21. AI Engineer Vihaan Nama on Privacy, Practice, and Empowered Learning

Season 2, Episode 10 of Kinwise Conversations · Hit play or read the transcript

• Read the Transcript

  Lydia Kumar: Today we're visiting Duke University to meet Vihaan Nama: an AI engineer, graduate researcher, and teaching assistant. At Duke, he supports courses like explainable AI, AI product management, and managing AI in business, breaking down big complex systems into ideas students can actually use. If you've ever wondered how to make AI education more human, or how student learning data could bring something more meaningful than dashboards, Vihaan brings both clarity and care. Let's dive in.

  Lydia Kumar: Hi Vihaan. Thank you so much for being on the podcast today. I'm so appreciative of you being here to bring your perspective and your expertise on AI to our audience. To get started, I want to give you a chance to talk a little bit about your journey and how you ended up working in the artificial intelligence space.

  Vihaan Nama: Perfect. Thank you Lydia, for having me on this. I'm quite excited as well to start talking about it. About my journey in AI. So it all started in undergrad when I was in a small project with a couple of my professors, very nascent stages of my learning of AI. And I was kind of going through the whole research field and wanting to publish my work. Wanted to get my voice out there, but I didn't know where to start. And I was very young, very early into it. I was like 18 years old at that point. And my professor told me, like, "Hey, why don't you build a small AI system that tells you the difference between the advantages of using AI for a problem statement versus doing the same problem statement without AI and seeing how good AI actually works in helping you and does it actually make sense." And I did. I thought this was an amazing idea for me to just get my feet wet, just start working on it. And it was a small little project on sentiment analysis and I decided whether a rule-based system was better or an AI system was better. And in this whole process, I noticed that the AI far outperformed a rule-based system. And that's when my interest in AI got sparked and I was like, okay, this is something that can really help us on a larger scale. I was still 18, so it was such a small, tiny, little problem, but I could envision how big this could get. And that's when my interest started. And as my career progressed, I worked in multiple places. I've worked in JP Morgan, I've worked in Samsung. I'm currently working at a company called PSNS as an applied AI engineer. And I'm just going ahead, diving through the field. And throughout this whole time, I'm also continuously working on research. So I worked in multiple research labs in my undergrad. I'm working at the Duke Trust Lab currently right now, and I'm figuring out where AI can take us. And I really feel like AI is the next big thing, and I'm grateful to have to be able to speak with you today about this.

  Lydia Kumar: It's interesting because when you talk about that first project, you were kind of looking at what is human versus machine and how this works and where to use them. Is that right?

  Vihaan Nama: Yeah. So the way I like to think about it is that traditionally all computer programs were written in a particular way where like the humans would actually write the code or in layman's term, define the rules. And then this code would be generated and the user would give it an input. Based on his or her rules, you would get an output for the system. But the way I think AI is different is that when it's AI, it makes the rules. You give it the data and you give it the expected output and you tell it, "Give me the rules. I'm not great at understanding what rules should be made for this data. You figure it out." And that's where I feel AI is different because for the small problem of sentiment analysis, basically it was tiny, small movie reviews and I was trying to analyze whether they were good or bad. I went through a particular process of doing it, but then I realized the AI had other ways of thinking about it. And that was my first look into the field.

  Lydia Kumar: That's really helpful for you to kind of flesh out what that means and what that looks like. It reminds me, Ethan Mullick posted recently about the Bitter Lesson and how if you just feed a lot of data into the machine, the computer is going to be able to solve the problem better than when we try and code all of these rules into how it works. And you learned that lesson pretty early on, just when you were 18. Yeah. So that's kind of incredible that you were thinking about that so early in your professional career.

  Vihaan Nama: Yeah, it was amazing. After that, I think all the professional experiences just strengthened that, that today, like data is so readily—there's so much data out there—and there's so little knowledge that's available from this data, that I think actually having machines that actually define these rules and generate patterns, understand certain ways of representing data and how a pattern should be recognized within these large purposes of information. I think that's the next big gold mine there is. And it's already being—we are already seeing that's what's happening. But especially when we look at like larger companies, like for example, at my time in JP Morgan, I noticed that a lot of data that was once considered obsolete, which had just thrown away into an archive somewhere, was actually being retrieved back. And everyone—they're pulling out records from long ago because there's so much information, data that they didn't want to sift through. But now that you have AI that's going to be able to help you go through that information, bringing it all back, and they were trying to, like, you know, what information can we gain from this? It's something we threw away a long time ago, but actually there might be a lot of gold in here that we can learn from and we can train these models on. And I think we're just at the tip of the iceberg of where it could be.

  Lydia Kumar: That's so exciting because I think a lot of our listeners are people in schools and in education, and education institutions have tons of data. And there's—without AI, there's a lot of interest in using, in something called data-driven instruction, where you are thinking about what students know and then you're trying to tailor your teaching around what they actually need instead of just teaching them, you know, word for word what's in the curriculum. You can kind of say, okay, I have this information and I'm going to tailor what I teach to the needs of my students, or these subsets of students. And so I think that's an example of how teachers have been using data in the past, but AI could take all the loads of data that exists—and obviously there could be some privacy concerns here—but you have all of this data in a school setting and you could learn something about students, the way people learn, the trends that have happened over time in a totally new way.

  Vihaan Nama: Yeah, it really just unlocks so many possibilities that just a few years ago, because of even just lack of funding in general, were considered like such far-fetched goals. But now that there's so much funding, there's so much research, there's so much stuff going on in the field, especially even in generative AI, which is the new buzzword that people are throwing out there. I think it is unprecedented and I think the amount of value that we can gain just from this previous amounts of information that we've just pushed aside, thinking that it might be useless, but actually there's so much information there. I think we can gain a lot.

  Lydia Kumar: What are some use cases that you see for businesses or for education institutions? I mean, you're at—you're at Duke right now, you're a student, and you're also working as a teacher's assistant and you have this business experience. So I don't know, what are some use cases that you see either in the, both in the business world and in the educational environment?

  Vihaan Nama: A really quick win that I've seen a lot of companies take after diving into this field of AI because the first thing that they want to do is that they want easy retrieval. So there's this concept called Retrieval Augmented Generation that was slowly becoming popular after the in—like after ChatGPT came out, and what it all basically did was have a very intelligent way to sift through tons and tons and tons of information—think of it in gigabytes or even multiple gigabytes of information—and bring out relevant content to the user based on a natural language query. And this is only possible because of the help of LLMs, right? Large language models helped us take this query that the user says, like, "Hey, can I have information about X, Y, Z?" And then it gets converted in the backend. It goes, the AI machine all goes through lots of data, brings back relevant chunks, and then is able to summarize based on the information that it retrieved and based on your query, is able to formulate a great answer for you and teach you more. So, I think one of the biggest changes that I'm seeing—so along with everything you said, like, I'm also working at a company called PSNS as an applied AI engineer, and PSNS is an architectural organization. Now, you would think architectural and AI is completely different, like, you know, they're on the different spectrums of engineering per se, but actually I'm seeing such a huge intersection is happening over there where they have so much data, so much information, because they're like a 70-year-old firm that needs to be brought back to the user in a very easy-to-understand way. So I'm building FAQ systems where people can just type in questions and retrieve answers for them. And I think this is the first big change that's going to come up, and it's already coming up. But with the field of education as well, giving access to information, especially your own information, not just the information that's in the LLM in general, but your own information, giving access to teachers, students, as well as people who are just partaking in the system is such a big win because, before this information was like, I would say—I've always said like—all this while data and information was in abundance, but knowledge was scarce. But after the advent of LLMs, I think knowledge is becoming so available for us, and it's just a matter of a simple word away.

  Lydia Kumar: When you were talking, I was imagining all this data is like, I was imagining the beach actually and all the sand that's on the beach and it's like so much tiny little grains of, of sand everywhere. But if you wanted to find that specific grain of sand that could answer your question, it would be impossible. And now you have an LLM who can—this intelligent technology that can identify the exact little grain of information that you need so that you can make, make a decision. I love that you just said, could you just say it again?

  Vihaan Nama: Yeah. Information and data was abundant, but knowledge was scarce. And I feel now that's completely changed.

  Lydia Kumar: It's very cool. I'm curious, you said something a minute ago about a student having, or a teacher or someone having access to their own information, and as if you were a student and you had access to your own information, how do you see that as—what, what does that mean and how do you see that as a value add? Or what, what kind of information might someone want?

  Vihaan Nama: So, as a student, like even looking at a micro use case of just a single student—me as a student, 'cause I'm currently a master's student—has so much notes available on my iPad, on my laptop. I have like, nearly a hundred PDFs every semester that are just stored over there. And this information is actually quite useful for me, especially 'cause when I'm in my job, since my job is so related to the field of AI and I'm studying AI, there's so much of stuff that I wish that I could, while I'm in my work, I'm like, "Hey, I just learned this last semester, but I don't remember where, I don't remember how, I don't remember if it was my own handwritten research that I did, like looking up stuff and writing it down. I don't remember if it was my teacher who taught it to me." But I just wish I had that, that instant. Like if I could just say like, "Hey, can you please tell me how to build this system based on my notes?" And my particular note came up because I wrote it in a way that I understood. And if that, that process became so much easier, that would be such a big win because I feel like, it's not just me. I think every student out there writes their own notes, writes their own pieces of stuff that they're writing down in the way that they understand it. For example, if I give my notes to someone else, they might not be able to get it as good as I do, but for me, that is everything that I've learned just over there. And having, even having that access, you know, being able to retrieve without having to go through, like going, going back home, looking through all my notes, spending the whole next few hours, like, "Oh, was it here? Was it there?" and bringing back, bringing it back out. Even having, imagine having a small, you know, chat interface where I could chat with my own notes, be like, "Hey, um, I spoke about topic X a few months ago. Um, I'm building this kind of a project. Do you think I could use that topic and tell me how to build a system?" And having an LLM in between looking, getting that information through various processes and then taking your data and taking your query and being able to generate a good answer for you. I think that would be such a big win in the space of education. And even when it comes to like revising before like, semester-end examinations or mid-semester examinations, I think having an AI study buddy kind to help you to go through your notes, to be able to quiz you, um, that really helps as a student. And looking into the lens from a TA, because I'm also a TA, I've TA'ed three subjects at Duke, which was managing AI in business, explainable AI, which is Dr. Ben's course, as well as AI product management. What I'm seeing in this field is that since AI is such a new domain, that it's even very hard to create educational information because the whole domain is so new. So having something to help you in brainstorming your ideas, having something to help you canvas out your presentations, to make it ready, just, you know, give it a prompt saying that, "Hey, I'm wanting a presentation about X, Y, Z. Can you please create a basic version for me?" And having that as a stepping stool to, like, you know, start from there as your base and move on. Having all that redundant stuff taken care of, I think both in the field of education as well as learning would be a big win.

  Lydia Kumar: Yeah, those are such interesting use cases, and it's like we, whenever you're learning, you get all of this information, you're kind of compiling it, but we forget so much. And so even if as you're going through school, you, you, some stuff stays with you, but a lot of the papers that you read, the facts become difficult to access. And so if you can build that as you go, especially if you're developing an area of expertise as you are in artificial intelligence, you could really kind of craft a tool that could be focused on that specifically. And the other use case I thought about was your, uh, if whenever you get feedback on something that you've done or if a student takes an exam and then you can feed that back into your AI tool, you could then practice based off whatever the areas you needed to, to improve on are. And so it allows you to be a lot more actionable with the feedback because you could just take your, take your exam and then upload the PDF and then have a, a whole new experience where you're able to, to practice and revise. I think it allows a learner to really take ownership of, of their education.

  Vihaan Nama: Yeah, I think it's really just transforming the way and we learn because I grew up in a time when there was no ChatGPT. I did my whole undergrad without ChatGPT. But now in Duke, I'm seeing a lot of undergrad students using ChatGPT for every assignment that they want, that they have to do. So I did my undergrad in India, and over there, a lot of learning was done in the form of memorization. I think that's how most of the country was doing learning up until recently. But before that, yes. And a lot of my examinations, my quizzes was literally like, "Define this, give me the exact definition." But now with the field of AI, even back then, it was not really required to give you the exact definition, but even with the field of AI, there's no more a need to explicitly remember certain terms and terminologies because you have access to it so easy. I feel that so much brainpower can now be used for so many other better things because the redundant heavy lifting can be done by the AI, and you can just focus on discovering, inventing, creating something new, something that people can use and people can love without having to worry about the technical complexities.

  Lydia Kumar: Right. It unlocks this whole, your brain can only do so many things at a time, and so it frees up this space that you may have been holding all of your memorized facts with. Now you can refer to your AI and instead you're thinking about how can I apply those facts to create something, something totally new.

  Vihaan Nama: Exactly.

  Data Privacy and Open Source AI

  Lydia Kumar: I'm curious 'cause as, as you know, I think all of this data is so useful, but on the flip side, there's this privacy element where if, if you are giving—I gave the example a minute ago about sharing your exam results with your AI assistant to help target your learning. But if you, if you do that, there's some data privacy that you're potentially giving up. And so I'm wondering about how you, in your studies or your perspective on data privacy, how, how AI uses that data, what risks are involved for people who may be interested in taking the, the information and, and finding insights about things that may be specific to individuals?

  Vihaan Nama: Yep. So for data privacy, I would say like, it is very—it's, it is a very known fact amongst the AI community that anything you put onto ChatGPT, Claude, Perplexity, any of these major services, unless you explicitly opt out of their retraining or data retention, they are going to retain it and they are going to train their model. Sometime in the future might not be immediately, but maybe even 10 years down the line, your data could be reused. Right. So it is a very known fact within the AI community, but outside the AI community, I think it's still not as well known, even though we are in the, like, in the space and we're like, "Oh, this should be quite obvious." But I don't think a lot of people do realize that their information is actually going to be used and the minute you put it onto their platforms, it automatically becomes their property. And if once it's their property, they have the right to use it when and how they want. Most of these platforms do have small check boxes, which are automatically selected. So you have to actually go and unselect them to say that, "Please don't use my data." But still, like, as much as we, we do that, we, we can't ensure what's actually happening on the backend because we are faced with like a curtain. We can only see as much as we are given. Even if we do unselect it, we actually—they say they're not going to be using it, but we don't know what's actually happening on the backend. And to fight this, I think the advent of like the open source field of AI has really started booming, especially with a lot of organizations—even like let's say the government, let's say organization with a lot of proprietary information—they want the power of AI, but they don't want to give out their information because their information is where their secret sauce is. That's where their business is running, and they can't afford for that to become public information. So now, nowadays, there's some stuff called open source AI or open source LLMs. And not just large language models, any AI, there's computer vision recommendation systems. There are a lot of models available online in the open source domain, which companies can invest in. Even education institutes, anyone can invest in that. And all the difference is, is that you have to set up your own infrastructure. And since you might have to buy some GPUs, you can have a little bit of infrastructure costs associated with it. You might need someone to manage that stuff, but actually doing this one-time capital investment, it's going to allow you to bring in these large language models or any kind of AI model locally, and then be able to train it and run it without having this information go onto the internet for the companies to retrain it, because it's all going to be within your own ecosystem.

  Lydia Kumar: For listeners who may not know what an open source, what open source means, could you explain that?

  Vihaan Nama: Sure. So, for open source basically means that there's a lot of companies who have put their models out for use for public. And they're—they don't just give you access to the chat URL, rather, they give you the actual model itself and they made it open. For example, Meta has done this with their LAMA models. Google has done this with their Gemma models. Mistral has done it as well, but they actually give you the entire model. And now you can download this entire model to your local machine and run it locally without even having access to the internet. You can bring it to your local machine, turn your internet connections off, and just start querying the model. And since the whole thing is hosted locally on your own machine, you have complete right to it. Your data doesn't go out, doesn't go anywhere. So open source basically means that they're just giving access openly to everyone and bringing it, making it more accessible.

  Lydia Kumar: Thank you for explaining that. When we think about the open source opportunity is this ability to run a more private and protected model. And I think whenever, like you mentioned, government, I think when we talk about student data, especially students who are minors, we have to be really careful about the data that is put out there. And so there's ways to, um, how making sure that we're able to protect those, that information is important. And so open source is one strategy to do that. And so thank you. Thank you for sharing that. If I was a superintendent or someone at a district and I was looking to invest in an AI tool, are there like a, an AI tool that was specifically built for education but had had an AI system that was running. What are things that I would need to pay attention to or questions that I should ask to make sure that my data would be protected in that system?

  Vihaan Nama: I think your first step in that field would be to do your own research, firstly into any machine that you're buying or like any subscription model, anything that you're buying. I think your first job would be to do your own research, look up online, see, because with all the machines out there, all the softwares out there, there are a lot of reviews online, there's a lot of information already available to you. So I would say first go through, go through those Reddit channels, go through the information, go through the reviews, and actually understand what the system is doing, not just from the customer point, but how they're using your data on the backend. Um, if you feel that this information, that these machines don't fit your use case, I think you need to either, number one, I think you should always have an AI expert on hand where you hire someone who's well versed in the field of AI because there's a lot of terms and terminologies that are tossed out there, especially in these systems that make it look all fancy. But actually all you could be doing is paying for something that is not as great as they're making it look. So have someone who understands these technologies. Make sure that they understand what is going on inside them. Make sure you explain to the person on your side as well as try to get a whole—most of these companies out there that have these systems available always have like a support page or a help page—try to get in contact with it, especially when they're doing stuff with businesses. They're always ready to like talk to you to understand and go through your questions. So have your people meet their people. Um, if it doesn't work out for you with this because of the data privacy regulations, because what we need to understand is that each company and each organization's values differ. So what I find important, maybe you don't and what you find important, maybe I don't, and that's just the way organization is. So what the AI companies find important might not, might not be what we find important. So if we're trying to build something and we have a lot of specific things that we want to pay attention to, that the AI companies are either after meeting their people, they're not able to accommodate your requests, or they're just saying that this is what we're going to give you and that's it. I would say again, open source is a great way to go because what happened then is that you're just getting the raw model. You're just getting the raw model. It is a bit of a capital expense by building a system around it, getting the infrastructure, getting the GPUs, but it is a one-time expense. But what that allows you to do is imbibe your own values into the AI system. You're able to build, like hire a couple of developers and tell them what you value. And you're able to build an AI system around this model that you pull down, which is open source, free to use. You're able to build an AI system that imbibes your cultural organization's values and what you require there to be.

  Lydia Kumar: Yeah, you don't always, when you build something yourself, you're able to have a better sense of what, what it is, if it's aligned with what you, what you value, as you just said. And I think, and, and there's the safety or data privacy aspect that's really important too. As you were talking, I think there's a reality that, um, a lot of education institutions won't have the capital to be able to build this kind of mo—their own, to take the open source and build their own model. But it is helpful to think about, okay, what questions should we ask? What do we need to understand? Going into those kind of conversations, what should you look for? When can you use a foundational model like ChatGPT or Claude, and when do you want something that's more tailored to the audience or to your use case? And I think all of these things are really important to think through because there's a tremendous amount of tools and things being made to, to help, but that doesn't necessarily mean that they're a value add to your organization. And so I think when you talked about values, just being really clear about what your values are, what you want to use this for, having some basic understanding of the technology is going to be really important to anyone who begins to make these decisions.

  Vihaan Nama: Yes, definitely. And when it came to like whether educational institutes could invest in them, what we're seeing with the scale with AI right now is that as time goes on, our cost for capital expense is dropping tremendously because the models are getting so much more powerful. And what that basically means right now—it used to mean that they're gaining a lot more information, they're becoming more knowledgeable, but we've already reached that point where you can chat with any model and it's going to be able to converse with you good. When an AI model is becoming more powerful. Now it also means that these large models are going to be able to be compressed into such smaller sizes that they're going to be able to run efficiently on easier infrastructure. That's also going to make it much cheaper for people to use in the end of the day. So I think as time goes on, we're going to see the, um, the amount that the CapEx investment to actually go down significantly. And all you're just going to have to do is maybe even just buy a single GPU, which is not that expensive.

  Lydia Kumar: If you're an innovative school district, thinking about doing, thinking like down the road, this is something that you're interested in. What are, what are tasks or things that school districts could be doing or preparing to do if they wanted to build something custom like this for their, for their institution?

  Vihaan Nama: I think number one is getting your data ready because so many times, like I've worked on project as an engineer itself, worked on projects where the person has such a grand idea and is, and like amazing forethought. And as an engineer I'm not so ready to start and I'm ready to get my hands dirty, but then the minute I'm there, we get stuck in a place where their data is not ready and our machines can't progress without their data. So I think the first step to do anything is to make sure that all the data is there and it's all in the right format. Um, and you're not going to be scouring for it, looking for it. After the engineers are on site, you're going to have everything ready to just give them. I think that makes number one and makes an engineer's life so much more easier because you're not dealing with that whole extra thing of like that layer of frustration where like, "I really want to work, but I don't know how to progress from here because there's nothing for me." Um, so I think having that is a number one thing. Number two, I think it would be like budget allocation. Um, speaking to people, understanding, um, even in different domains and fields, how expensive this is looking, and then trying to draw parallels. Maybe you're the first educational institute that's doing it. I'm just saying like, policy around hypothesis. Imagine you're the first. There are other fields that have done it, which have relatively the same size of data. So maybe you can talk to people from cross-domains and understand how they were doing it, what their costs were. And I think before you get started, it's always good to speak to three or four people who've already implemented their systems and understands where you need to learn from because you're not reinventing the wheel, you're not facing all those problems again, you can skip past those early stage problems that come up with a lot of stuff and you know where to get your ground running. Um, I think it's also important to have someone who is well versed in the field of AI available to you so that basically you can talk to them about your ideas and they can visualize out a system, because sometimes it's hard for us to visualize this out a system without having the technical knowledge of AI and stuff like that. But if you have someone who's able to do that for you, or even nowadays, there are many like LLM agents that are, that are there to help you along with it. You know, you throw in your use case and it tells you about what exactly you need. I think your planning should be really good before you start.

  Building Trust and Responsible Use

  Lydia Kumar: You've done a lot of explaining in our conversation today about how different things work and things people need to be aware of as they're thinking about implementing a system like this. As you have explained AI to people beyond just me, and you took your explainable AI course and—or you TA'ed for the explainable AI course—what have you learned about building trust and clarity for teams or organizations who are interested in using AI, particularly considering there are a lot of people in the United States right now who are very nervous about AI, the, the technology, and what that means for the future.

  Vihaan Nama: So when, when we talk about building trust and clarity, I think we need to understand that these AI systems should not be relied upon, because they can make mistakes. And for example, if you, if you open up ChatGPT, and you even search any query, at the bottom itself, like on one of the largest companies out there, it's literally written: "ChatGPT can make mistakes. Please check your important information," or something along those lines. And that's very important to know because AI is, in the end of the day, AI is a, uh, or like these models are a single file, which contains like the world's knowledge in it. These large language models, you can think of it like a single file that contains the entire world's knowledge in it. So if you're querying it, there could be chances where it gives you the wrong information or mixes facts up, mixes stuff up, and even in the end, even makes up, makes up wrong things. Like, you know, stuff that completely doesn't exist, just made it up. And this is known as like LLM hallucinations. Like it literally hallucinating. It didn't, there's no actual grounding. So, um, when it comes to people understanding and building trust, I think the first question you should ask, especially, uh, if you're putting your own information in there or you're getting data to feed to an AI or, uh, an AI in general, is just basically where is this information grounded in? Um, you need—AI currently has access like through a lot of softwares to the internet, to stuff. And you know, actually asking AI to cite its sources and you know, like telling it like, "Hey, you're going to give me this information, but I also want you to cite where you got it from," actually helps the model a lot. Picking out the right information and making sure it doesn't hallucinate or give you wrong facts because it's just, it is just shown through prompt engineering and through various techniques that this actually helps you bring out, um, the right information. But also if you're uploading, um, if you're using an already existing, um, online LLM, but it's, but someone else has fed the data to it or like has fine-tuned it, for example, and you're, you need to ask them where the data that they got was grounded. Make sure that any, if you're making synthetic data, if you're like basically since you're creating false, like even if you're creating false data to teach an LLM, it should be grounded somewhere. There should be a grounding source. And I think, um, that's where your trust is going to lie because you have to be confident in the data that it was originally trained on to be confident of your answers.

  Lydia Kumar: It like the trust comes down to understanding where, where does the information come from in the first place and, exactly. And right. And so if you're team, if you're introducing, using more AI with your team or your organization, really having maybe that as a norm of, we're always going to ask and try to understand where this came from. And then if we're post-training a model, we're going to make sure that we understand what the data is that, that we're feeding into it so that we're the, the humans in the room can have a conversation and be on the same page.

  Vihaan Nama: Yeah, exactly. Like, um, if we think about it like in, like as a restaurant, like, you know, we don't know where all the ingredients are from, but we somehow trust that it's put together, it's tasty and it's not going to poison us or it's not going to make us feel ill. But if we actually cared about where it's from, I'm sure restaurants can trace your entire supply chain back to the source. And that's important, having your own, making sure you can trace the supply chain of the data that was given to you. Like where was it got, how has it been transformed, what has been added onto it, the added-on information, is that grounded somewhere. And then we fed our models because at the end of the day, we don't know, like an LLM is a black box, we don't know what's going on inside it. There are trillions of connections and tiny, small things that are getting activated, not activated. And there's so much going on in there that it's basically impossible for us to quite truly understand what's going on inside. But what we do have control over is what we feed it, what we do have control over is what we ask it. So prompting it in the right way so that it can give out citations, it can give out. Um, not help it navigate to the right information for you. That's important, but also, as you said, while we're post-training it after, or fine-tuning it, that information that we're giving it, we have access to as well. So making sure that's correct, personally, factually.

  Lydia Kumar: It, I love your restaurant analogy. I am imagining, you know, if, if you went to a restaurant and you had gotten food poisoning before—you might be very cautious, or you heard about someone, you know, food poisoning being our hallucination analogy right now—and you're sitting there, if you, you might ask, "How long did this food sit out before I got here?" Or you could, exactly, you could ask that waiter questions and then you could feel safer in how you were using it, a more secure because you had the information that, that you needed. So the restaurant analogy is good, Vihaan, that that was very helpful.

  Vihaan Nama: Yeah.

  Lydia Kumar: Okay. I have a, I want to think now about responsible use of AI. So I think for organizations who want to adopt AI, what does responsible use look like? And then for students or people teaching students, what does it look like for a student to responsibly use AI?

  Vihaan Nama: Responsibility comes in many facets over here. Firstly, there is the responsible use in the sense that it should be used for only ethical use cases. Um, unethical use cases are as abundant as ethical use cases, and it's very easy for people to diverge onto that path. But having your—exactly, having your value set, having stuff done before, it's very important to make sure that this is all done beforehand. But there are so many other, uh, stuff when it comes to responsible use. For example, are we using it right for the climate? Um, because these AI systems access large data centers and they're using our electricity. So are we using it good in that front? Are we—does responsible AI mean that we are making sure that we're using AI, but we're also using our own critical thinking? We're not making our own brain idle away. While the AI does the heavy lifting, we don't become lazy in the process. There's so many facets of AI that needs to be done responsibly. So, like, if I start getting into it one by one, when it comes to the first point for ethical use of AI, I think it's important that firstly, as an organization or as a student or as an educational institution, we set our values straight. We are, we understand what is ethical and what is unethical for our use case. Um, and we make sure that we talk to not just the builders of the AI system about this, but also the users. The users should be as educated as well. Um, I believe afterwards, the best way that, that we need to do our best to prevent unethical uses from happening. So we can do this by incorporating various types of safety guardrails inside. Um, making sure if, uh, having question screening, if a student is asking to cheat on an examination—a normal ChatGPT would allow it because for them, that is not part of—cheating on examination is not part of their values as what they consider unethical. But as an educational institute, you would consider it unethical. So if the student is trying to cheat on an examination, if they use ChatGPT, it would probably give them the answer. But if we're giving them a system that we've built, then we as an educational student need to put in those safety guardrails saying that, hey, if a student's trying to cheat on an examination or ask questions that are trying to give them the answer, instead of giving them direct answer, teach them how to think, teach them how to go down this process of problem solving. So I think, um, that's very important as like to basically set your guardrail straight and also set your values straight. Um, when it comes to the climate, um, AI is as much as it's helping us, it's also destroying us because there is so much bad that's happening with the increase in increasing use of energy. Um, at my time in Duke, I have been a core team member of the—we have an energy and sustainability called club called Pratt Energy and Sustainability Club, Pratt being the engineering school of Duke. So I am one of the core members in the Pratt Energy and Sustainability Club. And we spent the entire time last year, nearly like a whole week and a half in the Responsible AI Symposium that happened in Duke, where all we did, we came into the into that symposium asking amazing leaders, amazing people in the space, what they feel about sustainability and AI. You know, it's such a question that is not asked because people often just look past it thinking that they just type question in and some magic happens and they get an answer, but they don't understand what's exactly happening in the backend. In the backend, there are mega data centers that are actually processing your requests. Just like you, there are millions of other users querying the system at the same time, so handling these requests become quite expensive. As a part of my research last year, in the Energy and Sustainability Club, we saw this particular document that was released by the Department of Energy. And they said that by 2028, they're assuming that 6.7 to 12% of the entire United States' electricity is going to be consumed by data centers. And this is a huge number. You know, 12% of our entire electricity produced is going to be by data centers. That's, quite frankly, like a lot. It's quite frankly, crazy, you know? And we need to understand the macro impact of this.

  Lydia Kumar: When you, when you were talking, I was thinking, Vihaan, just about how it's, as an individual, you are one person using these systems, but then if you're leading a large organization like a school or a business, you actually are contributing to a bigger impact when it comes to that energy usage. And so I think for leaders, there's a consideration that is a little bit different than for individuals. And so from your perspective and your research, how, how should leaders navigate wanting to capitalize on this technology and all the benefits, but also knowing that there's a big impact on our electricity grid and our environment? And then how do you lead accordingly? Like what, do you have advice or thoughts or perspective on that?

  Vihaan Nama: So I think the first thing we need to do is understand what is a machine and what is human. Right? So the reason I say this is because there's this funny little tweet that came out from Sam Altman, I think a couple months or a year ago, where he talks about how people are being polite to AI and that's actually costing tens of millions of dollars in the electric grid, where basically they're saying "hi," they're saying "thank you." And you know, even those small queries that you're sending back and forth to the AI, they're not required. You don't actually have to be polite to a machine. If you think of it, if, if you had a system that was, uh, like your washing machine after you're done, uh, with your laundry, you're not going to say thank you. Right? But people are actually querying these systems and saying, "Thank you for explaining this to me," and that's it. But, but what happens on the backend is that this actually goes as a query to the data center, to their server—OpenAI servers or whatever server in your, your query—and actually gets processed with the same amount of weight that any other query would be processed. Right. So that funny little tweet by Sam Altman was just explaining that, you know, you don't have to say hello to an AI system. You don't have to say thank you. All that's doing is just increasing the load on the electric grid. So when it comes to building these systems, I think the first step is understanding that, make sure your manners are there when you're talking to humans. But when you're talking to AI, be very efficient with the way you speak. Make sure your prompts, the stuff that you're telling it are to the point, concise and only what is required, because that's going to help the AI systems to get done with your query really quickly and give you the information that you need. So it's not just helping environment, it's helping you out because it's going to help you get precise information really quickly. Um, also when it comes to this, I truly believe, like if you're building your own AI system and you need to also do energy audits, you need to figure out how much energy systems are going to take up, um, how much extra strain are you putting on, and make sure that what value you are gaining from this AI system is much better than the strain you are putting on the environment. And also I think you need to keep updating these AI systems, as I mentioned. As time goes on, these systems are becoming so much more powerful. So with what you can do, with what you could once do with such a large model, you're now going to be able to do with a smaller model. As these AI systems keep getting updated, make sure that your, your organization is up with the trends and it also reducing the size of these models as they become more powerful because you're then going to not just bring your own electricity bill down because you're not going to be pumping in through a 70 billion parameter model that you're not going to be able to do in like 8 or 14 billion parameters. But you're also going to be helping out the environment as a whole.

  Lydia Kumar: Right. It's this awareness that what we do doesn't operate in a vacuum and there is an impact that's beyond ourselves when we are talking to, to these intelligent machines, to generative, to whatever generative AI model you're using, there is. It's not just magic. There's actually a data center. There's an impact. There's things that happen that you can't see, but that doesn't mean they're, they're not there.

  Vihaan Nama: Exactly. On the backend with all your extra queries that you send in, in the end, it's just fossil fuels that are being burnt and released into the, into this, the air that we breathe. So it's, yeah, you know, it's better to, yeah, it's better to just keep up to date. We, there are a lot of important people out there who are continuously working on and government officials continuously working in, investing money into making this smaller and more compressed so that you can fit it into your pocket, fit it into anywhere you want, and make sure that while that's happening, your organization is doing the same. Don't just develop it and leave it there and then make everyone use it, but rather keep updating and keep reducing your costs.

  Lydia Kumar: Thank you so much for sharing that and bringing us, bringing that ethical component to the forefront of people's minds. 'Cause I think it's, this technology is incredible and powerful and useful. Also, there's, there's environmental impacts and most things don't, most things are more complicated than what they see and have a macro impact. And so taking time to understand that, I think is, is really, really important. And my last question, Vihaan, which may be everything you just said, but I always end this show by asking if there's an idea, question, a hope, something that kind of stays with you when you think about artificial intelligence, maybe the thing that keep, that keeps you up at night. Um, and for you, what, what's on your mind right now? What do you keep thinking about in the AI space?

  Vihaan Nama: These models are improving, their capabilities are improving. And it's making us as human beings more redundant because it's going to be able to do a lot of the heavy lifting. Um, what got me thinking about this is that I saw this funny quote online where it was like, or not funny quote, sorry, funny tweet online where someone said it was like, um, as time goes on, I would like AI to do my dishes so that I could focus on art, but the way it's going right now is that AI is doing the art while I'm focusing on my dishes. And I think that's quite. Um, startling to me because, you know, there's a lot of stuff that people actually enjoy doing. For example, creativity, for example, inventions, discoveries, engineering, so many fields that AI can impact and can get done much quicker. So I think what keeps me up and makes me, what, what I think about is as this keeps moving on, as this field keeps progressing, how do I stay relevant? How do I make sure that I am collaborating with AI rather than, rather than being replaced by AI? And I think, I think the only conclusion that I have drawn in this statement—I think a lot of people in the world are taking the same thing as well—I think the only conclusion that I've drawn is that. Staying up to date on the technology, making sure that in whatever, whatever I'm doing, understanding is there a space where AI could help me? And if it helped me, what could I focus on rather? And, you know, having a clear distinction of making sure that AI is doing the redundant work so that you can focus on something new rather than AI doing the new work while you're focusing on something redundant. And that's what keeps me going every day and why I like to be so involved in the field.

  Lydia Kumar: It's really interesting because you are in the field of AI and you're worried about the thing that everyone is worried about. What does my work look like down the road? How am I choosing to use my time? What am I allowing AI to do for me? And then what am I doing in response to that? And so it's. I don't know. I, I really appreciate you sharing that. You have some of those same thoughts because we're all kind of in this space trying, trying to figure out what does it look like to live in a world with technology that can do so much of what we used to do as do as people. And you're, you're thinking that even as you work on these machines directly. So that's really interesting.

  Vihaan Nama: Definitely. Because there's, um, at the end of the day, the only people who know the true extent of what AI can be in our visible future is probably these larger organizations. And we need to understand that maybe their interests are not our best interest. So we need to make sure to stay relevant. Always make sure you're upskilling, make sure you're involving AI in your everyday workflows and understand and. Not have it surprise you later on, like after three years you decide like, "Hey, let me just start with AI." And it's got so advanced that everything you're done, that you're doing is now redundant, right? Rather have it as AI is growing incrementally, you grow with it and you know, you keep evolving with it because it's still growing. So your role as a human being in that particular domain that you're in is also growing. So don't just be surprised later on. Be actively involved and figure out how you can grow with it.

  Lydia Kumar: That's a wrap on our conversation with Vihaan Nama: AI engineer at PSNS, researcher at Duke's Trust Lab, and TA for some of the most relevant graduate courses in AI today. Vihaan's work shows how explainability, peer learning, and real world use cases make the difference between compliance and curiosity. If your school or district is navigating these same questions, check out the Kinwise AI Leadership Series. It's a five-session cohort that equips leadership teams with clear vision, guardrails, and a living AI roadmap. The next cohort starts January 26th. Visit Kinwise.org to book a free visit—visit Kinwise.org to learn more. And if you found value in today's episode, the best way to support the show is to subscribe, leave a quick review, or share it with a friend. Until next time, stay curious, stay grounded, and stay Kinwise.

• Connect and Resources

  • LinkedIn Profile: Connect with Vihaan on LinkedIn to follow his work at the intersection of AI, education, and engineering.

  • Personal Website: Explore Vihaan’s portfolio, project write-ups, and teaching philosophy in greater depth.

  • Google Scholar Page: Browse Vihaan’s academic contributions and research citations across AI, explainability, and systems design.

  • Pratt Energy & Sustainability Club: Learn more about Vihaan’s leadership role in Duke’s campus-wide effort to examine the environmental impact of AI systems.

  • YouTube: How GenAI Is Reshaping Education: Watch Vihaan’s insights on AI in the classroom from a Duke panel discussion, including open-source tools, student empowerment, and responsible innovation.

• Prompts inspired by Vihaan

  1. Student-specific Knowledge Retrieval and Synthesis: Access all my uploaded notes, lecture transcripts, and submitted assignments from the 'Explainable AI' course. Summarize the core differences between LIME and SHAP methods, and then generate five practice quiz questions, complete with answers and citations to the specific documents where the information was found.

  2. Addressing Learner Deficiencies: Analyze the attached PDF of my 'Managing AI in Business' midterm exam results. Based on the topics I scored lowest on, create a personalized study plan consisting of three detailed readings and five application-based case scenarios to improve my understanding of regulatory compliance in AI deployment."

  3. Educational Content Brainstorming and Development: I need to develop a one-hour introductory lecture on Retrieval Augmented Generation (RAG) for a class of non-technical business students. Generate an outline for the presentation, suggest three relevant real-world use cases, and propose a concise, non-jargon-heavy definition for RAG.

  4. Data Sourcing and Trust: I am preparing a policy on student data privacy for our district's new AI tutoring tool. I need to understand the current legal framework. Find and summarize all key excerpts from the Children's Online Privacy Protection Act (COPPA) and the Family Educational Rights and Privacy Act (FERPA) that relate to educational technology and minor student data. For each summary, include the source citation and document location.

  5. Ethical Guardrail Development: I am building a custom AI model for student homework assistance. Design three different safety guardrail responses for when a student directly asks the AI to solve a complex math problem for an upcoming exam. Each response should decline to provide the direct answer but instead use a different pedagogical approach (e.g., Socratic questioning, breaking the problem into simpler steps, or providing a link to a relevant instructional video).

• About Vihaan Namea

  Vihaan Nama is an Applied AI Engineer at PS&S, a graduate researcher at Duke University's Trust Lab, and a teaching assistant for leading AI courses including Explainable AI, AI Product Management, and Managing AI in Business.

  From his early experiments in sentiment analysis to his current work designing retrieval systems and open-source tools, Vihaan is driven by a passion for making AI both understandable and empowering, especially in education. His leadership in Duke’s Energy and Sustainability Club also reflects his commitment to ethical, environmentally conscious AI development.