Monica Tremblay

Show Notes

• What is DeepSeek
• What are the benefits of downsides of DeepSeek being open source
• What makes the DeepSeek founding company unique
• How DeepSeek differs from other LLMs
• How the release of DeepSeek has affected other LLMs
• What are prompts and how to maximize their effectiveness
• How students are adopting LLMs
• How to get started using LLMs

Transcript

Monica Tremblay

I personally pay the $20 for both ChatGPT and Claude because I think it's worth it in my work. It reduces labor in a crazy good way. But I wonder if now folks are like, hey, why do I need to pay $20?

Female Voice

From William & Mary in Williamsburg, Virginia. This is Leadership & Business, produced by the William & Mary School of Business and its MBA program. Offered in four formats: the full-time, the part-time, the online, and executive MBA. For more information, visit wm.edu.

Ken White

Welcome to Leadership & Business, the podcast that brings you the latest and best thinking from today's business leaders from across the world. Sharing strategies, information, and insight that help you become a more effective leader, communicator, and professional. I'm your host, Ken White. Thanks for listening. November of 2022, that's when the large language model, ChatGPT, was introduced. Professionals everywhere started to take notice of and adopt the technology. Now, just two years later, we have a number of LLMs at our disposal, including Copilot, Perplexity, Claude, Genesis, and DeepSeek. In January, DeepSeek R1 was introduced. Developed by a Chinese artificial intelligence startup, DeepSeek has quickly become popular across the world. Monica Tremblay is a professor of operations and information systems management at William & Mary's School of Business. She's a top scholar and researcher who was a systems analyst at ExxonMobil before starting her career in academia. She joins us on the podcast to discuss DeepSeek, what makes it different from other LLMs, and where the technology may take us. Here's our conversation with Professor Monica Tremblay.

Ken White

Well, Monica, thanks very much for sharing your time. It's great to have you on the podcast.

Monica Tremblay

Thanks for having me.

Ken White

Yeah, DeepSeek. Man, did it make a splash when it was first introduced? Not that long ago, really, just a few weeks ago. What is DeepSeek?

Monica Tremblay

So DeepSeek is a large language model, and it was founded by the same folks that the Baido researchers. It started in 2023. What makes DeepSeek interesting is that it has a pretty impressive performance, even though it has fewer parameters. It needs less umph than the normal large language models that we've seen so far. So, for example, DeepSeek has 67 billion parameters and is showing the same performance as ChatGPT, where ChatGPT has a much, much larger number of parameters. It's also open source, which I think is really interesting because it means that researchers and developers can look at the architecture and the training methodology, where for other large language models, those are black box.

Ken White

Yeah. Let's talk about the open source. Is that a bit of a gamble?

Monica Tremblay

You know, I don't know because this stuff progresses so quickly. There's pros and cons. When we say that an AI model is open source, it says that the model weights and the architectures and training methodologies are publicly available. What that does is it democratizes access to AI technology. It enables researchers like myself to participate. It also facilitates transparency and also scrutiny of how the models are being trained. It allows for innovation to accelerate more quickly, like anything open source, really, and allows for customization. Folks want to customize the large language model they can. But then there's questions of, are we giving away our intellectual property.

Ken White

Sure.

Monica Tremblay

I don't think they're worried about that because I think it's just changing so fast that they don't have to worry about this.

Ken White

I think you've hit on something that people find interesting. It is that attitude of it just seems like we're just going to go. One of the things that struck me was their employees. They're not the typical people you'd expect. Can you tell us about who works there and why they're a little bit different?

Monica Tremblay

I know that they have a very flat hierarchy, which is unlike a Chinese company. That's a little different.

Ken White

Yeah. They've talked about hiring non-computer science people and younger engineers. What's the advantage of that?

Monica Tremblay

You know a different point of view, right? Long ago, I studied a company that's actually out of Florida that does electronic health records. He specialized in. He went for the vertical. The very first electronic medical record he put out was dermatology. He actually was the founder of Blackboard as a student at Cornell. What he did is he hired dermatologists as programmers. He actually paid them to learn to program and to build the EMR, and the result is a massively adopted EMR for dermatology. Now, he's gone after other verticals. But I think that the advantage that that brings is that they know the context really well. It's probably harder to teach a programmer about dermatology than it is to teach a dermatologist how to program.

Ken White

Yeah, interesting. You mentioned parameters. What does that mean?

Monica Tremblay

Those are the weights that are needed as far as the algorithm. The more you have, the more processing power you need. AI is actually... DeepSeek is actually pretty lean compared to the other also, as far as how much training data it needs.

Ken White

We've heard in some stories about national security. There's some concerns. What do you hear? What do you think about that in terms of DeepSeek and national security?

Monica Tremblay

Definitely, that should be a concern because anytime we're training models, anything that we upload to an LLM, I constantly think, where's this data going? Is it becoming part of their training algorithm? From my perspective, am I giving away my IP by putting... Let's say I'm writing an academic paper. I'm like, help me this introduction. As I do that, is my personal knowledge now getting codified into the training model for this LLM? The same could be true if we start to put information that could be harmful, like stuff about our electric grid.

Ken White

I see.

Monica Tremblay

That's what I would worry because the minute it's uploaded, it's part of their training.

Ken White

You've used it, DeepSeek?

Monica Tremblay

Not yet.

Ken White

No.

Monica Tremblay

I really like Claude, and ChatGPT, and Anthropic for searches, though it changes so quickly because now ChatGPT has the search function. Claude not as good at searches. It doesn't like when you make it go look on the internet, but that's probably not going to last long. They'll come out with something that does that as well.

Ken White

How has the release of DeepSeek affected all of those other LLMs and companies?

Monica Tremblay

I wonder because ChatGPT, right around when DeepSeek came out, came out with that model; it was $200 a month. I wonder if it affected that. I personally pay the $20 for both ChatGPT and Claude because I think it's worth it in my work. It reduces labor in a crazy good way. But I wonder if now folks are like, hey, why do I need to pay $20?

Ken White

Yeah. As an academician or researcher, how do you use an LLM?

Monica Tremblay

Oh, my gosh. It's crazy to me to think that this has only been two years. I use it every day, all day long. Yesterday, I was trying to summarize how journals feel about special issue editors having articles in that same special issue. This would have taken me going to every single one of those journals, and I was looking at 12 journals, and looking at their web page and trying to find their. Well, instead, I went to Anthropic and asked it, and it took it a little minute, but in it's how I prompted it. You have to be careful on how you prompt it and asking the right questions. It made some mistakes. For example, one of the journals was JMIS, which is a Journal of Management Information Systems. It found something else that also has that acronym.

Ken White

I see.

Monica Tremblay

You've got to be careful that you check. Then, what I love that is fairly new is I now give you the source. In the past, when it hallucinated, I just didn't really know where the information was coming from. But now I can just click on the source number, and if it's taking me to the right journal page, I know that it's. I can also go read if it doesn't pass the sniff test. I'm like, that doesn't sound right. I can go read where it took it from. I use it to plan classes. I want to cover these five topics. I have an hour and 20 minutes. How would you schedule me? Make up an activity for me where we practice this, this, and this, and this is the learning objective that I have. I may not use it word for word on how it's telling me, but it sure is saving me time, and it's actually pretty creative. I've moved all of my classes. Pretty much, I don't have PowerPoints anymore because I teach a programming-intense class. I put everything in something called data camp workspaces, where I can put both text and code. It's all in one place. I used to jump back and forth between the two. But what I do is, okay, here is my PowerPoint. Give it to me in Markdown, which is a language that makes the text look better. With icons and everything, give it to me in Markdown so that I can put it in here. It gives me the code. I don't even know Markdown that well. But now my presentation looks professional. It looks like a book. I mean, it's crazy, and it saves me great amount of time. I mean, I could think of a million ways that I use even personal life. I had to go on a trip, and I had a crazy schedule where I was going from Sweden for a conference, Iceland to meet my husband for a vacation, and Cypress for another conference, and I was like, how do I pack?

Ken White

How do I pack?

Monica Tremblay

Yeah. I'm like, this is what's happening. My husband's meeting me in Iceland. It was like, tell your husband to bring you a suitcase with summer stuff and give him the winter stuff suitcase and switch. I mean, amazing. Amazing, right?

Ken White

Yeah.

Monica Tremblay

Summarize this document for me. What's the tidbit in here that I should be paying attention to?

Ken White

We'll continue our discussion with Professor Monica Tremblay in just a minute. Our podcast is brought to you by the William & Mary School of Business. Once again, the William & Mary MBA program has been recognized as one of the best in the world. The Financial Times recently released its global top 100 MBA programs, and the William & Mary MBA program is on the list. If you're thinking about pursuing an MBA, consider one that has world-class faculty, unparalleled student support, and a brand that's highly respected, the William & Mary MBA. Reach out to our admissions team to learn which of our four MBA programs best fits you: the full-time, the part-time, the online, and the executive MBA. Check out the MBA program at William & Mary at wm.edu. Now back to our conversation with Monica Tremblay.

Ken White

You mentioned prompts, and I'm hearing that's very important.

Monica Tremblay

Absolutely.

Ken White

Can you talk a little bit about that? How do you get good at writing the right prompts?

Monica Tremblay

Yeah, I think prompting is super important. The approach I have taken is just tell the LLM what you want it to be. You are an academic researcher, and you're writing a review. These are the issues that you have a problem with. Then I write the issues that. Let's say I'm reviewing an academic paper. Here's the issues that I have. But I write them pretty roughly without having to. Then it wordsmiths for me, but it's the prompts. I tell it, this is the role you're playing. You're trying to stay positive. These are the type of. Think critically. Don't use bullets because it loves bullets. I don't know if you've noticed that. It always wants to put everything in. Don't use dashes. But pretty much, you start to learn how to interact. I do notice that it remembers your previous conversations you had with it. Once in a while, I have to say, forget anything we talked about in the past because it tries to bring in. I'm like, no, I'm not talking about that anymore.

Ken White

How long did it take you to get to the point where, I know how to write prompts, I know how to interact?

Monica Tremblay

I don't know. I feel like I'm still learning.

Ken White

Oh, interesting.

Monica Tremblay

Because the tools themselves become more and more powerful, where before, I worried a lot more about hallucinations. I feel like that's not happening as much. I can't tell if it's me better prompting or the algorithm's getting better.

Ken White

Hallucination's meaning?

Monica Tremblay

Fake, like absurd.

Ken White

They just go off.

Monica Tremblay

I see it in my students. I don't discourage the use of LLMs in my class. I just tell them, be careful, because I've seen them go down these black holes where they keep prompting and the output just keeps getting worse.

Ken White

I see.

Monica Tremblay

Because they're not prompting correctly, they don't know what they're looking at. That's the key thing. You still need to understand the context of the question that you're asking. If you don't know the material very well. It's a model that predicts the next word that should follow. It could give you a bunch of gibberish and make it sound really well done. You're like, oh. But if you do know what you're talking about, you read it, you're yeah, that doesn't make sense.

Ken White

You mentioned students. How are you seeing them adopt LLMs?

Monica Tremblay

I think they're pretty useful. You got to remember these people grew up with Google. This isn't that far of a leap for them. They've always had something digital in their hand. This is an extension to me in a way of Google, getting information really quickly. But I think it's up to us to teach them how to think critically when they use these tools because I do worry that now, for anything, they're going to jump to an LLM tool without even thinking, is it even the appropriate use?

Ken White

Which is what I hear from many others, the critical thinking piece. We can all ask it to do stuff, but then what? Once it's done. The speed at which things are happening. We saw in terms of DeepSeek, they're calling it R1, the first version, and now R2 coming so quickly. Is this going to be the future? Faster, faster, better, better?

Monica Tremblay

It used to be more as long the technology. Now we've seen Morse law on like times 10 to the 20th with Avogadro. It's going so fast. I don't know. I wonder if there's a point where it starts to hit a steady state. For the longest time, we didn't have the processing power. Now we do. But that's also going to become, at some point, a constraint. You need water to keep those centers cooled. There's a point that we might hit a constraint.

Ken White

Someone who's not used it, and I'm not surprised, but I've run into many working professionals who have not. They've not tried ChatGPT, DeepSeek, or any of them. What's your advice and how to get started?

Monica Tremblay

Start with a small task. I actually encountered this with my husband. He had just a very menial; I've got this spreadsheet, and I'm trying to figure out how to allocate certain things. I'm like, try it. If you don't care about the material, because that's the first thing, I'm like, are you okay with that being out there? He's like, yeah, no, this is product distribution and sales. I don't really care. I'm like, try it. Put it and start to ask questions of the data, where, in the past, you might have had to write pivot charts. Now, you can just start to ask questions and start to gain confidence on the output that it gives you.

Ken White

Download, get the app, put it on your phone, use it on your laptop. What's the best?

Monica Tremblay

I use it on the browser, mostly. I do have it on my phone. Then I'm starting to get lazy; I'm like, oh, I've got to respond to this email. I'm like, here's what I want to say. I don't have time. Sorry about that. Hit me up in summer. Then it gives me a much better written, sounding less harsh type of email. I'm using it even on my phone. But I would say if you're just starting, probably the browser is the way to go.

Ken White

I don't think you're alone on the phone. It's the hottest app going right now. I think it hit number one on Apple in terms of the app. People are really going that way. What do you think for the future? For business, for academia, for research, what do you think is coming down the road next?

Monica Tremblay

For research, I don't know. Part of me is a little bit worried because are we going to generate any authentic research as people turn more and more to these tools to ideate? Are we all going to come up with the same ideas? Because I'll give you an example. GitHub is a place where programmers post code. If they have questions, you go to GitHub, and people answer you: hey, this is how you would fix that. That's why this piece of code is not working. Well, GitHub is seeing less and less information being uploaded because people are going just to the LLMs. Well, guess where the LLMs are training from? Yeah. Eventually, they'll run out of training data.

Ken White

Interesting.

Monica Tremblay

I don't know. I think you need the big crystal ball. But I think to your question, does it continue to grow at this pace I don't know. There's going to be a point where we hit be it competing capacity, but also government's trying to figure out how to regulate and keep under control and people worrying about intellectual property. I'm not sure. We're in this weird space because I don't think we've ever innovated this quickly.

Ken White

That's our conversation with Monica Tremblay, and that's it for this episode of Leadership & Business. Our podcast is brought to you by the William & Mary School of Business, home of the MBA program offered in four formats: the full-time, the part-time, the online, and the executive MBA. Check out the William & Mary MBA program at we.edu. Thanks to our guest, Professor Monica Tremblay, and thanks to you for joining us. I'm Ken White, wishing you a safe, happy, and productive week ahead.

Female Voice

We'd like to hear from you regarding the podcast. We invite you to share your ideas, questions, and thoughts with us by emailing us at podcast@wm.edu. Thanks for listening to Leadership & Business.