How Domain-Specific AI Agents Will Shape the Industrial World in the Next 10 Years

First, I really appreciate, just like Tim has pointed out, that a lot of you have come from very long distances. I'd like to get a sample. How many people came from outside the U.S. today? Wow, nearly half the audience. How many are from the Stanford community?

This year, we decided to also invite students, faculty, and staff, among others. So, about three or four? No, actually quite a dozen. We also extended the same invitation. This is a dangerous question: how many people are from Berkeley? See, that's why they're at the back.

The thing is, you're here at the invitation of the dean, Su Jae King Liu, who is a classmate of mine. As you probably know, her PhD is from Stanford. Did you bring the axe, though? Did you follow the instructions? No?

Okay. They have the axe this year. So, with that out of the way, I've cleared the mines about something that is actually a very hot topic.

But it's hot, but it's actually... People don't agree on what the heck an agent is. We've used the term "agent" for a very long time. However, more importantly, in the context of this conference, it's about industrial AI. So hopefully, by the time I'm done, you'll take away at least one thing: it turns out that generative AI is actually more important and significant to the industrial physical world, in my view, than it is to the digital world.

And some of you may know I have some credentials from the digital world. I helped launch Gmail, right? So when I say this is more significant over there, it's probably true. First, I want to start out by saying this is something that we all know, right?

But I want to put it on a slide. America today faces a deindustrialization crisis. This is the result of an economic theory that we've pursued for the last 30 years, right? Then one day we realize that, actually, we still need to make things. We can't solely rely on everyone else and expose ourselves to geopolitical risks.

But one thing that some of you may be very intimately familiar with, while others are not, is that we have an experiment in progress, right? That experiment involves TSMC. We actually have some guests or partners here from TSMC, so you know this story as well, or even better, than I do. We have TSMC launching new fabs in Kyushu with pretty much the same or even less budget than the one in Phoenix. I'm not going to spend the whole talk going over all the details of this, but I hope I don't offend anyone by pointing out that the right-hand side has been a failure, while the left-hand side has been a huge success. The only variable in this comparison is the people involved, right? So that says something. Some have referred to it as culture, but I think of it specifically as the last critical expertise. It’s quite interesting at an AI conference that I say what's really important is actually what's in your head, what's in your brain still.

And that's going to be true for a very long time, in my view. AI is not going to replace us; it's going to enable us so that our expertise becomes even more important. So that's the message that I want to convey here. You may have seen this wise saying: in Chinese, the word for crisis is the same as the word for opportunity, right?

Is it luck, right? So America also has a re-industrialization opportunity. In that sense, we’re not going to be bringing low-end manufacturing back; it’s going to be leapfrogging, right? We often talk about Asia and the developing world leapfrogging, going directly from copper lines to mobile, and so on.

So, America actually has an opportunity to leapfrog all of the manufacturing burdens, right? It has not taken advantage of AI. But AI can help here. So, I want to go through this; some of you may have seen me do this, right? If you look at the screen, it shows a survey that essentially measures technology optimism, right?

Some of you may have heard the term "tech bros" here in Silicon Valley, usually referring to techno-optimists, right? But technology optimism is a real phenomenon, isn’t it? I believe it's a positive thing. It's no surprise that I decided to take the data from this survey and plot it against country GDP per capita.

It turns out there's a strong correlation, right? Higher GDP countries... Oh, this chart is... Okay, this is AI optimism. I'm so sorry. The headlines are switched. Please ignore the headline.

Listen to me and look at the chart. Technology optimism increases with GDP per capita. That's no surprise because in wealthy countries like the United States, Google, Facebook, and other tech advancements can be leveraged much more effectively. Everyone's optimistic, but some people are more optimistic than others. The surprising outcome is that if you're more optimistic about something, you're more likely to apply it.

And if you're more likely to apply it, you're going to win more with it. So that's something important to keep in mind. I'm not going to go through some analysis of it, other than to share this interesting thing with you. I decided to take this chart and plot the residuals, right? So think of that as optimism, but with an extra dose of optimism.

You can see that some of the curves here are more optimistic than others, like for China, which is unreasonably optimistic — way up here. On the other hand, France is always below the curve. If you look at the residuals here, Saudi Arabia is particularly optimistic, as it is above the trend line.

And it's not surprising that France ranks this way on every chart in terms of optimism. That's okay. I grew up speaking French before I spoke English. I decided to correlate that, right? France is indeed like this on every chart regarding optimism.

That's okay; I grew up speaking French before I spoke English. We decided to correlate that, right? If you look at these economies that tend to have lower GDP per capita, it turns out that's because they've still been making things. They've been focusing on so-called lower value-added activities, as opposed to the advertising and marketing that we tend to do here in the U.S. So what do you think it would look like if we took that chart and correlated it against the industrial content of the economy?

This is a presentation about their GDP. Do you think countries that are still more industrial would be more or less optimistic about AI? How many people believe they will be more optimistic? What about those who think they will be less optimistic? Alright, let's take a look at the data.

Huh. The higher the economy is still making things, the more optimistic they are about AIs from the same data set. So that says something, right? That indicates that Vietnam, Japan, and China are going to be more optimistic and therefore more likely to adopt generative AI than we are here in the U.S. In fact, how many of you have seen some of the recent data where they survey businesses? For example, how many businesses are already adopting some kind of generative AI in their day-to-day workflow?

The number coming from China is 85%. It's amazing. I was going to say it's scary, but it's amazing, right? We may pride ourselves on creating things, but it's the people who use them, the people who scale things, right? The TSMC problem is not a science problem; it's a people problem.

We cannot scale that, right? Battery manufacturing, Panasonic—my Panasonic friends here—are working with the Tesla Gigafactory. It's a scale problem. You hear announcements of amazing battery technologies all the time, but none of them scale, right?

It's the scaling that matters the most to the economies, so let's keep this in mind. I decided to just, I like playing with things, so I decided to impute a model where for every percent of optimism above some level, that may give you, in this case, say 0.2% on the GDP impact. I simulated that out for 10 years to see what happens. If you can see the 2022 GDP here, the gray area where the U.S. is, it's still very large, much larger than China's.

The other data point is quite interesting, but we don't have time to go over it. However, due to this extra optimism, the 2032 output of China is expected to surpass that of the United States. This will likely lead to improvements in standards of living, technology, and everything else.

So anyway, I think it's something to consider, right? It's just the optimism and the adoption of technology alone. Okay. But let's get more specific, right? I think I'm putting forth some statements here that some of you may already be quite familiar with, while for others, they may come as a surprise. So I'm going to go over them one by one.

Generic LLMs, right? Something that people are starting to think about more, particularly in industrial applications, is that the very probabilistic nature of these LLMs is actually quite undesirable. You can run these academic tests in your paper and say, "Wow, accuracy 90%," and so on. However, these results are almost cherry-picked. In many industrial and business applications, you need the answer to be the same every single time you ask that same question because it's a business process.

You don't want creativity. You don't want the LLM saying, "Every once in a while, I'm going to mix it up." That leads to inconsistency and a shallowness in the answers. I like to compare a fresh PhD with someone who has 20 years of experience. A fresh PhD is very useful.

When you ask about dichlorosilane in semiconductors, they can tell you all the possible uses and so on. They ask, "What do I do with this problem? I don't know what to do." I would repeat everything I've learned. However, someone with 20 years of experience might say, "In that case, what you need to do is increase the flow rate to 200 sccm," and so on and so forth.

So think of these wonderful, amazing large language models (LLMs) as fresh PhDs, right? They're impressive in their own right, but they lack domain-specific knowledge and the experience that the people in this room possess. And that matters for the systems we build. Domain expertise is actually the key to success in the physical world. This is not just a conclusion I reached while preparing these slides; it’s based on my experience when I was helping to run Global Industrial AI at Panasonic. We encountered problems that AI could not solve without consulting one or two experts that Panasonic has in refrigeration.

And with that, we can actually solve the predictive maintenance problem. The opportunity lies in domain expertise. We always knew it was valuable, but it's been the reason we haven't been able to utilize it. Think about this: often, in technology, we implement something that turns out to be a bug, not a feature.

"Cyberspace is a bug, not a feature. The reason we had to coin the term 'cyberspace' is that our computers couldn't reach into physical space. Similarly, relying exclusively on data is actually a bug; it's because we couldn't use knowledge, as it's too difficult to encode."

But with Gen.AI today, suddenly we can speak to our machines, and they understand us. I think that's the biggest revolution. It's not that they're smart; it's that they know enough to listen and understand.

For the first time, we can easily interface with the analog world, right? Natural language, vision, and so on. That has been very difficult until now. So, that's the key opportunity. With that in mind, I want to share this vision with you.

Actually, it's a product that we're building, and it's going to happen pretty soon. Whoa, where are we? Okay. It's actually something that many of us work on, so I want to talk about this model of expertise in terms of capture and apply.

You capture the expertise, and you apply it. It sounds very obvious, but when you look across the landscape today, due to this data-driven perspective in machine learning, we don't focus on the capture as much. The capture is the difficult part, right? We tend to emphasize the apply part, the algorithms. But how do you then achieve the same thing in the knowledge age?

How do I capture a Samsung expert's knowledge and put that into my system? I am referring to someone who is actually about to retire, not figuratively. There is someone in the audience here involved in this project, so let's start talking. Okay?

Go to the next step. Okay, so we can have a process where I'm wondering about the etching process, which may fluctuate even a little bit. For the first time, I can get an expert to sit down with me at the machine and start discussing extemporaneously what we went through in this process. Then, with generative AI, we can actually say, "Okay, is this what you said?"

That's what I mean by the easy part. Okay, let's go to the next step. Yes, that's correct. Now, can we save that and structure it automatically?

In some structured form, in this case, it's a YAML file. However, you can imagine that it could be some other symbolic language. It could be Python; it could be anything, right? You can run through this process many times. This is what I mean by encoding the knowledge, okay?

And then you can go from there to another transformation. You can build a diagnosis program, okay? This is a simple example. You might think, "Yeah, I'm just going to hire a Python programmer to do that." But you can't do that for the free-flowing sharing of knowledge and rules and so on.

Do that, and you can simulate and see the results. Now you have one piece of knowledge that is operationalized. It can essentially indicate that the risk of non-uniformity is very high due to fluctuations in the plasma power. Okay, are there any questions or comments about this? You can repeat this process.

We can automate this now. I think that's the revolution. Okay. So with that, I want to talk about the parent-based diagram behind this. The way you can think about this is that, remember earlier I referred to the term "generic model," right?

Those are models that know a lot, but not specifically about my process. It's not that they're dumb; it's just that they're not specialized enough in my language, my system, my processes, and so on. What we want is an expert model. Earlier, Tim spoke of Semicon, and there's a booth over there, which is a result of AI Alliance work.

Of this expert model, remember that this expert model can ladder. You can have industry expertise, and then you can have company-specific, tool-specific, or even process-specific knowledge. That's the opportunity, but even that is not enough because models, as they go, do not have something that many of you may have heard of: planning and reasoning capabilities. This refers to the ability to loop over and reason through a process or a problem. As you'll see, this capability is emerging. I'd like to share with you a bit of detail about semi-con, which has been discussed. This is an effort under the AI alliance.

And hopefully, throughout today and tomorrow, you'll have an opportunity to talk with some of the AI Alliance representatives here today who advocate for open science and open development. The alliance has a particular methodology, or MO: we want to actually produce tangible outputs, not just sit and discuss concepts. So, Semi-Kong is one of those outputs, and it is built on language.

Semiconductors are very important. They play a crucial role in various technologies and applications. Additionally, there is an open model associated with them, which raises interesting considerations. What is particularly intriguing about building something on an open platform?

The slightly counterintuitive aspect is that, because it is open, you in the audience, who have very secretive and competitive processes, can take that open model and add your own knowledge without having to pay a fee to someone who has a proprietary model underneath. This openness is very important to innovation. It allows you to contribute back if you wish, and that is great. Thank you.

But if you need to build something proprietary to TSMC, Tokyo Electron, or Applied Materials, you are free to do so. I think that's what we want to promote. For those of you who don't come from the semiconductor industry, this is highly unusual for semiconductors. If you come from software, you might say, "Oh, that's a big deal. I've been doing this forever."

But in semiconductors, even the vendor and the customer have a clear wall between them. You may not know what I'm using your equipment for, right? And so having the ability to somehow share information would allow us to compete on a higher plane rather than at lower levels. This is an example of applying domain-specific agents to a semiconductor manufacturing process. In this case, it's a RAE etcher.

Okay, and here it's being used for, I believe, process engineering. I already talked about the AI Alliance. Take a look at the website, thealliance.ai. Today, there are over 100 members. I’m quite proud of the fact that it’s not just U.S. companies represented here, but also many industrial companies from Asia, particularly well represented by our Japanese members.

So, agentic AI—I mentioned planning and reasoning. I'll give you one example. This is not the only model, but it is one example of what we mean concretely, as you hear the term "planning" and "reasoning" all the time. In the context I’m using here, and you will see more and more people converging toward this in the generative AI era, we want to use the word "agent" to mean something that is goal-oriented. Essentially, it has the properties of planning and reasoning so that it can iterate over a particular problem statement.

And then it says, "Have I solved that problem yet?" Let's figure out how to iterate and get there. As they are today, LLMs (large language models) by themselves don't have that capability. To the extent that you see ChatGPT or others that seem to have this capability, it is actually because they have put a planning and reasoning framework around it. The models themselves don't do that.

One architecture that we use at Itomatic is called hierarchical task planning. Given a task, we break it down into subtasks. Then, we ask the system—not the model itself—if it can solve this task in one step. If the answer is yes, we go ahead and do it. If the answer is no, we break it down further.

So it's pretty straightforward regarding that planning idea. Within each task, there's a reasoning loop, right? You can use different paradigms again. This one is an OODA loop, right? How many people are familiar with the term "OODA"?

Great, okay, so I don't have to go over that. It was good enough for jet fighter pilots in deadly situations, so it is probably good enough for a lot of process engineering work. You go through this process of Observe, Orient, Decide, and Act. You observe the environment and the resources that you have available. Then, you orient and decide whether you have sufficient information and resources to solve the problem. Depending on the answer, you decide what to do next and then take action.

And then, now that one second or one day has elapsed, the world has changed. Let's return to that loop again. Okay, so that's the OODA loop. This is a very significant representation of planning and reasoning. I don't know if there will be a talk with a colleague of mine, Ving Leung, who is the leader of a project called Open SSA, which stands for Small Specialist Agents.

And so you can go to Open SSA, and there's an implementation of this. Again, it's open source, so you can download and use it. Here, I just want to highlight many of the guests who have participated in this revolution alongside Automatic Write, and these are real-world use cases that are either in production or in development today. I say this casually, but it turns out that in Silicon Valley and the digital world, there are actually a lot of startups struggling to find problems for which generative AI is a solution. I've been around long enough to say this is very weird.

The industrial companies are the first adopters of a technology. That's never happened before, right? Because industrials are slow—maybe from Silicon Valley's perspective, even stupid. I can say that because I was at Panasonic too. However, it turns out that's not the case. It's because the industrial companies have been working on much more difficult problems.

When I was at Google, I would make mistakes, and I’d like to say that when I made a mistake, it was just clicking on the wrong ad. But at Panasonic, if I make a mistake, somebody dies. So it’s very reasonable, it turns out, that these industries have been moving more carefully. There’s something about generative AI, and I think it lies in the fact that you can suddenly and easily capture all of that domain expertise and all of that pent-up demand with this technology. It's different this time, essentially.

And that's why the industrial world, I believe, will leap ahead in terms of the adoption of generative AI. Okay, so the call to action here is essentially: Don't fall behind, right? Lead the industrial AI revolution. Do it with us, do it with this paradigm, this base LLM model with the expert, and with agents above that. And if we do that, yeah, maybe—okay, so I basically did the same analysis.

I said, "What if the U.S. was equally optimistic about AI?" This is what the future is going to look like in 2032, right? Okay. That's the additional GDP growth that comes from AI-adjusted delta alone. So, I would like to leave you with this vision for those who care about the development of the U.S. industrial world.

So, what the world wants and what industrialization needs is to capture and apply domain expertise. Number one, it is very specific. We can use help from AI, right? And we can embrace a lot less fear and a lot more optimism. I would like to leave you with that thought.

Thank you, Dr. Wen, for the insightful talk on AI optimism across the globe and the French pessimism, specifically. That was definitely a takeaway for me. Before we let you go, we would like to see if anyone in the audience has any questions for Dr. Wen. Why are you asking about the domains?

Specific LLMs are not sufficient. Why do we need to combine them with agentic AI? I'm sure everyone has heard the question. Quite simply, it's because LLMs, or models in general, don't have planning and reasoning capabilities. More generally, you can think of models as one-path tools, right?

There's an input, there's a process, and there's an output. That paradigm alone tells you that it cannot solve problems. You still need something outside of that loop to say, "Okay, given that now the world has changed, what do I do about it?" There has to be a recurrent loop coming back, and that loop is a planning and reasoning loop. Now, if you build a system that has that built-in, then that system can do the planning and reasoning.

But LLMs, as they are built today, don't have that loop. So it's hard to imagine that an LLM is the asymptotic limit of what AI is going to look like. You can look at agentic workflows as basically a confession that you need a lot of crutches to make LLMs do something useful, reliably, and robustly. So, what do you think is going to be the next step after LLMs—post-LLM, if you will, or even post-agentic LLM? That's actually a very sophisticated question, so let me try to frame it in the way that I understand it.

In other words, all of this coding that does the planning and reasoning outside of the model seemed like a crutch, right? Because with these LLMs or with these models, it's magic. We just feed a bunch of data into it, and intelligence emerges. It doesn't feel good to add this extra layer to it. Let me accommodate that first.

And there are emerging recurrent models that have been around for a long time. The reason we don't utilize them is that they are too expensive, right? However, with more computing power and recurrent methods becoming cheaper, they will be integrated into models. In fact, there's a later model—I forget what it's called, WQKV—that has the property of being recurrent at a high level. I believe this is inherently part of the knowledge paradigm.

If we can have models that generate this code, then why not just use it at that higher level of abstraction? We don't have to go all the way down to every neuron all the time. That's my point of view. You presented an interesting example of overcoming the lack of domain-specific knowledge in large language models by providing an expert—in this case, a human expert who interacts via the language model. One of the challenges, and one of the problems I would be curious to hear your thoughts on regarding how to overcome, is that sometimes these experts come to different conclusions.

So, two experts may have differing opinions, or even if they reach the same conclusion, it could still be incorrect. How do you overcome this problem when testing requires interacting with a physical system, which is expensive and involves a very long time scale? That's a really good question that comes up all the time. It can be quite challenging, especially when it comes to humans. Will we be able to accept machines?

More readily. What I'm getting at is that machines can also reach different conclusions, right? But we don't worry about that. The reason is that we tend to think of them as much more deterministic, right? The different conclusions themselves; in these systems, we no longer think in terms of correct versus incorrect.

We think of better versus worse, right? So the difference is, first of all, in how many cases out of 100 that occurs. Let's say there's a 5% occurrence. That means the other 95% is not an issue, and we can still deploy that. There are different ways to handle those cases.

In other words, we can say, let's look at the two choices and have another human or another model come in and say, "What do you want to do?" So these are different options, right? In fact, that's built into the deployment. Today, we're not comfortable enough to go fully autonomous. There's always a human that is saying, "Here's the recommended next best action."

You want to push the button somewhere. Essentially, it's the recommended next best action; you want to push the button somewhere. Essentially, it's resolved at that level. Okay, you can send me the contact.

Dr. Erwey, the way we capture domain knowledge is through the Standard Operating Procedure (SOP), which is based on statistical knowledge accumulated over the years. So, there are SOPs and recipes developed that incorporate knowledge from experts as well as from machines. Why don't we use that as a way to capture knowledge instead of going back to the experts? There are already many SOPs readily available. Shouldn't we try to capture this knowledge in a more efficient manner?

Yeah, that's a good question. So the answer is not either/or. The answer is that we have been taking advantage of all those SOPs already. Cyberconducting companies are very advanced in having all of these things, right? But even that is not enough.

There are things that sort of escape these situations. For example, I had one conversation during a conference. When we sit down to conduct these interviews, we start out doing this manually, without any tools.

What I always say is that I'm not interested in discussing things that have already been documented. It's quite rich. But tell me about a problem or an incident that happened in the last five years where, for some reason— and I say this without embarrassment or humility— if you were not there, it would not have been solved. Tell me about that.

And really interesting information emerged. One example is a gentleman from a company facility in Phoenix who discussed yield problems that took a long time to trace back to pressure fluctuations in a chamber. They didn't initially understand why that was happening, but because he is a process engineer and also communicates frequently with the facilities staff, he recalled that there was a piece of equipment installed about a week earlier that used the same gas line. They had not been careful with it, which ultimately led to those fluctuations.

Would they have figured that out by putting 10 or 20 people together? Yeah, it may have taken a week or a month. But because that gentleman was there, in his mind, he said, "Okay, take a look at that," and it was solved quickly. That piece of knowledge was not in the standard operating procedure. Now we have a person who knows both facilities as well as the process.

Yeah, good question. And that's why I mentioned earlier that we have all these documents we're ingesting, but I think the grand opportunity lies in the information that is not already included in those documents.