Intro: You are listening to Behind the Ops presented by Tulip. 

Russ: Welcome to the Behind the Ops studio. We're back.

Madi: And today we have a special guest. 

Russ: The specialist of all guests. 

Roey: And the first guest. Hello. 

Madi: So it was, uh, it was so fun doing our bonus episode about ChatGPT and Santa, we had to bring someone who really knows about AI to help explain to our listeners: the pros, the cons, and the intricacies.

Russ: And you knew a pro in the space. 

Madi: I did. I did happen to know a pro. Roey, why don't you introduce yourself. 

Roey: Hi, uh, thank you for having me guys. My name is Roey. I'm head of AI at Tulip and I have been dealing with AI in the last 10 years at least, and quite a journey. 

Madi: Should we call you Dr. Roey, or is Roey fine?

Roey: Oh, this will be super embarrassing. 

Russ: Doctor it is!

Madi: Well, Dr. Roey, can you give us a definition of AI just so we're all starting from the same place? 

Roey: Yeah, definitely. There is no one, but seriously, I think that AI is not well, well defined in the sense. People think about AI as artificial intelligence per se, meaning a computer that can think like a human, as far as I can tell, that doesn't exist yet.

That's still a science fiction, but we are focusing there and what we currently see as AI is probably an algorithm, a statistical model, a black box if you want, that gets some kind of input, an image, a video, a voice. A signal, a vector, a set of numbers, and create some kind of prediction. And a prediction could be different things.

For example, here is an image. Yeah, that image is a dog image. That will be a prediction. It could be, here is a piece of text in Spanish and the output will be a text in English. So an AI, in my vague definition, is a box that get an input and provide an output, which is a prediction of something that is probably non-trivial, uh, but is easily to understand.

Madi: So in this kind of definition, you really need someone that's like, I guess curating the information going into the box, and able to interpret it after .This isn't like a Terminator Skynet situation yet and it's ability to just like make decisions and stuff. 

Roey: Exactly. So you know, there was a person, a data scientist, an engineer that was thinking about a very specific task for this piece of code, piece of algorithm to do so. It's not really different than any other software. While that is probably true, we do see models and more sophisticated black boxes such as ChatGPT, that their ability become very, very wide. And because it has many, many different capabilities, it can translate.

He knows a lot about the world. So it's, it's starting to look like it's maybe a human, but it's definitely not. And, and frankly, I don't think that it's significantly different than Google search. Google search is AI in the last, I dunno, 20 years. And it's, you know, using huge amount of data to give us results about what we just searched, right?

The input is our search key and output will be websites that are indexed, uh, in a, in a specific order. Okay, so ChatGPT took it one further, uh, step further and it give us text and we can reply and it'll fine tune the answer. 

Russ: I feel like ChatGPT t is a honorary guest host of the Behind the Ops podcast. It comes up at least every other episode.

Madi: I know. Maybe that's like something we need to be careful about. 

Russ: Madi, explain what ChatGPT is and how it, how it is in your life. 

Madi: I feel like in my home life, I would never think about ChatGPT. But at Tulip now it's, it's showing up all over and I think it's because, to Roey's point, it's really almost like the next evolution for search engines.

It's really changing the way, um, content is reviewed. Um, and as a marketer, especially as someone who works in, understands communications and especially digital communications, uh, I think it's the beginning of what will likely be a, a step change in how we think about those platforms and the content and all kinds of related things.

Russ: and the, the interface is you, you just type something in and then you get, you give a prompt. Right. And outcomes a response to the prompt. 

Madi: Yeah. You just give it a prompt and sometimes it, it can't give you an answer. So it, it really does depend on there already being enough content there that it can engage with the ideas.

So really novel or innovative ideas. It's either like just spitting out from one source or it's wrong. It's like only pulling from like preexisting, um, information so it could be dated. You have to be, you have to be careful. To Roey's definition about AI originally, it really depends on the quality of the information put in in the box, which in this case is ChatGPT, and the information that's put in is like what's publicly indexed on the internet.

Russ: I feel grounded now. 

Madi: So kind of moving into the industrial space.

Russ: That's where I was gonna go. I was like, I'm grounded in a definition of ai. I feel comfortable with the fact that there's not one consolidated, clear citable definition of ai. I like your short definition and now I'm ready to ask you.

But what about industrial use cases? 

Roey: So I think we are progressing very, very nicely in this direction. And specifically I want to speak for a second about cameras as a really cool sensor and why I'm in love with cameras is that it's this one sensor that can really combine the physical world and the digital world in a unprecedented way.

So it actually can understand what's going on. We can extract information from the physical world. You know, it's not like a, okay, a temperature sensor will give us what is the temperature. But a sensor can really look at people, see what they're doing, look at text, look at pages, look at objects, look at parts, and I think that's really opened up so many different.

and that's only probably 10% of usage of AI in industrial space. 

Russ: Okay, we gotta stop and take a detour into vision because I spent, prior to joining Tulip last year, I spent close to a decade trying to get data out of industrial equipment, which was all run by computers. So like it's a computer, you wanna get the data out of the computer that's running the industrial system, you wanna put into another computer.

And there was like a big frustration and a big joke around the fact that you, you have a computer, it generates all this data, but there's no way to pull it straight off the computer. So you'll do things like program a robot to press the start button on a on a cycle start button on on a machine, and it's like you're completely skipping the computer interface.

But that frustration led me to think about cameras in exactly the same way you're describing where it's this sort of like super everything sensor. And I'm like, well, if, if you can, if you can take a picture and you can get the data from all of these different contextual things that are right there, isn't a $90 webcam, like the cheapest sensor that you could use ever anyway for for the broadest, the broadest scope.

So is that the way you're talking about here or is it something else?

Roey: A hundred percent and I can give you a cool example. So I was working with, uh, with a customer a few weeks ago and they had a legacy system probably from the nineties. I dunno, it was a Windows 95 or something older than that, and it was just impossible to connect that system to anything else. Unconnectable. 

So we took the HDMI cable that goes out of the computer and connected it just, you know, did screen capturing. So we just take the signal that runs out of the HDMI cable, which is basically a camera in a weird way. And then on top of that camera we took, uh, object character recognition, the term is OCL usually, and just extract information out of the computer screen and suddenly we are able to connect to this legacy system with vision.

Russ: Did you feel like a mad scientist as you were doing that? 

Roey: It's, it's hacking. It's a workaround, but think about the industrial system that industry shop floors, whatever. You have so many legacy systems, so many old machines that are not connected well and you know, giving your, your point and story in the beginning.

Connectivity is so significant. If you cannot connect, you don't have the data and then you cannot build anything else on top of it. 

Russ: In addition to all the, all the already amazing things about that scenario you just described, one of the headaches with pulling data off an industrial controller is it puts a load on the controller that the controller wasn't designed for.

So if you're streaming out additional data, that potentially can disrupt whatever the primary intended purpose for that, for that computer was. But if it's, if you're hijacking what was already signaled that was designed to come out for optical, it was gonna drive a monitor, instead it's driving this genius.

I'm just gonna sit back and relax and bask in the glory of what you've done here for a second. 

Madi: It is wild that you're taking a picture of an old system and like parsing it in the same way that you would have a person staring at that screen and like manually transferring the information, but in a much more efficient and less mind numbing for the person that would've previously had to do that.

Roey: Right. And it, and it's not different from from papers, right? Basically, this, this system has become like a paper. This screen is a piece of paper that show information and we can take, you know, physical papers and, and bring them into computer and digitalize them in a similar way. By the way, Think about the person looking at something.

There are so many people that can use another set of eyes to verify what they're doing, to help them make less mistakes, to find things that they missed. And I think, of course, industries we see more and more vision use cases like that that, you know, uh, augmenting the worker with cameras. And I think definitely in, in industry and Manufacturing, it's very relevant.

Madi: Vision systems are not new to Manufacturing, but this iteration is definitely different than the previous iterations. How is what you're describing and, and the way that you're thinking about vision and AI in an industrial context, like building on what was previously known or maybe challenging the way that it was previously done before?

Roey: Can I take you into a historical, uh, journey for ten seconds? 

Madi: Yes. 

Roey: So 2012 was a peaking point in ai and that was the introduction of deep learning in, in its modern way. Deep learning was invented in the fifties and was really kicking in in the eighties, nineties, but it really got into a peak usability point in 2012 with a case where they took 1 million images that divided them into 1000 classes and deep learning was able to classify these images into classes outperforming by 10% or more any previous method. and that was the modern way of AI that was introduced. 

Everything that you are seeing now from ChatGPT through Alexa, Google Translate, and any other AI that you can think of today is basically based on deep learning. And when that happened, it started, you know, a very complex, algorithmic way to do machine learning and ai.

10 years later, 2022, I think we're starting to see democratization of computer vision, so, If you think about industry, we don't have millions of data scientists working industry and Manufacturing. They work in big tech usually, and suddenly AI become democratized. ChatGPT is a great example of democratization because my kids can play with ChatGPT as well and do amazing things, and we saw so many examples in the, in the last few weeks, but vision is in the same state. Training a model to do classification, object detection, things like that is a no code, no effort type of of things. It's few hours and you have a ready model that you can use for so many different applications, and I think that's a, that's a real game changer that can really change the world in a way that suddenly it can be adopted at scale and make the difference. 

Madi: Can I ask you a challenge question here? Based on what you said. What you're describing it, it sounds like super exciting and really revolutionary, right? This idea of democratized ai, that anyone can do it, that it's now accessible, this like idea of like, Low lift, high impact, no needing to know code or data science.

But it also feels like if that's true, why isn't it everywhere? Like if it's that easy and accessible, like what is the thing that is tripping people up or, or blocking widespread adoption of this across the industrial space? 

Roey: For me, it's a question, not about ai. It's about technology. And first of all, I think technology starting to be everywhere, but it's going back to the business problem. AI can be used for so many things, but if it's not ROI driven, if it's not solving a problem, if it's not, you know, moving the needle for this company, for this, uh, Manufacturing environment, there is no reason to adopt it. And along the years, I'm in AI startups environment in the last 10 years, I guess, and I saw so many companies trying to solve different problems with limited value to the world.

And if it's not impactful, It could be cool. It could be nice, but what I'm going to do with that? Back in my PhD time in grad school, I was working on style transfer, which I'm sure you are familiar with these Van Gogh, like images of your grandmother . And that's cool, but it's not practical, right? It's not solving a business problem.

And we've seen a lot of things like that. I think we are reaching a point where there are many things that can done with ROI positive with AI in Manufacturing, in the world in general, and that's exciting. So if you think about, I dunno, a Gartner hype cycle. we are passing this point where the technology is valuable enough to solve real business problem.

Madi: So it sounds like the most broken thing about AI isn't ai, but people's understanding of how to attach all the possibility with the tool to their work and business. 

Roey: I guess it's falling in love with the technology, but not closing the loop with value and business problems. You know, dot com, back in the two thousands probably was exactly the same issue.

Everybody suddenly was able to open up a web page to do something, and in 2000 we discovered not all of these websites are really valuable to humanity and to people. So I guess, you know, we are reaching similar point in AI and, and I think that's good. That's healthy, that's, uh, maturity of the, of the technology.

Russ: Do you keep a little black book of prospective industrial AI use cases? 

Roey: Yeah. Um, I can give you a negative example if you want. Predictive maintenance is a very common use case in Manufacturing for ai, and it could be amazing. It could really be a game changer in many facilities, but it could be so unusable.

For example, a machine that brokes up 10 times. How predictive maintenance is going to help you? It's going to tell you it's going to break tomorrow. I know that already. And the opposite way is the machine that will break, uh, once a year, and the ability to predict something that will happen once a year is very, very low.

In the middle there are many, many use cases of predictive maintenance that are amazing and can really be a game changer in terms of fields. 

Russ: Can we play a quick game of is AI gonna fix it or not? Because I have my own list I've been carrying for 10 years, since deep learning became a thing. But I feel like a tourist in AI land, and I don't know if these are things that actually have a direct application of AI or if I am completely misunderstanding what this, what it can do.

Roey: Let's do that. 

Russ: All right. Uh, shift reports basically telling like a plant manager what went on today.

Roey: Very hard. I'm struggling to see that happening. I think maybe it's possible, but the amount of constraints and amount of changes that are happening in this system are probably too specific, too big, too dynamic to really be practical.

Russ: All right. Material reordering. 

Roey: Probably possible. It's some type of, I guess, an optimization problem. If you can collect enough data with different orders, probably possible. 

Russ: Okay. Pre-populating a list of reasons for downtime. 

Roey: Um, so you are thinking about root cause analysis, something like that.. 

Russ: Yeah, something like that.

So narrowing the universe from, I could type in anything I want to, like we've, we've anticipated that if this guy, if he, if he said that the reason that that something was down was waiting on a tool that can kind of like be automatically grouped in with the tool wasn't here like that? 

Roey: Yeah, I think, I think it's possible, like even highly applicable.

The challenge could be maybe to collect enough data of examples from the history, but if you have enough data and it's things that happen quite frequently, it's definitely could be. And you can even add the cameras angle that I mentioned. Cause I think collecting video of these events and what's really happened and look at them and extract some information out of that could be very, very useful.

Russ: I've only got one more and it's, it's a doozy. Routing or scheduling based on matching a capability of a piece of equipment or a line or something like that to a requirement of a part or an assembly.

Roey: It's roughly solved with ai. Uh, so I, I would characterize more than a optimization problem, which, you know, AI is subtype of optimization problem, but I think it's more of, um, probably graph theory type of problems than, than, than the planning AI type of problems. 

Russ: So that's, that's gonna be, uh, mostly solved, but it's not ai. 

Roey: Yeah, I think it's mostly solved. Yeah. Like, you know, 10 rooms, 20 people that need to do hundred tasks, and you need to do a schedule.

It's, first of all, a very complex problem, but it is solvable. It's interesting, by the way, to see that it is not that hard for people to do. Uh, but it's, it's one of the most challenging problem for computers. 

Russ: Take that computers . 

Roey: Yeah. You know, things like ChatGPT and also computer vision models. And language models.

They're very much like an infinity memory type of of device. This black box just has a lot of memory and it's very strong ability to do query. You know, if you want to classify an image and tell if it's a, if it's a dog or a cat, basically try to memorize how a dog looks like and how a cat looks like from seeing a lot of images.

That's, that's the learning process. And when you are using ChatGPT, it's doing quite the same with textual thing, but the, the magic that is why it's not really memory, it's when it's tried to generalized. So, okay. I saw 1000 images of a dog and I will be able to classify a dog image, although I haven't seen it, but I seen similar images, and that's the generalization.

The issue is that if you don't have enough data, the boundaries will be quite strong and it will fail. If this dog will start to look a little bit like a cat, but it's still a dog, and oh, if that system will see an image of a giraffe that was never introduced during the learning, uh, stage, no reason that it'll know what giraffe is.

So it's really bounded within the information that it's seen. and the ability to generalize this information. 

Madi: I've definitely seen what you're describing happen with, uh, consumer focused ai. I have a very fluffy dog and she's been AI cartoonized as a very beautiful cat, so hasn't seen a dog like her before on their training model.

Russ: Has she ever been identified as a giraffe? 

Madi: Unfortunately not, but I would love to see it. Neck is too short for, for the AI to make that guess. I have one last question for you, Roey. Kind of like looking forward, um, for the next year, what do you think is one of the bigger accomplishments that you would expect to, to happen or milestones in the maturity of ai?

Roey: Big one. Um, I don't have a crystal ball, yet, but I think that suddenly something changed completely in the world a month ago. In my point of view, when I open my computer in the morning, I open the email, I open WhatsApp cause I don't want to use the phone during the day, and I open set of tools that I'm using, uh, occasionally, documents, sheets, maybe my writing code, IDE and ChatGPT.

Seriously, it's consistently open in my browser cause I'm using it to write an SOW. I'm using it to write a LinkedIn post. I'm using it to tune my email and it's became a very practical tool for so many things, and I can see huge amount of practical applications that are suddenly easily achievable on top of this technology.

So I think we are really going to see so many cool ideas coming over in the coming year, and I'm very excited. 

Madi: I love that as a story that you're sharing with us, because what I take away from that is, Because it's pulling from things that already exist. What you're doing is getting rid of the like repetitive kind of boiler plate part of your work and really giving yourself more time to be creative and level stuff up.

And if we tie that back to Manufacturing and also how we think about the value of bringing in more flexible tech and just digitization from moving from like lower value, not worth, kind of like the amount of the human resources of like thinking through problems and giving all of those folks more time to, to do the problem solving.

This is basically an extension of that. 

Roey: I love what you just said cause I'm really not a big believer in automation. I'm a big believer in augmentation and I think that AI can really augment. People, and it can augment, you know, another set of eyes to do visual inspection in the shop floor and can augment my work when I need to write a document and I need, you know, a structure so I can, you know, here's the structure of the document. This is how an SOW should look like. 

Now, you know, do your thing, be creative. Put the important aspects instead of, okay, another SOW title, you know, subsections and then normal, you know, I don't know, limitations, scope, whatever I need to write. And just as an example. And I think it's so, so helpful. 

Madi: Well, thank you for being our first guest. A plus, a plus, uh, starter guest for behind the ops. Uh, Russ, do you wanna take us out? 

Russ: This has been Dr. Roey for Behind the Ops. Thanks so much for coming. 

Roey: Pleasure to be here. 

Madi: See you next time. 

Russ: See you next time.

Outro: Behind the Ops is brought to you by Tulip. Connect the people, machines, devices, and systems used in your production and logistics processes with our frontline operations platform. Visit Tulip.Co to learn more. 

The show is produced by Jasmine Chan and edited by Thom Obarski. If you enjoyed listening, support the show by leaving us a quick rating or review. It really helps. If you have feedback for this or any of our other episodes, you can reach us at behindtheops@tulip.co