Episode 103

Human-First AI with Christopher Nguyen


November 23rd, 2022

42 mins 30 secs

Season 3

Your Host
Special Guest

About this Episode

Augmented reveals the stories behind the new era of industrial operations, where technology will restore the agility of frontline workers.

In this episode of the podcast, the topic is Human-First AI. Our guest is Christopher Nguyen, CEO, and Co-Founder of Aitomatic. In this conversation, we talk about the why and the how of human-first AI because it seems that digital AI is one thing, but physical AI is a whole other ballgame in terms of finding enough high-quality data to label the data correctly. The fix is to use AI to augment existing workflows. We talk about fishermen at Furuno, human operators in battery factories at Panasonic, and energy optimization at Westinghouse.

If you like this show, subscribe at augmentedpodcast.co. If you like this episode, you might also like Episode 80: The Augmenting Power of Operational Data, with Tulip's CTO, Rony Kubat.

Augmented is a podcast for industry leaders, process engineers, and shop floor operators, hosted by futurist Trond Arne Undheim and presented by Tulip.

Follow the podcast on Twitter or LinkedIn.

Trond's Takeaway:

Physical AI is much more interesting of a challenge than pure digital AI. Imagine making true improvements to the way workers accomplish their work, helping them be better, faster, and more accurate. This is the way technology is supposed to work, augmenting humans, not replacing them. In manufacturing, we need all the human workers we can find. As for what happens after the year 2100, I agree that we may have to model what that looks like. But AIs might be even more deeply embedded in the process, that's for sure.


TROND: Welcome to another episode of the Augmented Podcast. Augmented brings industrial conversations that matter, serving up the most relevant conversations in industrial tech. Our vision is a world where technology will restore the agility of frontline workers.

In this episode of the podcast, the topic is Human-First AI. Our guest is Christopher Nguyen, CEO, and Co-Founder of Aitomatic. In this conversation, we talk about the why and the how of human-first AI because it seems that digital AI is one thing, but physical AI is a whole other ballgame in terms of finding enough high-quality data to label the data correctly. The fix is to use AI to augment existing workflows. We talk about fishermen at Furuno, human operators in battery factories at Panasonic, and energy optimization at Westinghouse.

Augmented is a podcast for industrial leaders, process engineers, and for shop floor operators hosted by futurist Trond Arne Undheim and presented by Tulip.

Christopher, how are you? And welcome.

CHRISTOPHER: Hi, Trond. How are you?

TROND: I'm doing great. I thought we would jump into a pretty important subject here on human-first AI, which seems like a juxtaposition of two contradictory terms, but it might be one of the most important types of conversations that we are having these days.

I wanted to introduce you quickly before we jump into this. So here's what I've understood, and you correct me if I'm wrong, but you are originally from Vietnam. This is back in the late '70s that you then arrived in the U.S. and have spent many years in Silicon Valley mostly. Berkeley, undergrad engineering, computer science, and then Stanford Ph.D. in electrical engineering. You're a sort of a combination, I guess, of a hacker, professor, builder. Fairly typical up until this point of a very successful, accomplished sort of Silicon Valley immigrant entrepreneur, I would say, and technologist.

And then I guess Google Apps is something to point out. You were one of the first engineering directors and were part of Gmail, and Calendar, and a bunch of different apps there. But now you are the CEO and co-founder of Aitomatic. What we are here to talk about is, I guess, what you have learned even in just the last five years, which I'm thrilled to hear about. But let me ask you this first, what is the most formational and formative experience that you've had in these years? So obviously, immigrant background and then a lot of years in Silicon Valley, what does that give us?

CHRISTOPHER: I guess I can draw from a lot of events. I've always had mentors. I can point out phases of my life and one particular name that was my mentor. But I guess in my formative years, I was kind of unlucky to be a refugee but then lucky to then end up in Silicon Valley at the very beginning of the PC revolution. And my first PC was a TI-99/4A that basically the whole household could afford. And I picked it up, and I have not stopped hacking ever since. So I've been at this for a very long time.

TROND: So you've been at this, which is good because actually, good hacking turns out takes a while. But there's more than that, right? So the story of the last five years that's interesting to me because a lot of people learn or at least think they learn most things early. And you're saying you have learned some really fundamental things in the last five years. And this has to do with Silicon Valley and its potential blindness to certain things. Can you line that up for us? What is it that Silicon Valley does really well, and what is it that you have discovered that might be an opportunity to improve upon?

CHRISTOPHER: Well, I learn new things every four or five years. I actually like to say that every four or five years, I look back, and I say, "I was so stupid five years ago." [laughs] So that's been the case.

TROND: That's a very humbling but perhaps a very smart knowledge acquisition strategy, right?

CHRISTOPHER: Yeah. And in the most recent five years...so before co-founding Aitomatic, which is my latest project and really with the same team...and I can talk a lot more about that. We've worked with each other for about ten years now. But in the intervening time, there's a four-and-a-half-year block when we were part of Panasonic. So we had a company called Arimo that was acquired by Panasonic for our machine learning AI skills and software.

And I would say if you look at my entire history, even though I did start with my degree in semiconductor all the way down to device physics and Intel and so on, but in terms of a professional working career, that was the first time we actually faced the physical world as a Silicon Valley team. And anybody who's observed Silicon Valley in the last 15-20 years, certainly ten years, has seen a marked change in terms of the shift from hardware to software. And my friend Marc Andreessen likes to say, "Software is eating the world."

If you look at education, you know, the degrees people are getting, it has shifted entirely from engineering all the way to computer science. And the punch line, I guess, the observation is that we Silicon Valley people do not get physical. We don't understand the manufacturing world. We don't know how to do HVAC and so on. And so when we build software, we tend to go for the digital stuff.

TROND: Christopher, it's almost surprising given the initial thrust of Silicon Valley was, of course, hardware. So it's not surprising to me, I guess because I've been observing it as well. But it is striking more than surprising that a region goes through paradigms.

CHRISTOPHER: Yeah. Yeah. And it's a global trend. It's the offshoring of low-end, shall we say, low-value manufacturing and so on. And we're discovering that we actually went a little too far. So we don't have the skill set, the expertise anymore. And it's become a geopolitical risk.

TROND: Right. Well, a little bit too far, maybe, or not far enough. Or, I mean, tell us what it is that you're losing when you lose the hardware perspective, particularly in this day and age with the opportunities that we're about to talk about.

CHRISTOPHER: Well, I can talk specifically about the things that touch my immediate spheres. Maybe you can think abstractly about the lack of tooling expertise and manufacturing know-how, and so on. But as part of Panasonic, the acquisition was all about taking a Silicon Valley team and injecting AI, machine learning across the enterprise. And so we were part of that global AI team reporting to the CTO office.

And we found out very quickly that a lot of the software techniques, the machine learning, for example, when you think about people saying data is the fuel for machine learning and specifically labeled data, right? In the digital world, the Google place that I came from, it was very easy to launch a digital experiment and collect labels, decisions made by users. You can launch that in the morning, and by evening you're building examples. You can't do that in the physical world. Atoms move a lot more slowly. And so when you try to do something like predictive maintenance, you don't have enough failure examples to train machine learning models from.

So all of the techniques, all of the algorithms that we say we developed from machine learning that seem to work so well, it turns out it worked so well because the problem space that we worked on has been entirely digital, and they all fail when it comes to manufacturing, the things that you can touch and feel, you know, cars that move and so on.

TROND: I want to ask you this, Christopher, because the first company you helped co-found was, in fact, a contract manufacturer. Do you think that reflecting on this long career of yours and these various experiences, what was it that convinced you before others? I mean, you're not the only one now in the Valley that has started to focus on manufacturing and including hardware again, but it is rare still. What does it require to not just think about manufacturing but actually start to do compute for manufacturing? Is it just a matter of coming up with techniques? Or is it a whole kind of awareness that takes longer? So, in your case, you've been aware of manufacturing, acutely aware of it for decades.

CHRISTOPHER: I would say there are two things, one is obvious, and the other was actually surprising to me. The obvious one is, of course, knowledge and experience. When we work on sonar technology that shoots a beam down an echogram that comes back to detect fish in the ocean, it's very necessary, not just convenient, but necessary for the engineers that work on that to understand the physics of sound waves travel underwater, and so on.

So that education, I have long debates, and it's not just recently. When we were trying to structure a syllabus for a new university, I had long debates with my machine-learning friends, and they said, "We don't need physics." And I said, "We need physics." That's one thing. But you can concretely identify you need to know this. You need to know this. So if you're going to do this, learn the following thing.

The thing that was more unexpected for me in the last five years as I sort of sound this bell of saying, hey, we need to modify our approach; we need to optimize our algorithms for this world, is a cultural barrier. It's kind of like the story of if you have a hammer, you want to go look for nails. So Silicon Valley today does not want to look for screwdrivers yet for this world.

TROND: So you're saying Silicon Valley has kind of canceled the physical world? If you want to be really sort of parabolic about this, it's like software is eating the world, meaning software is what counts, and it's so efficient. Why go outside this paradigm, basically? If there's a problem that apparently can't be fixed by software, it's not a valuable problem.

CHRISTOPHER: Or I can't solve that problem with my current approach. I just have to squint at it the right way. I have to tweak the problem this way and so on despite the fact that it's sort of an insurmountable challenge if you tried to do so. And concretely, it is like, just give me enough data, and I'll solve it. And if you don't have enough data, you know what? Go back and get more data. [chuckles] That's what I myself literally said. But people don't have the luxury of going back to get more data. They have to go to market in six months, and so on.

TROND: Right. And so manufacturing...and I can think of many use cases where obviously failure, for example, is not something...you don't really want to go looking for more failure than you have or artificially create failure in order to stress test something unless that's a very safe way of doing so. So predictive maintenance then seems like a, I guess, a little bit of a safer space. But what is it about that particular problem that then lends itself to this other approach to automating labeling? Or what exactly is it that you are advocating one should do to bridge to digital and the physical AIs?

CHRISTOPHER: I actually disagree that it is a safer space.

TROND: Oh, it's not a safer space to you.

CHRISTOPHER: That itself there's a story in that, so let's break that down.

TROND: Let's do it.

CHRISTOPHER: So, again, when I say Silicon Valley, it is a symbol for a larger ecosystem that is primarily software and digital. And when I say we, because I've worn many hats, I have multiple wes, including academia; I've been a professor as well. When we approach the predictive maintenance problem, if you approach it as machine learning, you got to say, "Do this with machine learning," the first thing you ask for...let's say I'm a data scientist; I'm an AI engineer.

You have this physical problem. It doesn't matter what it is; just give me the dataset. And the data set must have rows and columns, and the rows are all the input variables. And then there should be some kind of column label. And in this case, it'll be a history of failures of compressors failing, you know, if the variables are such, then it must be a compressor. If the variables are such, it must be the air filter, and so on.

And it turns out when you ask for that kind of data, you get ten rows. [laughs] That's not enough to do machine learning on. So then people, you know, machine learning folks who say they've done predictive maintenance, they actually have not done predictive maintenance. That's the twist. What they have done is anomaly detection, which machine learning can do because, with anomaly detection, I do not need that failure label. It just gives me all the sensor data.

What anomaly detection really does is it learns the normal patterns. If you give it a year's worth of data, it'll say, okay, now I've seen a year's worth of data. If something comes along that is different from the past patterns; I will tell you that it's different. That's only halfway to predictive maintenance. That is detecting that something is different today. That is very different from, and it isn't predicting, hey, that compressor is likely to fail about a month from now.

And that when we were part of Panasonic, it turns out the first way...and we solved it exactly the way I've described. We did it with the anomaly detection. And then we threw it over the wall to the engineer experts and said, "Well, now that you have this alert, go figure out what may be wrong." And half of the time, they came back and said, "Oh, come on, it was just a maintenance event. Why are you bothering me with this?"

TROND: But, Christopher, leveraging human domain expertise sounds like a great idea. But it can't possibly be as scalable as just leveraging software. So how do you work with that? And what are the gains that you're making?

CHRISTOPHER: I can show you the messenger exchange I had with another machine-learning friend of mine who said exactly the same thing yesterday, less than 24 hours ago.

TROND: [laughs]

CHRISTOPHER: He said, "That's too labor-intensive." And I can show you the screen.

TROND: And how do you disprove this?

CHRISTOPHER: Well, [chuckles] it's not so much disproving, but the assumption that involving humans is labor-intensive is only true if you can't automate it. So the key is to figure out a way, and 10-20 years ago, there was limited technology to automate or extract human knowledge, expert systems, and so on. But today, technologies...the understanding of natural language and so on, machine learning itself has enabled that. That turns out to be the easier problem to solve. So you take that new tool, and you apply it to this harder physical problem.

TROND: So let's go to a hard, physical problem. You and I talked about this earlier, and let's share it with people. So I was out fishing in Norway this summer. And I, unfortunately, didn't get very much fish, which obviously was disappointing on many levels. And I was a little surprised, I guess, of the lack of fish, perhaps. But I was using sonar to at least identify different areas where people had claimed that there were various types of fish. But I wasn't, I guess, using it in a very advanced way, and we weren't trained there in the boat.

So we sort of had some sensors, but we were not approaching it the right way. So that helped me...and I know you work with Furuno, and Garmin is the other obviously player in this. So fish identification and detection through sonar technology is now the game, I guess, in fishery and, as it turns out, even for individuals trying to fish these days. What is that all about? And how can that be automated, and what are the processes that you've been able to put in place there?

CHRISTOPHER: By the way, that's a perfect segue into it. I can give a plug perhaps for this conference that I'm on the organizing committee called Knowledge-First World. And Furuno is going to be presenting their work exactly, talking a lot about what you're talking about. This is kind of coming up in November. It is the first conference of its kind because this is AI Silicon Valley meets the physical world.

I think you're talking about the fish-finding technology from companies like Furuno, and they're the world's largest market share in marine navigation and so on. And the human experts in this are actually not even the engineers that build these instruments; it's the fishermen, right? The fishermen who have been using this for a very long time combine it with their local knowledge, you know, warm water, cold water, time of day, and so on. And then, after a while, they recognize patterns that come back in this echogram that match mackerel, or tuna, or sardines, and so on.

And Furuno wants to capture that knowledge somehow and then put that model into the fish-finding machine that you and I would hold. And then, instead of seeing this jumbled mess of the echogram data, we would actually see a video of fish, for example. It's been transformed by this algorithm.

TROND: So, I mean, I do wish that we lived in a world where there was so much fish that we didn't have to do this. But I'm going to join your experiment here. And so what you're telling me is by working with these experts who are indeed fishermen, they're not experts in sonar, or they're not experts in any kind of engineering technology, those are obviously the labelers, but they are themselves giving the first solutions for how they are thinking about the ocean using these technologies. And then somehow, you are turning that into an automatable, an augmented solution, essentially, that then can find fish in the future without those fishermen somehow being involved the next time around because you're building a model around it.

CHRISTOPHER: I'll give you a concrete explanation, a simplified version of how it works, without talking about the more advanced techniques that are proprietary to Furuno. The conceptual approach is very, very easy to understand, and I'll talk about it from the machine learning perspective.

Let's say if I did have a million echograms, and each echogram, each of these things, even 100,000, is well-labeled. Somebody has painstakingly gone through the task of saying, okay, I'm going to circle this, and that is fish. And that is algae, and that's sand, and that's marble. And by the way, this is a fish, and this is mackerel, and so on. If somebody has gone through the trouble of doing that, then I can, from a human point of view, just run an algorithm and train it. And then it'll work for that particular region, for that particular time. Okay, well, we need to go collect more data, one for Japan, the North Coast, and one for Southwestern.

So that's kind of a lot of work to collect essentially what this pixel data is, this raw data. When you present it to an experienced fisherman, he or she would say, "Well, you see these bubbles here, these circles here with a squiggly line..." So they're describing it in terms of human concepts. And then, if you sit with them for a day or two, you begin to pick up these things. You don't need 100,000-pixel images. You need these conceptual descriptions.

TROND: So you're using the most advanced AI there is, which is the human being, and you're using them working with these sonar-type technologies. And you're able to extract very, very advanced models from it.

CHRISTOPHER: The key technology punch line here is if you have a model that understands the word circle and squiggly line, which we didn't before, but more recently, we begin to have models, you know, there are these advances called large language models. You may have heard of GPT-3 and DALL-E and so on, you know, some amazing demonstrations coming out of OpenAI and Google. In a very simplified way, we have models that understand the world now. They don't need raw pixels. These base models are trained from raw pixels, but then these larger models understand concepts. So then, we can give directions at this conceptual level so that they can train other models. That's sort of the magic trick.

TROND: So it's a magic trick, but it is still a difficult world, the world of manufacturing, because it is physical. Give me some other examples. So you worked with Panasonic. You're working with Furuno in marine navigation there and fishermen's knowledge. How does this work in other fields like robotics, or with car manufacturing, or indeed with Panasonic with kind of, I don't know, battery production or anything that they do with electronics?

CHRISTOPHER: So, to give you an example, you mentioned a few things that we worked on, you know, robotics in manufacturing, robotics arm, sort of the manufacturing side, and the consistency of battery sheets coming off the Panasonic manufacturing line in Sparks, Nevada as well as energy optimization at Westinghouse. They supply into data centers, and buildings, and so on.

And so again, in every one of these examples, you've got human expertise. And, of course, this is much more prevalent in Asia because Asia is still building things, but some of that is coming back to the U.S. There are usually a few experts. And by the way, this is not about thousands of manufacturing line personnel. This is about three or four experts that are available in the entire company. And they would be able to give heuristics. –They will be able to describe at the conceptual level how they make their decisions.

And if you have the technology to capture that in a very efficient way, again, coming back to the idea that if you make them do the work or if you automate their work, but in a very painstaking way like thousands of different rules, that's not a good proposition. But if you have some way to automate the automation, automate the capturing of that knowledge, you've got something that can bridge this physical, digital divide.


In the new book from Wiley, Augmented Lean: A Human-Centric Framework for Managing Frontline Operations, serial startup founder Dr. Natan Linder and futurist podcaster Dr. Trond Arne Undheim deliver an urgent and incisive exploration of when, how, and why to augment your workforce with technology, and how to do it in a way that scales, maintains innovation, and allows the organization to thrive. The key thing is to prioritize humans over machines.

Here's what Klaus Schwab, Executive Chairman of the World Economic Forum, says about the book: "Augmented Lean is an important puzzle piece in the fourth industrial revolution."

Find out more on www.augmentedlean.com, and pick up the book in a bookstore near you.

TROND: How stable is that kind of model knowledge? Because I'm just thinking about it in the long run here, are these physical domain experts that are giving up a little bit of their superpower are they still needed then in a future scenario when you do have such a model? Or will it never be as advanced as they are? Or is it actually going to be still kind of an interface that's going to jump between machines and human knowledge kind of in a continuous loop here?

CHRISTOPHER: Yeah, in the near term, it turns out we're not working on replacing experts as much as scaling experts. Almost every case we've worked on, companies are in trouble largely because the experts are very, very few and far between, and they're retiring. They're leaving. And that needs to be scaled somehow. In the case of, for example, the cold chain industry all of Japan servicing the supermarkets, you know, there's 7-ELEVEN, there's FamilyMart, and so on, there are three experts who can read the sensor data and infer what's likely to fail in the next month. So in the near term, it's really we need these humans, and we need more of them.

TROND: I'm glad to hear that even that is a bit of a contrarian message. So you're saying physical infrastructure and the physical world matters. You're saying humans matter. [laughs] It's interesting. Yeah, that's contrarian in Silicon Valley, I'll tell you that.

CHRISTOPHER: It is. And, in fact, related to that problem, Hussmann, which is a refrigeration company, commercial refrigeration supplies to supermarkets. It was a subsidiary of Panasonic. It has a really hard time getting enough service personnel, and they have to set up their own universities, if you will, to train them. And these are jobs that pay very well. But everybody wants to be in software these days.

Coming back to the human element, I think that long-term I'm an optimist, not a blind optimist but a rational one. I think we're still going to need humans to direct machines. The machine learning stuff is data that reflects the past, so patterns of the past, and you try to project that in the future. But we're always trying to effect some change to the status quo. Tomorrow should be a better day than today. So is that human intent that is still, at least at present, lacking in machines? And so we need humans to direct that.

TROND: So what is the tomorrow of manufacturing then? How fast are we going to get there? Because you're saying, well, Silicon Valley has a bit of a learning journey. But there is language model technology or progress in language models that now can be implemented in software and, through humans, can be useful in manufacturing already today. And they're scattered examples, and you're putting on an event to show this. What is the path forward here, and how long is this process? And will it be an exponential kind of situation here where you can truly integrate amazing levels of human insight into these machine models? Or will it take a while of tinkering before you're going to make any breakthroughs?

Because one thing is the breakthrough in understanding human language, but what you're saying here is even if you're working only with a few experts, you have to take domain by domain, I'm assuming, and build these models, like you said, painstakingly with each expert in each domain. And then, yes, you can put that picture together. But the question is, how complex of a picture is it that you need to put together? Is it like mapping the DNA, or is it bigger? Or what kind of a process are we looking at here?

CHRISTOPHER: If we look at it from the dimension of, say, knowledge-based automation, in a sense, it is a continuation. I believe everything is like an s-curve. So there's acceleration, and then there's maturity, and so on. But if you look back in the past, which is sort of instructive for the future, we've always had human knowledge-based automation.

I remember the first SMT, the Surface Mount Technology, SMT wave soldering machine back in the early '90s. That was a company that I helped co-found. It was about programming the positioning of these chips that would just come down onto the solder wave. And that was human knowledge for saying, move it up half a millimeter here and half a millimeter there. But of course, the instructions there are very micro and very specific.

What machine learning is doing...I don't mean to sort of bash machine learning too much. I'm just saying culturally, there's this new tool really that has come along, and we just need to apply the tool the right way. Machine learning itself is contributing to what I described earlier, that is, now, finally, machines can understand us at the conceptual level that they don't have to be so, so dumb as to say, move a millimeter here, and if you give them the wrong instruction, they'll do exactly that. But we can communicate with them in terms of circles and lines, and so on.

So the way I see it is that it's still a continuous line. But what we are able to automate, what we're able to ask our machines to do, is accelerating in terms of their understanding of these instructions. So if you can imagine what would happen when this becomes, let's say, ubiquitous, the ability to do this, and I see this happening over the next...Certainly, the base technology is already there, and the application always takes about a decade.

TROND: Well, the application takes a decade. But you told me earlier that humans should at least have this key role in this knowledge-first application approach until 2100, you said, just to throw out a number out there. That's, to some people, really far away. But the question is, what are you saying comes after that? I know you throw that number out.

But if you are going to make a distinction between a laborious process of painful progress that does progress, you know, in each individual context that you have applied to human and labeled it, and understood a little case, what are we looking at, whether it is 2100, 2075, or 2025? What will happen at that moment? And is it really a moment that you're talking about when machines suddenly will grasp something very, very generic, sort of the good old moment of singularity, or are you talking about something different?

CHRISTOPHER: Yeah, I certainly don't think it's a moment. And, again, the HP-11C has always calculated Pi far faster and with more digits than I have. So in that sense, in that particular narrow sense, it's always been more intelligent than I am.

TROND: Yeah. Well, no one was questioning whether a calculator could do better calculations than a human. For a long time --

CHRISTOPHER: Hang on. There's something more profound to think about because we keep saying, well, the minute we do something, it's okay; that's not intelligence. But what I'm getting to is the word that I would refer to is hyper-evolution. So there's not a replacement of humans by machines. There's always been augmentation, and intelligence is not going to be different. It is a little disturbing to think about for some of us, for a lot of us, but it's not any different from wearing my glasses.

Or I was taking a walk earlier this morning listening to your podcast, and I was thinking how a pair of shoes as an augmented device would seem very, very strange to humans living, say, 500 years ago, the pair of shoes that I was walking with. So I think in terms of augmenting human intelligence, there are companies that are working on plugging in to the degree that that seems natural or disturbing. It is inevitable.

TROND: Well, I mean, if you just think about the internet, which nowadays, it has become a trope to think about the internet. I mean, not enough people think about the internet as a revolutionary technology which it, of course, is and has been, but it is changing. But whether you're thinking about shoes, or the steam engine, or nuclear power, or whatever it is, the moment it's introduced, and people think they understand it, which most people don't, and few of us do, it seems trivial because it's there.

CHRISTOPHER: That's right.

TROND: But your point is until it's there, it's not trivial at all. And so the process that you've been describing might sound trivial, or it might sound complex, but the moment it's solved or is apparently solved to people, we all assume that was easy. So there's something unfair about how knowledge progresses, I guess.

CHRISTOPHER: That's right. That's right. We always think, yeah, this thing that you describe or I describe is very, very strange. And then it happens, and you say, "Of course, that's not that interesting. Tell me about the future."

TROND: Well, I guess the same thing has happened to cell phones. They were kind of a strange thing that some people were using. It was like, okay, well, how useful is it to talk to people without sitting by your desk or in the corner of your house?

CHRISTOPHER: I totally remember when we were saying, "Why the hell would I want to be disturbed every moment of the day?" [laughs] I don't want the phone with me, and now I --

TROND: Right. But then we went through the last decade or so where we were saying, "I can't believe my life before the phone." And then maybe now the last two, three years, I would say a lot of people I talk to or even my kids, they're like, "What's the big deal here? It's just a smartphone," because they live with a smartphone. And they've always had it.

CHRISTOPHER: They say, "How did you get around without Google Maps?" And then somebody says, "We used maps." And I said, "Before Google Maps."


TROND: Yeah. So I guess the future here is an elusive concept. But I just want to challenge you one more time then on manufacturing because manufacturing, for now, is a highly physical exercise. And, of course, there's virtual manufacturing as well, and it builds on a lot of these techniques and machine learning and other things. How do you see manufacturing as an industry evolve? Is it, like you said, for 75 years, it's going to be largely very recognizable? Is it going to look the same? Is it going to feel the same?

Is the management structure the way engineers are approaching it, and the way workers are working? Are we going to recognize all these things? Or is it going to be a little bit like the cell phone, and we're like, well, of course, it's different. But it's not that different, and it's not really a big deal to most people.

CHRISTOPHER: Did you say five years or 50 years?

TROND: Well, I mean, you give me the timeframe.

CHRISTOPHER: Well, in 5 years, we will definitely recognize it, but in 50 years, we will not

TROND: In 50 years, it's going to be completely different, look different, feel different; factories are all going to be different.

CHRISTOPHER: Right, right. I mean, the cliché is that we always overestimate what happens in 5 and underestimate what happens in 50. But the trend, though, is there's this recurring bundling and unbundling of industries; it's a cycle. Some people think it's just, you know, they live ten years, and they say it's a trend, but it actually goes back and forth. But they're sort of increasing specialization of expertise.

So, for example, the supply chain over the last 30 years, we got in trouble because of that because it has become so discrete if you want to use one friendly word, but you can also say fragmented in another word. Like, everybody has been focused on just one specialization, and then something like COVID happens and then oh my God, that was all built very precisely for a particular way of living. And nobody's in the office anymore, and we live at home, and that disrupts the supply chain.

I think if you project 50 years out, we will learn to essentially matrix the whole industry. You talked about the management of these things. The whole supply chain, from branding all the way down to raw materials, is it better to be completely vertically integrated to be part of this whole mesh network? I think the future is going to be far more distributed. But there'll be fits and starts.

TROND: So then my last question is, let's say I buy into that. Okay, let's talk about that for a second; the future is distributed or decentralized, whatever that means. Does that lessen or make globalization even more important and global standardization, I guess, across all geographical territories? I'm just trying to bring us back to where you started with, which was in the U.S., Silicon Valley optimized for software and started thinking that software was eating the world. But then, by outsourcing all of the manufacturing to Asia, it forgot some essential learning, which is that when manufacturing evolves, the next wave looks slightly different. And in order to learn that, you actually need to do it.

So does that lesson tell you anything about how the next wave of matrix or decentralization is going to occur? Is it going to be...so one thought would be that it is physically distributed, but a lot of the insights are still shared. So, in other words, you still need global insight sharing, and all of that is happening. If you don't have that, you're going to have pockets that are...they might be very decentralized and could even be super advanced, but they're not going to be the same. They're going to be different, and they're going to be different paths and trajectories in different parts of the world.

How do you see this? Do you think that our technology paradigms are necessarily converging along the path of some sort of global master technology and manufacturing? Or are we looking at scattered different pictures that are all decentralized, but yet, I don't know, from a bird's eye view, it kind of looks like a matrix?

CHRISTOPHER: I think your question is broader than just manufacturing, although manufacturing is a significant example of that, right?

TROND: It's maybe a key example and certainly under-communicated. And on this podcast, we want to emphasize manufacturing, but you're right, yes.

CHRISTOPHER: The word globalization is very loaded. There's the supposedly positive effect in the long run. But who is it that said...is it Keynes that said, "In the long run, we're all dead?" [laughs] In the short run, the dislocations are very real. A skill set of a single human being can't just shift from hardware to software, from manufacturing to AI, within a few months.

But I think your question is, let's take it seriously on a scale of, say, decades. I think about it in terms of value creation. There will always be some kind of disparity. Nature does not like uniformity. Uniformity is coldness; it is death. There have to be some gradients. You're very good at something; I'm very good at something else. And that happens at the scale of cities and nations as well.

TROND: And that's what triggers trade, too, right?


TROND: Because if we weren't different, then there would be no incentive to trade.

CHRISTOPHER: So when we think about manufacturing coming back to the U.S., and we can use the word...it is correct in one sense, but it's incorrect in another sense. We're not going back to manufacturing that I did. We're not going back to surface mount technology. In other words, the value creation...if we follow the trajectory of manufacturing alone and try to learn that history, what happens is that manufacturing has gotten better and better. Before, we were outsourcing the cheap stuff. We don't want to do that. But then that cheap stuff, you know, people over there build automation and skills, and so on. And so that becomes actually advanced technology.

So in a sense, what we're really doing is we're saying, hey, let's go advanced at this layer. I think it's going to be that give and take of where value creation takes place, of course, layered with geopolitical issues and so on.

TROND: I guess I'm just throwing in there the wedge that you don't really know beforehand. And it was Keynes, the economist, that said that the only thing that matters is the short term because, in the end, we are all dead eventually. But the point is you don't really know. Ultimately, what China learned from manufacturing pretty pedestrian stuff turned out to be really fundamental in the second wave.

So I'm just wondering, is it possible to preempt that because you say, oh, well, the U.S. is just going to manufacture advanced things, and then you pick a few things, and you start manufacturing them. But if you're missing part of the production process, what if that was the real advancement? I guess that is what happened.

CHRISTOPHER: Okay. So when I say that, I think about the example of my friend who spent, you know, again, we were a Ph.D. group at Stanford together. And whereas I went off to academia and did startups and so on, he stayed at Intel for like 32 years. He's one of the world's foremost experts in semiconductor process optimization. So that's another example where human expertise, even though semiconductor manufacturing is highly automated, you still need these experts to actually optimize these things. He's gone off to TSMC after three decades of being very happy at one place.

So what I'm getting to is it is actually knowable what are the secret recipes, where the choke points are, what matters, and so on. And interestingly, it does reside in the human brain. But when I say manufacturing coming back to the U.S. and advanced manufacturing, we are picking and choosing. We're doing battery manufacturing. We're doing semiconductor, and we're not doing wave soldering.

So I think it is possible to also see this trend that anybody who's done something and going through four or five iterations of that for a long time will become the world's expert at it. I think that is inevitable. You talk of construction, for example; interestingly, this company in Malaysia that is called Renong that is going throughout Southeast Asia; they are the construction company of the region because they've been doing it for so long. I think that is very, very predictable, but it does require the express investment in that direction. And that's something that Asia has done pretty well.

TROND: Well, these are fascinating things. We're not going to solve them all on this podcast. But definitely, becoming an expert in something is important, whether you're an individual, or a company, or a country for sure. What that means keeps changing. So just stay alert, and stay in touch with both AI and humans and manufacturing to boot. It's a mix of those three, I guess. In our conversation, that's the secret to unlocking parts of the future. Thank you, Christopher, for enlightening us on these matters. I appreciate it.

CHRISTOPHER: It's my pleasure.

TROND: You have just listened to another episode of the Augmented Podcast with host Trond Arne Undheim. The topic was Human-First AI. Our guest was Christopher Nguyen, CEO, and Co-Founder of Aitomatic. In this conversation, we talked about the why and the how of human-first AI because it seems that digital AI is one thing, but physical AI is a whole other ballgame.

My takeaway is that physical AI is much more interesting of a challenge than pure digital AI. Imagine making true improvements to the way workers accomplish their work, helping them be better, faster, and more accurate. This is the way technology is supposed to work, augmenting humans, not replacing them. In manufacturing, we need all the human workers we can find. As for what happens after the year 2100, I agree that we may have to model what that looks like. But AIs might be even more deeply embedded in the process, that's for sure.

Thanks for listening. If you liked the show, subscribe at augmentedpodcast.co or in your preferred podcast player, and rate us with five stars. If you liked this episode, you might also like Episode 80: The Augmenting Power of Operational Data, with Tulip's CTO, Rony Kubat as our guest. Hopefully, you'll find something awesome in these or in other episodes, and if so, do let us know by messaging us. We would love to share your thoughts with other listeners.

The augmented podcast is created in association with Tulip, the frontline operation platform that connects the people, machines, devices, and systems used in a production and logistics process in a physical location. Tulip is democratizing technology and empowering those closest to operations to solve problems. Tulip is also hiring. You can find Tulip at tulip.co.

Please share this show with colleagues who care about where industry and especially about how industrial tech is going. To find us on social media is easy; we are Augmented Pod on LinkedIn and Twitter and Augmented Podcast on Facebook and on YouTube.

Augmented — industrial conversations that matter. See you next time.