Juan Carlos Santamaria, Trimble. Physical AI On The Jobsite

We talk with Juan Carlos Santamaria about how AI in engineering has evolved from rule-based robotics to modern systems that perceive job sites and help machines make better decisions. We dig into Trimble’s push from AI perception to operator assist and what it will take for engineers and operators to trust AI in the field and in design tools. 
• Juan Carlos’s PhD-era view of AI as a multidisciplinary field 
• Planning versus reactive robotics and why brittle plans fail 
• AI for perception on construction sites using point cloud images and video 
• Turning recognition into jobsite semantics like cycles and bucket loads 
• The shift toward decision making with operator assist in the cab 
• Autonomy in mining versus the realities of safety policy and adoption 
• Why trust builds faster when operators can experience the system 
• Zero tolerance expectations for machines compared with human error 
• Point cloud segmentation today and what engineers want next 
• Natural language interfaces that execute software commands from prompts 
• SketchUp and 3D Warehouse visual search and AI-assisted edits 
• Interoperability across tools and the “Tower of Babel” problem 
• Planning for unknown unknowns when digging into existing infrastructure 
• How AI work gets prioritized across product teams 
• Why kids and professionals should learn AI as a tool for thinking clearly 

Transcript

Roopinder 0:07

Hello and welcome to FODES, the future of design and engineering software podcast. My name is Roopinder Tara. On the show, we will have guests that will discuss tools and technology that engineers will find interesting and useful. Hello, Juan Carlos Santa maria?

Juan Carlos 0:26

Yes, Juan Carlos Santa maria. Yes, thank you. Yes. Very good.

Roopinder 0:30

Nice to meet you.

Juan Carlos 0:32

Nice. Nice to meet you too.

Roopinder 0:34

Well, now do you normally people address you as Juan Carlos or Juan or Dr. Santa maria?

Juan Carlos 0:41

Juan Carlos is fine, yes.

Roopinder 0:44

Where are you uh calling from? Are you I see that you're in living in north of Atlanta? Is that where you are?

Juan Carlos 0:51

That is where I am, yes. Okay, it's in Wood, Woodstock, Georgia. That's does Trimble have an office there? We used to have an office, they closed it down after the pandemic, and now the people who were here are now remote. Okay. Half of my team is um is uh remote.

Roopinder 1:10

Uh did you go to school in the area as well?

Juan Carlos 1:13

Yes, definitely. Uh I went to Georgia Tech, which is uh here at uh in Atlanta, graduated in 1997, PhD uh in artificial intelligence, and then a master's in industrial systems engineer, also and and then I went to Houston for two and a half years.

Roopinder 1:33

We've talked to several people about AI, all of them experts, of course, right? Everyone's an expert at AI. You're the first person who has actually got a doctorate in AI, right? Computer science in AI.

Juan Carlos 1:49

So that is true, that is true. A lot of friends tease me about this. You say, hey, I remember when you got your PhD in AI, and I tell my friends, you know, I know a person who had an AI. Now, AI back then is not the same as the AI nowadays, but yes, it was AI.

What AI Meant In The 90s

Roopinder 2:07

So, do you find that when you tell people you have a PhD in AI and it stops all the conversations? Nobody wants to talk to you anymore.

Juan Carlos 2:16

Yeah, but not only in recent times. Before it was I have a BS in AI, and people say, How can I use that for?

Roopinder 2:26

That was a while ago. People that probably didn't know how to spell AI back then. Now it's the most popular thing in the world. So today's AI is all about language models. Tell me what you studied back then and what was AI.

Juan Carlos 2:40

Yeah, it's a multidisciplinary discipline for combining many different aspects from psychology, from brain science, uh, neurophysiology, and from computer science, optimization. There was natural language, but more about rule-based and all the theory behind language parsers using the old school techniques with uh defining rules, and it was uh robotics was about frameworks and how to create those frameworks to scale. Back then, there was this distinction between planning and reaction. Planning it was a traditional way you create a plan, and then once a plan is formed, you give it to the robot to execute, which was very brittle because as as soon as you start executing, the plan changes and the robot doesn't know how to adapt, then the reactive was all about reflexes and how to do behaviors or compose behaviors around reflexes. And that's why if you see the robotics 20, 30 years ago, it was all about more simulating insects or small animals, and the behaviors were very impressive, but were were not intelligent, right? Because it was all reactive. So that was very fun, very, but it was uh not as advanced as it is today.

Roopinder 4:14

The game has changed. Where did you originally grow up? Was that Venezuela?

Juan Carlos 4:19

Yes, Venezuela.

Roopinder 4:21

And you studied uh industrial and systems engineering at Georgia Tech.

Juan Carlos 4:25

Okay, so that was graduate, uh that was a master's. I did that also because I had uh this um interest in uh problem solving in the industry, but using innovative technologies. So that in order to satisfy my my desires, we're um always grounding it on in industrial systems. So that's why I also took the master's in industrial engineering.

Roopinder 4:51

I see. And it became a question for you: how would an insect do this?

unknown 4:58

Right?

Roopinder 4:58

How would an insect's brain handle this sort of industrial process? Let's flash forward then. All right, now you're much in demand, everyone wants you to work for them. Trimble, like every other company, people that I talked to from those companies, they're all about AI. Your CEO, Rob Painter, was that a dimension? Over two years ago now. We came on stage, he was very much talking about AI, it was center stage, and uh and I finished a lot has happened, and I want to go over what's what has happened since. But at that time, I talked to your probably your predecessor, Karoliina?

Juan Carlos 5:35

Is that oh yeah, Torttila? Uh-huh. Yes.

Roopinder 5:37

Yeah, yeah, she came.

Juan Carlos 5:40

She moved um from to another company. That's okay.

Planning Versus Reactive Robotics

Roopinder 5:43

Okay. Yeah, we talked considerably, but I want to catch up on her work and then how you took over and what things have been done. We're we're especially interested in what it does for the design tools. That's my area in CAD and SketchUp and uh the CAE tools, the uh engineering tools. Uh, but I want to cover in general what is Trimble's concept of using AI? How is it going to implement the modern version of AI and all its products and across all its lines? I asked a lot there.

Juan Carlos 6:15

Yeah, that is a very broad question, but very interesting question. So let me just cover a little bit about that history with Karoliina. The tools, the AI tools, uh four or five years ago, and still today, some of them are targeting about perception. So, how to interpret massive amounts of data like point clouds. So, Trimble has hardware devices that collect massive amounts of point clouds or cameras that are watching the construction site, and processing that information to be able to determine where are the trucks or what are they doing or how to count the number of cycles of a whole truck or an excavator. That technology for image processing and video processing is very difficult to do in uh traditional uh technology. But with the advent of AI, it became a lot uh easier to train um these systems to be able to recognize and segment the different objects in the construction site, like trucks, excavators, humans. And then you with that information also be able to do some meaningful uh extra meaningful semantics about the construction process, like cycles, how many cycles, how many bucket loads this excavator is doing per day, which is a crucial information for the performance or the site manager. So that's where Karoliina was um working, how to integrate all of this new technology for processing perception, a massive amount of data. Right, right. That that and that trend is still going. Yeah. Now the new or not not new, but the new emphasis that we are trying to apply AI is for decision making. So now it's not just about perception, it's about how to make intelligent uh decisions, optimal decisions. And that's where this operator assist team is starting to grow. So it's it's about helping the operators of heavy machinery in the construction side do their job more efficiently. Because you you always have a few experts that know how to do this well, but a lot of intermediate and novice operators that they can do the job, but they can improve a lot. So what we're trying is to enhance their performance by putting operator assistance in the cab that works very uh similar to a physical AI, right? So if you imagine this heavy machinery be like a robot, it's a different type of robot. It's not uh an Android, it's not a wheel, it's an excavator, it's a weird kind of robot. It has an arm with a bucket and has threads. But if you imagine that being a robot and you uh instrument that uh vehicle uh with the technology, then all the system is doing is uh suggesting the operator what is the next high-level action to execute next.

Roopinder 9:38

Okay.

Juan Carlos 9:39

So that way, for example, the bucket loads. Okay, if you if you are building a trench, there is a a good sequence of bucket loads to follow. Only expert operators know how to do this well. But if you uh instrument the machine and put a novice operator that follows the advice, then it behaves from an external observer as if it is an expert operator.

Roopinder 10:03

So you've taken all the knowledge of an expert and put it in the hands of any operator.

unknown 10:08

Exactly.

Roopinder 10:09

How much of that have we actually implemented here? Because are we at that stage now? Could can it be can I put a dummy like me in the in a earth moving equipment?

Trimble AI Starts With Perception

Juan Carlos 10:20

Or let me just give you a little bit of history. We we started with autonomy in construction site, and that was we're talking about 10, 15 years ago, but the industry was not ready for it, and you know, the policies and all the legal framework was not ready for it. It was a very difficult sale having autonomous, especially because they need to interact with human uh operators uh around it. So the safety and all that stuff. Very similar problem with the autonomous cars nowadays, right? The all the how to ensure safety and all the legal. We were ready to do this in dosers, and we demonstrated this with caterpillar using dosers, autonomous dosers with operators outside the cab, doing slow dosing for mining, for mining operations. Um, and it was uh in that system there was one operator monitoring five dosers, up to five dosers uh working autonomously, doing uh slow dosing, which is a very repetitive task, but it's also a very difficult task. Uh going back to the expert, knowing where to engage the blade in a slot is a crucial decision because if you engage the blade too late, then you miss the chance to push more dirt. But if you engage the blade too early, then you risk stalling the dozer because now there's too much dirt in front of you. So knowing exactly in every pass where to engage the blade along the slot so that you can move the maximum amount of dirt every time is a difficult operation. Only expert operators know how to do it, and even expert operators cannot do it eight hours a shift continuously. So uh that uh system, autonomous system, was uh super successful, but it was uh very punctual. So we Trimble wanted to uh to expand this to the normal construction site, not just mines. We had a difficult time trying to sell it because of all of these uh legal policy issues. So that's why we had to cast the technology as assist. So operator is in the cab, but it's like uh being enhanced their skills by the suggestions of the system.

Roopinder 12:50

Yeah, must be a little frustrating to know that a assistant can do it much better than a human, and yet humans insist on being in the loop.

Juan Carlos 13:01

I yeah, it's it's true, it's true. Well, not unlike not unlike the way I the analogy I always do with this is Google Maps. When you are in your car and you you put your destination and you follow the directions of the Google Maps or guided system, GPS guided system, very similar, right? And the system is observing what you're doing, and if you don't follow that suggestion, then it recalculates. Well, now that you're here, now this is the best thing you can do next. So imagine it's the same, but now for a doser operator or an excavator operator, but instead of driving, is moving the machine and engaging the implement, the blade or the bucket. So the suggestions are about those actions instead of turning left or right. It's about engage the blade here. Well, you didn't follow my suggestion, but now this is the best place to engage the blade. It's very similar.

Roopinder 14:05

That sounds like machine cutting tools, where machine cutting tools have to get the right speed and feed. And a human operator will oftentimes be very conservative in their approach and want to break anything. Whereas a technology, if you put AI into it, you can get the optimum cut. Uh, you can go deeper, you can go faster, and you'll be more optimized. But on the other hand, I totally see this happening with this fear of autonomous anything. In San Francisco both are all over the streets, right? And most people that I a lot of people that I talk to, even people who should know better, are afraid of this wave mode, right? Until they actually get in the car and they actually see it and they drive in it, then they're amazed, right? And and and I and I feel like that's that should be the way people experience the technologies actually see it and feel it and participate in it, and then they will lose their fear.

Juan Carlos 15:05

Do you agree? Yes, I uh I definitely agree with that, and um also um operators trusting that what the system is suggesting is um is the right thing to do because you know they have their ways. We have we have had this confrontation too. Operators saying, Why is he doing this? This is not the way, uh always 10 years doing it this way. And but it turns out that when you compare the long-term behavior, and we have the studies to prove this uh of an operator doing it in the way they have been doing it, and doing it the way the system is suggesting to do it, is an improvement. We have a a study in Dayton where we improve the time 20% from the operator, expert operator to where uh to how the system is doing it. As you say, we have to put it in the cab and let them experience it. Uh you'll be amazed.

Operator Assist And Physical AI

Roopinder 16:09

I was I was amazed when I saw the construction site at the dimensions, and I was with a whole group of press people, so it was very nice. Uh, you never actually see the machines in operation, I wasn't expecting to, these large machines, some of them guided autonomously, some of them excavating. One machine was excavating underwater. Do you remember that?

Juan Carlos 16:31

Yes, yes, uh, in uh in the offsite.

Roopinder 16:34

Yeah, yeah. And that was uh people were just amazed by this happening. And this was the press, like we should know what technology is capable of, and we were amazed that this is happening. So I totally understand what you're saying. Like, Trimble is so far in technology to the actual implementation of the technology, there's so much that can be done compared to how much is accepted. But how much of that is related to well, I'll call it zero tolerance. You know what they say about uh zero tolerance for mishaps in aircraft, for example. We cannot tolerate a single aircraft going down, right? There's zero margin of error. There's a considerable margin of error given to humans in cars. We tolerate 35,000 deaths a year in the United States. Tolerate is not the right word, but we somehow accept that as normal, and we don't tolerate a single fatality in autonomous driving vehicles. Do you think there's a similar fear of that in construction and in your industry?

Juan Carlos 17:43

It is, it is. Exactly, or doing something that there are some like walking behind a heavy machine without the proper signaling, or following the protocols. AI very good about doing this.

Roopinder 18:53

Yeah. So pe so people that are you know that are implementing technology and people that are covering technology, you know, we we tend to fall in love with it, but the pub, the rest of the public is very suspicious or wary or tentative about AI, right? There's a lot of uh mistrust, and yet we're able to forgive humans for those errors. And if technology was in place, this is an argument that everybody makes about autonomous vehicles, there'd be a lot less deaths and fatalities and injuries if you have autonomous vehicles, and yet we don't tolerate a single death or injury. It's it's just weird and frustrating.

Juan Carlos 19:34

It is frustrating, no, not only in terms of safety, I also saw a study of um engineer, a civil engineer making studies on traffic patterns. Yeah, if you inject certain autonomous vehicles, you have certain density of autonomous vehicles, it will actually help the flow of traffic and prevent that backtrack wave that forms on highways that cause accidents.

Roopinder 20:02

Right.

Juan Carlos 20:02

So even in that, even that is successful. But you know, it's it's I think it's a matter of gaining the trust.

Roopinder 20:10

I gotta go back to some of the things I started to talk about. You had mentioned that the work that was being done on shape recognition or object recognition, I think. And that's always been my hope for AI that it can look at look at a point cloud. Because when I look at a point cloud, at first glance, it looks like the stars in the sky. I can't tell what's going on. But I need AI to look at that point cloud. Let's say it's inside of a building. I want it to look at the inside of a building and know that that's a pipe overhead or it's a ductwork, or is it a valve, if it's a factory, or something. I need AI to discern that for me. So I don't have to. How close are we to that kind of world where we could interpret all those points and all that data and make them into objects and just kind of throw away the points?

Juan Carlos 21:01

Yeah, so that we currently have uh some of that uh in production and available to use. Point cloud, and the one of the applications we have, Trimble Business Center TVC, is able to segment the point cloud based on trees or uh for surveyors, right? It's very important to determine what is a building, what is a street, what is vegetation from the point cloud. So very similar to what you're saying, but in a different context. So that is already available. And the ability to train the system to detect customer specific uh groups is there and it's also available, is l uh is used less often, I believe, if because of the familiarity of the feature, but uh it is there.

Roopinder 21:56

It might have been there. I just didn't see it. When I sat through some of the sessions. Acer scanning. It was all about handling the points and the processing you have to do. And to tell you the truth, it sounds as a mechanical engineer, I'm not a civil engineer, I don't do this, but to me it sounded like it was a painful process. And I just sat in that one class, very detailed, very knowledgeable presentation about the process and how to handle it. And I think it was a facade of a building. And I just thought, wow, I don't do I really want to do all that. It's like why can't all this be automated? Why can't they recognize details?

Juan Carlos 22:34

I see what you're saying, and I agree. It's uh Trimble Business Center, it is uh an application that has been around for 10, 15 years, yeah, and it has the same workflow that is difficult for engineers and surveyors. One of the hopes that we are trying to do is augment these applications with the traditional LLM. So now the interface to the application is no longer your mouse and keyboard, it's just a natural language and the AI knows what are all the capabilities of the software and is able to interpret your request and execute those commands. It's no longer about an engineer manually selecting which menu option do I need to execute next, but just providing a prompt. Say, can you classify this point cloud with the vegetation floor and buildings for me? And then the application will guide you. Can you let me know where the files are? I'll take it from here now, and then it will do all the steps of data preparation and all that stuff that normally the operator or the user of the application will have to do, which is what you are alluding to. Like it's sometimes it's it's difficult to know how to prepare the data so that it can be classified so that I can have my answer.

Roopinder 24:05

Triple has, by virtue of its SketchUp acquisition, got have an enormous database of uh things, of uh it's called a 3D warehouse, right? So yeah, it has an amazing amount of 3D models. And I always thought, hey, why okay, if I if I can show, if I use my phone and show it, what have you got that looks like this machine or this part or this earth-moving equipment? So there's a lot of AI that I see, and I then expect it to be in my design products. I want it to be there. I'm sort of impatient. I can point my camera at a plant, for example, and this is an iPhone, and it'll look at the picture and tell me what plant it is, right? I can identify plants. Um so when I see that I think, whoa, that's pretty neat. Why can't I do that to a screw, right? Why can't I look take a picture of a screw and say, oh, this is a quarter 20 screw, right? And this is a quarter 20 nut. I want my engineering tools to be as smart as my consumer tools.

Juan Carlos 25:18

Yeah, true. True. I'm not up to date with the latest SketchUp release, but I've seen the feature. The the feature is there, visual search for the warehouse, SketchUp warehouse. Okay, yes.

Roopinder 25:32

So if it is karoliina promised me that she promised me that would happen, so it's there.

Trust Safety And Zero Tolerance

Juan Carlos 25:38

Yes, I I've seen it, I've seen it working. I don't know if it's already available for in the production pipeline, and also some other plugins because the same steps that I was describing about TBC or Dingo Business Center. In SketchUp, there is also similar things, right? Sometimes you have a model, but you need to change the paint or the walls or the bricks, and going through that process is a it's a step and a step and another step, and opening the material and changing the texture of the material is cumbersome. And I've seen AI uh uh plug-in in SketchUp where you just say, Can you change the brick from red to construction brick? And and it will do it for you. No, uh, you don't have to get involved on the on the minutia of uh following the process.

Roopinder 26:34

So you don't have to get in the minutiae of following the uh learning the language either. That's what I love about LLMs. I I understand it doesn't do good. LLMs aren't for shapes, they're for language, right? But I love it that I can interpret a my my language into can, for example, right? I love it that it'll eventually just tell it to draw a circle rather than pick a circle from a menu, right? Or make a wall or make a construction site. I'd love it if I could say, okay, make me this is a one mile by two mile site that I'm working on. Here's a terrain, now excavate or level or something, right? And it would it would start doing that sort of thing from my natural language. I don't want to learn 300 different commands of a site.

Juan Carlos 27:25

Exactly. Yes, that is instrumenting the good existing applications with a uh user interface that is AI-based.

Roopinder 27:35

Yeah, yeah. Because Trimble is it's not just Trimble, it's every every design company all has its own languages, and engineers find that they have to jump from one program to another. Like on your like on a Trimble site, those people may also be using Bentley. Right? And each and then they have to become, I say you have to become a Bentley engineer, you have to become an AutoCAD engineer, right? You don't, you're not an engineer, you're you're becoming, you're learning the language of your system you're using. I'd rather use my own language and have the program do my bidding from the natural language. I don't want to learn all its different languages. I find that to be a tower of babel sort of thing. I am good at certain CAD programs, but not all of them. How can you be? So do you see a world where natural language takes over uh and is able to perform, let's say, most of the type of commands?

Juan Carlos 28:32

I yes, actually that is happening. Most of the companies that have legacy products, if if we can call it like that. Um the focus, um, a major fraction of the focus of using AI is to actually add the AI as a UI interface. The model knows what are all the operations that the application is able to provide and understands their um the model of each of these actions, right? And then it understands what the user wants with the natural language and is able to do this mapping and create a plan and execute it. So that is um well, I think that's what we are talking about.

Roopinder 29:17

I'm glad to hear you say that. I I had a I'm not a conspiracy theorist, right? But I thought for a while that software companies weren't employing a natural language interface because it would make anybody be able to use their system and anybody be able to use another system. So if this caught on, they wouldn't be trimble, they wouldn't be trimble um experts. They you wouldn't need, you wouldn't take the pride and energy and learning a system because you could use any system. So people could move from one software to another. And I think the industry, I thought, oh, maybe the industry doesn't want that to happen because they want to retain the users, right? Because if there's no investment made in a in a CAD product, you have no reason to stay with it necessarily, right? You can migrate very quickly. So I'm sure you there's no such thing.

Juan Carlos 30:11

I I understand what you're saying. I can see some companies uh protecting their investment, they see themselves invest as an investment on the product, uh-huh, but I believe that interoperability is a feature that customers want, and it is in the interest of the companies to provide this interoperability. And AI is a way to prove that can happen, and they don't have to do that too much uh an investment in putting this layer on.

Roopinder 30:45

Um okay, so there's no industry-wide, let's say, directive or goal to make keep software hard to use.

Juan Carlos 31:00

I don't think that company will survive much.

Point Clouds And Natural Language UI

Roopinder 31:03

I'm glad just glad to hear you trying to dispel such conspiracies. There was an underground railroad that was being constructed in in the UK. It was going from one site to another, but it turns out it was along the site of a Roman road, thousands of years old, right? A Roman road. And as they were excavating, there were all these, they were digging up all these artifacts, and it became an archaeological site. I'm sure this progress was much slower than it was anticipated, but there's all these issues with digging on sites that are developed even more recently because this old infrastructure is underneath the ground. Do you see any case where machines are able to interface with records of infrastructure so that they can be knowledgeable of where digging might be dangerous or interfere with?

Juan Carlos 32:03

Yes, that that is a very interesting question. Yeah, currently the models for especially for the digging of the dosers and the excavator, they do consider different material densities and the distribution, like silk or sand, or you know, the different types of dirt with the different densities, and they need to adapt when you reach that layer to modify uh fine-tune the behavior.

Roopinder 32:33

But that's almost like they're feeling their way around. I what I'm talking about is will they be able to access a database which tells them that this is already there? Do you know what I mean?

Juan Carlos 32:44

Yeah, that would be like uh if uh if on a specific site there is a hard rock ore that is not evenly distributed, so it's it's higher in one end, and so you can work fine in this end because it's too deep, but uh here you need to uh be able to design according, taking into account that there is a hard rock layer just a few meters below surface. So it becomes a tool for the the engineer designing uh how to construct the site more more or less more than than when you are executing executing the plan and then you found something that was unexpected. Like uh the example I think you gave is belongs in this category. The people designing train did not anticipate there was a Roman buried there. But if they if they had that, then they could have created the plan accordingly. Yeah, it's a difficult problem. You know, it's planning for the uh designing or planning for the unexpected is possible, yeah. But uh, you know, it's it's um what are the levels? It's uh on unknown unknowns, it's things that you don't know that you don't know. That one is the most difficult type of problems to solve. But it is possible to when you know what you don't know, I mean you know the class or you know there's something there. I don't know exactly what it is, but there's something and then um then the plan can accommodate for that and then say, well, if the probability of having rock is 20%, then I the plan is this way. But if it's 80%, the plan changes.

Roopinder 34:39

Right.

Juan Carlos 34:40

Because you know that there may be rock, but you don't know exactly how where or how uh I mean at what depth, but you have some probability. So it's that is how that's how you capture that you don't know something that you know. But knowing and not knowing what you don't know, that's uh that's the most difficult part.

unknown 35:01

Yeah.

Roopinder 35:02

Yeah, it takes a combination, right? A skilled operator does this on a machine, they'll use the machine's feedback to to say okay, where to dig and uh you know where to how fast to dig and all that sort of stuff. And then also he'll probably have to what they call, what do they call it, call before you dig? Yeah, call and find out what's what's what's here, what's around here, what what am I interfering with? Is a gas main or water main or something, right? Need to know all that thing. So combining the both things, the in on-site feel feedback with off-site databases, right? Yes, to get that, and then it can be all done with one machine. So how busy is your team? Like, is everyone asking you to implement these things? Not just media people like me.

Juan Carlos 35:56

With the product manager, uh he's the one really finding the opportunities and ranking the opportunities, and then uh we just tackle the top tier, and you know that that continuously is evaluated every quarter, every every six months, every year.

Roopinder 36:16

We're more or less the service organization within a company. So there's some a product manager has a certain demand and he gets a crazy idea. Okay, now I'm sure AI can do this, and it falls upon your your team to actually make that happen, or you have to tell them uh, no, no, that's that's pretty, I don't know, you wouldn't say stupid. Yes, yeah.

Interoperability And Planning For Unknowns

Juan Carlos 36:40

The way we were trained, I mean, they we just give the product manager the the technical feasibility and cost of achieving, and then it's you know, you want to do that? Well, it may be possible, it may take two years and it costs gazillion dollars. Uh it's up to you if you want to chase that.

Roopinder 36:59

I see. Is it product managers all only or doing this, or is it like does Mr. Pater come and say, hey, one calls, let's do it do this for me, do it, do it by next quarter. I've got another dimensions to address, right?

Juan Carlos 37:15

Yes, no, it's a team. Uh, you know, but the product manager has the responsibility of ensuring that the opportunity is commercially viable, it is technically feasible, and it's it's collecting all of this from different expertise in the team.

Roopinder 37:33

Do you find them to be realistic expectations most of the time with when they make an AI demand? Uh if I could call it that.

Juan Carlos 37:42

So sometimes uh yeah, I hear crazy ideas, but uh I think um we are all learning about what AI can do. And yes.

Roopinder 37:53

Do you ever get the opportunity to suggest things to them?

Juan Carlos 37:58

Yes, definitely. When when uh I see that there is a long-hanging fruit about the use of the technology, I try to push it forward, and then it is their responsibility to find out if it's commercially attractive to pursue.

Roopinder 38:14

Is the natural language interface the the low-hanging fruit right now?

Juan Carlos 38:19

It's one of the low-hanging fruits. It's okay because we we have the existing products, applications, and we have the technology that we know it works already. So it's it's a matter of joining the two, and then there is there is technology that helps to do that, which is the MCP protocols.

Roopinder 38:38

Um But tell me this though, those the physical AI. Now that is a much more difficult problem, right? That's still, you know, the because we're dealing with physics and shapes, and and like you said, the uh volume of the earth that has to be moved and recognition of quantity and mass, right? That's very much a work in progress, right?

Juan Carlos 39:00

That's yes, yes.

Roopinder 39:01

Yeah, okay.

Picking AI Projects And Learning AI

Juan Carlos 39:02

And that that is exactly the focus area of the where the decision making happens. It's about physical AI, is about embedding the system in the environment and be able allowing it to execute, and for that needs to perceive, needs to decide, and needs to act, and continuously doing that in a loop. And that is what physical AI is about. They not the LLM there is a way to communicate with the physical AI, but the AI that needs is needed to drive that is not about language, it's about physical models of what happens if I try to load the bucket with too much dirt or with rocks. And that is what is involved, and the knowledge that is uh behind the physical AI. What is how those how is the world uh behaving under the influence of actions and by knowing that the system is able to contemplate different alternatives and pick the one that is the best to it's to achieve the goals.

Roopinder 40:16

One last thing. I see you have children, right? So, how how do you advise them? Do you think AI is they should all be learning AI or computer programming or I I think so.

Juan Carlos 40:28

I think so. I know it's challenging for education right now, the all the AI because it's transforming not just how well uh I think we need the ability to the kids learn AI and the teachers to teach how to use AI efficiently, but I I don't think it's something to be feared, I think it's something to uh it's a powerful technology that can be used for good or bad. Uh so all of that is still uh on the table, but I think we need to learn. I always remind them of the 60s or the 50s when the engineers used to use the slide rules, right? They always carry that.

Roopinder 41:11

Yeah, I know, I remember that. You see this? I keep this around. This is a slide rule, but it's circular. Ah, exactly. Yeah, but yeah, when I when you got into computing, when you got into the university, they are already they were into computers by that point, right? Yeah, I was on the vert on the cross, right?

Juan Carlos 41:37

Exactly.

Roopinder 41:37

When I started going to engineering school, people were just starting to use calculators, and then when I got out of engineering school, they just started to use computers. What a world.

Juan Carlos 41:48

Exactly. So I I see it the same way. Um it's it's a new tool that is enhancing the capability of your working environment. So I think AI is another magnitude, more than those, uh, and it may be challenging, but I we have to find a way of uh of doing it. And as I saw in one comment the other day, uh we are learning how to use AI, but in the process of learning to use AI, I think we are improving ourselves in the sense that we now need to structure our thoughts to be able to do meaningful prompts. If I if I say so, and that is is that is the AI teaching us how to communicate better.

Roopinder 42:38

I think well, Carlos, this has been great, and I've got to say a lot of fun. Thank you so much for opening up and talking about uh AI, human to human. Uh not just as a company, man, but I really appreciate meeting you and talking to you. I'll let you let you get back to work, but thank you so much for this conversation.

Juan Carlos 42:57

Thank you. Thank you for having me. And it was very fun to talk about AI too. It's my passion.

Closing Thanks And Listener Outreach

Roopinder 43:04

Very good. All right. I'll see you. Hope to see you soon. Bye-bye. Bye-bye. Thank you for listening to FoDES, the future of design and engineering software show, brought to you by ENGtechnica. I hope you have learned of a new application or technology that will help you with your job. If you have an application you think would be of interest to other engineers, please let me know by emailing me at roopinder at engtechnica.com or message me on LinkedIn.