How Domain-Specific AI Agents (DXA) Will Shape the Industrial World in the Next 10 Years
What this covers
Over the next decade, millions of domain-specific AI agents (DXA) will transform industries such as semiconductors, manufacturing, and logistics. These AI agents, designed to tackle complex industry-specific challenges, will surpass the capabilities of generic AI, driving innovation and efficiency across the industrial world.
As the US re-industrializes and invests in key sectors like semiconductors, these AI agents will play a critical role in optimizing production, improving supply chains, and enhancing competitiveness in global markets.
Follow Aitomatic CEO Christopher Nguyen on Twitter for more insights on the future of AI agents: [https://twitter.com/pentagoniac]
Stay connected with Aitomatic:
Twitter: [https://twitter.com/aitomatic] LinkedIn: [https://linkedin.com/company/aitomatic] Website: [https://aitomatic.com]
Don’t forget to like, comment, and subscribe for more updates on AI agents and industrial innovation!
#AIagents #DomainSpecificAI #DXA #SemiconductorAI #ManufacturingAI #USReindustrialization #IndustrialAI #AIInnovation
© Industrial AI Conference (IAC) Stanford 2024
Source description (no synthesized summary yet).
Industrial adoption of generative AI will outpace digital-world adoption because manufacturing domains have deep, hard problems that require capturing expert knowledge, and GenAI's natural language interface finally makes that knowledge capture and operationalization economically feasible.
- GenAI enables easy capture of domain expertise that was previously too expensive to encode, transforming how industrial problems are solved
- Countries with higher industrial content in their economies show greater AI optimism and adoption rates than service-oriented economies
- Agentic AI systems combining planning/reasoning loops with domain-specific models solve problems that generic LLMs alone cannot handle
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Open-source base models (like Llama 3) are strategically important because they allow industrial organizations with proprietary processes to build domain-specific models on top without paying licensing fees to proprietary model vendors, enabling competitive differentiation while maintaining IP control.
“what's the most interesting thing about building something on open right the slightly counterintuitive thing is because it is open you hear in the audience that have very secret very competitive processes you can take that open model and add your thank you I have that and add your own knowledge without having to pay a tax back to somebody who's a propri model underneath so this openness is very important to Innovation right it means you can just if you want to contribute back that's great but if you need to build something proprietary to tsmc or Tokyo electron or apply materials you are free to do I think that's what we want to promote”
Agentic AI systems (models with planning and reasoning capabilities) are necessary because models alone lack the ability to iterate, loop back, and reconsider problems—agentic systems add a reasoning loop that allows for multi-step problem solving and adaptation as the environment changes.
“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 the ability to Loop over and reason through a process or a problem so you'll see this emerging... in the generic in the generative AI era right we want to use the word agent to mean something that is goal oriented and essentially it has the property of planning and reasoning in order so that it can iterate over a particular problem statement and then it say I solved that problem yet I haven't let's figure out how to iterate and go there and models as they are today LM by themselves don't have that capability”
America faces a de-industrialization crisis stemming from 30 years of economic theory that pursued outsourcing manufacturing to other countries, and we are now discovering we still need to make things and cannot afford the geopolitical risks of over-reliance on external supply chains.
“America today we Face a de-industrialization crisis this is actually the economic theory that we've pursued for the last 30 years right and then we find one day that actually no we actually still need to make things we can't really just Outsource it to to everyone and suffer some geopolitical”
When experts disagree on conclusions or may even be wrong, the system should treat 'correct vs incorrect' as a false binary and instead think in terms of 'better vs worse'—deploy solutions that work 95% of the time and handle the 5% disagreement cases through secondary human or model review before final action.
“we no longer in these systems we don't think of correct versus incorrect we think of better versus worse right so the difference is first of all in how many cases out of a 100 does that occur so let's say there's 5% so the other 95% is not an issue we can still deploy that in those 5% there are different ways to handle them 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 he here's the recommended next bet action you want to push the button somewhere essentially it's resolved at that level”
Standard Operating Procedures (SOPs) and recipes in semiconductors already capture statistical knowledge accumulated over years, but they do not capture the tacit, problem-solving knowledge of experts—capturing expert knowledge via interviews about real incidents reveals insights that escape documentation.
“so already ready available make you the way you try to capture an knowledge easier yeah that's a good question so the answer is not either or the answer is we have been taking advantage of all those Sops already semiconductor companies are very advanced in having all of these things right but even that is not enough there's stuff that sort of escapes these things so for example I had one convers so when we sit down and do these interviews when we start out we do this manually without these tools what I always say is that I'm not interested or don't talk about stuff that's already been documented it's quite rich but tell me about a problem an incident that happened in the last 5 years when for some reason”
If you are more optimistic about a technology, you are more likely to apply it, and if you are more likely to apply it, you will win more with it.
“if you're more optimistic about something you're more likely to apply it and if you're going to more more likely to apply it you're going to win more with it so that's something important to put in mind”
The illusion that systems like ChatGPT have planning and reasoning is created by external planning frameworks wrapped around the model—the models themselves lack these capabilities, which are added by the system architecture.
“to the extent that you see chat GPT or others that that seem to have this it is actually because they have put a planning and reasoning framework around it right the model themselves don't do that”
Hierarchical task planning is an agentic architecture where tasks are recursively broken into subtasks, and at each level the system decides whether it can solve the task in one step or must break it down further, enabling structured problem decomposition.
“one hierarch or one architecture that we use atomatic is what called hierarchical task planning that is given a task break it down into subtask asks right and then you ask you can your system not the model itself but you ask the system says given that can I solve the task in one step and if the answer is yes go ahead and do it if the answer is no break it down further”
Relying exclusively on data (a data-driven paradigm in machine learning) is actually a 'bug' rather than a feature, analogous to 'cyberspace' being a bug necessitated by computers' inability to reach into physical space; the real opportunity is that generative AI now enables natural language and vision interfaces with the analog world.
“a lot of time in technology we do something and we don't realize it's actually a bug it's not a feature like cyberspace is a bug not a feature Cy the reason we create had to coin the term cyber space because our computers couldn't reach into physical space so same thing relying exclusively on data is actually a bug it's because we couldn't use knowledge because it's too difficult to encode but with Gen 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 to understand for the first time we can actually easily interface with the analog world right natural language vision and so on”
Generic large language models know a lot about many things but lack domain-specific knowledge and context, making them shallow and inconsistent for business applications that require deterministic, repeatable answers.
“generic llms right maybe this is the first time you hear the term generic because I'm going to talk about something that isn't generic but think of open AI uh Google and everything else as generic right there they know a lot about many things but they're broad but they lack domain specific knowledge and that's actually quite important we all know about how many of the responses can be hallucination and people are trying to fix that in the model themselves and more importantly people are trying to fix it in the systems that that that leverage these models and something that people are starting to think about more particularly again in the industrial applications is that the very probabilistic nature of these llms are actually quite undesirable right”
AI will not replace human expertise; instead, AI will enable and amplify human expertise, making domain knowledge and critical expertise more important, not less.
“it's quite interesting at an AI conference that I say what's really important is actually what's in your heads what's in your brains still and and that's going to be true for a very long time in in my view right AI is not going to replace us it's going to enable us so that our expertise become even more important”
Domain expertise has always been recognized as valuable, but in the data-driven machine learning era, it was not actually used because encoding human knowledge into systems was too difficult—GenAI changes this by enabling natural language interfaces that make knowledge encoding economically feasible.
“the opportunity is that domain expertise we always knew it was valuable but it's been the reason we don't use it think about this okay a lot of time in technology we do something and we don't realize it's actually a bug it's not a feature like cyberspace is a bug not a feature Cy the reason we create had to coin the term cyber space because our computers couldn't reach into physical space so same thing relying exclusively on data is actually a bug it's because we couldn't use knowledge because it's too difficult to encode but with Gen 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 to understand for the first time we can actually easily interface with the analog world right natural language vision and so on that's been very difficult until now”
Domain expertise is the key to success in physical/industrial systems, as demonstrated by Panasonic's inability to solve refrigeration predictive maintenance problems until they consulted with domain experts who possessed specific knowledge.
“domain expertise is actually the key to success in the physical world and this is not something that I just reasoned through these slides this is something that when I was helping to run Global industrial AI at Panasonic we actually ran into problems that AI could not solve without consulting with the the one or two expert that Panasonic has in refrigeration and with that we can actually solve the predictive maintenance problem”
Industrial companies are not inherently slow or less sophisticated (as Silicon Valley may perceive); rather, they approach technology differently because in industrial systems, mistakes have high consequences (at Panasonic, a mistake leads to fatality; at Google, it leads to a wrong ad click).
“because Industrials are slow maybe from Silicon Valley perspective stupid I can say that because I was Panasonic too but it turns out it's not it's because the Industrials have been working on much more difficult problems when I was at Google I make I like to say when I was at Google when I make a mistake you click on the wrong ad but at 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”
The planning and reasoning code written at high-level abstraction (not at the neuron level of models) should be viewed as part of the knowledge paradigm, not as a crutch; higher-level abstractions are appropriate for representing structured knowledge.
“I'd like to still speak for code I don't think of what we're doing with plan and reasoning in terms of doing in high level code is a crutch I I think it's inherently part of the knowledge Paradigm if we can have models that generate this code then why not just use that at that higher abstraction we don't have to go all the way down to every neuron all the time”
The value of capturing expertise from individuals about to retire is not just preserving their knowledge but accelerating problem-solving by transferring decision-making capability; without this transfer, problems that one person solves in minutes might take a team weeks to diagnose.
“would they have figured that out by putting 10 20 people together yeah it would may have taken a week a month but because that gentleman was there he's in his mind he said okay take a look at that and it was solved quickly”
Industrial companies are the first major adopters of generative AI, which is unprecedented—historically, digital/consumer technologies (like internet and cloud) were adopted first by digital companies, but GenAI's industrial adoption is ahead of digital adoption because industrials have harder problems that GenAI addresses.
“I say this like it's nothing but it turns out in Silicon Valley in the digital world we're actually there's a lot of startups that are struggling to find problems for which gen is the solution so so 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 perspective stupid I can say that because I was Panasonic too but it turns out it's not it's because the Industrials have been working on much more difficult problems”
A fresh PhD is useful for breadth of knowledge (all possible uses of a chemical in semiconductors) but lacks practical judgment, while a 20-year expert provides specific, context-dependent guidance (increase flow rate to 200 sccm) that solves real problems—generic LLMs are like fresh PhDs, lacking experiential knowledge.
“I'd like to compare a fresh PhD with somebody with 20 years of experience fresh be is very useful when you ask about dry choros silan in semiconductor they can tell you all the possible uses and so on then you say what do I do with this problem I don't know what to do I'm going to repeat everything that I've learned but somebody with 20 years of experience in that case what you need to do is increase the flow rate to 200 secm and so on and so forth so think of these wonderful amazing llms as fresh phds right they're amazing in it of themselves but they have no domain specific knowledge and they don't have the experience that the people in the room here do have and that matters to systems that we build right”
Expert models can exist at multiple hierarchical levels: industry-level expertise, company-specific expertise, tool-specific expertise, and process-specific expertise, allowing for layered specialization.
“remember this expert model can ladder you can have an industry expertise and then you can have company specific tool specific even process specific right and that's the opportunity”
The OODA Loop (Observe-Orient-Decide-Act) used by jet fighter pilots is a reasoning framework applicable to industrial process engineering, where the system continuously observes the environment and resources, decides if it has sufficient information to solve the problem, takes action, and repeats as the world changes.
“there's a reasoning Loop right and you can use different paradigms again this one is an UDA Loop right how many people are familiar with the term UDA great okay so I don't have to go over that it's if it's good enough for jet fighter pilots in in deadly situations is probably good enough for a lot of process engineering work but you go through this process of observe Orient decide and act You observe the environment the resources that you have available and then you Orient you decide is that do I have sufficient information resources to solve the problem and depending on the answer you decide what to do next then you take the action and then now one second or one day has elapsed the world has changed you go through that loop again okay”
TSMC's new fabs in Kushu have succeeded with comparable or lower budgets than Phoenix fabs, with the critical differentiator being the people involved, not capital or technology—indicating that expertise retention is the primary competitive variable in advanced manufacturing.
“we have an experiment of tsmc being launched new Fabs in kushu with pretty much the same or even less budget than the the one in Phoenix and 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 and the left hand side has been a huge success and the only variable is the people involved right”
Manufacturing and scaling are the critical bottlenecks in industrial AI, not scientific innovation—problems like battery technology and chip manufacturing have amazing innovations that fail to scale because people and process expertise are the limiting factors.
“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 we working with with Tesla gigafactory it's a scale problem you hear announcements of amazing battery technologies all the time none of them scales right it's a scaling that matters the most to the economies”
The problem of knowledge capture and application in industrial AI is distinct from knowledge discovery or research—industrial systems need to operationalize existing expertise, not create new knowledge.
“I want to talk about this model of expertise in terms of capture and apply you capture the expertise and you apply it sounds very obvious but you look at the landscape across the landscape today because of this very datadriven point of view in machine learning we don't think about the capture as much the capture is the dirty part right we think about the apply part the algorithms”
Semiconductor industry practice of maintaining strict boundaries between vendors and customers (each not knowing the other's use cases) is unusual and historically necessary, but open model collaboration creates opportunities to compete at a higher level rather than protecting secrets.
“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 so that we can compete on a higher plane rather than at the lower levels”
Technology optimism correlates positively with GDP per capita across countries—richer countries are more tech-optimistic—but AI optimism shows the inverse correlation: lower-GDP countries (China, India, Peru) are more optimistic about AI than high-GDP countries like the US.
“technology optimism goes up with GDP per capita that's no surprise because Google Facebook and so on we are in the rich countries we can take advantage of that a lot more everybody's optimistic but some are more optimistic than others the surprise perhaps is that there's a a correlation at all and you already saw the last slide when they did the same thing for AI but AI is the opposite that's weird right the correlation is even stronger 65% in terms of 65% of the variation is explained by the line alone that's the 2022 GTP per capita and so on you notice United States we're less optimistic or more pessimistic than China India Peru and so on”
America has a reindustrialization opportunity to leapfrog outdated manufacturing processes by adopting AI-enabled approaches, rather than replicating legacy industrial methods that other nations already use.
“America also has a reindustrialization opportunity and in that sense we're not going to be bringing lowend manufacturing back it's going to be Leaf frogging right so in the sense there's a we talk all the time about Asia and the developing World sort of Lea frogging going from copper lines directly to mobile and so on so America actually has an opportunity to LeapFrog all of the manufacturing burdens right of that has not taken advantage of AI but AI can help here”
The process of capturing expert knowledge for AI systems involves: (1) having an expert extemporaneously dictate their knowledge about a process, (2) using GenAI to structure that freeform knowledge into a machine-readable format (YAML or similar), and (3) generating operational diagnostic code that can be deployed and simulated.
“so think about this as a faul diagnosis AI system this is actually a real use case that we're working on right but the first step has to be capturing the knowledge of somebody who is actually about to retire... so we we can have a process where I'm saying wondering etching process fluctuate even a little bit this is what I mean for the first time I can get an expert to sit down with me with a machine and just start dictating just start saying extemporary extemporaneously... and then with Gen we can actually say I can say okay is this what you said right that's what I mean by by the easy part... and then you can now we save that and then we can now structure it automatically right okay we can put it in some structured form in this case this yamamo file but you can imagine that could be some other symbolic language it could be python it could be anything”
The core challenge for industrial AI is capturing and applying domain expertise through technology—this requires less fear and more optimism about AI, with the belief that domain knowledge can be effectively operationalized through modern tools.
“what the world wants and what industrialization needs is to capture and apply domain expertise number one very specific we can use help from AI right and we we can use a lot less fear and a lot more optimism and I'd like to leave you with that thought”
A real example: A yield problem in a Phoenix facility was traced to pressure fluctuations in an etching chamber, which were caused by a gas line installed a week earlier; the solution came from an expert who understood both process engineering and facilities infrastructure—this knowledge would not have emerged from standard documentation or typical debugging.
“one gentleman from company facility in in in Phoenix talked about yield problems that were took a long time to trace back to pressure fluctuations in a chamber and then they didn't know why that was happening but because he's a process engineer but he also talks a lot to the facilities people and then he remembers that there was a piece of equipment that used the same gas line that was installed like just a week earlier right and they weren't careful with it and that led to that that that fluctuations would they have figured that out by putting 10 20 people together yeah it would may have taken a week a month but because that gentleman was there he's in his mind he said okay take a look at that and it was solved quickly and that piece of knowledge was not in the standard operating procedure”
85% of businesses in China report already adopting some kind of generative AI in their day-to-day workflow, compared to much lower US adoption rates.
“some of how many people have seen some of the recent data where they survey businesses like how many is your business is already adopting some kind of gen in your day-to-day workflow the the number coming from China is 85% it's amazing I I was going to say scary but it's amazing right”
Recurrent models have been around a long time but were not used because they were too expensive computationally; as compute becomes cheaper, recurrence will be built into models themselves rather than as external frameworks.
“recurring models have been around a long time the reason we don't do it because it's too expensive right and so with with more compute with with recurrent becoming cheap that will be built into models right”
Multiple industrial companies (TSMC, Tokyo Electron, Applied Materials, and others) have real-world use cases of domain-specific agents either in production or in development today, demonstrating that industrial agentic AI is not theoretical.
“here I just want to highlight many of the the guests here that have participated in this revolution alongside with isomatic right and these are real world use cases that have been that are either in production or in in development today right”
If the US matched China's AI optimism levels, the resulting economic simulation shows the US would narrow the projected 2032 GDP gap with China, indicating that AI optimism differences are a material driver of economic divergence.
“I basically did the same analysis said what if the US was equally optimistic about AI this is what the future is going to look like 2032 right right okay so essentially that's the additional GDP growth that that comes from AI adjusted Delta alone”
LLM responses can be hallucinations, and researchers are trying to fix this both within models themselves and within the systems that leverage these models.
“we all know about how many of the responses can be hallucination and people are trying to fix that in the model themselves and more importantly people are trying to fix it in the systems that that that leverage these models”
Open SSA (Small Specialist Agents) is a project that implements planning and reasoning for agentic AI and is available open-source for download and use.
“I don't know if there's going to be a talk with what a colleague of mine Ving L is the leader in a project called open SSA stand for small specialist agents and so you can go to you look for op SSA and there's an implementation of this again it's open source so you can download and use it”
The AI Alliance is an open-science, open-development initiative with over 100 members (including many Asian industrial companies and Japanese companies well-represented) that focuses on creating and releasing real outputs rather than just discussing aspirations.
“for those of you that don't come from the semiconductor industry This is highly unusual for semiconductors... I already talked about AI Alliance take a look at the website the alliance. okay today over 100 members already and I'm quite proud of the fact that it's not just us companies over here but also a lot of the industrial companies from Asia and certainly well very well represented by our ja Japanese members”