amenable for action. So all the architecture that's involved, design decisions and new technologies navigating the landscape.

That is what we are here for Sounds great! Thats exactly what were going to be talking about. David your LinkedIn tagline says Pi system is my bread & butter but I enjoy a balanced diet. Great quote For those of our listeners who do not know if it possible at all please tell us What Is The Pi System?

Well, you gotta have a little fun with these things. You know? Okay so okay yeah the Pi system for I guess those that are uninitiated um i like to tell people it runs the world in the background because its install base is all of the foundational industries that make modern life possible. So if think of power generation transmission and distribution...so it monitors the large majority of the power grid manufacturing, of course, pharmaceuticals and really just all these different industries.

It started as well. I don't know how far back to go but essentially it can collect real time data from just about any industrial source in a speaks All your typical and atypical protocols so the ones you probably familiar with our OPC and Modbus and DNP three central archive. So, you know the term people often use as a historian has a lot of baggage associated with it.

now kind of seems out-of-date but It is The mechanism by which you bring in all this disparate data bringing into one common Place that you can then start making use of it. There's more layers on top of it But um at least let's say we could start with that. Yeah sure I'm gonna come to details A little bit further around.

Very good David, you've been sharing your knowledge or experience through videos on LinkedIn. Meanwhile understand on your website as well and I suggest that we make them available to a bigger audience by talking about the concepts of some of them here in our industrial podcast. so just let me start.

it's not about Just copying one-by word You know? that appeal to me and I'm sure will appeal to our audience as well. So, That's how we're just gonna be chatting about those topics. Me starting with...I think it was one of the final ones you did so far? No! It is not true. in April actually here And if there have been updates then share them too.

You talk about querying the Pi system with local LLMs Alternatives to send your operational data off to entropic and open the eye. What is that about? Right, so for one just I guess back me up for a second. The whole idea here with putting these videos out Is That Like you know These large language models they kind of came onto the scene.

You Know somewhat's you know everyone kind of realized it. suddenly With the chat GPT It just feels like humanity, you've got this new stick that landed in their lap. You know it's like. well what do we do with this thing? Kind of wave around and see what is good for... So I felt a time for everyone or at least myself to explore what the technology can be done on an intersection between what you are familiar with.

so the bi-system was natural place to start those explorations And thats sort of what these videos came out as sharing those findings. as I went and sorry the next part of that question was... Yeah, no sure. So the first topic is about where you have this video about querying the PySystem with local LLMs And your sharing two alternatives which are really in case You're not allowed to send your operational data off too for example Entropic OpenEI.

so what were the alternatives? Yes, right. What I was showing Was so I mean the default that you normally think about is your querying. this large language model has got to live somewhere Right? So typically it's on servers That are owned by Anthropic or OpenAI Or provider of choice and I Mean industrial environments Are very sensitive environments.

The data Is Very Sensitive. that parts of the network are very sensitive. Like things like go wrong at that network level can leave craters in the ground, you know? The PLC's and all that. so often those Those uh Parts to the Network Are Very Sensitive. It Often Gets The Data Often Can Live On The Corporate Network.

That'S Usually Where A data Archive Would Live For The Pi System. But Even Then You can't just send that kind of data off to enthropic or open AI without really thinking about what you're doing. So, that's one level you could work at if you have maybe some business agreement with them. If you also have a cloud architecture... ...if you've got Amazon and Azure... ...and your tenant who is comfortable with the operational data living in….

…you can actually host and run models within your own tenet. That would be good option you know, wanting to keep data within a certain boundary. So it's just all about finding the boundary that your comfortable with and The ultimate option is of course You can just bring the model to whatever part of the network you're comfortable With-you Can run an open weight Model?

You can buy the hardware And have It Run Locally. so in some Of the videos I've shown i showed those three different levels of where you can run the model and Just considerations That come along with this. Yeah, it's interesting. You know Robert and I we've been running this podcast for roughly eight years now And being based in Europe at the very beginning you know We talked a lot about that was The European German customers In the industrial setting not sending their crown jewels To the cloud and we always thought this was a very typical European.

Those Europeans are. they're so careful And I must say, we haven't talked about it for long time. The topic wasn't there for a long time. It is kind of interesting to see you back again. As you said There's all these different options. But in question... I was nevertheless going.. Is not really a question i think.

So as You Say If You Have a deal with, let's say AWS or you said Azure. Then that's the deal you have I mean. but in that case just to be clear You're still sending your data out To A partner That you have a contract With right? Right exactly i mean it's not really Your server It's somebody else's Server and thats The non-starter.

for alot of people Its Still very much the Case. In US too. A lot Of companies aren't Comfortable with doing that And then Thats okay. If thats What we need to work with Thats fine Yeah, right. But nevertheless you say there is the option today to run things locally. I guess it depends a bit on The people that make your connectivity too.

let's say AWS or Entropiq OpenAI or Azure whatever To an outside service. Are they the same people? Do that need the same capabilities from a technical perspective of running a local solution. You talked about running Olana on a mini PC, couple things. maybe we come to that later or do day. I mean would there be the same type of person's time... Same type of interest?

what did have different capabilities and knowledge? I guess it can open up some different use case, I would say largely the same. These are all just sort of swappable one for another. but let's say...I mean i would imagine it being pretty common where you're comfortable with some data like a breaker status tag or something living in the cloud very different.

if we were talking about intellectual property trade secrets type things Having a model on the local network that can work with data in context could definitely open up more use cases. That wouldn't otherwise be possible, In an AWS or Azure type of situation. Isn't it special as we're talking inference?

I mean if were talking you know wanting to do things knowing things humans agents We come to all at the point of infants where things are happening with a rubber hits the road. At that very moment, isn't it? You know even good thing to have. you know the connection within your own network so to say rather than needing to go outside with the potential there's going to be glitch whatever millisecond glitch that you cannot really use right.

yeah I mean you are contingent upon having a good internet connection. You know, it's these are still. I mean i like to say These Are The Early Days? So the full realization of the use cases Of This Technology Have Yet To Fully Play Out And At The Moment at least... I'm Not Seeing Like The Mission Critical Workflows That Must Occur Or Else.

There'S Major Problems. Those Aren'T Quite There Yet And that is a good point. If you are doing something along those lines, then yeah be resilient to connectivity outages which means the numbers have got to get crunched somewhere. so it has. I read about, saw your video. It was using cloud code to talk to the PI asset framework which is...I understand that by getting in contact with the model at a right time.

so i think this topic would like you spend little bit of time on together with you, I mean also moving into knowledge graphs in general. The need for a Knowledge Graph for the large language model. tell us what is your knowledge? Your experience there... Sure! So i'm in the general name of the game kind.

like you mentioned it's getting the right information In the Right Situation to the Model. that takes several different mechanisms based on situation two. I mean, these models can be very capable but they need the data to be available in a decent format inside the context window. And that's really where a lot of engineering and creative design decisions go.

so having right systems work with is important too. So when we were talking about the Pi system those extra layers are one left out at beginning as asset framework. that is where you can start to get the notion, instead of having in the archive just a big pile of tags. So like I mentioned break your status earlier or power being measured Just one sensors data and bunch of disparate names an asset framework.

You actually build out the notion of assets. so my mind always goes to wind turbines as an asset. They have attributes associated with them, so wind turbine will have power. It'll have wind speed going across it the oil bearing temperature and those. all those attributes point back to tags on the data archive.

So its a way of structure The information...the assets themselves. they're organized in hierarchy. There's notions that these turbines exist under this site. This site lives under this region. You start to create. You know, people use different words and ontology is semantic model. But really it's just the having data in a clean readable context for the model to us.

And yeah that makes all of difference. This not new problem either. We've needed this before the advent of LLMs Yeah, but we didn't. most of us do it before. Oh yeah maybe now where if we say we want to do AI or whatever We say you wanna Do? Maybe when I were gonna be forced To do. and the way that You already start describing this brought me as a co-producer Of The OPC Foundation podcast.

But It's not about That. there is something For example OPC UA does with an information model. you referred to it before. But we're not talking about that, but that's the way of thinking about logical winter binders. there is a foot under whatever and as I was up there on this piece isn't electricity and all know things like that now?

You didn't second video which goes little bit more in detail where you talk from tax to ontology. And that's why we talked about, I understand the importance of putting stuff into an ontology and always trying to find out as you say most of us haven't done it. those are our listeners who have already in their past understood that it's a lot of work to put stuff, for example into an OPCUA information model.

Would that be maybe sufficient? Or you say no because you refer to these three different terminologies on which the ontology I understand is big one knowledge graph ontology. Is there something missing then? And how do you get? even if we move to the large language models, what is it that they need today so they can really be optimally working for them.

Oh man yeah! There's a lot here... For one I did an article from Tag Dump To Ontology which was actually from my partner Rob Reisman that he wrote, it's a really good article. It's up on our website. but a key distinction in the article maybe helps kind of differentiate the idea. making an OPC UA model versus say something like and AF or asset framework are elsewhere is the concept early binding context late-binding context.

so I don't want to confuse the terms contexts as in that's sort of separate here, but like the actual structuring and organizing of the data creating that ontology yes can happen on what I would say is closer to the data source side. So OPC UA has definitely...I've seen how you could have a nice hierarchy of item IDs i guess?

And thats really great! That still looks at early binding of data. I'm not used to having that. So normally, i think of the uh delayed binding of data. so let's say an asset framework after you know...I think of OPC UA as one of many data sources and there are many instances in my experience too. when when data collection is set up it's done by different people at different times we all have a different idea what's the right way.

I mean, it's really hard too. You could have you can be collecting what ten thousand or more different streams from a single OPC UA data source and that is a lot to try and go through an organize. i think the argument of shifting now where makes sense To try and tame that data closer out to the edge?

I'm starting to come around to That but What I've been used to Is just accepting The fact All this data is born wild, and it just needs to be normalized at a later point when you actually are aggregating the state across several sites. You're not only collecting that process data but also getting third-party data like your pulling in weather information or market conditions supplier information.

That's where we start building enterprise asset model relating all of these stuff with each other And so that, the enterprise asset model when achieved properly. That is an excellent contextual foundation to large language models. but again it's something that has always been needed and a few people really have come around.

But is it easier these days to gather? So if we say, let's stick to the knowledge graph or the ontology. You choose it can you? these days as you refer to Claude As an example there are other ones. Can you just ask Claude To produce for And in, let's say maybe interactive way ask Claude Claude what do you need for?

You know I want to have this knowledge graph of XYZ. so Maybe interactively. You can define What it is for the complete company a division a subset A specific production line doesn't matter and Claude Is going to tell you okay i can Do that For you but you Know you point me To whatever It is That you know i'm gonna be Making it for us.

This now theory sharing here or is it, should we be using the tools that today have to? We could not imagine having like a year ago. Yeah I mean so for me that's theory territory because i haven't been quite hands-on doing that. but... That goes to suppose you can do something along those lines But just with everything its very useful tool.

Um, but you need to set it up the right way and we'll have limitations. Like there's just certain tribal knowledge that is just...that lives in the operators' and technicians head. That needs to get translated And they can be a useful tool In getting into facilitating that process. We actually do have a You know I'd say like a prototype product But something kind of help along those lines.

We call it The Intent Repository And the idea Is To Take These Streams And Provide A Governable workflow for having all these different people incorporate that knowledge into the streams and have it become more normalized, and traversable in LLM. That's what I would call the ontology... All our terms.

this is kind of new you know? We can go into that if you like. Let's move to the next topic here, which is very interesting for me as well. there was one about Pi vision displays in ten minutes. Is it something? You can share with us this and I think One is The faster way of doing things or producing these displays.

question I have there, who are we producing the displays for today still? For the same human? is that the same person. Is it a different type of person? maybe not to day but Maybe tomorrow. and does this person in The near future with the use of large language models? agents they'll need the same Type of information that we've been used too for the last whatever Twenty, thirty years.

you know all the charts maybe to red lights. The green lights. Does the agent need that information then? The agent or the LLM needs the access but they don't need To have it shown in that way I believe. so that's another topic and very interested if You can share your thoughts on that. Yeah Okay So yeah the pie vision thing just give context here again.

in PySystem, you collect data out by the sources where the data lives. You bring it together and an archive. You structure that data make it kind of give it that ontology at the AF layer And on top of that you can build visualizations for typically humans. It's meant to be very...it is a user friendly interface actually meant people who are not you know, technically inclined to just go in this drag and drop interface.

You can help yourself And that's part of the idea here is so much. This data Is not being fully used To its potential So the user That you Can make it to pull that Data out? And get It do so he can. Interface with your eyeballs is really important. so py vision is meant to be a very approachable slow learning curve type Of product.

but you Know there's one of The capabilities of large language models too. uh...you can produce data in like you can produce text and Very well, of course. Sorry is not the cleanest way to describe it but yeah The idea that I've kind a couple times in my videos It's really the pattern Is there is some object in system?

That is represented by text. so AF structures Can be exported xml files pi vision displays can Be exported to pdix files. essentially its just json under the hood. And So if you can get the model to work really well in that format, A lot of the work went into producing examples and also finding good, basically good north stars for it to use when following these requests.

And those were all wrapped up in a skill. I'm sure who might be familiar with that but essentially its just way conditionally reveal information. so you can kind of bundle this information up and it's very easily accessible in a conversation where we say, hey. Can you do me a Pi Vision display? It'll bring all the information into the context window.

that is part of having the right context at the time for these models. So basically with that You start producing objects Especially with the Pi Vision Display. Getting to show up on the system is easy because Pi Vision has an API And you can just publish that display through the API. So in that video, I type to it a little bit and then as soon as the model is done new display shows up which was pretty cool but like i mentioned PyVision's very user friendly type of application.

so its already easy-to use But kind need show natural language interface. And still, because you shared about the below level to Jason. You say all kind of things. if we talk about large language models maybe it would be talked with agents which don't go into details today. but no agent a piece or software.

basically that I'm asking. do something for me through the LLM. you know I'm asked in the Llm to do some thing. It will find its way thru an agent. The agent as long is has access exactly at lower leaven let's say Jason as an example, to the databases. The data or whatever kind of format they're in. that is what the agent will be happy with right now.

The agent doesn't need see it in a form humans would have needed. so do you consider changing something there if we are going leave more things? More doing things for the LLMs and agents Is the job of a person today looking at these graphs, is that job going to change? rather two words? I have no idea.

You know maybe confirming something that he agent comes up with and clicking on confirming or...I don't know. do you see it changing for this reason? I can definitely some changes coming along those lines. really what we're talking about? formats for ingestion. You know, the display is a nice format for ingestion for humans because we got our eyeballs and pretty amazing to depending on what you're showing.

so could be useful for large language model in I think more narrow situations. if there's like... If he have heat map or something that can put into multi-modal it takes both vision text But for the most part, yeah. I mean...I think you're right like if you have these JSON parsable results LLMs will be very happy to crunch on that and no need to bring a display in.

Okay, I've got a final topic here which was not part of your videos. Maybe you're going to do that sometime? You were suggesting are gonna continue with the. once you have. it's a LinkedIn post and there was about AI potluck if i pronounce that correctly. It is an open source stack. understand industrial environments where planned data can't leave this side.

It's maybe similar to the topic we had at the very beginning, Can you talk a little bit about what that is and what you've been doing with it or playing around? Yeah I mean i'll say its more of just like... What..I kinda want put out into the world Just amplify its presence. so hopefully be more of a thing.

You know, I haven't done any explicit work with it but in general i believe and having a viable easy to reproduce open source stack as the solution for In this case large language models. so I mean imagine like we were talking about company that For The reasons We mentioned want To have A local model running inside their own Network.

It's gonna Be. I Mean you Have to bring in let me for sure at the moment you have to Bring in some pretty decent expertise to make sure it's set up really well. It would be great if we can just snap these things together like Legos and that's kind of what seems this group is going after. so I'd just really liked the idea, you know?

It's similar with working data collection out at The Edge having a good open source solution for that. You know its terrible affliction that industrial companies get locked into a vendor. So when I see an open source like ethos, and one is want to see that do well. Sure there was the letter by Jensen and Vidya And most of the other companies except for some of the closed models ones i guess but sure i mean Open Source is a big thing.

it's not about choosing one or another. in the end each us, each of the listeners. Each of your customers. you know they will make their own decision but it's one of the options that we have at hand right now. I think you were sharing two specific. could've been other ones i guess open weight models and K three GLM five two.

I hadn't seen heard about that one. And then every Sunday, it's in my calendar when i've got nothing better to do and look at this the arena leaderboard just because I used to be doing that for many four as long as they exist on day. yes They are both into top ten still with a lot of clothe opus in the top.

It's not about going in there. But just an example yeah? I believe That's what you're. so within this AI potluck environment, I guess you can be choosing whatever open-weight models that you would like to then. Right? That will be the idea and yes it does. there's a lot of nuances in making sure they're set up just right.

when we go between different models which is... The part where they hope eventually we could streamline. but yeah those models you mentioned GOM five two Kenny K three its already old news There's. I mean, it's a man. And and i think we were kind of had a moment recently where the open weight models have just really picked up pace on close to gap with front gear.

And they're based in their punching way above their weight like my favorite example is quen three point eight twenty seven billion parameter model? I mean its trading blows with like with claude sonnet five. I mean, Clotsonic Five is really good. You can put that in a coding harness and get a lot done with it.

And it runs on a single GPU. so its kind of incredible. we have these options now like people say this the worst that will ever be. So as closing our final thought where are all those bringing us? It's part your work I'm part of your work, where you're helping customers around the PIE system. And this is showing a future for them or already today?

Maybe share with us what percentage are in that area and when we stand here... And the things that we talked about, videos you put out to topics large language models in an industrial environment. In the area of the pie system specifically where is it going? bring us the next five, ten years. What are changes less at technical level and higher?

Are we going to need people who do things you already suggested depending on quality of certain applications or different education? What is it going to be doing? To, you know, manufacturing we hear about here in Europe. Manufacturing United States Is there going to Be easier for that reason? what was a?

what are things That You see because of large language models AI In an industrial specifically pie environment? yeah well It's always dangerous to predict the future. sure better Better you do it than I did right? Yeah Thanks. And for one too, I mean you mentioned like how much of this am i currently doing?

and I'll be upfront. Like it's very a low proportion of the work right now but This is what? Yeah, and one thing too is that like if you freeze these models in place. If all development stopped people type Oh they're going to plateau. it's then maybe the? well don't know. but even if it stopped right now I mean we're just figuring out how to really integrate these things and work them into an environment.

so integration. You can't just drop a you know, if you drop a GLM five three or whatever like a big heavy model or something like that on your network and expect great things to happen, you got to integrate it. they have a simple like, a dump of information that it has to trudge through. I mean, It can reason over a pile of unorganized tags just about as good as you know.

so its matter having these things really well integrated and have the information represented in a capable way. And then there's all this strategies bringing into right context at the time We didn't even get touch on knowledge graphs or my version but basically having structured representation the information about a certain topic or process that can be easily traversable and you bring in only what's needed at a given time.

So all these things really still need to manifest themselves in an industrial operation, That's for them here. try help companies do. And there is a lot of upside too. it I mean i think There Is The ability To Bring In Essentially If It's Well Integrated You Can Have Basically A co-pilot operator, you know overseeing a process and understanding what's going on.

And What will need to be done? It doesn't sleep it is always giving full attention can work faster than You. so having it all integrated in one place I think as we're gonna see more of that amount, the quality of integration. Of these models will increase with the systems that are present at companies and The economic value of that we'll continue to be unlocked And I mean a more personal level.

i see That is super important because you know These. like I mentioned At the very beginning this Is what me? This is the foundation of What makes modern life possible. This is what gives us the prosperity that We have when we turn on the light switch in. lights come On. You Know that's Because this miracle of modern innovation behind it, that's made it happen.

And if we can use the information you know comes from those... That process to make it run more efficiently We have more and more sustainable prosperity. Yeah I think its a really fun area. Thank you David, huge potential I hear but a very clear message we got from you. It only works if it is going to be integrated.

well that's yeah i think its an important message. So I'm sure those people that haven't, most of our listeners have heard. Those who are new to the message and weren't really sure about what LLM can be doing in an industrial setting now maybe a stronger feeling. but before they think or believe yes i am just going to do AI tomorrow you say you need integrate it such as the LLmcan be giving you the optimum support.

Yeah, I mean what people start exploring it and we want to give people tools to do too. We have our MCP server for the Pi system publicly available The Community Edition. You can download. use it Explore. Great David! It was a pleasure talking with you. Keep up good work sharing your industry knowledge experience through videos.

We're gonna most certainly put those onto the podcast as well for people to look at. Keep it up, sharing your experience. Looking forward to seeing another video again! Yeah thanks Peter appreciate it. take care have a good day.

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin