Seng Lin Shee
Instead of debugging code, you're debugging AI logic
In this interview, Seng Lin Shee, an AVP of engineering at GM Financial working on AI, explains how a career that began in microprocessor architecture and electronic design automation shapes the way he sees generative AI: as one more tool in a long line of automation, and a non-deterministic one. His first question for any use case is whether the business is OK with it being wrong some percentage of the time, which makes adoption a risk management call that runs through compliance, governance, legal and security. He walks through buy versus build, when an agent workflow needs an LLM to choose the route and when an if-then-else will do, and why automation moves bottlenecks rather than ending them. He also covers LLMs on edge devices and why he sees AI as the last piece that connects people to decades of data infrastructure.
Seng Lin Shee is an AVP and engineering director at GM Financial, where he heads the Innovation Lab and generative AI platform work and leads enterprise architecture efforts.
Seng Lin studied computer engineering and worked on microprocessor architecture and electronic design automation before learning software from scratch at Microsoft. He later led solution architecture work at SVB Financial Group and was a technology lead and architect at Google before joining GM Financial in 2023.
“You should own something which is unique, and you should probably outsource things which are commodity.”
Seng Lin's core argument is that generative AI is one more tool in a long line of automation, and a non-deterministic one. Adopting it is a risk management call: own what is unique, outsource what is commodity, and accept that automation moves bottlenecks rather than ending them.
"Now this is the last piece, where you have human understanding," Seng Lin explains. "That's where we connect humans to all this infrastructure that has been built throughout the years."
The conversation
In this conversation: Josh Rubin (Host, CTO Studio) and Seng Lin Shee (AVP of Engineering, GM Financial).
Seng Lin, thank you so much for joining me here. There's a lot I would love to get into with you, but I kind of want to start with a little bit of your background. Right now you're operating as an AVP working in the AI space with GM Financial, but what was that journey? How did you get there?
My background is in hardware design. I did my undergrad in computer engineering and then ventured into microprocessor architecture design. When I first went into the real world, working, I started at Microsoft, and that's how I got into the software world. At that time we were more on the hardware side, building hardware in VHDL and Verilog, and when I started at Microsoft, that's where I learned from scratch how to use high-level languages. I started with C# and moved on from there. It got me interested. It's very interesting, and I think in today's world it looks like hardware is back.
Yeah, hardware is so hot again.
Yes, it's a thing again. But I guess the benefit I see is I get to see both worlds. Hardware's important, software is so important, so when combined, we make things happen.
Brain and body.
Yes.
So what kind of hardware product did you work on, once upon a time?
I dealt with the way of designing, rapidly designing, chips.
Like EDA style?
EDA, yes. Electronic design automation. So given a problem, an algorithm, how do we translate that into hardware rapidly, taking into account power, speed and area?
So from the early parts of your career, it sounds like you've operated by trying to find automatable efficiencies in incredibly complex problems.
Correct.
Have you always been problem- and solution-oriented? I mean, you're an engineer.
Yes, I am. I always look at things as: there's a problem, you have a limited set of resources, how do we use them efficiently? And I'll use a quote, right? A good engineer is always someone who's lazy. So now you're given something where you have limited time, limited energy. How do we use it to create something? I find it very challenging. I like to build. I like to see it repeatable.
So what gets you excited? Is it the destination, or is it the actual journey of everything that goes into the trip?
I think it's the journey, I think it's the process. But at the end of the day, you want to finish it, right? Get the satisfaction. I want to see the light at the end of the tunnel.
Okay, so you've got your eye on what you believe the outcome is or should be, but it is the figuring out.
I might not even know the outcome, right? But I will see that now it looks like it's going somewhere. Sometimes you might go into a rabbit hole and you keep going, and then all of a sudden you say, okay, where am I right now? Probably I need to spend my time. So you might get bored, but it's an interesting way, right? You say, okay, now we've probably run our energy down this section, but you get more energy to go in different routes. So yes, back to your question, it's probably the journey. But at the end of the day, maybe it's a personal thing, I like to see something, I like to hold something, I like to use something. I think that's probably the builder in me.
Well, which explains how you've pivoted into this AI architecture style of profession. What does being an architecture leader in this space look like? What is it like for you?
Different people see it differently. People use it as a tool to get to a destination quickly. As an engineering leader, I see it as a tool. It's yet another tool. I think it's a natural progression of what has been done previously, right? Early in my career, I started with an automation background, test engineering as well. And coming from an EDA background, I see AI as a tool to help us get there faster, to extend our capability in terms of memory, to hold more things and to comprehend, and now with GenAI, more general things, right? But we also need to be prudent. We need to be careful of how we use it. It can be used incorrectly, and then you might not know the consequences. But used the right way, it's like driving a car, right? It helps you get there faster.
It sounds like what you're also talking about is the identification of bottlenecks and the appropriate way to leverage this tool to help alleviate friction within those bottlenecks. I imagine at something like GM Financial, it's deterministic. The math has to math. Not everybody can be using a thousand different tools to do a thousand different things a thousand different ways. Is it your role to basically say there is a right way to apply these tools?
That is correct. I'll probably generalize it before that.
Of course.
It's also for any technology besides AI. Previously you had big data, now you have AI. When you put on an architect hat, right, you see it more as a capability. You think about buy versus build. You talk about what's important to the business, or what is the focus of the business, and then what strategy you want to get there. Should you be buying things to get you there, or should you be building it? There are consequences. If you build it, then you need to have the resources, the engineering to go through this. Maintenance, everything else. And if you buy, you get there faster, but then you have a plethora of technology. And also, in other words, it's sovereignty, right? Sovereignty, meaning now you are highly dependent on somebody else to do it. Same thing with generative AI. I'm not sure if anyone's talking about that, but if you rely on some capabilities, you're relying on the infrastructure. So back again: what's unique to you? You should own something which is unique, and you should probably outsource things which are commodity. Because if the underlying thing is taken out, now you don't have anywhere to go. I've experienced this in my previous career as well, where the underlying thing is taken away, and then you have a rush to rebuild it.
Well, and that's the fundamental issue that the frontier models are facing right now, is that the commoditization of their LLMs and their systems is...
And it's very relative, what is considered a commodity.
Yeah. I think about how everyone should approach this. We have these tools in front of us that theoretically give us access to more information and more ways of doing things than ever before. How should an individual or a technical leader apply their wisdom to say this is a good use of this or a bad use of this?
It depends on your requirements and your expectations as well.
So define the problem up front.
Yes. And also you have to understand, specifically with generative AI technology, it's non-deterministic, all right? And back to the question, it's all just to automate, right? You're automating something away. And if you are comfortable with that being non-deterministic, then it's fine, right? So it depends on the business.
And by that you mean if you're comfortable with there being a percentage chance that this is wrong.
Yes, it depends on your business.
Oh, that's a really interesting rubric to work in.
Yeah.
First question you have to ask yourself: if it goes wrong, is that okay?
Yes, it's risk management.
The CEO is not necessarily the best person to make that decision either. I'm just thinking about so many of the conversations I have. It is often the CEO and the board that come in and say, these new tools are out there, we need to use these new tools, push it, efficiency this and that. But it is the architect, it is the CTO that has to do the risk management side of this.
So, compliance, governance, legal, security, right?
No, and actually the governance and observability side of this thing...
Yes, exactly.
...is really important. So how do you observe? It's not just a matter of, if it goes wrong, are you okay with it? You also have to ask yourself, if you do this, can you observe it? Can you build in governance? So how do you observe and govern and trust these agentic systems and this technology?
Yes, and it's a thing right now. Observability is a very hot topic, and probably down the road, economics as well, right? How do you deal with cost? There are different frameworks out there, open source frameworks. When you build it, you want to understand, when you architect it, right? Now you have multi-agents, so you want to know which agents pick up the job. And when you design it, should the workflow or the orchestration be deterministic or not? That, in fact, comes back to architecture, right? Should you use an LLM to reason about what route to take? Or should it just be a basic if-then-else statement, and that's when it's more deterministic. So you need to decide what architecture to go down. All these things, there's nothing new. It's a very similar concept to before. You had big data, you had tracing, you needed to debug. So instead of debugging code, now you're debugging AI logic. It's the same problem.
The problems haven't changed. It's simply the way to get there, the solution, the path.
Yeah, the implementation's different, right?
So are best practices still best practices, I imagine?
Just at a different level, different technology.
Does AI shrink technical debt or explode it? Speaking of EDA, when I was talking to Juan Rey at Siemens, he said it shrinks technical debt and it's basically irrelevant at this point.
I think it's both. It depends where the bottleneck is and where the bottleneck has shifted. There are scenarios where you might reduce the technical debt, because now you have some automation tool to actually go through this vast set of data and process it, right? Be it ticketing, be it code conversion, you want to migrate things or build things, so you can go through all of this quickly. And depending on your process and your methodology, maybe there are some parts where it's not fully automated. There might be some humans doing it, and that's where the bottleneck is, right? Then if that also can be automated, yes. Then, back to the question, by automating it you're increasing the risk, right? So at the end of the day, do you still want to have someone eyeballing it? It really depends on your process, where the bottleneck is, because at the end of the day, there will be a choke point. So you are just shifting the bottleneck.
Right, bottlenecks don't end. You solve a bottleneck and that just introduces, okay, where's the new bottleneck?
Yes, unless everything's fully automated, unless everything is handled by agents and robots and everything, then that's where you'd probably use it.
But even then, the bottleneck just moves to the next analog piece of that.
Yes.
From your embedded chip days: there's a lot of talk right now, and I don't know if you even know the answer to this one, about just putting LLMs directly onto a processor.
Yes, yes. Which I think is fascinating. Yes, it can probably just pop up in my mind.
What are you most excited about in some of these edge techs that are being developed?
I think edge computing. You probably heard the recent release where now edge devices or embedded systems have enough computation to run LLMs, right? So you get a faster way to use it. I think there is a use for that, because I use it daily, and sometimes I get frustrated because the turnaround time is very long. Having the ability to process at the edge, you are solving a lot of things: solving speed, probably solving data privacy, because everything is all in there. So yes, there is a market for that. The downside is it's probably more expensive, probably more energy usage, because with cloud computing you're sharing resources. Now you're actually distributing the workload everywhere else, so you have duplication, redundancy, right? So it's a balance.
Maybe 2% of the country actually uses AI as anything more than just a glorified, terrible search engine.
Yes.
Do you think people need to be worried, or they just need to be thoughtful, or they should be excited? You're obviously excited by this technology and what it brings to bear and the opportunities that it creates. Is there anything about it that is causing you some kind of anxiety?
It can be both. It depends on how you see it. You see it one way, as I mentioned, it increases your capability if you're not already using technology. In my daily life, right? If you know how to use it, you can use it to go through your own data, because now you're doing a lot of things, right? So it helps you create solutions quickly. If you have a builder mindset, then yes. But even if you are not too tech savvy, if you use it as your search engine, I think the right way is just use it, right? Continue using it, and as you go along, you will learn things and you'll get more comfortable. I think you should embrace it, like how you embraced social networks when they first came about, right? People were a bit concerned about privacy, but then you get a little more comfortable. Of course, there are some folks who are a little bit concerned, and those are very valid concerns, no doubt. Very valid.
I think your greater point is, ultimately, it can give you superpowers, but for some people it could strengthen dysfunction. Anything can be used in excess and anything can be used wrong. You referenced social media. With social media, the promise was connection. There are downstream negative ramifications that have come from it, and we don't always know what the second- and third-order effects of a new piece of technology will be.
Exactly. And for me, I'm really excited for it. Things that you can't do today, you can do very fast. Again, we're using technology to... Remember, previously we had a bundle of data, information overload. Now you have a tool that's able to control or decipher or go through all these things. Today I probably have an overabundance of emails, and I'm able to use it to comprehend all those things. Previously we built all this AI infrastructure, deep learning, but now this is the last piece, where you have human understanding. That's where you connect humans to all this infrastructure that has been built throughout the years.
That is a beautiful thought, specifically because what it speaks to is: too much data leads to paralysis.
Correct.
But there's no such thing as too much context. Too much context potentially leads to nuance, empathy, understanding, and that's the opportunity.
Sometimes you get lost. Some folks are able to comprehend it, and they're a lot smarter than me, but it helps me extend, you know, helps me in my job, helps me in my daily life, helps me to learn as well. So I think it's just a tool, at the end of the day.
Thank you very much.
Happy to be here.
Appreciate it.
Thank you.
Be part of the
conversation.
Whether you're a CTO who wants to be featured, a company looking to sponsor, or an engineering leader wanting a seat in the room — there's a place for you here.