# Transcript

**Srikanth**: [00:00] Anand, good to see you from Singapore. Mr. Murthy, Anand is joining us from Singapore and he's been my collaborator in coming up with the AI transformation for NIE.

**Mr. Murthy**: [00:15] Oh, wonderful. Wonderful. Fabulous.

**Srikanth**: [00:18] He's got a wealth of experience.

**Mr. Murthy**: [00:21] Yeah, yeah. I know they have to solve it. Yeah, yeah.

**Unsure**: [00:28] Barthaidara? [English translation: Are they coming?]

**Srikanth**: [00:34] Good morning, sir. Just tell them Murthy avaru bandidare [English translation: Mr. Murthy has come]. Just tell them to the Management Committee, correct.

**Mr. Murthy**: [00:46] By the way, I met Rishi Sunak very briefly at the AI Summit a couple of days ago in Delhi.

**Srikanth**: [00:54] Oh, wonderful. Yeah.

**Anand**: [00:59] I'm not able to hear you.

**Srikanth**: [01:03] You're not able to hear me?

**Anand**: [01:06] I am able to hear you now. Now it's fine.

**Mr. Murthy**: [01:09] They wanted me to go there. Somehow there are zillions of fellows there. I just politely, very respectfully declined. Too many, too many people. You see, **in any of these things, you should not have more than one major speaker in the morning, one in the afternoon. In two days, just four speeches are more than sufficient.** Just zillions of fellows. I hope they come out with some book or something about what all these people said.

**Srikanth**: [01:58] They put it up on YouTube, I think, some of the talks and so on.

**Mr. Murthy**: [02:02] Who will have time to go through all those YouTube stuff?

**Srikanth**: [02:11] True. I told Rishi Sunak, the one thing in common I have is your father-in-law. We are working together. Then he said, "How come?" I said, "We come from the same alma mater," and he dragged me into the management committee. He said, "That is very much like my father-in-law!"

**Mr. Murthy**: [02:32] No, I mean, you are clearly the most illustrious student of NIE. There are very few people like you, so therefore I think it is an honor for me to be with you.

**Srikanth**: [02:54] It's been an interesting experience last couple of years working with NIE, but this is a pivotal moment with this AI coming in. So I wanted your inputs as we go along, so I thought while we are discussing this transformation that Anand and I have come up with, I thought getting your inputs would be valuable.

**Mr. Murthy**: [03:20] Definitely, it will be a pleasure and a privilege. But more than my inputs, I think everybody is very eager to listen to you because **some of the things that you have done for our country is so valuable at a macro level. We've all done at a micro level, but what you have done at a macro level is very important.**

**Srikanth**: [03:55] Thank you so much, but we have been inspired by your work at Infosys and so many other things. Thank you.

**Srikanth**: [04:06] Are they coming over? Okay.

**Srikanth**: [04:18] Do you travel these days often?

**Mr. Murthy**: [04:24] No, not really, because I am now running my 80th year. I don't travel much at all. But wherever there is an extraordinary institution and extraordinary individuals, I try and... they don't want to give me up. So AIM has an extraordinary Dean called Jikyeong Kang. She is a South Korean person, born and brought up in the UK. She did her PhD from Kellogg in marketing, and she was a very well-known professor there for about 35 years. About a decade ago, we invited her to become the Dean of Asian Institute of Management. I was on the committee, and she has done a brilliant job. So I had not been able to go there over the last two years, so I thought I should go at least this time.

**Srikanth**: [05:46] Yeah, so wonderful.

**Srikanth**: [05:55] Okay, looks like we have the management committee as well here. Can we get started, Ranganath, Uday? Okay.

**Srikanth**: [06:08] Well, welcome everybody. It's wonderful to be here at NIE always, my alma mater. And it's always special to be here. We have spent the day at the management committee board meetings since morning, and we have seen the wonderful constructions that are taking place. Girls' hostel, the two blocks, A and B. Fantastic facilities and comparable to anywhere in the world, so it's wonderful to see all the progress.

**Srikanth**: [06:35] Today's meeting here is a request that came from the management committee. Uday had written to me saying there's AI transformation taking place, and since I've been involved in some of these things at the central level and so on, he said, "Can you do a workshop on AI transformation and what we need to do as NIE to keep up with the AI transformation that is taking place?" That is the context.

**Srikanth**: [07:02] And I immediately contacted my friend Anand S, who is the founder of this company called Gramener out of Bangalore. It was sold to a company called Straive. He is an AI psychologist. He's one of the most brilliant data scientists I know. And he very rarely does presentations; he literally codes in front of the audience to explain the various concepts. So I really wanted him to show us things as I talk today about various things. So we are delighted to have Anand here, and I'm trying to also gradually pull him into helping NIE as well as we go along.

**Srikanth**: [07:44] So with that, a couple of questions came up at the management committee. We have this group asking, "**Is AI hype? Is there reality to AI or is it just hype?**" And so on. Can we all sit down, please? Can we all sit down? Yeah, just... yeah.

**Srikanth**: [08:10] So I wanted to address the initial aspects of AI, my understanding, as well as Anand is going to jump in. Mr. Murthy, please jump in with the rich experience and interactions you are having on this.

**Srikanth**: [08:30] Firstly, what I wanted to address was, Ranganath here was asking us, is this hype or reality? And that's a good question. So let me just take you quickly through the history so that we can appreciate what has happened.

**Srikanth**: [08:45] Actually, **AI work started in the 1950s with Alan Turing, the Turing Test**, and so on, trying to see if machines can do what humans do. And there was the idea of the perceptron that was way back in 1952, inspired by the brain. The idea of an artificial neural entity which was sort of like what our brain cells process was something that came up. Let me pause... excuse me? Stop talking, please. It's just disturbing. If you keep moving... please sit down. Wonderful.

**Srikanth**: [09:37] So a lot has been happening since the 1950s itself. When I went to college in the US, **ELIZA was a program that was there at MIT. It was a psychiatrist, but these were expert systems that were built out of rules and symbolic language.** They would say, if someone asks this, then respond like this. It was an expert system based on rules that human beings constructed. If you need to be a psychiatrist, this is how you respond to questions that come up. If you're a civil engineer, this is the expert system, and so on. They were carefully constructed rules using symbolic logic. And it was believed that AI could be powerfully built on these kinds of logical rules.

**Srikanth**: [10:23] It turned out that it didn't go very far. And in fact, when it says "AI Winter and the Disappointment," that is because of that approach that didn't take us far. **The approach that actually took us far was the connected approach of this neural networks**, the perceptron that came up in 1952.

**Srikanth**: [10:43] Let me very quickly give you an example, Ranganath, since you were asking, and Srinath, you were also referring to this. For example, let's say we just want to classify cats and dogs by an AI. If I give a photo of a cat, it should say this is a cat. If I give a photo of a dog, it should say it's a dog. In the earlier approach, we would literally code rules and say: cat is furry, it has four legs, it has whiskers, and so on and so forth. And we would say similar things for a dog—barking, it also has four legs, and so on. We would try to describe it through symbolic logic and hope that it matches these rules looking at the picture and then says this is a cat versus that is a dog.

**Srikanth**: [11:25] That approach did not work very well. If you took pictures of cats from the top or if a cat is under a carpet and only the tail is sticking out, you can't find out all this because those rules don't fire. Whereas human beings, we are able to say even if you see a tail coming out of a carpet, you know this is a cat. How is it that we are able to do it but AI systems through these rule-based expert systems were not able to do it? This was the AI winter. I'm abbreviating this enormously, but just to make a point.

**Srikanth**: [11:54] Whereas in this neural network-based system that you see, a mesh of neurons, those orange ones, green ones, and red ones that is representing a deep neural network... the input you would give an image of a cat or a dog that came in. And output, there's only two outputs. If the first one fires and says the probability is very high that the first one is true, that would be a dog. And if the second one comes out as one, maybe that's a cat. Let's say that's how we built it.

**Srikanth**: [12:28] **The way this system worked is feature engineering—that is, talking about legs, whiskers, the furry things—all of those features that we humans had to embed in AI, none of that went in. Simply millions of pictures of cats and dogs in various positions with the correct answer was given to train this deep neural network.**

**Srikanth**: [13:00] Initially, the weights—the smarts of the brain of that deep neural network is in the weights, those connections between the neurons as well as the biases in those cells—that's all it contains. Initially, it was randomized. That means it had no intelligence of cats and dogs. But as you give an input of a cat and it predicts something nonsensical... let's say the below output should have been a one and a zero on top, but it gave 0.3 below and 0.7 above. You would take the error, you know that the below has to be one, instead it gave 0.3. The difference error is 0.7. You propagate that error backwards through something called backpropagation using gradient descent. Once you started doing this a million times, it figured out what a cat is in the true sense. Meaning, even if you took a picture of a cat from the top or just the tail or a dog, it would classify.

**Srikanth**: [13:54] So suddenly things had changed from humans having to clearly say this is how AI expert systems should behave to...

**Anand**: [14:09] Hello? I just got pinged getting to know... Hello? We lost you. We lost you.

**Anand**: [14:40] I think he's still carrying on with the audience. Okay.

**Unsure**: [15:08] Hello? We can hear you now Srikanth.

**Srikanth**: [15:11] Okay, wonderful. So the various things happened, but **suddenly something happened in 2017 that completely changed the trajectory of AI. This paper was written called "Attention Is All You Need."**

**Srikanth**: [15:32] These neural networks that I'm talking about were working pretty well in the vision space. Convolutional neural networks (CNNs) were able to do very well in detecting cats and dogs, even breast cancer detection, and so on and so forth. In the vision space, it was doing very well. Whereas when it came to language, human language is incredibly complex, right? The so-called RNNs, recurrent neural networks, that was built with that technology prior to 2017 were fairly slow. And they did not have an attention that went back into the context if you had a long sentence or a paragraph. It could only keep a short amount of information alive.

**Srikanth**: [16:11] They came up with this interesting paper where if you gave a sentence such as, "I live in Mysore, I can speak _ fluently," you all know what the answer is. I don't have to tell human beings what the answer is. The answer is Kannada. How do you get the machine to predict this next token accurately? It turns out this paper showed us that in that entire sentence, only the words "Mysore" and "speak" are important. Everything else is not important. So if that word you're looking for, Kannada, all you have to look at is pay attention to "Mysore", you have to pay attention to "speak", then you will have a high probability you will get the correct next token. All other words you can pretty much ignore, you will still get the right answer.

**Srikanth**: [17:05] I'm of course abbreviating this quite a bit for this talk. This "Attention Is All You Need," and the lead author was an Indian, Ashish Vaswani, and it came out of Google. This changed the entire course of AI. **This Transformer architecture started doing what we see in ChatGPT today.** When it first came out, we saw BERT of course came out of Google, and GPT-3 (175 billion parameter model) came out in 2020. But by 2022, when ChatGPT came, we were all shocked that a machine could talk like a human being, could respond to us. When we asked questions, it seemed to understand us, even our emotions, and so on. How did all this happen? It was this paper that talked about the fact that if you pay attention to the right words, it sort of understands the meaning while it's predicting the next word, and it starts making sense. And it turns out that that was the key.

**Srikanth**: [18:07] So everything changed. Of course, many other things happened beyond that, not just text token prediction, but also voice tokens. The same idea, if I break up audio into tokens and give it, it is able to predict the next token as well. That's how ChatGPT speaks back to you and streams so quickly. It happened in image generation, of course, diffusion models came in and so on. So if you look at it, there is a huge change suddenly since 2017 because of the Transformer architecture.

**Srikanth**: [18:43] With that, I just want to talk about today, **what are these LLMs capable of doing? All India Rank 1 in IIT Joint Entrance Exam.** I wrote an article about this. A lot of pushback, people didn't like it. But can you imagine? It is considered one of the hardest undergrad final exams in the world. It is right now beating the first ranker at IIT JEE.

**Srikanth**: [19:13] **International Math Olympiad. It is able to solve four out of six questions, which is almost at the gold level.** I mean, we had our neighbor, this boy Pranav, at eighth standard in Bangalore, he was writing books on algebra. He was that good at math. He went to Stanford and he did math and computer science and so on. He couldn't even make it to the International Math Olympiad as a representative of India. The standards are so high. So LLMs are today close to the gold level to win the gold of IMO.

**Srikanth**: [19:48] **In medicine, USMLE is a tough exam in the United States to become a doctor, to be licensed to practice in the US. It has passed USMLE 1, 2, and 3.**

---

**Srikanth**: [00:00] ...at 70 to 85th percentile. The bar exams, in about a year and a half ago, it was at the 10th percentile, I think, or the bottom 10 of the bar exam. Within less than one year, by the time GPT-4 came out, it's at the 90th percentile of the bar exam. So **not only is it smart, it is getting smarter very, very fast.** In many areas, it's doing better than human beings. So this is what's happening.

**Srikanth**: [00:36] I wanted to address this hype versus reality. This is reality, this is borne by fact. But of course, there is also a hype going on. We have to be able to distinguish the two and be able to look at what is the hype here. One of the ways in which you look at this hype is through the Gartner Hype Cycle. Gartner actually publishes this. This was published in 2025. They show the different innovations in AI, where they are in the hype cycle. The way you read it is, on the x-axis, you see time. This is a pattern for all emerging technologies. You can study whether it's cloud, crypto, mobile, SaaS, and so on.

**Srikanth**: [01:24] You initially start with an innovation trigger, some bright idea saying there is a new field. Then there is a huge peak of inflated expectations. A lot of cheerleading, billions of dollars in the VC industry, everybody says this is the greatest thing since sliced bread, it will solve hunger and poverty and everything else. Peak of inflated expectations. Then there is a trough of disillusionment, when it does not deliver to the promise of these expectations that have been set. Then slowly things settle down, a slope of enlightenment where people actually understand that these aspects are really good, not everything we thought would work actually worked the way we imagined, and then hopefully it settles down to a plateau of productivity.

**Srikanth**: [02:11] They've actually just put down the various aspects. If you ask me, things like Artificial General Intelligence are probably still a long time away. It might be a few years, maybe 10 years, maybe even 20 years away, where it can solve any problem, the same model. Whereas there are so many other aspects, agents and so on and so forth here, multimodal, model distillation, and so on, which have actually gone closer to stabilizing and very soon they are going to be used in industry as well.

**Srikanth**: [02:45] So just to summarize, from a hype perspective, there's certainly a lot of hype. Some of the key labs out there, their CEOs are saying GDP growth will be at 20%, 30%, and so on, which sounds very much like hype. There will be abundance in the economy and so on and so forth, which certainly seem like peak of inflated expectations. But at the same time, there are things like All India Rank 1 in IIT JEE, USMLE pass, Bar Exam pass, International Math Olympiad at Gold Level, and so on. I've just given you a few of the capabilities. There are so many more out there.

**Srikanth**: [03:29] So there is hype going on, but there is also reality in what's happening. We are already seeing it. In some of the companies where I am working closely with them, **we are seeing incredible productivity gains on programming, on our ability to plan. Today, our finance department is writing apps by themselves. They are not even going to the programming department, they are not even asking developers.** They are writing quickly programs that they want, they are taking Excel spreadsheets, connecting it to Claude Code from Anthropic, and solving problems themselves. Product managers are building products, at least prototypes, rapidly. We are seeing 2x, 3x jump in productivity, and I'm talking about direct companies where I am looking at closely. So there is a reality to this and we as NIE have to pay close attention to it.

**Srikanth**: [04:22] There's also another reason that I'll come to why it's very important, which is the impact on the labor market. We produce graduates, they get great jobs, parents and students are happy that we are able to place students, we train them well, and that is the cycle that keeps good educational institutions going. If AI is able to program so well, if we don't need junior engineers, everybody from Satya Nadella to Sundar Pichai to Jack Dorsey to so many of them, Jeff Bezos, have all been laying off tens of thousands of people, programmers. They are saying, well, we don't need junior engineers because AI can do the job of junior engineers.

**Srikanth**: [05:07] Now, **that begs the question, what happens to engineering institutions? This is not an NIE-only problem, this is not just an India problem, this is a global problem, and we all have to grapple with it, understand this, and hopefully today's discussion will enlighten us how to move forward.**

**Srikanth**: [05:25] This report came out just two days ago from Anthropic, one of the major AI labs that puts out this amazing model called Claude. The Anthropic report, based on their client usage, is showing what are the jobs that are most at risk, which sectors are most at risk. It is saying management, business and finance, computer and math, architecture and engineering. **Basically, a lot of the engineering sciences, STEM sciences, are under threat of big change.**

**Srikanth**: [06:04] Of course, if you ask me, do these jobs go away? I don't think it goes away. It transforms into different new kinds of jobs, but we have to ensure that our students are prepared for those new kinds of jobs or even create those new kinds of jobs. This just gives you a feel for legal, for instance. Whereas there are other places such as construction, agriculture, production, transportation, which are almost unaffected by AI because these are manual work. Unless maybe robots come and start doing that as well, but right now, a lot of those seem to be untouched by whatever AI models that have come out.

**Srikanth**: [06:45] Anand, you wanted to share some ideas here? Let me do that, I'll share my screen. If you have any requests, Srikanth, you may need to accept that.

**Anand**: [06:51] Let me do that. I'll share my screen. I have a request, Srikanth. You may need to accept that.

**Srikanth**: [06:56] Just give me a second. Let's see how I got through this as we go along on my screen. It shouldn't be an issue.

**Anand**: [07:03] Oh, in fact, sorry.

**Srikanth**: [07:05] No, no, that's okay. Yeah, yeah. No, no. This works. This works.

**Anand**: [07:08] Okay. Once you accept the share...

**Srikanth**: [07:14] Shall I stop share?

**Anand**: [07:16] Yes, please. I'm going to share my screen. Let me know once it's visible.

**Srikanth**: [07:25] Yeah.

**Anand**: [07:27] Okay, perfect. Actually, sorry, let me do one thing. I'll stop sharing and I'll share another screen.

**Srikanth**: [07:32] Can we pull on here? Just say yes. Perfect. Great.

**Unsure**: [07:44] [Kannada: Idu nillisalikkaagutta? Idu nillilla.] _(Translation: Can this be stopped? It hasn't stopped.)_

**Anand**: [07:54] Let me know when I'm good to resume, Srikanth.

**Srikanth**: [07:58] Yeah.

**Anand**: [08:00] All right, good. One of the things that Srikanth mentioned a short while ago is the kinds of capabilities LLMs are building. To give you a feel for that, let's talk about the state of models today. I'm going to show you a chart where the x-axis is the cost of models. This is the situation that was there as of March 2023. There were three popular models out there: Claude 1, Claude Instant 1, and GPT-3.5. Cost of Claude 1 was $8. GPT-3.5 was half a dollar to process the entire King James Bible or all the Harry Potters. So Claude 1, $8.

**Anand**: [08:50] Now, the level of intelligence is the y-axis. You can start with a level of intelligence of roughly a high school entry student, let's say class 6. These were, in March 2023, about 8th standard level. They were not even at 9th, 10th level, certainly not college, masters, and so on.

**Anand**: [09:11] Let's move forward a little bit. In as early as December 2023, we had GPT-4, a huge leap, which was a college-level capability, college-level intelligence. $10 is what it would cost to process the entire Harry Potters or the King James Bible. Let's fast forward a bit. Then in September, just 9 months later, we had a master's level student capability with o1 as a model. $15 cost. Fast forward a little bit now, in February 2025, just 5 months down the line, we now have GPT-4.5 coming in as a PhD level capability. A little bit further, in June, just 4 months later, we have Gemini 2.5, which is as smart as a tenured professor.

**Anand**: [10:04] Think about it. **In two years, it has covered what humans would have taken 10, 12 years to cover. In other words, Artificial Intelligence is growing up faster than humans are growing up in their intelligence.** Any human.

**Anand**: [10:19] But the other factor is the cost. See, if we take for instance a model like GPT-4, $10 for the level of intelligence of a college junior. You go out today and try and hire an intern from college, or let us even take somebody smarter, let's take a college graduate, or even a master's student. Let's say I hire an MBA. In September 2024, that would have cost me $15 per million tokens, which roughly means that if I had given them analysis of, let's say, a thousand books equivalent, to do all of this analysis, even then it would have cost me only $15,000. Not an unreasonable rate for that volume of work.

**Anand**: [11:08] But let's look at what happens over time. For instance, when DeepSeek-R1 was released just in Jan 2025, that $15 fell to half a dollar. 30 times cheaper. And with Gemini 2.5 Flash Review, that fell to 15 cents. 100 times cheaper for an even better intelligence. Today, it's at 10 cents. 150 times cheaper.

**Anand**: [11:35] So imagine if somebody two years ago had said, I will give you a college graduate or a master's graduate, an MBA student, and that person will cost, let us say, $15,000. Today, I come back to you and say $150. Or one lakh two years ago, 1,000 rupees now. For the same budget, I will hire a hundred of you.

**Anand**: [12:00] **The economics of this is changing dramatically. While the AI capabilities are increasing faster than humans, their cost is falling roughly 10 times a year.** Which is phenomenal.

**Anand**: [12:17] The other thing that Srikanth mentioned is the impact on the labor market. OpenAI had also come up with a survey. What they did was, they asked experts a series of questions. Then they had experts evaluate the answer. They also had AI do the same answers and evaluate, and had the experts evaluate it. Where are the world's experts better? Where is AI better? This is the answer.

**Anand**: [12:44] This diagram shows the various professions. The size of the box represents the amount of total salary across the United States. So for instance, one of the biggest boxes out here is General and Operations Managers. Every year, the US is paying General and Operations Managers almost half a trillion dollars. Or software developers, quarter of a trillion dollars is going into software developer salaries.

**Anand**: [13:16] The color represents how much better AI is or how much better humans are. Red means humans are better. Green means AI is better. So if I take software developers, AI is beating the best software developers in the world. But if I take accountants and auditors, humans are beating AI.

**Anand**: [13:37] If you look at the kinds of questions, for instance, and these are detailed questions. You are an auditor, part of an audit engagement. You have to test the accuracy of a reported anti-financial crime risk matrix. So they are given a spreadsheet which has all the details, and what they have to do is perform an audit based on this and create a new spreadsheet called sample in which all of the details are there. This is an example of a real-life task.

**Anand**: [14:00] On these real-life tasks, today, AI is better than the world's experts in some areas, not better than the world's experts in some areas. For me personally, here's how I use it. I look at this and then I ask, should I hire an AI? Should I hire a human? For instance, accountants and auditors, it said clearly humans are better. Great. I hired a tax auditor. For personal financial advice, it said AI is better. So my next investment, I went straight to ChatGPT, asked it, "Here are my risk preferences. What else do you want to know about me?" Research, tell me which fund to invest in. And it told me, "Here is the index fund that will probably have the best performance." And I know that it is better than the average, for sure, and probably even the world's experts.

**Anand**: [14:48] In that context, the way in which things are changing, and Srikanth, I'll hand it back to you in a few seconds, the way in which things are changing, particularly in terms of the impact on the labor market is so real that at least I as one person am using AI as an alternative to hire humans in areas where clearly it is both cheaper and better. With that, Srikanth, back to you. I'll come to the examples whenever you're done.

**Srikanth**: [15:15] Wonderful, thank you. By the way, all of those Anand has built himself. This is what he does. He builds these applications. In fact, if you go to his website, you will see amazing applications from understanding Indian elections to AI transformations. Okay, it was not some website that he went to, these are all things that he has himself built.

**Srikanth**: [15:38] Okay, so here I wanted to touch upon... now all this is fine. But on engineering education, what is the impact? What are we dealing with? We were talking in generality about the market, about what the capabilities are, and so on. Now we're going to get closer to NIE, closer to engineering education. My feeling is it's not as though this is the end of the world and there are not going to be jobs and our education is not as valuable, but we have to change quite rapidly because **what Anand is showing, it is growing so fast, from a sixth grader to college grad to master's to tenured professor in two years' time. Right? This is not one person, this is like the entire world being able to move at that speed. So we need to move fast to figure out how we transform our institutes to keep up with this.**

**Srikanth**: [16:38] So what can AI do? We talk about it. We are going to... we see programs. Let's see if I can just switch off my phone. That way, this will certainly not happen. Okay. You know, it can write programs, reports, documents. We have all used it to write documents already, I'm sure. How many of you have used it to write, you know, whatever, essays, documents, letters? Pretty much all of you, right? We've seen that. You know, support, customer support. A lot of companies are using it for customer support, debugging, design analysis, and so on and so forth. Of course, what is under threat is repetitive stuff, but also reasoning abilities, as we saw with new models using reinforcement learning. They are able to not only predict the next word, they are also able to take complex, let's say mathematical problems, try out a million different permutations and combinations of solutions, figure out what all did not work, and the few that actually worked. That reasoning capability has also come through. So those areas are also under threat.

**Srikanth**: [17:59] Okay. So now, what becomes valuable? So what do we create? What should our students be good at? Strong engineering fundamentals. One thing that we were all worried about when we were discussing is, wonderful, AI will come, they will just ask ChatGPT all the questions. They will stop listening to you people in class. They haven't understood anything. They think it's enough to ask ChatGPT questions, that's all they need to know. Obviously, that is not going to get them jobs, right? That is not going to make them useful citizens creating new products and services.

**Srikanth**: [18:36] **They need strong engineering fundamentals for sure. We have to ensure that we could use AI to get them those strong fundamentals.** Problem framing and systems thinking. See, when calculators came, calculators came about the time I started engineering, and there was this big debate. "Oh, these people are going to become dullards. Now we have given them calculators. They're not able to do even simple arithmetic. What are they going to understand?" Years later, we don't talk about that at all because it moved from arithmetic to these kids doing algebra, or from algebra to calculus. So we said it's okay if arithmetic is, you know, there's a machine that does that. But they sped up and they are doing something more complex.

**Srikanth**: [19:18] So if we look at it from that framing, AI could be a tool that allows our students to solve very complex problems. If today we are using them like we used to, you know, our students used to do arithmetic. If we are using them more for memorizing and coming up with answers to questions that we have asked them from memory, instead, that memorization is simply not required. AI will do all that, but understanding the concepts and solving a higher class of problems is sort of what we have to aspire for. Problem framing and systems thinking.

**Srikanth**: [19:54] Judgment and trade-off analysis. Okay, these are still not something that we can just give to an AI and say, you do yourself, decide. We can't. The kinds of discussions we had in the management committee today, AI is not capable of doing. They are incredibly complex and a lot of them are sort of human-related and so on. So some of these—teamwork, communication, ethics, and so on.

**Srikanth**: [20:17] But the other thing is, often people think, oh, this AI is a computer science related field, it's a mathematics related field. I am doing mechanical, I am doing civil, I am not affected by this. Not true. It is going to affect every branch out there, right? So all the branches need to start looking at, what is the AI intervention? How do I integrate AI into my curriculum?

**Srikanth**: [20:42] So what colleges must do: teach AI foundations, embed AI into every discipline. So we'll come to some of this detailing, I have a next slide on this. But our new graduates should not just be coders who understand the syntax of Java or Python or Julia or whatever language. Not just users of AI, but AI native engineers. Right? So what we did... Okay, Anand, do you want to share your ideas on this transformation relevant to engineering education?

**Anand**: [21:08] Yeah, I'll do that. What Srikanth shared was a perspective from the...

**Srikanth**: [21:17] Anand, we can hear you, but I cannot see you on Zoom. Okay. Now. Go ahead.

**Anand**: [21:22] Cool. What Srikanth was sharing was, how can we AI-transform engineering education, and with a strong student perspective, which is important. I teach a course at IIT Madras, and I have a bit of a faculty perspective on this, which is: how can AI make my life easier? Life is hard enough as it is. Making students' lives better is certainly valuable, but I also want to make my life easier.

**Anand**: [21:51] And for this, some of the things that we've been exploring as part of this course are... things like for instance, copying is allowed. I have no problem if people copy, but I want to know who's copying from who. So what I did was asked AI to do an analysis. Each of these little circles is one student submission. So this is student roll number 23f200, whatever. And okay, that particular codebase is no longer there, but we have the details of each of their submissions and can review that.

**Anand**: [22:27] With this submission data, what AI did was found out who are the people whose submissions are most similar. If I said give me 100% similarity, what it does is finds out who are all the people who are submitting the same code. And out here, there is a cluster of 32 students. All of these 32 have submitted exactly the same piece of code. Nothing wrong, because I've explicitly allowed copying. But here's the thing. Only 32 have... belong to this cluster. Okay, so this is one group of friends. They've just taken the assignment and copied it. Now who's copied from whom? Because I have the timing of the submissions, I know that this person in green, who has submitted the first assignment, is the original. Then there is this person in yellow, who is the second submitter, the first copier, so to speak. And then the rest have all copied from one or the other. So this is one network. Here's another medium-sized network of 10 students who copied. This is another medium-sized network of 8 students and so on. Very good.

**Anand**: [23:29] But this is only if they've copied exactly. Now AI is smart enough to tell me, okay, approximate copies. So you'll notice that there are two clusters that merge at almost a 95% level of similarity here. So what is happening here is one of these students, particularly this particular student in green who belongs to this cluster, has made a change, some non-trivial but not particularly large change from the original, and then a bunch of students have copied from him or her.

**Anand**: [23:58] So using this, it constructed the full set of large mega clusters. I put it at, let's say around a reasonable level of similarity, and these are the clusters that are forming. But here is the thing. Even at this 30% similarity, 20% similarity, etc., I was shocked to find that nearly half the batch is not copying. I've told them they can copy. Why aren't they copying? I don't know. Maybe a certain sense of "No, I must do it by myself."

**Anand**: [24:26] So I did the second analysis. Or rather, I had AI do the analysis. I just gave it all the data and said, tell me who scores the highest. I have four colors here. There are the greens, who are the originals. They have created the first answer. There are the yellows, who are the first people to copy. There are the reds, who are copying late. And there are the grays, neither are they copying nor are they allowing anyone else to copy from them.

**Anand**: [24:56] Now here is the thing. The obvious one is, the greens are probably scoring the highest. High originality, high initiative, letting others copy from them. Yes. But I think also because others are copying from them, somebody will say, "Oh, but I tried this, it is not working for me," they would have corrected it. So they're getting feedback. Good. Something that we want to value, encourage.

**Anand**: [25:19] But what was interesting was that the people who are copying first are scoring worse than the people who are copying late. The yellows, the person who's copying first from the original actually gets lower marks. The person waiting till the end and copying gets higher marks. Maybe because they have more choice to copy from. Maybe because they're able to copy from better submissions. So even in copying, there is strategy and you have to apply your brains for it.

**Anand**: [25:46] The worst, however, are the grays. Statistically significant, these are students who are neither copying nor allowing anyone else to copy, and without collaboration, there is far less learning opportunity. This was a shocker to me. But here is the thing. All of this I learned because AI was doing the analysis.

**Anand**: [26:05] I said, okay, great, then as a faculty, let me use AI more and try and understand what is happening with the students. There's another course that we are running, this is on Python. So I said, look, find out what are all the kinds of behaviors students have when they answer. For instance, when taking any exam, I can... they save their answers. And this is the Python code, you don't need to go through this, but effectively step-by-step, I can see exactly what they're typing. Every 5 seconds, every 10 seconds, and so on. So I can very granularly see the code, see what students are doing, and have AI guess, what is this person thinking when they're doing that?

**Anand**: [26:40] So it guessed that. And it said, look, here is the difficulty of your questions. You have one YouTube video engagement question. This is very tough. Only 5% are passing. And some of them are giving up. A reasonable number of... are giving up. A large number of students are not getting it well. But then you have very easy questions also, where about 90% of the students are very comfortably getting it. This is your mix of questions.

**Anand**: [27:04] But the way in which students are solving it, you can group into seven types of students. First is the wanderer. What these people do is, they build some kind of mental rule, and then that passes the easy test cases. They try out the code, it's kind of working, but then they miss something fundamental, and they systematically get wrong answers. They keep repeating, they keep repeating, effectively wandering around.

**Anand**: [27:29] Then there is the mimic. The mimic is a student who manages to solve just that kind of problem that we are asking by hard coding. They are not able to solve the deeper questions, which when we test later on and try their code against broader test cases, it is failing. Then there are the ghosts, the quitters, the confused, the crashers, a whole series of different personas.

**Anand**: [27:54] So I said, look, all of this is fine. Give me examples. Which it did. It created this website where it said, if you want to, for instance, take one of these personas, this is a wanderer. So what this program is trying to do is find out, if I give you a bunch of numbers, give me the square of the numbers in reverse order. So what this student did was said, return squares of the array in reverse order. But there is no squares function. What is the... this is a human hallucinating. Then after about 30 seconds, he tried something else. He said, I will square all the numbers and assign it to the list. But you can't assign a number to a list. Conceptual error. Then he said, okay, fine, then I will take these numbers and assign a single number to that list. But you're using a reserved keyword, that is another kind of mistake.

**Anand**: [28:43] So step-by-step, the system is giving me the walkthrough of what this particular student did, with enough detail for me to tell the teaching assistant, "Look, take this student. What he or she needs is an understanding of the basic concepts of Python."

**Anand**: [28:59] And **give me just the 10% of the teachable students. See, there are a bunch of students who are doing great. There are a bunch of students who are very hard to teach. I want to intervene in the best possible way. Who are the 10% teachable and what should I teach them?**

**Anand**: [29:14] So it did its analysis and said, look, here is a profile. There are about 6.6% of students who don't even come to your class. Low engagement. Leave them. Then there are 7.2 students who are totally chaotic, they are severely stuck, they are hopeless. Don't worry about them. There are about 57% who are doing okay. You don't need to intervene there. And there are 19% who are so varied, I don't really have any solution for them, I give up.

**Anand**: [29:39] But the 10% you should look at are: look, there is about 1.7%, they need to learn syntax. And here is the names of all of those students, the IDs. This is the specific kind of mistake they are making. There are 4.8% who are having trouble debugging. Here are these students, call them all into one session specifically for debugging. These are the kinds of debugging errors that you should tell them. They get a runtime error.

---

**Anand**: [00:00] runtime errors, wrong output, etc. Then there are 3.1% of students... here is a list of those who are struggling with the logic. Have one logic session for them. Now, this is golden information for the teaching assistants because they can straight away take exactly what this particular student knows, pull them whether they like it or not, put them into the class, and teach them what they need. **That to me is transformation of education.** Back to you, Srikanth.

**Srikanth**: [00:28] When they showed me this data about a couple of months ago, I was shocked. I always thought those really serious ones who'd say, "I will do it myself. I will learn myself. I don't need to talk to anybody else." I really thought, okay, those are probably the brightest ones. **It turns out the ones that collaborate end up actually doing better.** Very interesting. We all, I mean, this is my own cognitive bias, okay? But it seems like, so how is it that he was able to find it? With data. So we'll come to that. We as a college, if we entirely digitize our process and track our student progress like this, we will be able to actually see what is going on and truly be able to help them as opposed to the intuitions that we have. You know, Daniel Kahneman and Amos Tversky won the Nobel Prize for this on behavioral economics, and **many of the errors are because we feel our intuitions are so accurate**. And we want to jump into our intuitions and say, "that boy is really bright" or "this one is not that smart" and so on, as opposed to gathering more data before we make any of these kinds of decisions. But I just love this example when he gave it because it was completely counter-intuitive to what I thought is probably what is happening in my class. But anyway, now we've talked about all these things more as a preparation for three quick ideas that Anand and I have put together for NIE. It's not specifically for NIE, it's for any engineering college that has to go through this AI transformation, right? So the first idea, as all of you would have sort of already expected, is a new AI-integrated curriculum. Right? So, like I said, we might all assume what is AI? Largely mathematics, probability, statistics, linear algebra, so on and so forth. And some computer science ideas of complexity, algorithms, whatever, transformer architectures, so on and so forth. Even though the basis of AI is mathematics and computer science, every department is going to be affected by AI. Not just applied AI, fundamentally it's going to change too. Right? So firstly, we need to ensure that the curriculum of all departments, we move forward and start integrating AI into it. Okay. That is the first idea.

**Srikanth**: [03:00] The second idea. You know, this has been something for the last 15 years I've been sort of, not seriously studying, but observing. You know, when the MOOCs sort of products came in. edX at Harvard and MIT, OCW, OpenCourseWare, again I think maybe at MIT, we had Coursera, we had the NPTEL from IITs, and so on. Amazing content came out from around the world. Best teachers beautifully explaining concepts. I still remember Gilbert Strang's videos on linear algebra, and I was blown away. He wrote the landmark book on linear algebra and also he had this class that turned up on a YouTube video and it was fantastic, because I didn't learn, at my time we didn't have linear algebra. So we already know the content is there. Amazing teachers, beautifully curated content, very deep in engineering subjects, in STEM subjects are already there. But. The but is, MOOCs have not succeeded. **The completion level of MOOCs courses is 6.5%.** 6.5% of the students actually sit through and finish all the courses, take the tests, and get themselves certified, or even just complete everything, forget about certification. How come if we have the best teachers in the world giving the best content, it is beautifully explained, you can sit and watch it any time you want, how come people, more people are not using it? Because probably that self-motivation and discipline to sit and watch this and continue to learn is not there in the average student. Not just in India, not just in NIE, so we have to take that as a reality. That the classroom environment where students all hang out together, have fun, discuss, debate, teacher answers questions, there is a test, there is an exam, there is a little bit of competition, there is some ranking. All of these I think are necessary for the learning process. **It is not enough if we say, "Oh there is wonderful content, there is AI, ask ChatGPT, you learn yourself." I don't think it works.** NIE has provided Coursera to all the students. When I was talking to the HODs last week, and my guess was it's not like they are themselves learning with Coursera. You were all nodding and saying absolutely right. Only the brightest of the bright will start learning themselves using self-motivation, but anyway those students are going to do well. You don't have to worry, NIE doesn't have to worry about the bell curve on the right side. They will do well anyway. Right? What about the, you know, the bulk of the students in the middle? If they are not, they don't have that self-motivation, and I don't mean it in a negative way, this seems like a global phenomenon, right? How do we solve the problem? So in the companies where I, you know, work with, what has worked is we literally, the manager, so-called teacher in that context, and a bunch of young developers, they literally sit and watch videos. It could be from Coursera, it could be from edX, wherever. And then they discuss. And then he or she gives problems to them related to what they learned, but what the company needs. Okay? So we had to go through some iterations to get into a place where they actually are interested, because it is an engaging discussion, they're learning something without huge cognitive load, and at the end they are also learning how to build something, and all this they did without themselves lifting the heavy weight. Right? So can we get that combination of physical and digital? Can we get digital content, okay, you HODs, you faculty, you can figure out what is that best content for your particular course. Find that curated content, okay, and integrate it with your classroom education. Okay? I'm not talking about you have classroom education and then if you want to do digital, you tell the students you can take whatever course you want, we'll give you credits for it, as long as it is part of this list. No. I am talking about actually integrating it, where you have the best content, everybody maybe sits and watches it for 30 minutes, and then there is a Q&A, discussion, debate, projects, and so on. This I think would help us supremely. Also the reason why I bring this up is, this AI transformation at NIE is not going to be easy. Overnight, I don't think we can change our entire way of teaching and completely change our course curriculum and bring all this amazing stuff ourselves. So I feel sort of using this amazing digital content as a crutch might help us to transform fast, and also maybe provide some of the best content that is there. So that is idea number two.

**Srikanth**: [08:04] Idea number three. You know, this has been something that's been bugging me for a long time. And if you all remember, I spent the last fall at MIT, the last three months, and I took a course and I talked to you all about this course 'How to Make (Almost) Anything' by Neil Gershenfeld. And after, you know, some 35 years, maybe more, I went back to college and I actually took a course and worked incredibly hard. That course is all about making and doing things. Literally you have to make things. CNC machines, laser cutters, embedded controllers, Raspberry Pis, whatever, so on and so forth. But the realization to me was, oh my god, I understand something 10x better by doing, as opposed to just reading about it or somebody talking to me. Okay. I still watch my YouTube videos and so on, educational videos, and I do that almost on a daily basis. But nothing like trying to do something, make something. It just breaks in every way possible, and you learn so much more deeply. So here the idea was, we need to transition. And I'll tell you more than just because of my experience, I'll tell you why: **the industry values makers**. When we hire people for programming jobs, if somebody tells me, "Oh, why do you want to look at my projects, let's go to my GitHub repo, I will show you what I have built." Right? Very, very interesting to me. I would love to see what you've built. Show me how it works. "Oh, you built all this? Where did you get the idea?" "Oh I got it when I was doing engineering itself." Already you're like half the way there. Okay? So, to me, and generally in industry, **doers are given a much more value than just theoreticians**, right? And so on and so forth. So we'll come to these three ideas. So the rest of this will be about these three big ideas. Anand, over to you.

**Anand**: [10:02] I'll just share one example, showcasing how we can tie some of these concepts to this new idea. Supposing tomorrow I have to teach a class, let's say logistics, transportation, something or the other. We know there is a war going on. So I want to create something new in the curriculum using AI, that I can use in a physical classroom, and I'm going to apply this in practice. Let us try. I'm going to open ChatGPT and dictate. "I'd like you to create an interactive HTML presentation. This is intended to teach students some of the key principles in logistics. And I want you to use the current context of the US-Iran war as the backdrop to teach this. I'm not really sure what is the right set of news items to pull in to make this relevant, I'm also not sure what would be the most important and relevant concept to teach here, so I want you to think about all of these. Do your research. Figure out what are highly teachable elements in the supply chain and logistics elements, and using that, create a four-page interactive presentation. But keep in mind, we don't want too much text. We want some SVG illustrations in this, preferably even animated, to keep it lively. At the end, I want you to have three thought-provoking questions that we can use for a classroom discussion. And give me a toggle where I can switch between English and Kannada on every slide." This is the prompt. I'm just coming up with this on the fly. We can pick anything else that you want. Now I'm not going to paste this into ChatGPT. I'm going to paste this into Gemini. Why? ChatGPT has the best dictation option, Gemini has a decent slide generation option. Claude is better, but Gemini is faster and I don't want it to take too much time. After about 5 minutes when it is done, we are going to come back to this and see what it is creating. But I wanted to give you this as one out of two examples of how we can apply something like this in practice.

**Anand**: [12:18] The second example that I want to share is something that one of my college interns did. I said, look, take ChatGPT, Gemini, something or the other, feed it a textbook. Have it go page by page and find a mistake in the textbook. The specific textbook that he chose to feed was the NCERT History textbook, Class 12. And his name is Varun. On 20th Jan he generated this, and he tested only about some 15 pages or 20 pages before he ran out of credits. But it identified 45 claims in the textbook and said there is one factual error, two precision errors, and two questionable claims. Okay. What was the error? That is what was particularly interesting. It says "Only broken or useless objects would have been thrown away." This apparently is quite wrong. It turns out that during the Bronze Age in particular, and in several other ancient societies, people were taking intact objects for rituals, and offering them in the temple, or they would... okay this is somewhere else further down maybe or inside this particular research paper from Cambridge... where people because of migration, would leave things behind, and let it go. There was also a fair amount of recycling of materials as a result of which you would often find intact articles. So this whole thesis that only broken and useless objects would have been thrown away is not true. People are throwing away objects because they have to go somewhere else. People are throwing away objects because they have to give it to the god or whatever spirit they are praying to. So many reasons why we are finding these kinds of things. Factual error. So **if we can use AI to find errors in other people's work, and certainly our own work, that is a pretty useful way of applying this.** Partly from a curriculum perspective, what could we improve in the curriculum that is new, therefore no longer true. What could we do to for instance evaluate student's work to give them ideas on how things could be different. So many opportunities. Let's go back and see how far this has gotten. My guess is it would be less than halfway there... it's still working out the logistics principles. But no, it seems to have completed something. Let me see if I can put this in full-screen view. I'm going to click on share, and it's creating a link, and I'm going to copy this link, paste it in a new window, go to full-screen slide show. See how lucky we are. Okay. So it says there's a choke point vulnerability, a single point of failure can halt a global network. And the US-Iran conflict is demonstrating that. Yes, we have one particular spot where all the oil is getting choked. But this has ripple effects. So since logistics is fiercely interconnected, ships are getting rerouted through the Cape of Good Hope. It adds 3,500 miles. And this is absorbing a lot of the global vessel capacity that even trade routes that are far from the conflict there is a shortage. Interesting. So I can teach the theory of constraints using this. The next prompt that I have in mind is now go back and explain the principles of theory of constraints based on the book and start including the equations, ideas are starting to come to mind. Okay. And this shows the cost of just-in-time failure. I'm not going to go through the content of course. But let's switch to... Oh ok, for some reason it picked Tamil instead of Kannada, maybe my transliteration was wrong, but okay. So it is giving me the content in Tamil. Next is... okay I think the reason it picked Tamil is because I constantly ask it for content in Tamil, so... let's switch back to English. And it's saying "Here are some questions. If you are a supply chain manager for a global smartphone brand today, how would you redesign your network?" Fair point. It's a real question for many of our pharma clients, for instance, they are asking exactly this question, how are we going to get some of our supplies? And use this as a mechanism, partly to engage the students physically as well as digitally. To get them to practice in real life what is happening out there in the world, and for us also to practice AI on the side. So creating a new curriculum, see, four slides in what, four minutes or less? It took less time than that. 40 slides will take 40 minutes at the worst case. **The bottleneck is just our imagination.** So a fair bit of this is less work than we might think. And even if it is more work, what is the big deal, you just do a little bit of it, if it doesn't work, give up, move on to something else. Back to you Srikanth.

**Srikanth**: [17:01] Wonderful. So on the new AI-integrated curriculum, which is the key change at NIE that is being suggested. Now let's look at what the ideas are.

**Anand**: [17:18] You want to present Srikanth, that might be better? I'll stop sharing screens.

**Srikanth**: [17:22] Oh, okay. Let me get my window. Okay, wonderful.

**Anand**: [17:35] You may need to share as well.

**Srikanth**: [17:37] Oh, yeah, sorry. I will do that. Share the desktop. And you don't want to see my WhatsApp. So, there we are. So, you know, before we launch into the curriculum for the different branches of engineering, we felt there is a common AI foundation that we need to teach. Don't worry too much about this three-course and, you know, how much time it will take, how many credits, less important. We'll come to the implementation details later on. The directionality is really what we are trying to get across today, right? So AI literacy for engineers. You know, what is, you know, machine learning, LLMs, what can they do, what they can and cannot do, capabilities, failure modes. Hallucinations! Everybody talks a lot about hallucinations of LLMs. We need to really learn about these. Can we create reliable systems even though these probabilistic models are hallucinating at times? Right? That's a very important question. A lot of people ask that in industry, they keep asking, can I use these systems? Because it doesn't have 100% reliability, but yet it is so useful. So how do you... and so on. Prompt design and evaluation. Anand teaches an entire course on prompt design and use of tools with AI at IIT Madras. An incredibly interesting course. His courses are all like this. Hardly any static slides and so on, it's full of, you know, examples and literally coding examples and students are all with their laptops trying out things, right? Which is really the way I think we learn a lot better. Coding agents. I don't know, we were talking about, you know, of course, well at lunch we were talking about how coding is becoming so easy. Claude Code is just amazing. The companies that, where I am closely working with the programmers, their productivity is going through the roof. Earlier, you know, it could give you small functions, you know, it could auto, sort of tab-complete, and then it would give functions. Now you give it even complex systems and it is writing the entire code. There are a set of autonomous agents that are sort of collaborating and so on, and now there are, you know, the agent architecture that's come in and of course the Claude bot and what not that's happening that we were discussing earlier. So a lot is happening. Of course, voice and multi-modal is happening. Vision and image generation. We need to teach all of these as foundational bits because these are needed not just in computer science but in mechanical and civil and electrical and electronics and so on and so forth. Data and experimentation. In fact, Anand talked a lot about the use of data to find out who's copying and it turns out that copying is not that bad as long as they collaborate and they actually learn, right? And a lot of, you know, I'm working... you know, I'm at the apex committee of the Government of India creating AI Centers of Excellence. And the one in healthcare is in Indian Institute of Science, and we are building an oral cancer detection model that runs on the mobile phone. Where you just take photos of the lesions in the mouth. But our training data, there is a huge imbalance because the number of cancer cases is much smaller than the number of people without cancer, so there's a data imbalance. And you all know how we... that can't do well to train good models. So how do you create that balance in data? So again, AI use for, you know, synthetic data with the distribution that comes out of the reality out there in the field and so on. All of those areas of data and experimentation we need to teach. Again these are basics, fundamentals, not related to just computer science. Okay? AI engineering. There's so much going on. Oh gosh... I hope we lost them... Okay, I think... did we lose... Anand are you there?

**Anand**: [21:50] Yes.

**Srikanth**: [21:52] Okay. Alright. Good.

**Anand**: [21:53] You can't see the screen though.

**Srikanth**: [21:55] You can't see the screen. So well, you can't see the screen either?

**Anand**: [21:59] No. We can't.

**Srikanth**: [22:00] Oh. How is it that... okay. Let me just re-share the screen and...

**Anand**: [22:09] It's visible now.

**Srikanth**: [22:11] Okay. I'll just, I'll continue without the slides, maybe oh looks like it's all coming back up shortly. And AI engineering, right? Okay. Alright. This is all about actually putting together systems, AI systems running over on the cloud, integrating with OpenAI, you know, containerization, so on and so forth to get together large-scale systems to work, right? On a maybe a cloud environment. And in fact, we need to maybe offer an AI literacy for faculty and AI literacy for engineers as well. Just to sort of come up to speed at a basic level, foundational level, across that is useful for all departments. Okay? So this is an important thing before we get into the department specific stuff. Okay. Now, you know, we have put together some sample, more of a guidance on the specific curriculum for computer science and related fields. More than the specifics of what we have written, I would urge you to look at sort of the categories. One is the curriculum additions to the existing courses. Things like problem solving with AI, context engineering, using agents to automate workflows, applied ML, applied MM, LLM engineering, and so on. But also the research and project areas that can be taken up by students and faculty. The capstone templates, these are the final year projects that our students will do related to CSC and other ISC, CSC, and so on, and the key tools that are necessary. Whether it is GPT, Claude, Copilot, Codex and so on and so forth. LangChain and so on. Okay? I don't want to go into the details of this, but we came up with this categorization for each department and I met with the HODs and faculty of each of these departments in the last week. Okay? Because frankly they have a lot of experience, been teaching for many, many years, so they have a good idea of what this new curriculum has to be, where we have to integrate AI, and so on. So in fact they have edited these slides with their inputs. So I don't want to go through entire details, but you know on ENC, Electronics and Communication Engineering, there's ideas on that are being implemented currently, proposed lab...

---

**Srikanth**: [00:00] Lab enhancements, proposed theory course additions, and so on. I'll just go through it quickly and then maybe some of you, if you have remarks you want to make, please do. AI-based research activities at Electronics and Communication, what they're already doing, deep learning techniques for detection and mitigation, quantum, fault attacks, AI ML attacks, and so on.

**Srikanth**: [00:30] It looks like they have come up with a fairly detailed year-by-year applications in ECE, expected skills, AI tools that need to be introduced, and so on. This is fantastic. Sorry we won't have the time to go through it in detail today, we just want to go through the larger ideas, but thank you for updating these slides and putting in your inputs which are truly valuable because frankly, you are the ones who know the subject in detail as well as what our students need to learn in your classes.

**Srikanth**: [01:10] Electrical and Electronics Engineering. Again, there are a bunch of updates that have come in here. Time series for forecasting, optimization with ML, reliability and safety, and so on.

**Srikanth**: [01:23] Mechanical Engineering. Actually, quite a bit of detail here on ML-based manufacturing, computer vision, robotics, so many AI-related fields in mechanical. And there are two slides with lots of detailing on the proposed curriculum, AI-based research, lab projects. Mechanical and Civil, some of these core courses have been extremely strong at NIE for decades now. So obviously, a lot of inputs here on the AI curriculum.

**Srikanth**: [01:58] Civil engineering. Balaji and team gave a lot of inputs on this as well. Geospatial with ML, structural health monitoring, construction analytics, and so on.

**Srikanth**: [02:10] So a lot of the branch-specific curriculums you have taken a look at, and you have also given your inputs. Maybe it's initial inputs, maybe further you have to fine-tune and so on, but that is fine. But it was important for me to share with you the structure and see if this works, if we need to make some modifications and so on because at the end of the day, you are going to implement it. So you need to take ownership of this project. **It's better at the get-go you start getting integrated into this and give your inputs and take this further on the curriculum parts of it.**

**Srikanth**: [02:48] Now, of course, Civil engineering, there's a lot of detailing here. I am just looking at this today, I had not seen this slide earlier. It looks like Balaji and team have put in a lot of work on the curriculum additions, research, and so on. The font is really tiny, but I'm sure there's a lot of good work out there.

**Srikanth**: [03:09] Moving on. So I think we need to go through the common foundation, we need to go through the branch specifics, and then the capstone, the final year projects and research as well. Now let's get to... again, we're still in the from three big ideas, this is just the first idea. The whole curriculum, pedagogy, learning part.

**Srikanth**: [03:31] Next is research. **As we know, NIE's NIRF ranking has to improve considerably. We are still at the 150 to 200 range, and once we get below 100, there are lots of benefits.** In fact, there are huge programs of the Government of India we're not even able to take NIE to those levels because of our NIRF rankings. And Professor Nagaraj and myself have been sitting with a lot of you for the last couple of years now, and it seems that improvement in the research publications will give us a huge boost when we look at the formulas of NIRF and so on.

**Srikanth**: [04:13] So the whole question is AI-assisted research. How can it help us improve the research? Today we are barely hitting one paper per faculty per year. **I have seen productivity levels in our companies which are 100, 200, 300% improvement, and I can't see why engineering colleges cannot do the same.** In fact, I am sure every college is thinking the same. So this is going to help us. This is not about running faster, but really running smarter.

**Srikanth**: [04:44] So there are many areas where AI can improve research. Faster literature review and problem discovery. AI can scan, summarize, and organize large volumes of papers quickly. If you are focused on some area like cybersecurity, do the literature review much faster. Tell it to summarize. Don't even type, like Anand says, just talk to it to give you back quick answers.

**Srikanth**: [05:09] Acceleration of data analysis, coding, and experimentation. Coding is so good now with Claude 4 and other models. Data analysis, Anand showed us many examples of data analysis. You can create all that really rapidly. Improved quality through reproducibility and verification. Have it create quick prototypes. See if it works. So you can reproduce your results.

**Srikanth**: [05:33] Creation of new domain research opportunities. So this is one area I think we need to improve. So some of the ideas I had was one is, if we are doing AI courses, AI-infused teaching in the class, the course itself can inspire research. Students can become interested in certain things that they have learned and work with the faculty. Our faculties are working in specialty areas. Of course, you will be inspired and you can use these AI methods to quickly come up with really interesting topics.

**Srikanth**: [06:06] Final year research projects. Students are spending a lot of time. They have all the capabilities of the smarts, the thousand times human brain smarts that we are seeing, they have those tools with them as well. So their final year projects can definitely turn into several papers.

**Srikanth**: [06:26] Industry project research. This is one thing. In the US, I used to work in the Silicon Valley, we used to work incredibly closely. I was in chip design. While we were developing the SPARC chips at Sun, we were working with both the professors at Berkeley and Stanford who came up with the SPARC architecture. That needs to improve because the practicality of those ideas will be much higher coming from industry, and imagine we can do deep research, which many times industry doesn't have the time to do. They might find it hugely beneficial to have this collaboration.

**Srikanth**: [07:02] So these are some ideas that I think... I'm going to go a little faster.

**Anand**: [07:07] Just a time check, how are we doing for time, Srikanth? Okay, for the...

**Srikanth**: [07:11] Yeah, we have about 30 minutes, so do you have something to share here, Anand?

**Anand**: [07:15] Yes, let me do that. Maybe you could stop sharing your screen and I'll end with this. Let me share my screen. Tell me when it should be visible. Now.

**Anand**: [07:32] Okay, so I thought I'll show you an example of how this works. Meaning, how can AI actually be used in the kind of research that we are talking about here? So let's take this idea: can I discover problems or can I discover new research opportunities with AI?

**Anand**: [07:54] I'll tell you what I did while Srikanth was sharing this slide. I have a little tool called Ideator. I take notes, and as part of my notes, I categorize them into a variety of buckets: LLM related, things I learned, blah blah. And I randomly can browse through any of my notes. And I have little buttons here where I can ask the tool for what I want. Maybe a web app, business idea, new research idea, I don't even know what all I clicked here. And then I can click on one of these AIs. What it will do is take the two ideas that I have suggested, randomly even, try and combine it into a research idea.

**Anand**: [08:35] So I did that with two random ideas a short while ago. First was Magnus Effect. That is basically how the cricket ball spins. The spinning ball actually moves due to air drag. That's something that I learned in 2023. Another idea that I learned again in 2023 was—not idea, but something that I noted for myself—is to start prayers. I started two years ago. And it has a lot of power but it's something that I am blind to.

**Anand**: [09:04] So I told Claude, new research idea from these two concepts. And that is exactly what this prompt is. The prompt says, "You are a radical concept synthesizer hired to astound even experts. Generate some big useful non-obvious idea." This is just me generally giving it some motivation. But the ideas here, Magnus Effect and explore prayer.

**Anand**: [09:25] I've also given it some ideas based on principles of research on how to come up with new ideas and what format to put it in. So it's thought about a whole bunch of ideas and it's given them scores. The one that scores highest is prayer as a physiological oscillation or ANS deflection.

**Anand**: [09:47] What it's saying is the spinning ball doesn't curve because of magic. The rotation creates an asymmetric pressure in the medium around it. Maybe prayer works the same way. Not metaphysics, but because the repetitive rhythmic prayer is actually a psychological spin mechanism that creates a kind of pressure in the body.

**Anand**: [10:09] So what can you do? It's saying the spin is basically the rhythm or the repetition. So for instance, the Catholic rosary, you breathe about six breaths a minute. The Sufi dhikr chanting is about 12 cycles a minute. Buddhist mala varies, etc. Now you measure the autonomic nervous system impact. You can measure, for instance, the heart rate variability, cortisol suppression, inflammatory cytokine reduction. Great!

**Anand**: [10:33] But here is where I have to tone myself down and ask it, "Boss, find all the previous research related to this. I don't want to sit and redo what others have done." And it did. It says the idea is real, but it is half discovered, only half.

**Anand**: [10:48] The core mechanism, that is rhythmic prayer leads to autonomous nervous system deflection via respiratory resonance is not just plausible, it is proven. And there seems to be a peak for this. That is, six breaths per minute has been established as the best resonance peak.

**Anand**: [11:06] But what's not been done is nobody has mapped this frequency across prayer traditions. Is Islamic dhikr, the Jewish davening, the Buddhist nembutsu, the Sufi chanting—do they cluster away from this resonance point or do they cluster towards it? And it goes on and finally gives me a series of specific areas. It says look, don't try all of this. This kind of research people have done. This is partly researched, maybe you think about it.

**Anand**: [11:35] But here is stuff nobody's done: Cross-tradition rhythmic frequency. Second, rhythm-controlled prayer frequency. So nobody has taken a single prayer and had subjects recite it at four beats per minute, six breaths per minute, eight breaths per minute, etc. Does it make a difference? Nobody's tested whether it makes a difference whether you believe versus you're agnostic. Take a believer, take a non-believer, give them the same thing, and does that make a difference? And so on. How does it align if you take the person from the same religion who has practiced the same prayers for a long time and measure this?

**Anand**: [12:07] And it gives me a full-fledged research proposal: week one, week three, etc. **This happened in about five minutes with me not even doing any work in this area. It is doing all the work. Now I'll give it to a bunch of PhD students, master students, undergraduates even, and say do your research, here is idea number one. Like this, I have 50 ideas and I'll happily co-author and put my name on it, I have 50 papers.**

**Anand**: [12:33] The other part is the review. Stanford organized a conference. This entire conference was allowing people who had used AI to write papers to submit. And they had AI review it, partly to see how well AI can review papers. Here are the findings, and again this presentation is AI-generated, I just had it read the paper and get the results.

**Anand**: [12:56] So what they did was they took about 315 research papers, and three AI systems reviewed every paper, plus humans reviewed about a quarter of the papers and compared it. So what we are finding is that the AI systems are wildly disagreeing with each other. For one particular paper, one AI reviewer said this is fundamentally flawed. Another one said outstanding quality. Same paper, two different reviewers. And another one says has merit. Okay.

**Anand**: [13:30] And averaging it doesn't help. The variation is very high. So for instance, Reviewer 2 happens to use very big numbers and a very wide range. So if you start taking Reviewer 2's rating into it, that drowns out everything else. You can't simply average either. Worse yet, all the AIs were claiming that they are 100% confident. How can all of them be 100% confident? I mean, that makes no sense.

**Anand**: [13:54] **But what is interesting is, however, that what the humans and AI were reviewing seemed to be very different things. And in particular, the AI seemed to be particularly good at catching some obvious problems.** Fact-checking, basic mistakes. You have cited a paper that does not even exist. The humans all missed that. AI went systematically, checked every one of those and said in all of these cases you are citing papers that don't even exist. Here are arithmetical mistakes. That sort of a thing it is fantastic at. This is as of about eight, nine months ago.

**Anand**: [14:26] So what they were saying was it's not about AI replacing the human reviewer, but acting as a fantastic complement to a human reviewer, which means that the quality of review also goes up. Which means that naturally, when we start doing research, that improved quality through verification certainly is what I just showed, and obviously also reproducibility because you can have it write the code starts jumping up.

**Anand**: [14:49] So these are examples of how AI actually is starting to get used in practice for research. Back to you, Srikanth.

**Srikanth**: [14:59] Okay. I will wrap up shortly. And in fact, let's...

**Srikanth**: [15:13] So I'm not going to go into too much detail. This whole issue of, is allowing students to use AI, is that cheating? Will they learn? What should we do about this transformation? By the way, when I was at MIT last fall, there were lots of these debates going on as well. So everybody is just asking these questions.

**Srikanth**: [15:35] There seems to be at least some broad directionality right now, in saying that you know, the proper use can be literature review, hypothesis generation, code writing, data visualization, and so on and so forth. A lot of the examples that Anand has given, for instance. Whereas the misuse is if you are submitting work and claiming it to be your own while it is AI-generated.

**Srikanth**: [15:58] So in fact, even the classes I was taking at MIT, besides what I was teaching, they would say, please feel free to use whatever model you want to use, but put down your prompt. What prompt you used, right? And so of course, Anand has built tools that can figure out who copied, when did they copy, how deeply did they copy, and so on. So of course there are detection tools.

**Srikanth**: [16:21] **But I'd like us to think of this more as an enabler for students, not we as cops have to catch them in the process of using AI.** Because at the end of the day, if they are able to solve, instead of just arithmetic they are doing algebra, instead of algebra they are doing calculus, well, if they solve, you know, if they cure cancer, it's okay if they use AI. Curing cancer is much more important than hey, you cheated and you actually used a calculator, you know?

**Srikanth**: [16:51] So in some sense, I think we have to reorient ourselves. And it's not easy, I'm saying it much easier, you are the people who have to implement it. But I think that general direction of misuse, let's put out an opinion as to how we would like to treat AI use at NIE. And a lot of people are thinking about it, so it's not as though we are lost in the woods. But here is a framework, I don't want to go into details, but I just wanted to address this. So there are various... there's more ideas around this, I'll skip this for now.

**Srikanth**: [17:28] On infrastructure, very quickly. Here's what I felt. Today, Uday took us around the amazing infrastructure projects that are happening, the building A, building B, fantastic 32 classrooms, state of the art. Loved it. The girls' hostel and so on. So one thing I feel is we need to completely transform all our classrooms into smart classrooms with of course projection, recording, broadband connectivity everywhere. Every classroom should have it, okay?

**Srikanth**: [18:02] And people should not struggle. When the professor and the faculty come in and plug in, projection, all of this should be easy. They can't spend 15 minutes trying to work around these things. So it's important that we build out, and I was pleasantly surprised that we are planning and doing the wiring and LAN networking and for projection and all of those good things, but let's make sure that it works. This is not a side edge problem anymore, this is central to us. We have to get completely digital.

**Srikanth**: [18:31] Campus-wide Wi-Fi digital learning backbone. Everywhere, we need students to be able to sit, wherever, under the tree, in the library, in the classrooms, in those beautiful meeting rooms and so on that you showed us, they should be able to connect to the Wi-Fi, be able to do things. **Everything that I am doing today, most of the useful things I'm doing today are with LLMs. Okay, day in and day out.** This, all those diagrams you saw on a banana, all those beautiful pictures in that sketch format for this slide deck came from Gemini's Nano Banana. Most of our work is using LLMs today. Of course, we have to give it the ideas, we have to come up with what we want to create. But campus-wide Wi-Fi network broadband is super important.

**Srikanth**: [19:19] Digital exams and assessment systems. We should completely move into digital. Okay, it's not about let them write on pen and then I will scan it and I will OCR it and all. No, entirely on laptops. All students, I'm sure NIE students are already on laptops and so on. If we go completely digital, you can do the kinds of things that Anand is showing us. Okay, in fact, there's a whole bunch of things that's going to make your lives easier. Even grading is going to become much easier for you if we are digital because LLMs can help you, right? I think Anand alluded to that.

**Srikanth**: [19:54] Now on the GPU infrastructure. As you know, these graphics processing units are the backbone of AI. Basically matrix processors. And we need our students to really understand it. It's not enough if they use LLMs and ChatGPT and get some answers, they need to become the future creators of LLMs. They need to understand GPUs. They need to understand how to train LLMs or fine-tune them and so on and so forth. So I would actually say let's work on creating a GPU cluster. We don't have to spend a heck of a lot of money, but at least we need our faculty and our students to start getting exposed to that.

**Srikanth**: [20:35] Institutional LLMs, Copilots, software subscriptions, super important. I have every subscription. I pay for it personally, whether it's Claude or OpenAI or Gemini, because the return on investment for me is like 100x. Maybe more, right? And that will be the same for the college as well in financial terms. Okay, not just in learning ability, in financial terms it's going to affect us. So I would say let's be generous and ensure that our students and our faculty has full access to these models. They are giving up productivity gains that are so large that they will show us how we can be a much more effective organization.

**Srikanth**: [21:16] Edge AI, robotics, sensor labs for applied work. Like we said, doing, making, maker labs, the Alpha Lab that we've created, all of those. Amjad is here, he's working on the entrepreneurship center to do startups and so on. Super important, we have to invest in those, right? AI-enabled LMS, lecture capture, and so on and so forth. As you can see, we need to invest in all this, and we have invested quite a bit on infrastructure. Now it's time for some of our AI infrastructure to kick in as well.

**Srikanth**: [21:48] But I'll tell you one thing, HODs, Deans, all of you are here. Today's management committee meeting, all of them are supportive of this idea. Okay, in fact they were asking me, tell me how much is it going to cost, we will provision this money. I didn't have the answer. I have not worked out the details. We will work out the details in about the coming weeks. I'll work with all of you, but the management committee is entirely supportive for this transformation, so it's not going to be somehow we scrounge and make this happen, okay? So that's on the infrastructure part.

**Srikanth**: [22:21] Phygital, physical plus digital, I was just going to say, you know, this is all about how we deliver the content. I don't want to belabor this point, it's basically using digital content in your curriculum and ensuring that you can deliver all these AI courses. Okay, so I will skip through.

**Srikanth**: [22:45] But I want to... so lastly, let me just say on the _doing_ part. I just wanted to say, course projects, I would like at least our new AI courses to be project-oriented. Do things like Anand is showing. Your students are going to learn a heck of a lot more, right? They will remember all of these concepts because they have actually built something, right?

**Srikanth**: [23:09] Final year project, let's not only get the final year project with AI assistance done, but also papers that come with it.

**Srikanth**: [23:16] Internships. Our new principal has written a book on internships. His experience in BITS Pilani has been fantastic about how to make internships really useful, not just for our students, but also for industry. It has to be symbiotic, otherwise internships don't work. We can't keep begging and saying please take our students. They should also say I want NIE students because this is the cutting-edge AI center of excellence in Karnataka, in India, so I want your students.

**Srikanth**: [23:42] Internships, industry projects. Work with industry. By the way, there's so many industries and companies and government projects I work with who would love to farm out a lot of the research. Right? But we need to have the capability to pick it up and actually do useful work. And of course startups.

**Srikanth**: [23:59] All of these are examples of doing, and I'm saying we need to **transform NIE to a doing, maker place which will really enable our students to get the jobs in the future**, given that the theoretical understanding is no more sufficient.

**Srikanth**: [24:16] I will stop here, but I wanted to just ask Mr. Murthy, who's here. Sir, you have enormous experience and you've just come back from a conference. Just give us some... what are your thoughts? What is the direction you would like us to move forward in?

**Anand**: [24:38] You're on mute.

**Mr. Murthy**: [24:41] Sorry. Yeah. Srikanth and Anand, this is one of the finest lectures I have ever heard on any platform. And much more so, the finest lecture I have heard in...

---

**NR Narayana Murthy (NRN)**: [00:00] Almost 60 years at NIE. The last lecture that I heard that was similar to this was by my teacher, Dr. N Krishnamurthy, who spoke on analytical thinking in structural analysis. He took three known methods. I am not a civil engineer, I am an electrical engineer, but even then he was such a wonderful teacher, we attended it. He took slope deflection method, moment distribution method, and column analogy method. These are the three well-known ways of analyzing structures. So after that, **this is the finest lecture I have heard.**

**NR Narayana Murthy (NRN)**: [00:49] What was very impressive about this lecture is the fact that **you focused on the need to teach our students problem identification and problem solving.** Systems thinking. You said these are the things that will remain with human beings. In 1975, I defined a term called learnability. **Learnability to me is the ability to extract generic inferences from specific instances and use them in a structured manner to solve new problems.** That was the basis for the training program at Infosys. I don't know what they do now, but as long as I was there, that was the focus.

**NR Narayana Murthy (NRN)**: [01:50] What you people have gently encouraged NIE people is to focus on learnability. That is, the challenge is really on how do I use a certain body of structured knowledge that I have, and see what combination of that will help me make progress in an unsolved problem. I think that is what you people have helped our wonderful teachers at NIE. I think we should all be very grateful. Srikanth, I know you are an alumnus of NIE. I don't know about Anand. Anand, are you also connected with NIE?

**Anand**: [02:50] No, no.

**Srikanth**: [02:53] He studied at IIT Madras and I think IIM Bangalore if I'm not mistaken.

**Anand**: [02:58] That's right, yes.

**NR Narayana Murthy (NRN)**: [02:59] Wonderful, wonderful. Both of you did a brilliant job. I think the question really we have is, there are all these bags of tricks. But as I pointed out in my definition of learnability, how do you use this bag of tricks, in what combination do you use this bag of tricks to take a stab at a new problem? That to me is the biggest issue that we have. As you know, we derived all our revenue from defining new problems, solving new problems in information systems for our customers. So I think to cover so much of what you did in this lecture is just extraordinary. And number two, some of the examples that Anand took were mind-blowing. **And if it had been done by McKinsey or somebody, it would have cost an arm and a leg.** So I think we should all be very, very grateful.

**NR Narayana Murthy (NRN)**: [04:25] But the key question that our professors will have to think is, what is analytical thinking? What is critical thinking? How do we bring analytical and critical thinking from a known bag of structured tricks to take a stab at an unsolved new problem? I think both of you are just extraordinary in your ability to communicate, in your ability to take examples. So if someday, if you can—there is no hurry—but if you can help NIE, and even all of us, on how do you approach solving a new problem from whatever has been taught.

**NR Narayana Murthy (NRN)**: [05:28] There were wonderful things. You talked about LLMs, you talked about Python, you talked about various tricks that we have to learn in AI. But still, that question is, there is this unknown problem. Are there any methods? Are there any indicators? Are there any directions in which we can choose a set of these tricks and then solve the problem? Because if our youngsters can be taught by our teachers on doing that, I think that is... But unfortunately, that is not the problem of engineering colleges alone. It should have started from primary school.

**NR Narayana Murthy (NRN)**: [06:20] For example, first learn physical things around us. What is the meaning of saying sunrise rises in the east? What is the meaning in saying that when the motor car goes away from me, the intensity of sound reduces? Why do we see a rainbow? All these should be explained using formal theory of physics to the children of first year, second year, etc. Then I think they would be automatically trained in the kind of thinking that you people have mastered and that we are all trying to gain. So that's what I would say. I don't know, in some way if you people can help, I think in some way you are our best hope. Even though late it is, because by the time students come to engineering college it's too late. But in some way, if we can have two or three or four courses as part of the engineering thing on analytical thinking and critical thinking in problem solving, I think it will be very, very useful.

**NR Narayana Murthy (NRN)**: [07:44] I am deliberately requesting you people to go beyond just AI because AI is another technology. It does help, there is no doubt about it. It has a lot more impact than most technologies that you and I have seen. But **the key message there is very clear: as long as we use these technologies in an assistive manner and we remain the masters, this is a safe world.** As long as we create guardrails, then it is a safe world. But that big question of how do I relate what I know today to make an attempt at solving an unknown problem... that to me is the huge challenge. And I think our hope is that extraordinary people like you two should help us, and it will be very, very useful. Thank you very much for this opportunity. **The best lecture that I have sat through in the last so many years.**

**Srikanth**: [08:56] Thank you, sir. That is very generous of you. And we will certainly look into what you said. I think it's a very thought-provoking idea of using critical thinking and analytical thinking to solve new problems. Is there general methods? Is there methodologies to teach our students? That's really thought-provoking. I'm going to throw it open to questions. I know this was a long lecture. Sorry for that. We tried to cover a whole bunch of things. Any questions you have, comments, thoughts?

**NR Narayana Murthy (NRN)**: [09:36] Srikanth, sorry to just butt in again. In my definition of learnability, I said learnability is the ability to extract generic inferences out of specific instances, whether it is slope deflection method, whether it is power distribution, it doesn't matter. But our students must be taught generic inferences and show them how it can be used in many situations. That I think is perhaps the first step.

**Srikanth**: [10:12] Absolutely, absolutely. Thank you, thank you so much. Any questions, thoughts, comments? You want to tell us a little bit about the syllabus that you thought about for your different departments? Srikanth, did we do new justice on AI the question that you had asked in our preview? Can you just give the mic to him?

**Audience Member**: [10:41] Yes. When Anand was presenting, Anand spoke about the stages of AI or the evolution. First generation, generative AI, which you called broadly now. The second stage was reasoning AI, where AI systems would become somewhat mature, they would be able to design things from whatever data. And the third, of course, is getting to the rest of the world. In that context, can you explain what these AI stages mean? How long do you think it will move to the second and third generation?

**Srikanth**: [11:36] Well, if you look at the large milestones that I see, one of course is the simple transformer model that was able to predict next tokens. But suddenly with the help of making it a human assistant for answering questions, it was super useful. That was the first step. And GPT-1 came out, and we could ask questions, ask it to write letters, ask it to write an essay, write a poem, whatever. And we have all seen that, we have all tried that.

**Srikanth**: [12:04] I think the next big leap was reasoning. When using reinforcement learning, they actually got the models to reason and think deeper on deeper problems, as opposed to just answering one question zero-shot. "I ask a question, you answer." No. You ask a question, I try many answers, I try many methods. I see that some of them are not good. I myself check whether the answer is good. Eventually, I figure out, "You know what, I've actually solved your math problem. I've proved the something that you've asked," through many trials, like a human being does. That is a huge step, reasoning. And reinforcement learning brought that. It's not the most efficient way to do it, but it works. Human beings are a lot more efficient in reasoning, but maybe we also have the big edge of evolutionary knowledge that is in our DNA, gathered over eons, whereas these transformers have to learn from scratch. But that was another big step.

**Srikanth**: [13:06] Now the third step I'm seeing is these autonomous agents that are collaborating. All these agents are the same LLM, but they are wearing different hats. I am a programmer LLM. I am a QA tester LLM. I am a verification engineer LLM. I am a release engineer LLM. And all these engineers and LLMs are talking to each other and collaborating and building amazing things.

**Srikanth**: [13:35] And the next thing that we're seeing is computer use, tool use in general. The ability to answer your programs by not just the transformer giving next tokens or even reasoning, but using tools like Python programming, the internet full of information, other APIs through MCP and other methods of integration. All of this is bringing enormous capability to these agents, way beyond what they themselves can do by either token generation or reasoning. They are able to do much larger tasks.

**Srikanth**: [14:13] So this sort of is the progression. Of course, the holy grail everybody talks about is AGI, artificial general intelligence, which is the same model being able to solve pretty much every problem, like human beings are. We are not just able to answer questions or write poetry or do math or whatever. We are able to do all of the above, right? So can it get to AGI? I think we are still far from it, but people talk about it could be anywhere between two years to 20 years or whatever, that AGI will come.

**Srikanth**: [14:40] But already, the way I look at it is, we don't have to wait for this holy grail. **Already the overhang of capabilities that we are seeing today in LLMs has not been completely used yet in the form of applications.** Whether it is NIE transforming itself into an AI university and getting its students to do all these amazing things, or companies like Infosys using AI to build software very fast for its clients, or detecting breast cancer and oral cancers and so on, which we are doing at the Indian Institute of Science for this AI center of excellence. All of those, I still feel that we haven't even scratched the surface of the applications of the existing capability that LLMs already have. So it's very unusual that we are in a situation where the overhang of capabilities is so large right now through LLMs, and they are underutilized in my mind. There is so much we can do in so many different areas. So I see already we can move ahead. We don't have to wait for AGI and super intelligence and so on and so forth. So that's my thinking along those lines. Yes, Ramnath.

**Ramnath (Audience)**: [15:58] Thank you very much for an inspiring lecture. I think partly reality, but I just want to tell you that the students especially, and the faculty should not just get carried away by saying, yes, I should use AI for everything. Then they will stop thinking and they will become dummies [inaudible/slaves]. So I feel like Narayana Murthy said, **you should be a master and not a slave.** So I think the design is going very much to an AI tool. Just my comment. Thank you.

**Srikanth**: [16:34] Absolutely. See, that thing about, it is actually a thin line. We as educators have to figure this out. It is not an obvious answer. If AI can do everything, of course our students can also become lazy and say, "Why do I have to even learn all these things? Anyway AI is telling me so quickly. What is the need?" Right? So between that and actually teaching the fundamentals and making them 100x more productive using AI, that's the balance that we have to figure out. I don't think the deck shows you how to do that. But as we all practice, we'll figure out how to use this better. Yeah. Amjad, you had a...

**Amjad (Audience)**: [17:16] Thank you very much for a very inspiring lecture. I have two basic questions. One is about our mental models. Mr. Narayana Murthy was making a lot of profound statements. What is taking place is, societal expectation is changing. And our identity itself, that students are not going to college for learning, but for socialization. And over a period of time, people won't want to work. Employment will be totally redundant. Everything will be available in abundance because of amazing AI. That's number one.

**Amjad (Audience)**: [17:55] Second question is about introducing AI in first year. The integration of AI in first year. As I see in our time, giving that in the first semester itself, it's like asking a child to learn tables giving them a calculator. Giving the calculator first. So naturally people will take the easy path. How do you see that it impacts?

**Srikanth**: [18:25] Okay. Well, on the first question, it's anybody's guess how it's going to roll. As in, it's very hard to guess whether abundance is going to happen, suddenly the GDP jumps, or suddenly we'll all be jobless. It's very hard to figure out because it's almost like the laws of physics are suddenly being changed on us. So it's hard to predict what the economy, how this is going to impact really. So it's hard to take those things seriously that suddenly we're going to have enormous abundance and so on. Let's see. I mean, hopefully it will be a great day, there'll be great abundance and no poverty and so on.

**Srikanth**: [19:03] Leaving aside that, with respect to starting at the first semester itself on AI, and Anand, do weigh in on this. My feeling is it's not too early. In fact, our kids are going to start doing AI in school, and they will be AI-native by the time they come here. If we are still the old guard holding on to traditional methods, they are not going to be pleased about it. So I actually think that first year starting AI courses is too early is far from it. Anand, you want to weigh in on that?

**Anand**: [19:39] Yeah, it's like learning how to use the internet, I think. How soon should we teach people how to use the internet? The earlier the better, it is going to be the norm. Using the internet has... Google searches have become like reading and writing.

---

**Anand**: [00:00] And there was a time not too long ago when Google did not exist, the internet did not exist, and that's the background that we come from. So for a technology that in our lifetime has become that popular, seeing that, where we are at right now, I think means two things. Whoever gets there early is going to have an edge. When they get into work, they are going to be competing with others, some of whom have early exposure. Long exposure leads to more experience, more experience leads to more building of capabilities on just how to use AI. And if we want our students to get that exposure too, it is going to be important. I see a lot of risks in the use of AI. But right now it's a new technology. More often, I'd say 95% of the people are underusing AI. There is probably a 5% that is overusing it and that has a different kind of risk. But I would ask myself, see if AI is going to be growing, the use of AI is going to be growing, all of the competition for the students are going to be using AI more and more, then maybe **the bigger risk is underuse of AI.**

**Srikanth**: [01:23] Yeah, actually on the lines that he just mentioned, **I am actually quite worried that the divide between the haves and have-nots will increase with AI, not decrease.** Because the knowledge asymmetry will worsen, okay. But one of the things India has done really well is digital public infrastructure. We have ensured that our infrastructure works for the poorest Indian, whether it is UPI or Aadhaar or DigiLocker or whatever. And in fact, one of the things I spoke at the United Nations about this is, if we use the same playbook, the digital public infrastructure playbook, and ensure that our benefits of AI and AI interventions, AI apps reaches the poorest Indian, I think we are probably the best if not qualified, the best experienced I would say, in diffusing this. Because our experience in Aadhaar, UPI and so on, the whole world marvels at the fact that it works even for the smallest banana vendor or chaiwala [tea seller] is using UPI. Same also with Aadhaar and DigiLocker and so many innovations. So this is the reason why I joined the apex committee for the Government of India, when they asked me to join this AI Centers of Excellence committee, Ministry of Education, to ensure that all these innovations actually go to the smallest village. Okay, it's not just a few people who can benefit from all these amazing things that are happening, but there is a strong danger, there is a very eminent danger that the divides will increase with AI, not decrease. Because it is so mathematical, so involved, so intellectual in some sense, it has to be brought down and simplified on a mobile phone, people talking, asking questions and, you know, pretty much getting the benefit. Any other questions, thoughts?

**Unsure** (Audience): [03:24] Good point sir, because any problem statement we frame, the problem statement are different domains. And start giving that problem statement from the L1 or L2 to the students. And group them, classify them based on their ability. Because they cannot apply this to the complete audience. They cannot take all of, say, all five parameters at the same time. So we have to reclassify them, and then allot the projects or the statements so that they can start applying slowly from the lower level tools to a higher level of tools. Because normally we do research in corporate base, you know... on the side. Try to teach them to the students, especially that lack technically they can level up. So that may happen many times. This is what my observation as a teacher... whatever you've noted that. But definitely we have to mold that practically on the different segments. Otherwise I don't feel I'm able to deliver to all of them. So there is still a connection needed on this, and absolutely as an individual we have to fill that gap.

**Srikanth**: [04:45] That's a very interesting observation. One thing, you know, I mean one of the things I'm working on is AI for education with IIT Madras. And there's a famous Bloom study that says that instead of one teacher teaching 30 students or 50 students or whatever, and every student is going at a different level, what pace should the teacher teach? Do you want to teach the smartest kids because they are very quick at picking up? Or the slowest ones or in the middle and so on. **And that study shows that if there is one-on-one tutoring of students, there is an enormous jump in learning levels. Two standard deviations jump in learning levels.** This is the Bloom study. And so the obvious thing is we were like okay, maybe AI can be that teacher, right? So what I'm saying is we could have... we as NIE could build assessment tools for the entire class. They take their test on their smartphones, on their phones or laptops or whatever, and like Anand showed, we can classify them as to what they have learned, what they have not. In fact he showed very good examples of what they have actually learned, what concepts they didn't learn, and the LLM is doing all that hard work for you. Right, so that's on the assessment part. Now, even on the active learning part... sorry adaptive learning part, that is oh, that particular student has not understood this concept, and she has not understood that particular concept and so on, we can use LLMs to fill those gaps as well. So which is the adaptive learning part. So in some sense, I think... and these are good projects for our students and faculty to take up. Okay, I was at MIT largely to do some research in this area on assessments and adaptive learning. Of course, for FLN, for foundational literacy and numeracy, to teach young kids language, that's all. Because we are doing terribly as a country on that. But it is useful in engineering education also, because you have the same problem, you have a large class, what pace should you go at, right? Anything, Anand, you want to share on this?

**Anand**: [06:44] No, I think that's great.

**Srikanth**: [06:47] Okay, Dr. Nagendra, please, our Principal.

**Dr. Nagendra** (Principal): [06:50] It was a very good session. Thank you... Listening to you, I feel that here the key consideration entirely is integrating AI into the engineering curriculum. How is that... what kind of... how can you restrict the structural map for that structure... that framework what you are giving... One area... entirely I am not supposed to dictate to a team. So if we have a proper outline, the first phase what we have deliberated is largely on what you have given. And then as an institution we need to draw that institutional strategic intent to take up AI. We need to draw out the explicit structural map... And another thing is that you know you also mentioned about basics to strictly to introduce you to right? Now, the basis what I see is that, instead of superimposing the AI curriculum... right... so practically that gives us our answers. So what I feel... myself thinking deep over this... see, instead of AI core branch what we are opting for... see, from the existing... start AI-ifying the existing branch structures, the existing lines what we are operating now. Okay. We as an institution must be knowing about the outcome-based assessments what we're working on... structural components to outcome base. So along these bases... then, what structural changes of the existing metric... so structural changes, that is what the curriculum what we are adding, how we can use derived tools, these software... into this. That will directly fit into standard core subjects. Say for example... directly you apply to heat transfer or fluid mechanics. Design essentially a process maybe capstone project, minor projects, capstone projects, programming lab, data analysis. So these are some areas where we can bring AI tools... how we can blend... in that way we can bring...

**Srikanth**: [08:44] Absolutely. No, no, it's a great idea. **You're saying let's bring AI tools for existing curriculum first.** Then maybe let's look at AI curriculum and so on. See, that is the reason why I'm saying eventually, you are the educators, you've been teaching. So I think you will have really implementable ideas. Concrete ideas, like you are bringing up. So I think, a very valid point. So that implementation detail I wanted to stay away from that because frankly it should not come from me or well Anand at least has been teaching, I've not even... I've been doing that for only one semester and that too sort of guest lectures which doesn't even count. So I think your points are very valid. And I think the faculty should take a call on how this implementation is done. Okay, it should not be top down, because you know frankly I'm not an expert. Maybe Anand can speak about his experience, but certainly, one thing I just wanted to show, I put down some thoughts on the next steps. So one is, **come up with an implementation plan with timelines and owners.** Okay, like what you are mentioning. Okay let's come up with a strategy, so that we don't disturb the existing batch that is going out, but let's teach them how to use AI tools of the same curriculum and so on, let's move forward. **HODs work on branch-specific AI curriculums** and maybe overall the Dean Academics can oversee all that work, you know, and ensure that they are all working on the same template and so on, like our website project that we did, we had a common template and so on. By the way, just suggestions. I mean, the principal and the faculty, you all should decide how to go about it. But since I knew these next steps would come up, so some just quick thoughts I put down. Digital course template. See, if we are going to do physical to digital, once again, it comes into sort of academics, but we are bringing in digital content into our classroom lecture. How is that done? What is the template? Can we try it out? And so on. Not everybody can go and start thinking from scratch and say I'll do it myself in a different way. Can we create templates. Again, maybe the academic team can look at it. AI assisted research, Dean Research, some ideas came in. What inspiration, from where do we get and what are the uses of AI in research? AI infrastructure tools, planning and budgeting. I don't know, somebody has to work. I mean, we can help. But eventually the college has to take control of this. Right? Estimate the financial impact of AI transformation. You know, I was talking to Likith about it a lot. I said, why don't you look at it? I can give you some inputs, some ideas and so on. Theory versus practice. You have worked on internships at great detail. Okay, whether it's internships, whether Sanjeeth is on startups, final year projects all the HODs and faculty work with them. So that is another very important area. So I just put down some thoughts. This was just before I spoke, when I was getting ready for the lecture. So don't take this, take this with a grain of salt. There's not nothing great about this. But I would say, if you can, Dr. Nagendra, lead the charge on the implementation details along with your team, along with everybody here, the HODs and the Deans, that'll be wonderful. And anybody won't. Anything else before we close?

**Unsure** (Faculty member): [12:05] Ah, heli [tell me], Balaji.

**Balaji**: [12:08] Sir, a question for Anand, sir.

**Srikanth**: [12:12] If you could be a little louder, that will help, sorry. Why don't you come here? It looks like he can't hear you. Illi banni [Come here].

**Balaji**: [12:30] Sir, I have a question for you. It's related to how about student's well-being and cognitive knowledge in use of this technology in day-to-day life? There is one thing... I think in NIMHANS there is something called as a SHUT clinic which gives a service for healthy use of technology. Because that is a new one, past two years they are running. Those who have adapted to these new technologies, maybe use of gadgets or tools like this. Because without this thing we cannot implement this AI-based view. One part we need to implement, we need to develop our curriculum or any other related activities. Along with this, parallelly we need to see the well-being of students also. Well-being of students as well as faculty members. Because you are an AI psychologist, maybe you can bring some views on this.

**Anand**: [13:30] I think that is a beautiful question. So I thought I'll share from two angles. One, just the use of technology, when we encourage students to do it in a good way and people see who is being appreciated for it, it creates role models. It creates pockets of adoption and from there it spreads. If we are able to play some small role in identifying who are the people who are using technology well, highlighting them, allowing people to use them as role models, and I think that should not just be students, faculty also, and researchers, teaching assistants, whoever is involved in this, that is a powerful thing. Secondly, like with social media, AI has the power to spread people apart. I find that I am confiding in AI a lot. Whenever I feel distressed, I go to AI, I ask for emotional support and advice. And it does a very good job of it. I find that I am inspired by AI's politeness. When I talk to my colleagues, I get very irritated, and then I find that when ChatGPT... I ask the same question to ChatGPT, it replies in such a nice way that I am inspired now to be a better human trying to imitate that machine. It works in so many ways. So **I believe that there will be disruption. Some of it good, some of it bad. We definitely do not know how it is going to be disruptive. But in the process, having... or even teaching students to support each other... or I should just say, creating an environment where students, faculty, staff, everyone supports each other is important.** One of the ways I was trying this in my course was, I have one question. Every student gets a secret agent number. One of them is 007, one of them is 023, etc. And they all get a password. The assignment is to find out the passwords of three other students. They're only given the student's agent number. They don't know who the student is. So they have to reach out to a lot of students, find out their agent number, get their password. The other student can decide to give, not to give. All of that is there. But this is effectively a way of getting students to talk to each other. Now there is that 20% who have no problem with this. But there is also the 50% who, as you saw, won't even copy from each other. So there is a lot of merit in encouraging. We will learn new problems that emerge. We will also have to along with it learn new solutions for it. That's what I think.

**Srikanth**: [16:17] Wonderful. Anything else? Okay, well thank you very much! You've been a very patient audience as well as, it's been wonderful to work with Anand. Anand, thank you so much for spending so much time not just at this discussion but also for days that we have collaborated on coming up with this. Thank you so much.

**Anand**: [16:38] Pleasure. Thanks everyone.