In this episode of The Parlor Room Presents: Hello AI, host and Harvard Business School Online Creative Director Chris Linnane speaks with HBS Professor Willy Shih about AI, disruptive innovation, and the future of competition.
Shih explains why AI isn't inherently disruptive, explores how organizations can use it to improve productivity and supply chains, and shares how leaders can identify innovation opportunities while avoiding the pitfalls of overserving the market.
Resources
Learn more from Shih in the HBS Online course Disruptive Strategy
Catch up on previous episodes of The Parlor Room, featuring Shih:
You can also watch The Parlor Room Presents: Hello AI on YouTube.
Transcript
Editor's Note: The following was prepared by a machine algorithm and may not perfectly reflect the interview's audio file.
Willy Shih: That's an example I would call gratuitous.
Voiceover: Willy Shih is a professor of management at Harvard Business School.
Willy Shih: No technology, inherently, is disruptive or sustaining.
Voiceover: Understanding what truly causes disruption is essential.
Willy Shih: Areas like that, I think, offer tremendous opportunities for innovation.
Voiceover: Recognizing opportunities for innovation will be crucial in the coming years.
Willy Shih: Some of the most important steps I've taken in my career have involved huge step backs in pain.
Voiceover: Welcome to The Parlor Room Presents: Hello AI.
Chris Linnane: When you think back to before AI dominated all the conversations that are going on right now, what was the state of manufacturing and supply chains?
Willy Shih: Well, I think you raise a very good point because AI really had started dominating the conversations with the appearance of these large language models—ChatGPT, in particular, when it came onto the scene. Now, that's all anybody can talk about. But actually, if you look back at what was going on in manufacturing and a lot of the sciences, machine learning applications have been coming along for many years already.
And in fact, I would say, in a lot of manufacturing systems, particularly in tech products like circuit board inspection, for example, when I am looking at solder balls, and I want to see, "Is there a void in there so I don't have a good circuit connection?" Stuff like that, that's been an application for machine learning. That really probably started about 10 years ago, right? So that's been coming for some time.
Chris Linnane: So before we go too far, help me understand the difference between AI and machine learning.
Willy Shih: Well, machine learning is one technique within AI. And machine learning basically says, "I'm going to feed the system a lot of data, and I'm going to try to extract patterns within it." So, I think it's kind of a broader categorization of it, if you will.
Chris Linnane: So what did the rise of AI change in how you look at competition and disruption in regards to these systems?
Willy Shih: I think what AI is doing is it's enhancing a lot of the capabilities and processes that we already know in terms of providing better precision or extending our ability to really digest and leverage huge amounts of data that were very hard to do before. A lot of people say, "Well, AI is going to be disruptive." If you think about disruption in the Clayton Christensen disruptive innovation theory, there are a lot of the characteristics of AI systems that make them a candidate for this type of disruptive innovation.
One of the key features in the disruptive innovation theory is you see an improvement curve, right? So you see a technology that initially may not be quite good enough for a lot of applications. But as you go to larger models, you ingest more data, the systems continually improve. So you see a lot of these characteristics on systems that employ AI that you would say make them good candidates for disrupting something.
Now, what Clay used to always say is, "No technology, inherently, is disruptive or sustaining. It depends to a large extent on how it's used, how it's going to be framed, how it's going to impact the business model." I think what we're seeing is a lot of people using new AI tools in very sustaining ways. In other words, "I'm going to use this new tool in my existing processes, but I'm going to use it to give me more performance, to give me a better product, to give me more precision, or to do it less expensively."
So I think we're seeing a lot of productivity-enhancing financing applications of AI. Especially when you look at some of these automated inspection systems and things like that, it's like, "How do I get to higher yield faster?" And a lot of that is, "How do I digest all this information from all these sensors that are coming at me and then make a quicker improvement on my existing process?"
If you go look at Toyota's manufacturing line, if you look at the new setup they have in their factory in Georgetown, Kentucky, they have put in a lot of vision systems, and they're using machine learning. And one of the things they're trying to do is they said, we want to present to our workers more information so they can correct problems more quickly, or they can make improvements more quickly. So I see a lot of people using these tools in kind of a sustaining way. It's like, "How am I going to make my existing product better?"
Chris Linnane: Now, when I talk to you, I talk to you a lot about disruption. And that word, we're using it a lot so far. So it might be good to frame up the true HBS definition of disruption, because that word is used in lots of different ways that might not be accurate to how we're using it.
Willy Shih: Right, and I always tell people, "Do you mean the Clayton Christensen definition of disruptive innovation, or do you mean disruption as in 'causing chaos everywhere?'" Right now, the Clayton Christensen model of disruptive innovation really looks at the problem of when I have a new technology or capability that initially may not be so good. But it's only good enough for the low end of the market, but it's cheaper.
So it's less expensive. It has maybe limited capabilities. It's going to serve the low end of the market. But then this technology or capability improves over time, so it can serve progressively more and more of the market. And in the Christensen model of disruptive innovation, what happens is incumbents who will look at this technology initially, and they say, it's not good enough; their business model is built around serving their best customers on the high end.
So they'll have a tendency to neglect that technology because their business model motivates them to make sustaining improvements to better serve their existing customers, and ignore this kind of low-end technology that might initially be not good enough. And so they usually leave that to new entrants, who will then take that technology or capability and improve it over time so that they can serve a larger and larger part of the market.
And by the time incumbents see it, it's too late. So that's what we mean by "Clay Christensen's disruptive innovation." The way a lot of people use disruption nowadays as more, "This is kind of a turning point where a new technology becomes available, and it kind of changes how we do things." As Clay said, "Innovations aren't inherently sustaining or disruptive. It depends on how you frame them."
Maybe it's because a lot more people understand the disruptive innovation model now. But I see a lot of people saying, "This new technology brings me a capability, which I can apply in a sustaining way. I can use this to make my existing products and services better. Or, I can deliver them less expensively." Now, there are also cases where we will see AI assistants displacing humans. And there's a very strong tendency here to say, "Well, that's disrupting human beings."
Well, in some sense, what you're saying is, I would argue, what we're doing is we're going to use these new tools to improve human productivity. A good example of this would be, let's consider a law firm that is preparing for trial. And they have to go through discovery. And these days, especially with electronic documents, in the old days, I would hire a whole crew of paralegals who would go through millions of pages of documents, looking for a few words here and there, or looking for an appropriate phrase.
I've done that in my days. We've gone through six million documents, looking for—these days, it's like, that is a great application for machine learning, which I can now feed millions of documents into my AI system and say, "Look for these types of phrases." That's, again, applying enormous processing capability to sifting through huge amounts of data. Now, will that reduce the number of paralegals that I need? Probably. Will it eliminate the need for me, the counsel, to review what I find and verify that? No, right?
What it really does is it improves your productivity. Now is that disruptive, or is it—in some sense, it's a productivity enhancer. So you could say a sustaining innovation. For the paralegal who got replaced, it was probably causing some amount of chaos in their life. But in some sense, it's really improving the productivity, improving the capabilities of the people. And whenever you improve productivity, that also means you need less people to do the same level of work. Unless you're going to do much more sophisticated, much more complicated work in the process.
Chris Linnane: It's fascinating. Every time I talk about AI, and every time I talk about the advancements, there's a moment where I start to feel kind of excited, like, "Oh yeah, that's great." And then when we start to talk about the part of how it affects people, I start to get a little nauseous. And I start to feel badly for the paralegals out there that are potentially losing work to these machines.
Willy Shih: Well, but that's been the case since the beginning of the Industrial Revolution. People used to spin cotton by hand, or they spun yarn by hand, and they used to weave things by hand, and then along came automated equipment that did those things. And what ended up happening is cloth became much more widely available at a much more affordable price. Did it require less people to produce more quantity? Absolutely. I would argue we're going to see the same thing here.
Chris Linnane: I hope so, because there are other jobs for those people. And I hope we have more jobs.
Willy Shih: Well, I think what it says is that it says a lot about managing your career and managing what kinds of skills you need to have. Another area where I think we've seen a huge amount of activity just in the last couple of months has been in coding, writing computer code. And one of the systems, Claude, from Anthropic, and a number of other systems, have really broken fresh ground on the ability to write usable code for all kinds of applications.
And what we see in the world is so much is driven by software right now, just for everyday applications like controlling the lights. Software is everywhere. So there, we have another example of a huge base of code that these AI systems could examine and look at: "Here are some examples of well-written code." So that can provide a foundation for how to write all this additional code for the machine learning system and AI system to generate code.
OK, now the question in systems like this is, "Where's the creativity going to come from?" So I saw one study where a company had deployed an AI chatbot replacing humans responding to chats that came in from customers. And it actually compared, if I had a real person working with a chatbot, with just a real person responding to chats versus the chatbot alone. And the chatbot alone, or the person alone, produced the best results. It turned out the chatbot and the person combined was actually less efficient than either one of those alone.
So, for a chatbot that you can tolerate some errors and stuff, it costs a lot less. So that's an attractive value proposition. Although one of the things that the people who did the study recognized, they said the reason this worked is because we had all these chat logs with all these responses that we could feed into the system. So, if we don't have a human inputting that, then we are necessarily limited into what our responses can be, because there would be no new insights that could ever go into that.
So, there are limitations associated with this yet. And I think there's still work to be done on that. But the way I would look at it is: this is not that different than a lot of other productivity enhancers that we've had since the beginning of the Industrial Revolution. The whole question of, "Can I use automation versus do I use capital or do I use labor?" And we've seen a lot of automation replacing a lot of manual tasks, which I think the world is better off for.
So for example, if you look at automotive vehicle assembly, just about every modern auto plant these days uses robots for the so-called body-in-white part. The body-in-white part is when I stamp and weld together the unibody for the car, and then I paint it. Those are dangerous jobs. Those are dirty jobs. It's like, "Yeah, automation, I'm all for it. Let's let the automation do that."
When you get back to the vehicle assembly, when I'm going to stuff the interiors, I'm going to put in the instrument panel, I'm going to put in the seats, I'm going to put in all the fittings on the doors and stuff, a lot of that is still done by workers because the workers are more flexible. So we'll see more automation as we get better vision systems and things, which will improve productivity.
If you go to most of these factories that have already been pretty automated, the big change you'll see over the last five years is they no longer have people driving forklifts or people driving tractors, towing bins of parts to the side of the line. Most of those are being replaced with automated guided vehicles. So, a lot of this is all about improving productivity.
Chris Linnane: We've talked about disruption. And it got me thinking it would be good if we could play a game to connect some of these things to figure out if they're disruption or not. So I've got five questions.
Willy Shih: OK.
Chris Linnane: And what I want to do is I describe what some companies are doing using AI, and they're calling it disruptive. And I want to read these scenarios to you. And you say to me, "I think that is disruptive, and here's why." Or, "That's not disruptive, and here's why." There's no right answer, I think. But I'd love to get your opinion on some of these.
So, a company uses AI to automate internal work so it can sell the same product cheaper to the same customers.
Willy Shih: See, I think that would be well within the definition of using it as a sustaining innovation. I'm going to take this new way of doing things. And I'm going to make myself more productive; sustaining innovation.
Chris Linnane: Perfect. Scenario number two: A company uses AI to offer a simpler product to people who could not use or afford it before.
Willy Shih: That has the potential for being a low-end disruption. Now, one of the things that I think is new news there is that Clay's model of disruptive innovation is out there now, right? So the real question you have to ask is, "Are incumbents motivated to flee?" Or, are they going to look at this kind of low-end, not-good-enough application, and it's like, "That's going to eat my lunch if I don't pay attention?" I think these days, more people are going to say, "That is potentially disruptive. I need to pay attention."
Chris Linnane: Oh, OK. So let me try something with you and see if I'm getting this right. So I remember working with Clay and talking about low-end disruption of the Honda Super Cub in relation to Harley-Davidson motorcycles. So Harley-Davidson motorcycles, and maybe he was talking about some other high-end motorcycles as well, they were going pretty high up, serving a big need.
And then Honda basically took a bike and put an engine on it. And that was the low-end disruptor, because it was cheaper and it didn't have all the bells and whistles. Are you saying now that because of this, because of AI, that it's easier for someone like Harley-Davidson to go back down and make a competitor for the Super Cub?
Willy Shih: I'm not quite saying that. What I'm saying is, these days, Harley is going to say, "Huh, I've got to think about whether this is a threat to me," right? Now, the Super Cub is this cheap bike, and it's not good for road touring, right? It's good for maybe riding around town. I think if you look at, I think, Clay's view of what Honda did, it's like they had a line of motorcycles, and the Super Cub got them a foothold in the market.
But then they also produced better, larger bikes as well. Once they got a foothold in the market, they were able to sell some of their larger bikes. And really, the thing that was innovative for them was also opening up different distribution channels, not the traditional dealers. And so that was more a distribution problem. I think these days a Harley would be perhaps more conscious of "What could these guys do?"
Now, to look at the Harley example, the one they didn't pay attention to was the Chinese motorcycle makers. The Chinese motorcycle makers started out counterfeiting Japanese parts. But they got better. And they've put a lot of innovation into their bikes. And they have put a lot of tech into their motorcycles as well. So they started out as, again, the big road bikes, competing with the Harleys that were cheaper and not as good, but they have improved dramatically. I think these days, if you look in the motorcycle sector, who are the competitors that you have to watch out for? They are the Chinese-made ones.
Chris Linnane: Ready for scenario three?
Willy Shih: Sure.
Chris Linnane: All right. Scenario three: A company uses AI to strip features and automate support to serve the least demanding customers profitably.
Willy Shih: Well, if they're taking features out, that means they're making an inferior product. One of the things that I think is very interesting when you have software features, I always viewed this trend of replacing hardware function by software as a huge driver of disruptive threats. Let me give you an example of replacing hardware functionality with software.
Many years ago, I had a Volkswagen. And this Volkswagen was made in Germany. And it was great. When you closed the doors, your ears would pop because the seals were so good, right? That was always a joke about Volkswagen. So a couple of years ago, maybe 10 years ago, I was in a rental car lot. And you could pick whatever rental car you wanted. I said, "Oh, here's one of the new Volkswagen Beetles," the rounded one.
I said, "I'll rent that." So I hop into the car, and I close the door. And just as the door was closing, the window would roll down to equalize the pressure. And then as the door slammed, the window would roll back up. Now, if you think about it, in the old days when I had power windows, I'd have all these switches, and I have these motors and pulleys and things like that. And it would all be mechanical switches.
If you wanted to have that function to roll down the window when the door closed and roll back up, that would be impossible to do in hardware. It'd be very complicated. You'd have all kinds of linkages and so on. Once you have just a stepper motor and a little microcontroller in the door, to do that function is just a couple lines of code. And so what we've seen over the last couple of decades—we've seen a huge amount of replacing hardware complexity with simple hardware and software to do these more elaborate functions.
And anytime you see something like that, what it means is I'm going to have a much lower cost of performing these functions, much lower cost of delivering these features. So if you look underneath it, it's like there are these technologies like sensors and stepper motors that you can control with the microcomputer, and shaft encoders, which tell you where the position is. But you have all this technology that allows you to replace hardware with software.
And so what you're talking about now is the same thing. It's like, "Let me replace some of that functionality with simple software." And you said initially, "It's not good enough." The key thing you have to say is, "Where is that going to go?" Because improving software, once you have something working, it's very easy to improve software and come up with more functionality, and so on. So that means there's going to be an improvement curve, which can be quite dramatic. So it may start out not as good, but it's not going to stay there.
Chris Linnane: I wonder if that's what happened. I had a car in college. Maybe they were—I won't bring up the brand because I don't want to say anything bad, but the back seats, they had these fancy locking mechanisms that always broke on me. It couldn't just be a standard pull-the-switch-up. It was all these buttons, and they broke. So, in order to keep the back doors closed, because they wouldn't close anymore, I had to tie a bungee cord between the two handles and use the tension between the two doors.
But if I took a turn, one door would open, and it would slap closed. I'd take another, and it'd slap closed. And all I wanted was a standard hardware working system, but it tried too hard to make something fancy happen. And it broke because of that.
Willy Shih: Yeah, well, I think the key is, these days, everybody is always looking at, "How do I reduce cost?" And if I can replace complex hardware with simple hardware and a little bit of software, that's a great way to elaborate more complex functionality at a lower cost. So I always look at these kinds of—people call them cyber-physical systems, where I'm mixing hardware and software.
I mean, whenever you see this mixing hardware and software, that means that's also an opportunity for an AI or machine learning to come in and improve that functionality by just looking at, I'm going to collect some data and just look at the behaviors; how do I improve the behaviors? So that's another thing to keep in mind.
Chris Linnane: Let's go to number four: A company uses AI to coordinate suppliers so it can scale without owning assets or building deep operations.
Willy Shih: I think one of the things people are trying to apply AI to is just understanding supply chain visibility. We are well along the way to companies really producing products and services without necessarily owning the assets for manufacturing, or even developing some of those products. If you look at a typical notebook computer from any of the big brands, the big brands don't do those—they don't design them anymore. They don't manufacture them anymore.
Those are all done by so-called original design manufacturers. Generally, most of them are Taiwanese companies, who will give you the complete turnkey design, and then what they'll do is they'll buy all the critical components from other people. And those ODMs, as we call them, will buy components from tier one suppliers, who will buy them from tier two suppliers. So you take any complex product today, and you'll have components that are coming from all over the world from hundreds of different suppliers.
Now, the problem is, people don't have the visibility to, where are all those parts coming from, or where are there potential disruptions? Especially when you have longer supply chains that are geographically distributed. So, for example, you might have shoes made in Vietnam, where some of the materials might have come from China, some might have come from Taiwan, some might come from Korea, some might come from Europe. And so I'll have this geographically dispersed supply chain with lots of links in between.
We really learned this lesson during the pandemic when you have a disruption, because the factory shuts down. And it's like, "I didn't even know." So one of the areas people are trying to apply AI to, in particular, is this whole area of supply chain visibility. How do I see who's in the various tiers on my supply chain?
How do I know when an earthquake and tsunami hit a microchip factory in Naka, Japan, back 15 years ago? And all of a sudden, half the world can't get microcontrollers to make cars anymore, because they didn't know it was that one factory. So, what a lot of people are trying to do is say, "I'm going to take all this data that is out there and try to mine it and show me where all my vulnerabilities are in terms of supply chains."
Now, that's a hard problem because most companies don't want to disclose their bill of materials. They don't want to tell you who their suppliers are. And their suppliers don't want to tell you who their suppliers are. So, what a lot of people are doing is they're mining customs data, or they're mining import data that shows a lot of these types of parts come from this country or so on. So you get a broader sense on the vulnerability.
So that's a big application area for AI. It's a classic application area because, again, you have a lot of data in a lot of different forms. A lot of it is in paper documents, by the way, in bills of lading or customs declarations, and stuff like that. So, just taking all that data, ingesting it, and trying to make sense of it, there are so many startups going after that application right now, just to try to give me more sense of, where are my vulnerabilities, where are my exposures?
Chris Linnane: Onto our last scenario: A company uses AI to simplify how a physical product is designed and built.
Willy Shih: A great example of that is building software code, where a lot of people are now saying for a lot of these simple things, let me have the AI build the code, right? Because, to the extent that the software is a black box, where I can define what the inputs to it are and what the outputs to it are, this is actually a great application for that, right? And so that's going to change a lot of things. Because on the one hand, it should dramatically improve the productivity of people who are writing the code and designing those products like that.
It potentially can also open up the field to people who didn't have a lot of those capabilities before, just in terms of making simple blocks of code available at a much lower cost. That should enhance the number of functions and features that we get delivered in software because it's going to make it easier to do.
Chris Linnane: Great. That concludes our little questions there. But I do have a question that came from that. With the tailwind of AI going on right now, are there industries that you think are more ripe for disruption?
Willy Shih: So, I think we're seeing a huge investment wave in AI right now, especially in the United States. We're seeing massive investments in data centers and scaling data centers. And a couple of things to think about here: the first one is, do the people who are leading the charge on these large language models have a business model that allows them to earn a return on some of these massive investments they're making?
And I'll put that out there. At the same time, I look at how some Chinese companies, like DeepSeek, have come up with a less expensive approach that is less resource intensive that is almost as good. Are they going to improve? I think so. In fact, I'd say you bet they're going to improve. Are they lower cost? Yes. Now a lot of what motivated them was access to technologies that they were restricted from, and they couldn't get access to the most advanced technology.
So they had to make do with something that was not as good and lower cost, but it's improving. And they've applied a lot of creativity to their approach. So I think there's going to be a lot of interesting things to watch there. Another thing that is going in parallel with that is the open-source movement on large language models. And we see, for example, Meta Platforms is advocating open source. And a lot of people say, "Well, it's not good enough right now. It's not as good."
But if you're trained in thinking about disruption, you are always looking at, "What's the improvement curve? And at what point does it get to serve a larger and larger part of the market?" So, I think there are a lot of very interesting currents in the market today which we should really be paying close attention to. It's going to be a very interesting time.
Chris Linnane: Now the groups that are at the higher end, are they there not looking around because they're overconfident?
Willy Shih: I think what they're doing is they're chasing after the high end of the market. They're chasing after artificial general intelligence. They want to be able to deliver systems that really have a range of capabilities, where you can replace humans for a lot of stuff. Not all the applications in the marketplace need that yet.
Chris Linnane: Yeah.
Willy Shih: In fact, what is striking to me in terms of what Google did with DeepMind in this AlphaFold program, which looked at protein folding, it's like, they got it so it's good enough. It's good enough. And now, there, they really used physics-informed machine learning. It wasn't just pure machine learning. It was physics-informed machine learning, and brought in a new capability, which I think will dramatically improve what drug researchers can do.
But some of these enormous investments that are being made, chasing after the high end of the market, there are a couple of risks there. It's like people are starting to ask questions about the business model. "How are you going to pay for this?" It's not quite the same scaling curve that we saw in the broader general semiconductor market, where we would just get more performance every year. And that drove the improvement curve in so many of these Clay Christensen-identified disruptions.
In these large language models and this race for the artificial general intelligence, as you put in more complex queries, it's costing more to produce those, as well as costing a lot of compute cycles. It's costing a lot of energy. Now, those costs are coming down, but the question is, "What kind of return are you going to earn on these gigantic investments that are being made?" That's still an open question in my mind. That should be an exciting question.
Chris Linnane: Exciting, not a scary question?
Willy Shih: Well, it depends. Scary depends where you're sitting. For me, it's exciting. Probably for some of the investors, I would say it should be a little scary.
Chris Linnane: Well, that brings me to a point. If you're running a group that's currently on the high end of things, what's a question you should be asking yourself right now as you consider AI?
Willy Shih: I would always be looking at, "Is there a cheaper way that is good enough, and what is the improvement curve?" Because we tend to underrate the importance of the improvement curve. We tend to think processes don't continue to improve over time. Especially if we're trained at a particular point, we kind of get frozen, and here's what the capabilities are.
I look at so many areas, and going back to the Clay Christensen example, the steel minimills—even today, if you look at the last 40 or 50 years, there's an improvement curve in steel manufacturing technology. Even today, the latest machines are better than ones that were produced 10 years ago in terms of capabilities. People in production tools continue to innovate, right?
So, that's something that I think you always have to be conscious of is, like, the underlying technology, how is it improving? Am I frozen at some point? Or am I locked out from investing in some of those new technologies because the stuff I have hasn't been paid for yet? That's how so many people got in trouble in terms of Clay Christensen's disruption.
Chris Linnane: Speaking of Clay—we've talked about him a lot, and we all miss him—what do you think he would think of what's going on right now? Because he didn't get the opportunity to be in this most recent wave of AI.
Willy Shih: Well, I think he'd be looking at that improvement curve. Because you see it in the LLMs, right? You see it in so many of these technologies, right? Every time you see this improvement curve, and you see people chasing after the high end, I suspect he would be looking for, "Where are those people who are going for the good enough?" He'd be watching the Chinese.
Chris Linnane: So that's really where it breaks down most of the time? It's people just going for the overserving and not keeping an eye out for what's happening?
Willy Shih: People forgetting that an awful lot of the application space would be satisfied with good enough and satisfied with something that is a lot cheaper.
Chris Linnane: Yeah, that happens all the time, it seems.
Willy Shih: Happens all the time.
Chris Linnane: Yeah, I can tell you right now, the cameras that we're filming this with have around 75 percent of the buttons I will never touch. And most people won't.
Willy Shih: And the cameras you're filming this right now, because I have an earlier, more expensive version of that camera, that replaced something that was five times as expensive. And the camera that was five times as expensive replaced something that was five times as expensive before. And you know what? The cameras you're using here are good enough for a Netflix video because they're certified for Netflix.
Chris Linnane: Are they really?
Willy Shih: Yes.
Chris Linnane: OK. Well, let's try to find a place for this video then. If you're looking toward the future and you're thinking, "What is the greatest opportunity?" We've talked about this a little bit, but I'd like to get examples or just a succinct question and answer of the future. Where is the biggest opportunity for disruption in relation to AI?
Willy Shih: I think there are great opportunities in producing cheaper, good enough solutions. Right now, you see this rush to build high-end data centers with high-end GPUs, high-bandwidth memory. I think people will be looking at when I can build something that is maybe not quite as expensive. GPU technology was designed for computer games and animation. It wasn't designed for AI.
It happened to be well suited for machine learning because it was massively parallel processing. I could have 256, 512 cores. And it was well suited to the structure of the problem. What you're seeing is people saying, "Well, maybe I don't need the full word length of those GPUs. Maybe instead of 64 bits or 32 bits, maybe I need less than that."
And the answer is, "Yeah, you don't need that much. And you're carrying a lot of silicon overhead for capabilities you don't need." And so you're seeing startups; they have to overcome the lock-in that companies like NVIDIA have developed because of their software. But we are already seeing diversification of that competition with people saying, "I'm going to design chips that are more suited, that are more cost-effective or more efficient." Or, "I will take approaches like DeepSeek did in China that require less resources and are simpler and more cost-efficient." I think that's a huge threat to the incumbents.
Chris Linnane: Question for you about good enough: What are some examples of things that went well beyond the good enough into the kind of absurd over-serving in the past? One thing I can think of, and maybe you don't agree, is a refrigerator with a screen on it that tells you what's inside the refrigerator.
Willy Shih: Oh, yeah, that's a great example. It's like—
Chris Linnane: It's ridiculous.
Willy Shih: Well, that's an example I would call gratuitous innovation. I really didn't need that type of function. I just needed it to keep the food cold. There was a group that gave awards for the stupidest products they saw at the Consumer Electronics Show this past year. I won't go into some of the ones they awarded. It was pretty funny, though, because—but that's the nature of the world we're in today.
You have a lot of innovative people, a lot of people thinking about, "How can I differentiate my products?" producing some of these gratuitous innovations. On the other hand, there's still room for doing very innovative products that I think require sophisticated design and require innovative thinking as well.
Chris Linnane: Yeah, so if you're thinking of ideas, if you're a part of a company or an individual coming up with ideas, how do you ride that balance? I mean, you'd hope people have instincts to say, "That's overserving." Just because I can make it doesn't mean I should, but—
Willy Shih: Well, I think the challenge, and a good place to look at this problem, is, go look at the personal care products aisle in a grocery store, or go look at the toothpaste aisle, right? I don't know how often you buy toothpaste, but what I find is that a lot of people just change things constantly because their job is to come up with new products. I didn't actually need this new combination of flavors in the toothpaste or whatever. But there's this drive to do that.
But that's a dynamic that results from competition in the marketplace, as somebody else is doing it, then I need to do it. But having said that, there's a lot of room for genuine innovation. I always hold up the example of Chobani and yogurt, right?
Chris Linnane: I love that brand.
Willy Shih: OK, now the founder—I once met the founder, and he would always tell people, "It's like, here's this like 3,000-year-old product." Hardly, you would think there's room for innovation in it. But they took the market by storm. They completely transformed the market, and all the incumbents were caught flat-footed.
Chris Linnane: Because it's giving us access to so many things, we're going to come up with some of the worst ideas easier, but also potentially some really great new ideas because of that.
Willy Shih: Well, I think there's real opportunity for AI. What I'm really keen on is any type of problem where you collect a lot of data. And we're collecting a lot of sensor data now and all kinds of data. And somewhere in the data lies insights. And what AI will allow us to do is have better access to those insights than if we had to trawl through all of that on our own because some of these problems are just too big or they're too multidimensional.
Or we talk about multiphysics problems, where there are different types of physical models that I have to deal with at the same time. And those are very hard to comprehend within a human mind. I don't care how smart you are. Areas like that, I think, offer tremendous opportunities for innovation. That's why I'm very keen on these AI systems that are physics-informed.
In other words, if I can put in some of the physical models into the weightings and train them so that they can recognize the importance of the physics, things like weather prediction—people are doing that—protein folding, drug design, chemical reaction kinetics, things like that. I think there are great possibilities for innovation there.