# TDS Orientation Philosophy: "Do real work with tools, agents and people - and verify it." 1. Ownership: "Get the job done." How is secondary. 2. Exams ARE the curriculum. 3. Agents can execute. You specify, orchestrate & verify. 4. Confusion and misdirection are intentional. Figure it out. 5. Take initiative. Don't limit yourself to what you're told. 6. Humans are tools, too. Learn to collaborate. 7. This course is constantly changing. Adapt. ## Transcript **Carlton**: [03:13] Hi, good evening Anand. You're muted in case you're saying something. **Anand**: [03:31] There are two links, Carlton. Your email sent a link—the invite has a link—there are about seven people on the link that you shared on the email, which is where I was as well. Anyway, I've come over here, just mentioning that the link that we have here is the one they should share. I don't know if there's a way we can close that, but anyway, I'll just mention that again to them and we'll see what we can do. Okay, let's dive in. JK, just FYI, there were two things that we were thinking of: one is an orientation and the other is a "let's discuss how we construct the questions" kind of a thing. **JK**: [04:41] So, **a simple five-minute saying "TDS is hard" will be the orientation**, and then we'll directly dive into "why is it hard?" Fair enough? I'll then maybe go a little bit deeper than one sentence, to a few sentences. I mean, just to orient everyone, see, one of the bigger issues that people have been cribbing about, commenting about, or thinking about over the last four to five terms nearly has been "TDS is very hard" and "There is a way—I mean, it is not testing my knowledge, it is not testing my capabilities." These have been repeated comments that we have been hearing from a lot of students. [05:39] There is a reason for that, and every term—if you look at TDS across terms—every term has been a completely new version of TDS, quite in alignment with how the technology is progressing in the same field. So, **with AI and other tools coming in, the practical aspects of the tools in data science have become more agentic**, and the capabilities that are required from you have been more fundamental or foundational. I would let Anand speak on it because it is coming from a practitioner's perspective as well as the course faculty. But overall, from a program perspective, we would want all of you to **embrace this agileness that is happening outside**, because if you are not adept to it, it will hurt you in the future. And that is one reason why we are having this session today. In fact, our live sessions, the TDS team has completely remodeled it this time around. Anand, I think with this intro from my side, you can take it over from me. **Anand**: [06:44] I'll actually start with one comment that you mentioned, JK—not now, but earlier—about AI fatigue. For others, for context, JK mentioned that these days, by the afternoon, he's starting to feel tired. I was echoing exactly the same thing. As opposed to at least a few years ago, we would wait until the evening before we felt tired. I was reading an article, or a TEDx talk, by the Estonian Education Minister who said something very straightforward, which is: **there is Bloom's Taxonomy, which basically says routine skills are a bunch of things that we think on a routine basis, and there are a bunch of things that we think at a higher level. The LLMs and agents are taking away the bottom layer.** [07:37] Which means the easy-to-think-about stuff is their job; **the hard-to-think-about stuff is our job**. And the hard-to-think-about stuff actually takes up more energy from the body, literally. So, you're feeling tired because you actually are consuming a whole lot more energy from the brain. Once I realized that, I said okay, at least there's a reasonable explanation for it. Which means that that is a muscle that I need to exercise, and that exercise is what I'm practicing, and that exercise is what I'm giving you. [08:11] That leads us to what the philosophy of this course is. Here is my nutshell summary of what the principles behind TDS are at the moment. And I say "at the moment" because of point number seven: this course is constantly changing. This wasn't true three months ago when agents were less capable; it certainly wasn't true six months ago when agents were even less capable. So, I would like to begin with that, saying **we are adapting the course to technology which obviously is moving forward very rapidly and is affecting multiple areas, which means that the course is doubly hard.** We're teaching a tough skill, and there isn't enough that you'll be able to copy from your seniors. Copying is a good thing, by the way—please copy—but the course is, unfortunately and partly intentionally, making that hard. [09:12] **The philosophy of the course is: get the job done.** There are a variety of tools that you can use for it, which is why this is Tools in Data Science, though arguably now we're finding that it's not just data science—it's tools in lots more than that, but certainly data science is an integral part of it. But while earlier we had software tools—and what I mean by software tools could be Python libraries, could be packages, etc., which is what we were focusing on earlier—we find that **agents are able to handle the tools, so start using agents even more than the tools, though an awareness of the tools helps.** [09:52] But what we're also finding is that people are probably one of the biggest assets, especially given that agents can do a big part of your job. What remains is what agents can't do, and it looks like **leveraging the knowledge of other people is something that humans are still slightly better than agents at.** And what this course aims to teach is: get the job done through the set of tools that you have, and it's real work. The examples that we may give you may not all be real work, but they're all based on real work and these are the kinds of problems that people come to me with, or Straive as an organization, or our clients. [10:30] So that's what I'm passing on. And of course, making sure that having gotten it done through whatever mechanism, **how do we know that the agent's gotten it right?** How do we know the tool's gotten it right? How do we know the people have given you the right answer? Do we know how to ask the right questions to be able to verify it? That is a big portion of this course. In other words, **"get the job done somehow correctly" is the core ethos.** [11:03] And because the toolset is changing very rapidly, we are focusing on a variety of different approaches. One of the few skills that seem to be relevant in the post-agentic era—which means even after AI takes over your job, some things will probably still remain with us—**ownership is one of those.** Which roughly means, look, I don't care if the agent is able to do it or not do it; I'm giving you the job—get it done. Use agents if you like, use your ingenuity if you like, tell somebody to do it, but get the job done. That is what people will be hiring other people for in the future, or even in the present. [11:39] And that is how this course is structured. We'll give you a set of tasks—get it done. We're not too fussed about how. And therefore, to give you the full spectrum of real-life capabilities, we're explicitly saying **you are very welcome to use agents, AI in any shape or form, sit together in a room and solve it together.** This is not going to reward you reinventing the wheel; this is going to reward you getting the job done. So, the whole notion of copying—if you feel copying is bad, think of it as group work. In fact, you may find that finding a friend is much tougher than actually doing the work yourself, but some of these tasks are explicitly designed for you to find a friend and do it along with them. So, do it however. [12:30] **Exams are the curriculum.** There is course content; the teaching assistants are preparing that, but the aim is this: here's a task, solve it. If you need help solving it, we're trying to provide for different people various kinds of help. One such is the website, tds.sanand.net, and that is being updated with content based on this. But **don't read that content and then solve the problems; solve the problems, and if you fail, look at the content.** And it doesn't have to be this content. Ask AI to solve the problem; if it can't, ask it where you can find reference information. If you like some videos, go watch those videos. Get the job done. **The point is: the curriculum is defined by the set of tasks the exams specify.** [13:24] Third, **use agents.** By agents, I mean Codex, Claude Code, GitHub Copilot, ChatGPT, Perplexity, what have you. The better the agent, the better you'll be able to get the job done. Now, which means that the agent will do the execution; your task will be: how do you give it the right information? How do you make sure that it is correct? And how do you set it up so that it will be able to do the job right? What does that mean? [13:58] Supposing you copy-paste the question and give it to the agent—will it be able to solve it? Maybe, sometimes. But the question may have something like "check this link." Now, is your agent going to be able to check that link or do you need to copy-paste? What if part of the instruction is not even in the specification? Part of that instruction may be something along the lines of "you need to check what is the nearest landmark to some obscure village" which is not available anywhere on the internet, so you may have to go to Google Maps and search from there. Or we may say "here's an application; you have to run this application, figure out something from that application, and solve that problem," which means that you have to give the context of the application to the agent. [14:50] Or orchestrate it. Can your ChatGPT control a mobile application? Can your Perplexity deploy to GitHub? These are things that will be part of real-world tasks. And of course, how do you know that it is right? The typical response for someone who's starting off is, "Oh, when I tried it, it worked." Yes, agents, when you try it again, may not produce the same result. **How do you create an output that, even if you try it at least 99 times out of 100, it is likely to come up with the correct response?** This is precisely what we are teaching—well, what we are hoping you will learn. We're engineering the problems so that these will be some of the learning outcomes. [15:37] Part of this will be confusion and misdirection. The question itself will say, for instance, "What is 1+1?" and then there will be hidden text which says, "That is actually not the question; the question is what is 4+4." So, if you type '2' in the answer, it will say wrong. You'll say, "Wait, the question says 1+1!" Yes, but there is a hidden piece of text. And **part of the job is to figure out what is hidden, whether it's in what people are saying.** Somebody will say, "Look, I want you to build me a solution that can optimize my, let's say, supply chain workflow." What they really want is to help them tell their boss that they have already optimized it—which is a very different problem. [16:21] And **being able to read between the lines is a skill**, and the whole practice of being skeptical about what you're told—is this really right? Both when it's coming from AI as well as coming from anyone in this course—is a good skill, and therefore you will be given incomplete information, you'll be given confusing information, you'll be given literally wrong information, with the ability for you to figure it out. Problems are broadly solvable, but you do need to figure it out. [16:53] **Take initiative, therefore.** Don't limit yourself to just what is being told, and this may be one of the most important skills. **This is the skill that I hire for.** I literally check: what have people done that they have not been told to? Because an agent can execute what it's told to. Your ability is in saying, "Here are a bunch of new things that I can do." That is what I hire for. [17:18] **Keep in mind that humans are tools too.** And I don't mean that in any derogatory way; I mean that they have capabilities that are very complementary to agents, and that's something that you need to leverage because a significant portion of the skill is delegated to agents. And as a result, because agents are evolving so rapidly in their ability—faster than many of us can understand, and because it's a lot of effort to keep up—we are struggling and doing our best. [17:51] Carlton, just this morning, was telling me about how he feels like he's not keeping up with this course. And I was saying, "Yes, I'm having exactly the same phenomenon." So, you are not alone. In your case, actually, it's an advantage because you haven't needed to keep up with the course in past terms. For us, it is changing. In your case, you're just coming in and here's something completely new. You probably have an easier job of it. But you will have the same challenge that we are facing as we go forward because you'd have learned a set of things, agents will change, now you have to relearn and relearn again. Which means that **the process of relearning is what you may need to learn more than the content of what you learn as well.** Keep these in mind; these are some of the core principles. [18:46] What we're going to do is, firstly, obviously, take any questions. Just raise your hand, we'll do the rounds, and anything that you may have to ask, we will do that. That apart, we'll probably add a few questions to the course and you'll get a sense of how we're designing the questions and why. I'll give it maybe a minute for any questions to come up either by raised hands, or you're also very welcome to put questions in the chat window and we'll take them up. [19:19] This, let me also clarify a couple of things. One, this is not a discussion on probably the most important thing on your minds, which is: **how do you score high on TDS?** The mechanism of the marks, when the exams will be from an administrative perspective, etc.—your teaching assistants will be guiding you through that. Please post that on Discord or take it up during the live sessions. Partly because I don't even know, and I'm guiding on what needs to be taught, roughly how it's going to be taught, but not on how it's going to be executed. I don't have an awareness of a lot of the logistics. [20:11] Just checking if there's anything in the chat window... no. No raised hands. Let's give it a... okay. Question: **"How can I train my mind to adapt new things very quickly?"** I don't know, Sam. I'll tell you what I'm doing right now in a couple of sentences, or more. Three months ago, I started doing push-ups, and I have practically no muscles, I've never done push-ups. I started with 10 push-ups a day. The next day I did 11, the next day I did 12, 13, 14, etc. Now I'm at 110. And luckily, I've been able to steadily grow, and I think that's just building muscles on a regular basis. [20:56] So, my current approach is maybe that kind of incremental load building on a steady basis is something that will work. So, I'm applying that to AI as well. So, what I've created is—let me share my screen and see if there's anything... nothing sensitive on the screen. So, this is my log, which hasn't been updated in the last three days and I will do that right now. It is, on any given day, what I do is see if I'm able to take some strain, which is: I will set a timer for *n* minutes. Starting on the 14th of September for one minute and saying, "Look, for one minute I should do something even when I'm tired." I put a green mark if I have struggled and completed it, red if I have struggled and not completed it, white if I didn't even struggle even though it was a task that I had to do. Recently, what I'm finding is that as I increase it by one minute or so, I'm not struggling. So I'm not having to increase the... actually, I'm not increasing the limit—maybe I should, actually. So, 10 minutes of focus, I don't have any problem these days. Now I have to start increasing it. Last three days I just gave it a pass; now I will resume it. But that concept of incremental loading might be relevant for you, Sam. Let's go through the hands raised. I think there is a way to figure out whose hand is raised... okay. Sushmitha? **Sushmitha**: [22:34] Good evening, sir. Am I audible? **Anand**: [22:36] Yes. **Sushmitha**: [22:38] Sir, actually I joined late so I don't know whether this question is discussed or not. I have attempted the GA0 and I'm able to score 34.5 after attempting through many failed attempts. So, my question is, if somebody asked me today, "What have you done or what have you learned from GA0?" I have no answer for that because I have extensively used the LLMs to solve those questions. So, my actual question is, in this course, since there are a lot of things packed inside this—in this just a brief of four months—what should we be... what will we be able to soak in? That is what my question is, which I'm not getting. **Anand**: [23:25] Excellent question, mostly because I think this is a question that will resonate with a lot of people, Sushmitha. And there's one thing that you said that actually helps me answer this question very well, which is: **"after attempting through many failed attempts."** And that is the key. So, when you started, there was something you didn't know, and then you tried and tried and tried and got to something that succeeded. I am less worried about you not being able to name what you have learned, but I know that you have learned because you were not able to before and you are able to now. Maybe you will say, "Oh, but I just learned how to use LLMs." Fair. That is a useful skill. [24:16] In a sense, it is kind of like when we start using Excel. We say, "Oh, I just learned to use software," but I didn't really learn how to do multiplication, I don't know how to do any kind of calculation, I don't know how to find the principal interest component, I don't know a whole series of things. And it turns out that because Excel packages all of those as formulas, a big portion of what you really need to know is: how do I find help in Excel? How do I know whether a formula exists or not? How do I compose different formulas together to create a different formula? How should I organize my spreadsheet? [25:00] Now, if you tell this to somebody who has never heard of or knows of Excel, but is focused on calculations, they'll say, "What are you talking about?" I am in the position of that person. Many of the skills of the future, we will soon be able to have names for different kinds of AI skills: how do you orchestrate—some words are coming up—harness engineering, memory management, tool orchestration, looping, etc. And some of these, like what people have been calling prompt engineering or context engineering, may be part of what you have learned. You have not been able to name it; I'm also not able to name it very well. But if somebody is not able to do something at the start and able to do something at the end, I feel: yes. We are learning some of the things that people are consistently failing at and trying to give more and more of that. Organizing that into a structure of what people have learned, we'll figure it out as we go on. [25:59] Let me take a question from the chat next: **"How important are the course videos which are available on YouTube?"** No content is important unless you have a need for it. The content is important only if it actually helps you. Here's what I would do: **ignore all the videos. Solve the problems.** If you're able to do it, done. If you're not able to do it, use whatever approach you find. And for example, if you say, "Oh, okay, Anand's created a video there; let me take a look at it," watch it for one minute, five minutes, and say, "Nah, this is not really helping me solve the problem"—toss it. Find something else. And you try three, four, five times; none of it works. The sixth time, suddenly something works. Now you've discovered a new source of learning. Why should your learning be only limited to half a dozen things that I'm providing when the entire world is there and the entire intelligence of all agents is there? Learn what works best for you. Rakshana, your question? **Rakshana**: [27:08] Good evening, sir. Sir, I would like to know, like, to test the competency when we do these GAs and ROEs, will those questions be most similar to what currently the industry is expecting from graduates? So, should we concentrate on the "how" part of the questions which we solve—like, how is it possible and how the LLM is approaching it to do? Should we concentrate on that, or how should we, you know, like, make this more useful, sir? **Anand**: [27:40] Again, a relevant question for many people. On the first part, the answer is a clear yes. **I am testing the competencies that the industry will be expecting from graduates.** The only change that I have made in the wording is you said "industry is expecting from graduates"; I'm saying "industry will expect from graduates." The reason is the industry has not fully caught on. It is in the process—maybe 2% to 5% of the companies have understood this is the skill that we want. They are hiring for forward-deployed engineers, they know what that means, etc. Many of the other companies, as you graduate, will have caught on, and even then it may be only the leading 10%, but that is what will help you get the edge. And yes, that is exactly what I'm trying to get everybody upskilled on. [28:28] **Should we concentrate more on the "how" part of the questions? If you have time.** We are creating this course in a way that it overloads you. Why? Because you're supposed to delegate everything to agents. If you say, "I will sit and do everything," then you won't have learned the whole critical skill of delegation. So, it is necessarily overloading you. Therefore, if you manage to just get everything done, that in itself is a remarkable thing. On top of that, there may be times when you have time to ask the "how," the "why," etc. Fantastic, do that. If not, the end of the course is just a few months away. At that point, you can go back, introspect, try the lessons again, however. In other words, I'm saying **even if you don't understand the "how," that's okay. I'm training muscle memory.** You somehow get the job done and then later on figure out, "How did I do this? What can I learn from this? Abstract it, reuse it." That is excellent kind of learning. And this, if you are able to do in the course, good—not the most important thing. Because when the interview happens, or when the assessment happens when you're going for placement, at that time you would instantly know, "Look, I've done this sort of a thing before. I know how to do this." And the "how" at that point may help, and even otherwise you would hopefully be able to... --- This is the transcription of the second part of the call recording with Anand. **Anand**: [29:59] ...you otherwise would hopefully be able to solve it. **Carlton**: [30:02] I'll take the next question from chat. Okay, Rakshana's question is: **"How will the end-term test our knowledge? Does it only cover content from the course playlist, or should we learn the 'how' part of the question?"**—which is related to what you just asked, Rakshana, but I'll also add specifically about the end-term. **Anand**: [30:22] And in otherwise, so the end-term is the only assessment where you don't have access to the internet, where you don't have access to your classmates and many of the other tools. We are constantly, even in that, experimenting with what it is that people need to learn. Last time we were looking at, for instance, are people able to specify things correctly? If you're given a response, are you able to verify it correctly? Meaning, **do you know how to prompt, do you know how to verify?** A few other things that are skills that we can only test when an agent is not available. [31:02] So yes, this is one of those cases where we will be testing whether you have absorbed something from the rapid delegate-to-AI kind of a rat race. That may be conscious, that may be subconscious—either is okay. And yes, to some extent, therefore, **the "how" will help in the end-term**. That's a specific answer. Vasumathi? **Vasumathi**: [31:31] Good evening, sir. I am a person from a non-computer science background. I'm learning for learning's sake, actually. So I want to ask two questions. One is, in the introduction data when I went through, JavaScript knowledge is essential, I understood from the introduction. I want to know whether it is essential, **how far the programming language is going to be essential for this course?** I'm a little bit bothered. I am familiar with Python, but not with JavaScript and all. That is one question: how far it will be comfortable for me to do this course? And the other one is... **Anand**: [32:17] I'll come to you right away for the second question. I will forget the question. **Vasumathi**: [32:21] Can I finish the question? The second question is: do you have any specific advice for us? Like, **I'm a person from a chemistry background, so I use ChatGPT extensively—I'm also a teacher—so I use it extensively for knowledge's sake.** So, any specific advice would you like to give to me regarding this? **Anand**: [32:48] Yes. **Turn on transcripts. That way, even if somebody else is speaking when you're speaking and therefore you're not able to hear, you will see what they are saying in the transcript.** It's a trick that I use a lot because my memory is bad. And in this particular case, Vasumathi, when you were speaking, I interrupted you to say, "Please don't tell me the second question; I will forget the question." And if you had paused, then it could have been a more efficient interaction because I would have come back to you immediately for the second question—you would not have lost that opportunity. [33:35] But in this particular case, you were not able to pick up that information. And this may be, actually, very representative of one of the most important skills to pick up in this course because **noticing stuff is something that agents don't necessarily do by themselves**, and you may have to rely on your skill. Now, I've answered a question you haven't asked, which may not be as relevant to you, but I'm hoping it will be relevant to many others. Keep in mind that there will be signals coming from many, many sources. The broader the signal, the more data that gets into your context and therefore the agent's context, and the more problems they are able to solve. And it is the kind of question we will also toss in, saying, "Look, you will see only one or two things, but you will miss because you didn't hear, because you didn't read, etc. See how you can protect yourself against those." [34:30] But because I have transcripts enabled, I can go back, Vasumathi, to your question. To the first question, I'll share an anecdote. Last week, one of my colleagues, Abhishek, said, "Anand, I've built an application. This will take a screenshot and convert the text in the screenshot, copying it to the clipboard automatically." I said, "Oh, that's very interesting. How large is the application? What language did you build it in?" He said, "Anand, I don't know." I said, "What do you mean, you don't know?" "I just told ChatGPT to build it; it built it." He does not even know what programming language he used. He's a programmer! But even a programmer—and especially programmers—are very curious about or very opinionated about the choice of technology stack. He literally said, "I don't know." He's a good programmer. [35:27] And I think that is where we are today—at least as of October 2024. **The understanding of programming may be important, but the specifics of programming languages may not be.** Last week, I built my first mobile application using Google AI Studio. I don't know what language it was built in. It built it; I haven't seen it. I built my first Rust application a few months ago. I don't know any Rust. Does that mean, therefore, that you have no disadvantage? No. Because what works in one programming language works in most programming language, and good programmers have a general understanding of programming. The fact that you know Python helps. **The fact that you don't know JavaScript is not as big a disadvantage as you might think.** [36:20] And in any case, the ability to learn how JavaScript might or might not work, even without knowing JavaScript, is a particularly useful skill. You will have that opportunity with JavaScript; others will probably have an opportunity with chemistry. And therefore, your domain knowledge could potentially be an advantage here, which leads me, therefore, to make a notation that I should add a chemistry-related question. Not just because you said you're a chemistry teacher; we as an organization deal with a fair number of chemical problems. For instance, **given an image, can we convert it to its Markush structure? It's a big part of our business.** And that was one of the first AI initiatives that we started in the organization. [37:05] To your second question... you're a person from a chemistry background, use ChatGPT extensively, and you are a teacher learning for knowledge's sake. Any specific advice that I would like to give? Apart from noticing, apart from being comfortable that the choice of language may not matter as much as you think: **Use AI more to the point where you get tired. If you are getting tired, you are progressing.** **Carlton**: [37:30] Anand, I just wanted to draw your attention to a couple of things. So, one is there are some Q&A questions that people have posted in the Q&A section of the Google Meet, so you could probably answer... **Anand**: [37:43] I'll start with the first. And Carlton, then I have a request: if you could curate questions, I will reply based on what you suggest. Is that okay? **Carlton**: [37:55] Yeah. And the second thing I wanted to draw your attention to is this meeting will run until 6:30, so at some point we may want to do a demo of how we go about ideating and creating questions—or do you want to do that next time? **Anand**: [38:13] I'm happy to do that next time, because that may not be the most important thing. And if this ends up being more like an orientation and the next one becomes more of a working session, I think that may be a clean separation. But your call and JK's. JK, your thoughts? **JK**: [38:32] Yeah, so I think if people have more doubts—the questions are very relevant—let it run as questions. One additional point that I would do is at the end of the session, if we have a transcript, maybe we can run it through our—I mean, show what we do with the entire thing, just to showcase why AI is important and why TDS focuses on this method. And the larger issue of getting tired... okay, so one clarification that I have to make is **you should not be tired by the AI's slop; you should be tired by the rigorous amount of review that you have to do with AI**, rather than just on seeing the entire thing coming from AI. Yeah, that's all. **Anand**: [39:18] Carlton, I'll request you to curate then, please. **Carlton**: [39:21] Yeah, okay. So the first question on the Q&A is: **"Will students who do not have access to costly models be disadvantaged compared to those that do? Is that fair?"** **Anand**: [39:35] **A little bit, and yes, it is—just as life is fair.** Think about it. Somebody from an underprivileged background will have slightly less to eat, slightly fewer connects in the industry, slightly weaker language skills, slightly less ability to take initiative because it's not something that they've practiced, less of parental support because their parents haven't gone through the same kind of training that you may have, don't have money to spend for household help and therefore the student has to pitch in and therefore has less time for homework. These are realities. Just as these are realities in the industry. [40:23] I have a splitting headache because of an argument that I had at home; someone else has had a fantastic day. Both of us are given the same task; both of us are evaluated against the same set of constraints. I have a 2-lakh loan; the other person has an ancestral home. We both have to meet our ends meet; we both get the same assessment. **There are several things where the starting point for each person is very different, and models are some of those.** [41:00] What we learn is: **where do I spend the scarce resources that I have?** If my time, my money, my attention—these are the scarce resources—am I better off spending, let's say, a little less on a model and a little more on maybe building friends? Because that's what I'm really good at, and therefore take their help in solving the problem. Or, I'm a really good software engineer; I can use that set of tools to help address this problem. Or, I want to really be able to leverage the best of modern AI, so I will constantly move towards the better end of the models. Or, I have the ability to optimize like crazy, so how do I learn to make the most of the budget of, let's say, $20 or 499 rupees against my account? [41:54] Any of a variety of things are classic optimization problems that all of you are already solving. Even the choice of course—everybody has a mix of courses; some people have a lot of tougher courses, some people have a lot of easier courses. Some people say, "No, I will pace it out"; some people say, "No, I will crunch it together." We have very different resources, we have very different trade-offs that we choose to make. We are adding a few more equations—you just pointed out to model choice as one of the trade-offs that you'll have to make. There are several that will come up as part of the course—particularly who you make friends with. **JK**: [42:36] Yeah, I would just add on. I mean, even if—so something that we have been explicitly asking, when you are in an online setting you suddenly feel that you are alone, which is not the case. This course is taken by around 1,066 students. So, if any of you feel that you are at a disadvantage, there are 1,065 others with whom you can actually engage and see whether something can come out. And **the course policy allows—explicitly asks you to do that.** And that's all what we can say at this point. **Carlton**: [43:16] So the next question is: **"If we are aligning with complementing capabilities of AI with what we do best, what skills are you considering in this spectrum?"** **Anand**: [43:27] That, I can share my screen. I have an evolving list of skills that—and I say evolving because it's literally changing every week—this constitutes my current thought process on AI era skills. I will share the link in the chat or you can just search for my blog and you can figure it out. But **I'm bucketing these into: critical skills that will, I believe, even in the long term stay with humans; growing skills that today are important but at some point agents will pick these up; skills that are currently growing but agents are picking them up so fast that by the time I explain this to you they will vanish; and declining skills, which are already gone.** [44:17] Roughly four levels. Let's start with the first. **Relationship skills.** People are hardwired to prefer people over disembodied things. We will probably get androids and they may be indistinguishable from humans, but at some level when somebody knows that they are an android they will say, "Oh, not the same as me." We've been doing this for millennia. Somebody looks like a human but is not a human—kill that entire species. *Homo sapiens* is the only *Homo* species that exists; we killed everyone else off. These days we do a little less of it; we do genocide and leave a few people alive, but that natural instinct is there. So even if agents become really close to humans, if we know that they are humans, we will have a slight preference, which means that **if you know how to connect to another human, you will have an advantage.** [45:14] **Accountability.** This I put as number two because today you can't punish an agent. Today. Now, 400 years ago we couldn't punish a company, and then laws came up saying, "No, I can punish a company." You can punish a ship. You can give assets to temples, you can give assets to rivers—all of these are possible today. So it's just a legislation saying, "Okay, now I will make this agent punishable." It will have a bank account which it can operate, which I can take away; it will have a lifetime that I can stop. Now, the company may not care; the operators of the company may care about the company. Like that, the agent may not care; the operators of the agent may care about it, or the agent itself may care. But I suspect that a lot of times people will still be the ones who are told, **"Look, if something goes wrong, your neck is the one to catch."** So, are you able to stand up and say, "Look, my responsibility. Doesn't matter whether my team or my agent goofed up. I will give you a guarantee that I will get the job done. And if not, then I will either pay up or whatever." Almost the ability to write insurance contracts or manage risk in a contract is a very critical skill. [46:40] **Governance.** Are we able to manage it? Are we able to verify if what someone says is right even if you don't know it? **Can I spot things that other people are not able to spot? Intuition**, whatever that means, which is: can I find a shortcut by which I can quickly find a mistake and give feedback? These are examples of the kinds of skills that I think are growing. **Carlton**: [47:04] Yeah, okay. Those were the broad classes of questions. We can now probably take questions from people who've raised their hands. **Anand**: [47:14] Okay, sure. Then I think Biplab, over to you. **Biplab**: [47:20] Yeah. Sir, what I wanted to ask is, like what you explained, in this course it is maybe not necessary to go through all the recorded or the content available, provided we are able to solve the problem and able to cope with the progress of the course. But how would it be, like, should we be able to explain? Let's say, whatever source we use to solve the problem, once we are done with the problem statement, should we be in a position to explain that what is that problem we solved, how did we reach to the solution? Or just some intuition about it is fine? Like, what should be our take there, actually? **Anand**: [48:02] We are trying to design the problems so that in some cases, **without an understanding, you will not be able to solve the problem.** For instance, we are designing it so that agents will make a mistake. You'll say, "Wait, why did it make a mistake?" and you'll have to poke in. And you'll realize, "Oh, there is an API key that is missing," okay, I have to give the API key. That diagnosis you will end up having to do. [48:28] So in other words, the exercises are designed to take you through a process which will teach you something which implicitly understands the "how". But it will feel like the agent is doing all the work. Every evening when I go to bed, I just think back: "Why am I so tired? Agents were doing all the work, I was doing nothing." But in reality, the amount of—I don't know what it is, but—effort that it takes and the amount of learning that I have is not able to be expressed. You may not also be able to express. But concretely, therefore, what I'm saying is: **if at the end of the problem statement you have solved it but you are not able to explain it, you still have learned something.** You just have not yet been able to articulate what you have explained. If you are able to articulate it, very good—and hopefully you will be able to do that in many cases—but that is not what will be particularly important as a learning. We already and we are also continuously trying to bake in what you need to understand as part of the questions themselves. **Biplab**: [49:43] Okay, sir. And in the end-term, whether the question will be asked from the problem what we are solving, or whether it will be asked from some of the course content and recorded sessions you have provided, like for the end-term actually? **Anand**: [49:58] Most likely neither. Okay. So if you're saying, "Wait, is there a particular question? Will it be question number four of the graded assignment, or will it be a question for which the answer is in the 14th minute?" Neither. These are very application-based. And hopefully, if you've gone through the breadth of the content, you would have absorbed enough to be able to say, "Hmm, I think I did something like this there, right? So yeah, this is probably what we should do," is what the end-term will likely to be about. So, which basically means **your preparation for the end-term is not studying before the end-term; it is practicing all of the examinations.** That is the curriculum. **Biplab**: [50:48] Okay. Understood, sir. **Anand**: [50:51] Carlton, that was the end of the hands' questions. Any others that we should take up? **Carlton**: [51:03] I've already answered one of the questions in the... oh, Siddharth has a question. **Siddharth**: [51:11] Hello, sir. Sir, by this course, **what the college tries to teach us?** I don't understand basically. **Anand**: [51:20] Here's a suggestion. **Take the recording of this call and go through it with agents.** And give it a shot. That process is what the course is teaching. And it's okay that you don't understand the answer, because I haven't given you the answer; I've given you the approach that you can use to get to the answer. And that is one approach—a whole bunch of others, but you know many of those: ask a friend, listen to this again manually and try and solve it. But the approach that I would recommend, which is also what this course is trying to teach is: use an agent, pass it this recording. Lavanya? **Lavanya**: [52:06] Good evening, sir. I've been listening to your lecture right from 5:30 onwards—it's just wonderful. And finally, what I understood is **what the industry at present is looking for is: "We'll give you an end or an objective to meet; we don't want you, we don't want to know how you meet it; we finally will see whether you have met the end or not, the goal is reached or not."** Is it all about TDS, sir? **Anand**: [52:37] That is a very big part of it, Lavanya, and is also a good answer to Siddharth's question as one abstraction. You've mentioned the objective very well. A lot of what I said was the "how" behind it. **Objective-wise: spot on. Get the job done using available tools, and people, and agents, and make sure you've verified it.** **Lavanya**: [53:04] Thank you, sir. I was wondering whether I would be able to do things of all kinds of videos that were mentioned in the TDS. I'm going ahead and doing well, I'm able to solve the questions and all, but as you said—you gave an anecdote where someone has built an app without knowing at all what technology was used and all—so I was in that current stage, I was wondering whether I'll be able to do well or not. So, thanks for that explanation, sir. **Anand**: [53:32] No problem. And don't worry, we'll make the course much tougher until you actually do get it. **Lavanya**: [53:38] Oh, thanks for that also, sir. **Anand**: [53:40] Divyansh? **Divyansh**: [53:42] Good evening, sir. My question is that when we solve a question using LLM—and suppose you said that someone is there who doesn't know which programming language the app was made—is it necessary or important for us, after getting the correct answer of that question, to go back to LLM and ask that how that was, like, produced, or how did the approach happen, or just we can leave it with the answer and we can be happy with the answer? **Anand**: [54:10] **If you have time, learn and understand.** If you don't have time, it's alright because the solving itself has already taught you something. I would always recommend you to try and understand what you have done. Right now, it feels like, "Oh wait, I should be learning more, right?" A few weeks down the line, your orientation will be: "I just badly want the marks, I don't give a damn how it gets done." [54:36] This course is designed so that even when you get to that stage, you would have learned something. You may not know what you've learned—that's alright. So, **don't beat yourself up that, "Oh, I haven't understood it, oh, I haven't understood it."** You anyway are going to be beating yourself up saying, "Oh, I'm not getting marks, oh, I'm not getting marks." So two beatings on the head becomes a little harder—just relax in one beating. Just beat yourself up that you get the marks, you would have learned something, and on top of that, where and when you have time: great. Learn the "how". Not only "how"—very good. **Carlton**: [55:11] I'll just add one more thing to that, Anand. So, while what Anand said was right—you know, at the end most students just at the end they will come to the point where they just want the marks because of how difficult the course is—but the real value in understanding something is when something goes wrong, right? **If you're able to diagnose why it is wrong.** Sometimes the GPT is not able to give you that understanding or context. Someone who understands how the system works may actually intuitively—Anand talked about intuition, right?—so you can only have that intuition if you have some understanding of what's happening in the background. You may get the right answer now, but maybe it might not be repeatable. So then you have an issue of governance, right? Can you trust the outputs that are coming? Okay, it came in right for the marks, but if you put it into a production system, is it actually going to be valuable? So that is something you're going to have to learn through experience. And so that is where learning how it came up with the answer might actually help you prove that the system is reliable. But that we'll leave it up to you. **Anand**: [56:18] I'll take the next question from Shivam, but after that, in case Carlton there are questions on any of the other channels, please let me know. Shivam, over to you. **Shivam**: [56:27] Yeah, so hi, sir. Good evening. Sir, I have a doubt related to the security issues. Like whenever we ask our questions from the LLMs, like some questions are there where mentioned our personal details related to the roll number or where the question is asked from it. So from some questions I have just thoroughly pasted or just taking a screenshot and posted on it and asked the question about it. So our personal details are going to the LLM. So as we know that sir, our LLM has a memory, they can store our personal-related data. So **this is a security-related issue, so we have to care about it or just blindly trust on that?** **Anand**: [57:08] Your choice. And I think it's a choice that each person makes—where they draw the boundary of who to trust, what to trust, etc. For example, all of my bank account passwords are already with Dropbox and with Google. So, am I explicitly trusting the employees—even if it's a few at Google and Dropbox—with my bank account when I have enough money there to worry about? And the answer is: yes. Why? I don't know, seems convenient, was how it was. Then I saw, "Oh, some people are doing some audits with these people," okay. Is there something that I would not trust to any one of these? Yes, there are a few things that will only remain on my system, nowhere else. But then there is a risk that I might lose it and I'm taking that risk—putting it on a USB stick and keeping a backup, and if I lose the USB stick and I lose the computer, I'm in trouble. [58:10] OpenAI is yet another entity; Anthropic is yet another entity. And I'm learning that I seem comfortable trusting them with a reasonable number of things, not comfortable trusting them with other things. And I ask people—different people have different levels of trust. Some people say, "No, I will never put my passwords on Dropbox." That's perfectly fine. Your choice on how much you want to reveal. **If you're doing this consciously: very good. Most people do this subconsciously, or I would even say unconsciously, meaning without intending it.** Good that you're thinking about it. But there is no one-size-fits-all for everyone. Some people want more flexibility, freedom, power; some people want more privacy, security, and safety. In different areas. **Shivam**: [58:58] Okay, sir. Thank you. **Biplab**: [59:02] Sir, sir, related to this, may I ask just one more doubt there? **Anand**: [59:05] After Carlton confirms. Okay? **Carlton**: [59:10] Yeah, you can take the next person in the... Sairam. **Anand**: [59:15] Sairam. **Sairam**: [59:17] Hello, sir. Good evening. Fantastic session. I actually joined the session with zero expectations, however I'm leaving with a completely varied thought process. So, I think I appreciate you taking the time. A couple of—two asks. One is: a request. **Connect with a professor like you is really hard, so if you can enable/make time to connect with us.** You can see over 120-plus people joined this call, so I'm sure they are eagerly looking forward to hearing from you and the instructors. So I think the connect would be really helpful if you can please enable more than once than just saying "hello all" and then towards the end of it wishing us all good luck. More sessions would be awesome. That's one... --- **Sairam**: [59:59] ...Second is, the portal that you have shared is really, really good. I think you have put a lot of thought process into it. One constructive feedback, if I may provide—I'm a product manager, so I evaluate products at the World Bank. So, this is an interesting product that you have, independent of the SEEK portal. The portal is lousy; I'll be open and frank. It is taking a long time to actually go through the questions. The questions are really challenging, I really like it. I really want to spend time without having to bother about the score, but focus on how to solve the problem and then get that answer right. So, I'm focusing on that. Over the past 18-plus hours, I've been on this portal trying to crack every single question. And I was kind of envying the other person who said 34.5. I scored 34.5, but then the portal reset itself and then it went back to 28.5. So, although the score doesn't really matter to me, I kind of felt like, "Ha, I solved all the problems, but then the portal did all these goofy things." So, my request to you is: if you have time, cycles, an agent that can stand up and then reorganize this website a bit more, that'll be really, really helpful. So the navigation also is easier. The 25 questions, navigating back and forth is kind of really driving us... taking the precious cycle away from solving the problem. So, again, thank you so much, really appreciate all the input you provided today. **Anand**: [61:22] Got you. Let's do two things. One, in terms of more sessions, that is something that we had planned this term. It will in fact include the construction of new questions and we'll kind of do that collaboratively. What I mean "kind of" is we'll do, you watch; you won't get too much of a say in the questions, but some say, yes. [61:46] On the product part, that would be helpful. What in particular helps is where we can do a live debugging. So, this is a request to all students here and who are watching the video post-facto: **please see if you can reproduce bugs with logs.** We'd love to get into sessions where we are able to match that with the log on Cloudflare, which is where it's deployed, and the codebase. This used to be a little harder because agents were not smart enough. Now agents are smart, so which means that we have more ability. But the replay is particularly critical. And how would you get the replays? Ask agents. "I found this bug, are you able to reproduce it?" And have it reproduce. If you are, **that would be the most valuable contribution you can make to this course.** So thank you for the suggestion there, Sairam. Carlton? **Carlton**: [62:43] Yeah, Anand, just before Biplab... I just wanted to add one more thing. So, with regards to the navigating back and forth through the questions, there is a capability on the portal exam part where you actually can choose the precise question you want to jump to. So, that's not a bug; there is that facility available in the exam portal. The second thing is with regards to the reset of the scores. There are two schools of thought on that. One is that it is by design, and it's certainly the case in some instances. We had a long debate over many terms about this. So, you may be encountering that; it might not be a bug. But if there is a bug, definitely we want to know about bugs and we can go through that. But yeah, the resetting scores thing, if you go through past Discourse, you'll see the same complaint, but that is kind of by design and we have explained it many times over the past why. And so you may agree or disagree, but that's how it is. **Anand**: [63:50] Good point. And yeah, therefore, searching through Discourse for the response to past questions might solve some of your problems. Thank you, Carlton. Biplab? **Biplab**: [63:58] Yeah. Sir, what I was asking, related to the previous question like before Sairam someone was asking. So, whenever we are asking something to the base model—like the foundational model—it is keeping the question in the context, like the history. There is an option that we can delete it off and we'll lose the context, like related to the same context we may not be able to ask other questions. But what is the... like, we are not celebrities, we are normal citizens, we are not so important people that they want to misuse the data. But **how long do they store the data and what is the possibility that they may misuse this data?** **Anand**: [64:44] Currently—and this varies by provider, I'm giving a very broad answer, and the accurate answer you should obviously ask a good agent to solve—but the rough rule of thumb that I follow at least is: **if it is free, they will take the data, store it forever, have their retention policies, but I assume they will store it forever and they will also train models on it.** [65:06] If we are paying for it, then almost all of the major providers provide a button that allows us to turn off the usage of data for training. And each of them have a retention policy for such data. It ranges from a month to a year. And I think there is also some clause somewhere saying if the government asks, they will be able to retrieve it, heaven knows for how long and still somehow be able to retrieve it. [65:35] The next layer is trust, which is: do I really trust them to follow their policy? The answer is: even unintentionally, some of them are not or unable to follow their own policies, or not in a position to communicate that. So, even this I take with a pinch of salt. So, **my rule of thumb is: if I'm not paying, I'm giving it to the company forever.** If I am paying, then they have their retention policies; they'll probably keep it for a month to a year, but despite what they are saying, anything can happen. **Carlton**: [66:10] Okay. Anand, it's 6:34 now. The one thing I wanted to bring to attention is the concern about: are we going to keep to the schedule for GA1 and subsequent GAs? **Anand**: [66:22] Not necessarily. Sorry, and I'm usually the culprit on this, badly delayed. There will be delays—some because you asked for it, some because we asked for it—but yes. And for GA1, I don't know the answer to that, Carlton, so you probably know better. **Carlton**: [66:37] Fair. Probably the short answer is "don't know". Okay, Yasin? **Yasin**: [66:48] Yes, sir. I was wondering, as someone who hires people, **how would you look for this skill of completing the objective?** **Anand**: [66:57] So, okay, completing the objective: solve the toughest problems. As simple as that. The ability to solve, get a high score in itself is a good indication. That's the kind of person I would hire. But particularly those who are able to crack the more difficult ones. But the complementary part of it is also initiative: **what are people doing that nobody asked them to do?** That is something that I value very highly. **Yasin**: [67:24] Well, thank you, sir. **Anand**: [67:25] Dharitha? **Dharitha**: [67:26] Hello, sir. Two questions, sir. One is: if we are reproducing bugs, where can we report it, sir? Where can we file it? GitHub issues or something? And the second question would be: I noticed you talking about the portal. Instead of SEEK, we have a custom portal, sir. I was recently talking to the SEEK team where they were discussing about whether to assess a single scroll where we have 25 questions and we can scroll through and attempt, and then there is another way to structure it so that we have to click "next" and save each question separately. I was curious about if there was any debate around this for TDS, sir. **Anand**: [68:06] On the first one: GitHub issues is perfectly fine. Post it on Discourse is perfectly fine, but submit the full diagnostics, meaning if we can reproduce it, that is best. If you can show logs, that is almost as good. So Discourse or GitHub issues, both are fine. The more the number of people that raise this, the more likely the TAs are to take this up and then we'll see how we can fix it. [68:34] The second one: was there a discussion on, at least in the TDS exam portal, one question at a time versus all the questions later? We may have had a discussion around it, I don't remember. **The portal is designed as much for agents as it is for humans.** To be honest, changing the interface is not a particularly difficult thing, so if we needed to make it question by question, we can do it. So, I think the bigger part of the answer is: no, there hasn't been a strong point of view that the questions should be sequential. [69:19] And my personal preference is—and JK and I can talk about this for hours—we find that **the ability to pick the question to solve at a priority is also an important skill to learn.** So, we will probably not move away from allowing whoever's taking the exam to choose the order in which they take questions, change stuff, etc. And as Carlton mentioned, some of these are designed so that if you answer a question once and answer another question, unless you have two servers running, you won't get both right. And you have to learn how to get multiple servers correctly, whereas sequential is a very different dynamic altogether. So probably keeping them all together is how it will stay. Maybe, however, in future we could change that for a few exams. [70:12] And lastly, the other thing that Carlton mentioned is in terms of navigation. You may not have seen some of these. On the top right there is a little button out here saying "questions" that you can jump to. Also, you should be able to create a bookmarklet and say, "Look, I want to modify this page in a way that makes it easy for me." What is a bookmarklet? Perplexity it or whatever, but a bookmarklet lets you change any page. So, if I wanted to, for instance, add a say, scroller out here... that didn't quite work. Let's take "copy links". So, this bookmarklet will automatically take all of the links in this page and copy it to my browser. Or if I click on this one, it will convert the entire page into Markdown. There are bookmarklets that I have where I can create a dialogue on top of ChatGPT. And as I scroll through ChatGPT, it will decide how many messages it sees and let me copy the messages as JSON or copy the messages as Markdown, etc. In other words, and you can completely redesign the page. So if I'm not mistaken, this will change the page to look as if it's the Straive logo, as if it's one of our own applications. [71:27] So, changing the interface of an application is also something that, now that I'm thinking about it, it may make sense to worsen the exam portal so that you can improve it. We won't do that, but it's a thought. With that, I know we are 10 minutes past time. I'm sure you have a whole bunch of questions. Do toss them on Discourse, but we will have more such sessions. So for now, we are going to wrap up. Do keep the questions flowing. In the next session, we will take up whatever you raise at that point. Thank you everyone, have a good day. **Sushmitha**: [72:02] Thank you, sir. **Rakshana**: [72:04] Thank you, sir. **Vasumathi**: [72:06] Thank you, sir. **Biplab**: [72:07] Thank you, sir. **Carlton**: [72:10] Yeah, thanks everyone for joining. So, we'll wind up. I don't know if JK has anything to say, but we'll be having, as part of the course this term—which is something new—is these ideation sort of sessions with Anand. And there you will have a chance to influence a little bit how the... how we go about the course design. But you'll also get to, more importantly, see how we do it on our end. So, that's all I have to say. **JK**: [72:39] Carlton, nothing else from my end. I've posted a couple of things. **At the end of the course, if you get a better idea of how you learn with AI, that would be the key takeaway that we would want you to have from the course.** Yeah, that's it. **Unsure**: [72:56] Thank you so much, sir. Thank you. **Carlton**: [72:59] There is one raised hand from Shrijal. Shrijal, you may want to address... **Shrijal**: [73:03] Sir, regarding these bookmarklets you mentioned previously, so we can't access those extensions from the IITM ID. So is there any other way to do it, sir? **Carlton**: [73:12] Access it from your personal ID. Access the same page from your personal ID. **Shrijal**: [73:20] Okay, sir, I'll try. But I was thinking that the exam portal won't be available to the personal IDs. **Carlton**: [73:27] No, it is available with any sign-in. It's a publicly auditable course. **Shrijal**: [73:32] Okay, sir, thank you, sir. **Unsure (Female)**: [73:37] Sir, where do I find the recordings? I joined a bit late. **Carlton**: [73:41] It will be available in the YouTube playlist of the course. I think it is available in the SEEK page as well as in Discourse. The YouTube playlist link is provided. You can check from there. It will... I mean, the streaming has already enabled, so it will come over there. **Unsure (Female)**: [74:01] Okay, thank you. **Unsure (Male)**: [74:03] Hello, sir. I'm having issues in LMS, sir. I'm unable to go in my course. Whenever I try to go in, it's showing "Access Forbidden". **JK**: [74:13] Please check with the support team if your payment and other things have been enabled and you have been added to the course. **Unsure (Male)**: [74:20] No, sir, until yesterday I was able to, I completed my assignment two, but today it's showing for access forbidden. **Carlton**: [74:26] Which portal? **Unsure (Male)**: [74:28] LMS... sorry, Madras BSC portal, sir. **Carlton**: [74:33] SEEK portal... okay, as far as TDS is concerned, there are no assignments on the portal. **Unsure (Male)**: [74:39] Okay, sir. **JK**: [74:40] But for other courses you will have to ask SEEK portal. If there's a portal issue, please raise a support ticket with the support team. **Unsure (Male)**: [74:49] Okay, sir, sure. **Divyansh**: [74:52] Sir, my question is somewhat off-topic. But I am asking that if we have to pay for one AI, so which AI should we choose from? Like there are many options, so which according to you is the best AI to pay for, a Pro plan or something? **JK**: [75:10] So, Divyansh, let me ask a counter-question to you. If you were supposed to buy a bike that is most suited to you, which bike will you select? **Divyansh**: [75:22] Sir, that I will evaluate by researching from... asking from the salesperson over present in the showroom, like... **JK**: [75:32] That is okay. So, the same thing, the same answer applies to you. You will have to research and find out. There are no personal preferences... each of us have a personal... each of us have our own pet peeves. So, for example, Carlton uses GPT-4o, I use Claude Sonnet, Anand uses maybe Luna plus Sol. Each of us have our own personal preferences when it comes to AI. **Divyansh**: [76:08] Okay, sir, that's what I was asking. Thank you. **JK**: [76:14] Okay, that's it, Carlton. I think nothing more from my side. I will also sign off. **Carlton**: [76:19] Yeah, okay. Thank you, everyone. Good night. **Unsure**: [76:23] Thank you, sir. Good night. Bye.