Is AI putting an end to traditional factory jobs?
In this episode of Your AI Injection, host Deep Dhillon sits down with Chris Pickett, CEO of MASS Group, to explore how AI-driven manufacturing is rewriting the rules of industrial production. With automation already transforming warehouses, Chris explains how AI-powered execution systems can optimize workflows, improve quality control, and even detect defects in semiconductor chips before they become expensive mistakes. But what could this mean for the future of factory workers? Will AI simply replace routine, repeatable tasks, or is a fully autonomous factory in our future? Tune in to uncover the economic realities of AI-powered manufacturing and whether human oversight is here to stay.
Learn more about Chris here: https://www.linkedin.com/in/chris-pickett/
and MASS Group here: https://www.linkedin.com/company/massgroupinc/
Check out some of our related podcast episodes:
Get Your AI Injection on the Go:
xyonix solutions
At Xyonix, we enhance your AI-powered solutions by designing custom AI models for accurate predictions, driving an AI-led transformation, and enabling new levels of operational efficiency and innovation. Learn more about Xyonix's Virtual Concierge Solution, the best way to enhance your customers' satisfaction.
[Automated Transcript]
Deep: People often like underestimate the economic calculus with automation. And I think that's the same thing that's gonna happen with AI, that at the end of the day, there's a lot of tasks that even GPT already exceeds the cost of getting a decent human to do the job.
Chris: Like you're spot on. I give you another example from my past, an automated fork truck, called an ATL automated truck loader or automated, vehicle. Five years ago. $250,000 a piece, Way too expensive to replace an hourly worker doing warehouse work, Now, last year, two years ago, they're under a hundred thousand dollars, right? And so now the economics changes.
CHECK OUT SOME OF OUR POPULAR PODCAST EPISODES:
Deep: Hello. I'm Deep Dhillon, your host. And today on your AI Injection, we'll be diving into the intersection of AI and manufacturing and traceability with Chris Pickett. Chris has an MBA from the Haas School of Business at Berkeley and is the CEO of Mass Group.
We'll discuss how Mass Group is modernizing manufacturing software and explore AI's potential in the semiconductor and manufacturing [00:01:00] industries. Chris, thanks so much for coming on the show.
Chris: You're welcome, Deep. Great to see you again. Um, and good to be here.
Deep: Awesome. Well, maybe let's just get started by, tell us what a typical customer for you is like, and what do they do without your solution?
And yeah, it's different with your solution.
Xyonix customers:
Chris: That's great. So first, a little bit about us mass group. we provide a product called traceability made easy TME, which is, by the classical definition, a manufacturing execution system or MES. And so, what does our average client look like?
They're a discrete manufacturer, with, 50 to 500 employees. So we're targeting the small to medium business. But very complex, needs from a manufacturing perspective. So our largest segment is semiconductor. The second after that is aerospace and defense. And that's kind of where our, our largest customer sits as the intersection of semiconductor and aerospace and defense.
that's a little bit about our clients and what they're doing without [00:02:00] our solution. There's kind of one of three things. Either they haven't really started digital transformation at all, and they're using paper and, simplified documents to run their production process.
Or they have, implemented an ERP and started their digital transformation for the business side and high level macro planning. but they haven't penetrated the shop floor yet, or they're trying to use their ERP for some shop floor management tasks.
And, that's, where we usually see them coming from. So either they're working off paper and spreadsheets, they're using their ERP for some manufacturing controls, or they have a competing dedicated MES solution,
Deep: I'm gonna jump in here on the acronyms and like force you to expand them all and define them.
maybe just Tell us a couple of sentences about what the ERP approach they're using looks like. what are they actually doing there? what is an ERP?
Chris: Manufacturing, you know, ERP, it's enterprise resource planning, right? And it's usually the first step a [00:03:00] manufacturer takes to bring a big external software product in to help them run their operations.
it helps them with their annual planning and their monthly planning, but doesn't go down to the production, minute, hour, that MES does. And so the enterprise resource planning, yes, that's very important, critical for the business, but usually not suited for running manufacturing operations.
Deep: So it's like helping them with high level stuff, on the financial front, on the human resources front. Yeah,
Chris: exactly. And even some inventory management, but it starts to fail when it gets to down into the transactional level. I should say operational level of the operation.
Deep: so you guys are focused mostly on the actual operational floor, like the manufacturing floor. That's correct. you maybe just walk us through and describe a typical manufacturing floor, what does it look like when you walk in that room? How big is it?
What are the machines? How many of them are they, and maybe, walk us through and help us set up the [00:04:00] problem that these folks face.
Chris: For sure. Absolutely. So I'm going to stick to our kind of client base. So imagine a, let's say a hundred, employee manufacturing company, right? not huge, but also not just getting started.
I'm going to explain the manufacturing flow in a logical perspective. the first thing you need to manufacture a good is raw materials that you're going to use to make that good. So all of your raw materials receiving. and when they're brought into the building, they need to be organized and controlled and monitored.
That's done by the MES, the inventory management component of MES. Then you've got to be able to monitor and analyze all of your assets. All of your equipment is either connected to or monitored by your MES so you know whether a piece of equipment is running or not. And then you've got all your maintenance of your equipment that you need to plan and execute in a very organized fashion to keep everything running.
And those kind of things all add together to make a manufacturing execution system. Now there's much more than just receiving raw materials, putting them [00:05:00] together. And then shipping them out. But that's a simplified view of what it looks like. one of our, clients might have 10 or 20 pieces of equipment, say semiconductor equipment, that a process is going through.
Traceability Made Easy, our product, tracks every step of every process and every quality characteristic of the product so that all the decisions can be made by the leadership team on what to do on the shop floor.
Deep: Okay, so maybe we'll walk through one of the semiconductor manufacturing scenarios just to make it really crisp and visual.
for folks that are listening and for me, I like to be able to imagine stuff. so what's exactly, are the raw materials that enter the room? and then what exactly are the machines? Maybe walk us through the,
configuration scenario, like, cause that would probably help us understand what it is that you're setting up to interact with and trace,
I think it would be fun to walk through, like, the raw material, floating through the system, and then we'll call out all the AI issues along the way, or potential.
Chris: absolutely. A [00:06:00] semiconductor, it's funny because they've got one of the more simpler, raw material, needs you could have as a manufacturer.
It's different in every industry, but their number one, raw material is the wafer. And let's just say it's a 300 millimeter wafer, which is The largest industry standard is used now, they're putting that in at the front of the process and starting, wait, wait,
Deep: just a second,
I want to stop you, describe what a wafer is, what's it made out of,
Chris: generally, wafers are made out of silicon because it's a semiconductor and that's what most chips are made out of these days.
We actually have quite a lot of clients. They're using quartz wafers, which are glass, and they're used for different applications for making optical materials, optical, you know, computers or other devices. So it's either a quartz wafer or a silicon wafer, and they look very plain, just that material. then they're going to go through a number of machines that do all these semiconductor processing steps.
Starting with, chemical [00:07:00] vapor deposition, for example, where you deposit chemicals onto the wafer. Then you create a photomask where you, Put incident light onto the wafer and create a pattern with the light, which then changes the properties of that chemical and allows you to remove where the light hit and not where the light didn't hit.
And so these are all complex processes. One machine And the whole,
Deep: the whole point is you're actually constructing the core of the chip. all the Boolean operators, all that stuff is represented in this. process and execute it against the wafer.
Chris: Yeah, it really is. Now that you mentioned, it's like additive manufacturing before real additive manufacturing came along because you're building this wafer up layer by layer.
And you do this on a big 300 millimeter wafer at the end, You cut it up into all the chips, right? one wafer can become hundreds and hundreds of chips. that wafer had a certain quality characteristic at the end that can be measured and [00:08:00] impacts how much yield they get of chips out of that big wafer at the start of the process.
Deep: Awesome. So I'm going to stop you here because I really want to understand, understand this process. So at this point, in comes, a stack of wafers, and then, what exactly does your system do so far? Like up until the wafers have entered the office, like you've configured maybe who's sending you the wafers, what type they are.
size, dimensions, attributes of them, and then do you know when one's been picked off and put onto a, I'm envisioning a conveyor belt of sorts?
Chris: Yeah, we can do the purchase order right from the beginning to purchase the materials, know when they're coming in, receive them, store the supplier, the batch number, the lot, the date, expiration date, other quality characteristics that might come with it.
Then you know that you have the material you need for the production before you do it. Then when you, start the production, the first thing you do is consume the wafer in our system. that wafer is now going into a production [00:09:00] process the tools are building the layers on top of that wafer to create.
Okay,
Deep: I'm going to, I'm going to stop you for a moment because I want to understand this too. So how does your system know? That a wafer went from a pile to being on the conveyor belt and moving. Do you have cameras? is somebody staying in a terminal and entering it in like
Chris: everything from somebody entering it in to, a API or, web service to connect directly to the machine.
you have things like, recipes as well, communication back and forth between the machine to tell the machine what to do, what process to run on that wafer itself.
Deep: so your system can interface at the machine level. Each machine has a varying levels of, interrogation, like information you can interrogate, and some of the machines let you control them, so your system can integrate at that level too, it sounds
Chris: yeah, we have fully automated customers where the system really does everything. and then we have very manual customers where they still have their team members putting in the data and telling [00:10:00] the machine to start, et cetera.
Deep: I get it. So, when you said you had an array of customers, everything from, manual and paper processes, you were being quite literal.
they might've in the olden days, like had like a log book and somebody logs in there Hey, we just took wafer XYZ, stuck it on the conveyor belt. that got upgraded in the digital transformation phase to maybe a spreadsheet or tailored application, but again, requiring human input, then that got upgraded to network accessible machines that could take care of that process by itself that you can integrate and you have operate in all scenarios.
Chris: We have to operate in all scenarios. We're targeting, small to medium size manufacturers. we're not going after the big enterprise, which is a different game. in the small to medium size, we're going after all those people. And it's this natural evolution as you get more sophisticated as a manufacturer it's this trajectory they follow as they develop and grow.
Deep: Okay, awesome. So let's go back to the wafer now. So now The wafer Is on that conveyor belt [00:11:00] and it went into the first machine. what is that machine and what's happening to it?
Chris: Not quite a conveyor belt. Usually wafers are moved in what's called foops.
you might have 25 wafers in a big cartridge called a foop. and that cartridge is loaded into a piece of equipment. then it takes one wafer at a time this equipment might be the size of this room or it could be a small benchtop tester for testing the wafers.
it might spend hours in one piece of equipment. And this is where semiconductor again, it's very different than other manufacturing. A wafer can take, 30 to 60 days to make into chips. And that's because it has to go through, process after process, hours and hours of all these steps.
And there's probably
Deep: like drying time or something, Exactly. To that latency.
Chris: And so when it goes in, it depends on the type of machine, but generally these are very specialized machines that have one specific function and then they just get run through. 30, 40 times to do it over and build those layers.
Deep: you have a different mask for every layer so that you're creating a different, subsystem of design components on the [00:12:00] chip for every layer. And that's what makes your end device in the end.
I don't know how deep I want to dig into semiconductor, but I, I have a, just so you know, I was in Portugal not too long ago, and, and toured a, this is, the manufacturing mental model I've got, this is, It's like circa 1700s, but basically like income, the raw materials, which were sardines, fish.
Deep: And then there's a bunch of people and your foop, was, they had this big tray with, like a bunch of these fish in there. And then, and there was all these discrete steps, stations, sets of people chopping and dicing. there's a physical floor, stuff's moving from station to station, and each station there is an operation that takes place, manual or mechanistic, then data that needs to be pulled into your system at every point in time, and then you're going to operate on it.
Chris: That's a hundred percent correct. So I think you hit on the key point there, which is in manufacturing, the real traceability and data that you need is. Every step of that process. What is being done by [00:13:00] who and the data being collected. in M.
E. S. We're tracking every single step, and that's all their data to be used at another time for, decisions on the product for understanding what happened. Why was yield higher or lower? Why was throughput higher or lower? And making sense of that data can be very challenging, right?
Deep: Okay, I think I got the high level, landscape.
maybe let's talk a little bit about the data that you actually are pulling in. and a little bit on the configuration process. I'm going to just try to Say to you what I think happens and then you tell me where I'm wrong or don't quite have it. So I think if you go into a, I'm going to guess how you interact.
you sell to one of these small to medium sized manufacturers. The first thing you probably have to do is, get on their floor, walk around, and see what they do today. you're going to have a sense of, what their equipment is, and how you're going to interact.
you probably have to build a whole model of it so that you've got a diagram of the discrete steps. Then I [00:14:00] imagine you have to, like, interface either with their paper process, Digital process or the machines, systems directly you probably have, folks that show up on site figure out what's going on, you probably have written adapters for, all of the. normal sets of types of machines that are in your space, and then you're able to, install them, fire them up, connect with the machines. You probably run into a list of problems there, where some machines maybe have stuff, but you have to physically access the information because they predate the networking age.
Other, times, you have everything from paper to spreadsheets to tailored applications that might have APIs that you can get at. And so you have to figure out how to get to it all. And then once you get it all, you're basically after what are the inputs to that stage, what are the outputs from that stage of the manufacturing, and then you probably have to have some kind of observation or interpretation on the outputs.
so that you can look at an [00:15:00] image of a produced wafer and, identify defects or problems
Chris: Yeah, spot on. I mean, going back to the beginning. you mentioned the implementation. what we do as we're working with a prospect is really look at where is the value in their operation and how can we help them extract and collect that value.
And if we can't, then we don't have any, business doing business with them because we want to create value for them. So we start with a value creation, investigation with them before the sale and during the selling process where we can align on.
Where is the value? Is it avoiding unplanned downtime? Is it better quality? Is it traceability and guarantees to your customers down the stream? And how do we extract that? A lot of the time you mentioned integration. So first, you're right. we need to manage equipment that is 40 years old.
just stamping a piece of metal. what's the data that we want an operator to put in from that? But we also have these 10 million semiconductor pieces of equipment that can put out, huge quantities of data every [00:16:00] second on a huge amount of process characteristics. we've got to live in both those worlds.
So we integrate it. With, our own web services with APIs, we have partners that have platforms for, certain types of integrations. we need to deal with typical manufacturers, which usually use a protocol called O. P. C. there's the, Internet of Things.
technology, MQTT, and the sex gem interfaces, which is semiconductor equipment. so you have to live in this world with all these interfaces we do those plans in advance and have that built into the implementation plan. You know, the intersection of information technology and operations technology is really important to us, too.
So we spend a lot of time talking about the integration points with other systems. We already talked about the E. R. P. Side. And if they already have the high level plan in the E. R. P. And they're creating sales orders in the ERP. We want those to come straight to our MES to simplify the communication, the coordination of the whole floor and the whole company.
that needs to be spelled out in the [00:17:00] implementation plan that this is a critical integration. And there's other integrations like that too, throughout the operation.
Deep: So my first thought is there's a lot going on here for a given customer, like there's a lot of customization, that maybe isn't the first time that you're doing it, but it seems like there would be a lot of value in getting good at a particular sector So that you can have already seen the machines they're dealing with or already seen the software and already have adapters for it as opposed to like having to do the heavy lift of Interacting with each of those pieces separately So can you talk to us a little bit about what's your go to strategy here?
Like you mentioned, semiconductors and aerospace and defense. I'm guessing the aerospace and defense is probably where you're doing more of the custom stuff and the semiconductors is maybe where you can get a little bit of that repeatability, but talk us through some of the challenges you guys experience.
Chris: Yeah, you're right. It's this challenge of, being general and horizontal versus vertical and specific. what's interesting in manufacturing, versus other markets is, it changes pretty drastically, as you grow to the enterprise level in manufacturing.
At the small to medium sized business, I would argue doing the basics of manufacturing. Is much more important than doing some specific vertical need of your industry. most of the time receiving raw materials in every manufacturer is pretty similar, right?
It's not that different. Well, You put things in a warehouse where you put them storm safely organized, how you do your monthly inventory controls, your annual inventory controls. as you get really sophisticated and you grow to an operation of thousands of people, that's, massive.
Now you need really specific vertical features. So we're much more focused on manufacturing for complex, discreet goods in the small to medium size. range of customers. the beauty of TME is it's very configurable and doesn't require customization on implementation.
So back to your implementation question, we work with the client. On their first production process build out in TME when we build their production process and digitize it into TME and after that they can do the rest themselves because it's all through the UI and the design of the workflow in our software
Deep: so it's not like you're Sending out your solutions experts to implement everything on behalf of your customer, that might be like a white glove service or something, but for the most part, they're doing it.
does that come down to them writing adapters for particular machines to get particular data outputs if needed? can that be shared with other clients? how does that.
Chris: Yeah, integrations is something we need to continue to get better at ourselves. we've been developing new APIs for future integrations.
Historically, we've done things one off based on the piece of equipment, right? And the customer. And to your point, that's not Scalable when we got our partner earlier this year, symmetrics on the sex gem side and really all integrations, they make it much easier for us to go in, even with them as a partner to help with equipment connections.
So if we have a big client that wants to integrate 30 pieces of equipment, we're gonna bring our partner in because it's such a big lift and shift.
Deep: Okay, let's start talking AI here. So when these, Customers come to you. You mentioned that they have like different types of problems and you're going to try to, speak to the key pain point they're having.
what's the typical number one thing they have issues with? Is it traceability? is it trying to optimize throughput and they're just not getting the output they need? Is it that they're tossing away too many stuff and they have too many faults? Is it some combo of all of the above? ultimately it seems like they would have to come to you because they're frustrated with the inefficiency of their current process.
Chris: Yeah, I think the answer is what you said, which is it's a combination of all of the above and one of our challenges. Every prospect that comes to us a different one is at the top of the list.
it's probably going to be one of three to five things, their production throughput and how much they're actually producing. They're not satisfied with. They're having quality or traceability issues. They're having too much unplanned downtime and they don't know why their equipment's shutting down.
And that's kind of the First three, I would give, but there's so many underlying issues they could be going after that are all very relevant and can be business defining things. that's what's really cool about NES is we help with everything inside the four walls and even outside, but it also makes it hard to connect with that client until you get a little bit into the sale and understand what their needs are.
Deep: I imagine some of them are coming to you with a really. big and time sensitive problem on a particular machine or a particular stage of the process, but you're really after getting their entire process into your system. So there might be some friction where you have to implement a really surgical solution, but you're really after.
getting them into the full system. It does that happen? Do [00:22:00] you sell something like, where they say, Hey, we don't have traceability in this one system. Can you give us some views into that? And then over time you start pulling. into the adjacent systems or something.
Chris: one thing we've realized is a competitive advantage of ours is our ability to, be used as part of the MES. MES, Manufacturing Execution System, is made up of inventory management, maintenance management, asset and equipment management, and workflow management.
The only one of those that you have to have the others for is the workflow management. if you want to do workflow, you got to have the other pieces because that kind of ties them all together. But we have, clients already in our customer base, prospects we're talking to that just want inventory management now, or they just want maintenance management now.
And that actually allows us. to deliver one of our key value propositions, which is we're very scalable with our growing client base. And so if they're at 50 people now, and they're like, okay, I got my ERP, I want to do inventory now, but I don't want to do maintenance yet, or my [00:23:00] workflow and my production process.
Let's just do inventory this year. And then let's come back and do maintenance. when we go into 2026 or 2025.
Deep: Got it. I'm going to propose that we take a stage of the manufacturing and then maybe talk about where AI could be used to like help and whether it even makes sense or not.
And then we'll move on to the next stage Does that sound like a reasonable? Yes. Absolutely. Let's start upstream, raw materials come in, that wafer came in, that's interacting with your inventory system at that point, right?
The first thing that comes to mind is, assistive types of AI techniques. Whatever that configuration process is, making that quicker and faster, feels like maybe some low hanging fruit. And then maybe some forecasting hey, how many wafers are going to be coming in, on a time series forecasted basis, like a week out, a month out, or whatever.
Yeah. Am I in the ballpark or are there totally different things? Yeah,
Chris: You're asking all the questions I would expect somebody in your position to [00:24:00] ask. I would say so far what we found in our investigations, a lot more opportunity for AI help with demand planning, on the forecasting side than there is on the plain inventory management side.
we have intelligence on inventory management. expiring products, shortages of products, excess products, But I haven't found a lot of good fit for AI on just plain inventory management. Now, when you get into maintenance management, quality management, the production and efficiency of the production process, huge opportunity for AI.
Absolutely.
Deep: So when you say demand planning, do you mean? models that try to predict, based on like, economic factors, micro and macro, how many people are going to want to buy, X chips, therefore you can Kind of know, how many wafers you have to order.
And maybe I imagine you order more in bulk, you get better rates, all that kind of stuff. Is that what you mean?
Chris: Yeah. So let me go into it. I think it's not that, important for our existing clients, I don't know how [00:25:00] they could forecast the demand of very specific ordered products but let me give you a view into my past life.
So I worked Granheiser, Bush, Imbev the large global brewer for many years. enough years ago now that it's, it's not a competitive thing We looked at AI for demand planning of how much beer is going to be sold. Let's just use an example. How much Budweiser is going to be sold in a given gas station, you know, down the street.
And we found that yes, using artificial intelligence, you can pick out so many other correlating variables at the time we did this research, this proof of concept, you could correlate beer sales to how many employees were working in that gas station, that corner, because if it was more busy time, more people are buying, they're going to have more employees.
And yeah, it's a good proxy variable, like economic data from, just labor included. and there was many other factors, of course, temperature is huge in beer. that's the easiest one that's been in time series forecast. Forever temperature correlations, but the pattern matching of a I [00:26:00]
Deep: imagine seasonality effects are a big deal however many months before you need a chip to wind up in a map to show up under a Christmas tree, you know, in a phone or in a, a machine or something, all that stuff.
You rewind in time and that model should be able to Latch onto those exactly but
Chris: where? we focus on using ai in tme and researching moving towards is the Quality aspect maintenance and the production process. There's so much complexity And data that's hard to connect to root cause.
That's where the power of AI can come in.
Deep: Maybe let's jump into that quality question a little bit. Because I know, you and I had a conversation, I feel like it was maybe six months or a year ago about this. And I thought it was a really fascinating one. But going back to the semi conductor scenario, the wafers in the chips are getting made, you were talking about some kind of sensor that, detects dust particles, and, other things that can screw up the actual, manufactured chips of the way.
can you maybe Walk us through how that, happens today and how you envision some [00:27:00] machine learning or AI happening, there.
Chris: Absolutely. So what you're referring to is our wafer defect mapping feature. And this was I think at the time I talked to you, I just visited two of our customers with Paul, our CTO.
that's what was being asked of us, by those customers. the key is at the end of, the wafer when it's done with all of its processing and ready to be cut into chiplets. one of the final steps is to test perform material or contamination on the wafer at the end.
the way we do that is we allow the customer to digitally represent the wafer in our software, break that wafer down into as many chiplets as they want, and then track defects, by position on the wafer, with as much fidelity as they want. that allows them to know, okay, Let's say I got a 70 percent yield out of this wafer.
That means only 70 out of 100 chips were usable, and the other 30 weren't. That's a big inefficiency. now you know it's these 30 on the upper right portion of the wafer. And then you have in the system the data where your [00:28:00] quality issue is on the wafer, and you have every single step that process went through.
Which machine did it go through? through, which asset, who did what, what were the quality characteristics, the process parameters, the temperature, the pressure, the flow rate, et cetera, et cetera. And now maybe with AI, you can correlate the defects to various steps or problems on the manufacturing floor.
Deep: in some machine, for example, is like screwing up the wafer in this. Like physical region of it, and you might not have ever known that because before they would just chop up the chips I don't know how they assess whether they work or not, but at worst case scenarios they fire them up But yeah run through a bunch of software tests but if you find out then that you tossed 30 percent of the chips, but you don't really know Where or why but if you know now that it's physically happening on this region of all the wafers coming through Then you can, and if you have a central understanding of all the machines, all of the processes, anything that could [00:29:00] potentially implement a wafer spatially, then maybe you can start to build, some observability and ultimately fix the machine that's causing the problem, something like that.
That's exactly right.
Chris: So they have the data now on the wafer defects. They just have a hard time making that manual, jump, that conclusion, that root cause of what is causing this issue because it's not as simple as you make 24 wafers and they all have this issue No, it's spread out throughout.
It's this huge range of, continuous data and making sense of it is the hard part, right? Yeah,
Deep: that's a tough problem, right? Because from a machine learning standpoint, you have this understanding at the system level. And then you have, this up there's different parts where the machines, machine learning can help out.
On the wafer detection, maybe there's like, visual aspects of the wafer that are indicative of a problem. So then you could, I guess, start to instrument those kinds of, health checks along the way. If to the extent that you know them, [00:30:00] but to the extent that you're just presented with a particular scenario hey, 30 percent of our chips went bad, we don't know why it's not like enough data.
To build a training model to be able to automatically go back to the machine. it feels like a multi step process where you're taking, each input output scenario, putting in an iterative process for improving the health check on that output. And then elevating that to a dashboard or something. as you improve the quality of that health check, then you'll start to understand a scenario. So for example, you find out that I had a spatially contributing problem in a particular machine, maybe you didn't have a thing that automatically detected that before, so now your team goes off and you deploy a new model for that particular scenario.
And you catch and prevent a whole future line of potential errors.
Chris: Yeah, something like
Deep: that.
Chris: What I'm very interested in is this concept of, do you have any domain and, specific knowledge of the process or can you use, general models as [00:31:00] well? because I think that's in manufacturing and being able to scale solutions.
That to me is very, influential on how well we can do this for our clients. I can see the hope for a general model with a baseline of a whole lot of, manufacturing data from all the steps that it went through. that's just kind of based on a step by step consecutive process, and data in there, mixed in there when they do quality checks between steps, and that one final outcome at the end with the yield, having all that data.
to make a general model on defect detection and root cause, I think could be possible. but when you get to what we talked about earlier, the vertical nature versus horizontal of manufacturing, and you get into these much more specific needs of manufacturers, I think you'll need more domain knowledge and specific model solutions to solve problems.
Deep: so one of the things I like to do on the show is we basically try to cover three things. [00:32:00] the whole show's about AI. One of the things we like to do is talk about the what. What is it that you guys, Are building. and, what's the role of a I in that process? I think we have some sense of that now.
I have a vision of my head of your product. You've got this manufacturing floor. You've configured it. You have varying levels of, observability and control, depending on the customer sort of, the reality of their machinery on the floor, there's, a world where you can Improve the detection of health, throughput, all these operational parameters that can also not only be shown back to, in a visualization dashboard, but can actually be used to, to like impact and make decisions over time.
are there, so that's the what. we also like to go into the how. I think we've definitely covered the how on the configuration side and how you think about the problem. but maybe we can dig in a little bit more on the AI pieces that you're thinking of.
And then the final thing that I want to make sure we have time for, is the [00:33:00] should. . There's a narrative out there that like A. I. is, is making the world an amazing place and there's a narrative out there saying A. I. is destroying the world and we're all going to die Terminator style.
We try to just talk about the reality of it. one of the scenarios I like to play through on the should. Is if everything that you think happens, like all of your dreams are wildly successful, what's the second order effect that can cause problems that you're probably not thinking about.
So the example I like to give is, Facebook in the early days, you know, they're just trying to get people to talk to each other and to share and like communicate, and there's a lot of excitement in it. they ended up being wildly successful. The thing grows like crazy. they have no idea that there's going to be these major social transformations that happen and ultimately, you know, the suicidal ideation rate amongst 14 year old girls skyrockets.
Or there's a massive massacre in Burma, they make a mild business decision, seemingly mild, at the time. Hey, we're just going to deploy globally. [00:34:00] Seems like a reasonable thing. But not sitting down to think through we don't have anyone fluent, in Burmese, in Myanmar, so we're not going to be able to actually track all these rampant rumors that result in hundreds of thousands of people being murdered.
Not saying that anything like that's going to happen here, but what do you think would be wild success for you And what would be the inevitable ethical issues that might arise as second order effects from that?
Chris: Yeah. and you're saying success looks like for me as the CEO of mass group or for, yeah, let's
Deep: say mass group, let's take it to its end conclusion, You get all the small to medium sized manufacturers. Everybody's totally instrumented. Everybody's got complete observability into the manufacturing pipeline. Everybody's able to make products that have, minimized deficiencies and defects. they're able to maximize their throughput.
So the throughput goes up, they're making more stuff, There's a number of things that could potentially already be an issue, right? [00:35:00] There's carbon consumption that just went up because you were wildly successful. You can make stuff a lot cheaper, faster, better.
but what kinds of things, do you see as being more immediate as opposed to more abstract?
Chris: Yeah, I think, if I succeeded, we would make manufacturing software a lot easier and the manufacturing facility would be a much more unified, operation under one umbrella, one kind of tech stack without, dispersed systems out there unintended consequences is funny because this has been a challenge prior to the Internet, even with mechanical automation, right?
Going back to the Henry Ford days, as automation comes, what are the unintended consequences or what are the human aspects of that and things that can't be foreseen? And I think some of that will continue with AI and manufacturing. automation was, pretty much widely, loved by, Teamsters and managers alike for improving safety and [00:36:00] reducing wear and tear in bodies.
And I think AI where it can help with. quality or safety or, other things like that, all good, consequences. One area where I have a specific example, it wasn't with AI, but it was with automation and technology that you've got to be careful of unintended consequences. I didn't see this coming and implemented a project in a manufacturing facility, improving warehouse telemetry of fork trucks, this was a project to minimize the amount of time a fork truck operator moves without, something on their fork truck.
So anytime you're moving a fork truck without something on it, it's wasted travel, right? So the idea here is you minimize wasted travel around the warehouse. The warehouse management system, optimizing the workflow. So you go from point to point where
Deep: you're always picking something up and dropping it off.
Yeah.
Chris: And so it seems great. And we could show, man, it's saving us 5 percent of labor. What's the problem with that? The problem in the unintended consequence here was you took [00:37:00] somebody that had a job of responsibility, a purpose every day when they came in, and now their purpose became, I just do what the machine tells me, and I don't care about the quality impacts or the customer service impacts of what I do.
Not that I don't care because they're good people, They always cared, but they couldn't control it. Well,
Deep: you just put a really big piece of cheese in front of them and tried to put them into the algorithm. So they have to chase their cheese. it's the same thing that happens with the Amazon truck workers.
they're in that thing. That's why they're driving down the street faster. That's why they're peeing in water bottles in their trucks. It's because of, it's This is what happens.
Chris: Except ours wasn't driven by financial motives.
they had no change to their financial motives. It was just what we were asking them to do. before one guy would load one truck and he was responsible at old truck. Now you had 20 people and you didn't know if that truck was going to be properly loaded because they all did 20th of the work.
in regards to artificial intelligence, unintended human consequences is what I would be worried about the most. I think with our client base, I don't [00:38:00] envision any of our typical client base, 50 to 500 employee manufacturers being fully automated. it's going to be so rare, that, you're going to need the people to interpret the data to also act on the AI.
to operate some of this very sophisticated equipment still need somebody to monitor and push some buttons every once in a while. That's what most of our clients are like. They're pretty complex. They're lower number of employees, but, those people are always going to need to be able to understand that AI and, take action based on the recommendations of the algorithm.
Yeah,
Deep: So there's a couple of examples in I guess, recent history. So the two that I think of that I want to relate back to the scenario posing and see what you think. One is when, anti lock brakes first came out. pre anti lock brakes, I was, a pretty avid snowboarder and skier at the time.
Driving to the mountains, it was quite a skill to be able to drive a 1975 Ford Maverick with bald tires in high school up to the mountain. you got really good at pumping brakes and doing all this stuff. [00:39:00] And then anti lock brakes come out and there was a period of time, a good period of time where.
All cars didn't have ABS. And when you go into a skid without ABS, you must pump the brakes. When you go into a skid with ABS, you must not pump the brakes. All of a sudden, there's a brain issue here what do I do? Do I pump? Do I not pump? And it turns out, if you have ABS and you do pump, depending on who's the driver at the moment, it causes a problem.
Now, a more recent example is Tesla with a self driving car in it. those things are actually getting, like, pretty good now. Like, you know, I've spent a fair amount of time, checking out the self driving. So good, in fact, that people keep one finger on the steering wheel and fall asleep.
one of the fears I have is that we're asking humans to step into a scenario where machines are doing an incredibly good job the vast majority of the time. But I wonder, today, does a radiologist, overseeing the output of an [00:40:00] MRI analysis, still have the skills to know when it's right and wrong.
I would argue def yes, because they're actually, doing it all the time. But maybe in 20 years, is there a fear of the Homer Simpson phenomenon where there's a nuclear plant and he knows there's buttons but the thing just runs itself? does that kind of a scenario come up where we get so much automation, so much fault detection happening by the systems that the humans don't get exposed to the problems enough to learn how to intercede.
They get bored because the machines are doing everything and it's no longer fun to just be the supervisor, are there some risks on that front? I mean, it feels to me like. We've lived this scenario in other arenas quite a bit.
Chris: I was, the first one I'll come back to is another anecdote from my past that was similar.
We developed the capability at some point, many years ago at this point, to be able to remote into manufacturing equipment. an engineer can start a piece of equipment from their house [00:41:00] and the operators understandably were like, wait, that's not safe. I could have my hand in the piece of equipment and they can't start it from home.
That's not safe, right? And it took time to go through processes, business processes and policies and procedures to set things in place where you feel comfortable operating in that environment. But the same concern would arise for an AI controlling a piece of equipment is at what point do you have the kill switch, that says it can't do something.
There's something huge in manufacturing called lockout tag out. When you work on a piece of equipment. You actually physically lock it out with a lock. You disconnected from power, and that's kind of one of those kill switches physically that you have. But do you take the same mindset with say a quality parameter?
If our quality I don't say it's a valve for your heart, right? Is being made and below this certain quality characteristic. It just doesn't work anymore. and patients will die. How do you have that built as a hard limit that I could never violate, versus some other [00:42:00] quality characteristic that may have some area where, you can fluctuate around a parameter and be okay.
So I think, there's always gray area manufacturing in quality in production and a range of results and interpreting those, could get dangerous with AI where you need that last common sense. Checkpoint decision. No, you're more the AI guru than I am. Maybe you can say, Hey, I should be as good at common sense as a human should.
Deep: think like people who aren't. spending a lot of time in the kind of commercial sector. A lot of folks have this sort of negative like when stuff goes wrong, I don't know if you've seen this, but I've noticed that when something goes wrong a self driving car accidentally, kills somebody or, an Alexa, starts recording people's conversations and accidentally replies to everybody in the email.
I think there's a societal and maybe news driven tendency to, assume malicious intent. I've been working in the private sector for 30 years. I've honestly never witnessed a malicious intent scenario. But I have witnessed [00:43:00] scenarios that are so complex. Where there's a lot of stuff going on that the processes to, make decisions didn't always account for deep thought in potential scenarios that could go awry.
and integration into those processes, this is why I think a lot of people go to these weird doom scenarios with AI, because it's very Hollywood, right? You're making a two hour movie, do you really want to get into like, well, you know, there's this, competing bureaucratic processes, there's like a a lot of people, and like, sometimes something just slips through the cracks, that's not a very good plot line for a Hollywood movie.
Going back to the machines, when you introduce automation, somewhere in that process needs to be the human experts that are really thinking through Each stage, not only in the upfront automation process, but in the iterative refinement process of stuff that can go wrong, so that we make sure that the knobs and dials and controls we [00:44:00] give to these automated systems are really thought through, like the example you gave is a terrific one of you know, these, the engineers wanted to be able to remote start a machine from home and somebody saying well, yeah, my hand can be in there.
But you could imagine a world where there aren't people walking around in the shop floor anymore, that it really is already automated but every once in a while, say once every, I don't know, 30 days or something, an inspector or somebody walks in. You could imagine in that scenario nobody thought to instinctively say that somebody's hands could be in the machine.
Because it was not a normal part of the scenario. it feels to me like that could be where the trade off is between the need to make money efficiencies. thinking these safety scenarios through. things like security, safety, ethical considerations, they're generally seen as costs, not as, functionally leading.
that seems like the real risk that rarely gets talked about because we're too busy fantasizing about utterly unrealistic Terminator scenarios.
Chris: Yeah, I do think [00:45:00] that whereas automation has been driving this discussion in manufacturing for decades, this whole question, I mean, you see it even in the port strikes, right?
The port strikes are a lot about automation, replacing the port workers, and that's what they're negotiating around. So AI now coming to the picture, it's like already done that in manufacturing to a certain extent with all the, the technology and the automation that's been there.
but it's just different and I think it may not be as big of a challenge as it is in other areas. Like If you're developing weapons systems and now you got drones that are AI, right? And they can kill people. That's probably a little more, severe. So maybe some of the growing pains that manufacturing has already gone through.
Deep: I think that's a really good point. I was actually chatting with my son a while ago. He's a Finnish carpenter like the analogy he gave me is he says people assume that just because something can be automated. It makes economic sense to automate it, but he used the example of a particular tool that you can buy fairly cheap, maybe like under a thousand bucks, and this [00:46:00] tool, can save, a lot of time he said, nobody uses this tool, even though it's readily available, it's everywhere. And he says it's because it takes 10 minutes to pull out of your truck and set up, and the alternative takes a few minutes less. we got onto those, Tesla Optimus robot things, like this whole idea of humanoid robots like showing up You know in the construction arena his argument is like, okay, so i'm dropping 30 35k on a robot That can walk around I gotta charge it.
I get jealous stuff might make sense in like highly repetitive scenarios but People often underestimate the economic calculus with automation. And I think that's the same thing that's gonna happen with AI, that at the end of the day, there's a lot of tasks that even GPT already exceeds the cost of getting a decent human to do the job.
a lot of my GPT responses, for systems we're building, can be upwards of 50 cents a buck a response at that kind of cost, you know, you do. 15 of them in an hour, you got [00:47:00] yourself a minimum wage worker, you get yourself up to 2025, you got yourself a salaried worker.
Like you're spot
Chris: on. I give you another example from my past, an automated fork truck, you know, called an ATL automated truck loader or automated, vehicle. What is it? A C B. five years ago. $250,000 a piece, Way too expensive to replace an hourly worker doing warehouse work, Now, last year, two years ago, they're under a hundred thousand dollars, right? And so now the economics changes. I think everybody knows we're in some sort of AI bubble, and I suspect a big part of it is exactly what you're talking about, which is this realization of the cost. And then as we start to realize the cost of AI and all these trials.
Great test. We learned a lot, but not really production ready I don't have any idea how big the bubble will be, but I think there's going to be some pullback. and to your point, your son's point. why I made the point earlier in the conversation about our clients will never be fully automated is exactly [00:48:00] the economics and the lack of repeatable work.
Yeah, they're doing repeatable work. They might be doing 30, 40%. Every day is the same But then the next 15 minutes, they're doing something completely brand new that they had to do only today because it's a small company and they're dealing with lots of things going on, right?
Deep: yeah, I think A lot of times people talk about job replacement. I feel like that's the wrong word. It's, more like task replacement. as you start replacing, easily repeatable tasks, like well defined ones, then people kind of walk up the stack. Like one of the examples I use is, I built this, we built a system to automatically detect heartbeat anomalies from audio signals.
this is, Something cardiologists do. Physicians and cardiologists do a lot of things, right? They don't just sit down with a stethoscope, stick it against your heart, and get a diagnosis. it turned out that really wasn't, even though we could do it better, faster, Cheaper than cardiologists who cost an insane amount of money.
It still wasn't an obvious [00:49:00] sell
Chris: That's amazing, right I love your wording there It's replacing tasks and not jobs and I would 100 percent agree with that and I think that's very similar to the automation I'll give one more funny anecdote. when I started my career in manufacturing, I joined in a space of the warehouse, there was a person pulling empty boxes off of a conveyor and putting them onto a pallet.
ten feet away, there was a person taking the empty boxes off the pallet and putting them back onto a different conveyor. Within a year, both of those tasks were replaced by a big storage machine that would automatically. do that.
Deep: Yeah.
Chris: All the other positions in the brewery remained and you know, automation slowly eliminated tasks you went from a production line having 10 people to having nine and then eight and then seven.
So I think we're going to see the same with AI. It's going to bring it efficiencies, productivity. Absolutely. It's just hard to find those jobs you can replace entirely. That's the low hand fruit that's already being captured, right?
Deep: Yeah. And I would [00:50:00] also argue that it. Depends, but when easily automatable tasks get replaced by machines, and this story goes back, a few hundred years, people come up with other things to do, and they can often be higher stack, like they require more cognitive load, more emotional load, machines are not going to be carrying the emotional load very soon.
I'm sure there's plenty of companies raising money on this idea, but nobody wants to go to a bar with a bunch of machines, giving them a drink. Like Nobody wants that. It just doesn't want it. But I could see a world where like, trained psychotherapists or bartenders, like I could see that world.
Deep: One last question that I always end on, let's fast forward five to 10 years out, all your dreams, come true. What's going on? What is the world of, mass group look like at that point?
Chris: It's a world where, like I said, we are making manufacturing software easy, and we're making people's digital transformation effortless, allowing them to make the United States the global leader in manufacturing again, right?
We know it's national security these [00:51:00] days, and, Mass Group just wants to be a part of that, helping the U. S. reclaim its rightful throne in the manufacturing kingdom,
Deep: Thanks so much for coming on, Chris.
This has been an awesome conversation. I think we're a little bit over on time, but thanks a ton for coming on. Is there anything that we missed that you feel like, you'd love to address,
Chris: No, I think you know, we've obviously talked a lot about AI and building that into our, future.
We also are always looking for partners out there that are interested in the manufacturing software ecosystem. So if anybody wants to reach out and talk, I'd love to talk, always willing to network and hear what people are up to.
Deep: awesome. Yeah. And so for anyone listening, Chris's, info will be all in the show notes. Thanks so much for coming on, Chris. This has been awesome.
Chris: Thanks a lot, Deep. Pleasure. Great to see you.