# Transcript

**Anand**: [00:11] 4:30 PM IST. Let's get started. This is a session where you can ask me anything, and I have lots to ask you as well. Before we start, there is a link that I put in the chat, which is simply forms.s-anand.net. That contains a few questions that I'm curious about, but also where you can ask your questions. So, just feel free to go to this link—you can just go to forms.s-anand.net, which will take you to this link—and put in whatever questions you have for me out here and submit, edit, whatever. Even if I can't answer now because we run out of time, because it's stored here, I will be able to draft a reply to you. **You don't have to limit yourself to just TDS; you're welcome to ask anything about AI, careers, course design, or literally anything.** Whatever I know, I am very happy to share.

**Anand**: [01:30] I certainly had a few questions that I'd love for you to answer, which is: What contributed to your learning in TDS? What contributed to you getting marks in TDS? And a bunch of other things. Take a shot, see what you can answer—these half a dozen questions. But primarily, like I said, this is for you to ask me anything and for me to also toss a bunch of questions at you. I'm still waiting for somebody to ask something on this form.

**Anand**: [03:16] Okay, I believe we have one question—two questions, maybe. One of the questions is: "Are we relying on AI for all assignments? As we are relying on AI for all assignments, what's the expectation from the end term exam in person?" **The end term exam will be very simple.** The expectation is that you will be able to remember all of what you have learned and be able to cover it. **The end term in-person, no internet exam is a statutory requirement from the Senate**, so it is there because it has to be there. I actually don't know what it will cover. Please ask Prasanna; he sets the end term exams, and it is always simple.

**Prasanna**: [04:06] I can take that answer, Prasanna here. So, the end term basically looks for the logical reasoning. So, it doesn't look—I mean, we won't be requiring any specific syntax-related questions or anything to do with building applications. **It's more to give logical reasoning based on the knowledge that you have gained through these different courses, the different set of sections that are covered.** It will be a multiple-choice question, and there could be some subjective questions as well.

**Anand**: [04:44] Thanks for that, Prasanna. And now that gives me an idea. We will probably increase, Prasanna, if you're okay with it, the subjective-type questions a little bit. Because a) we have the ability to use AI to evaluate, and secondly, the ability that we are looking for—there are some things that we can now test that I'm curious about. **One of the major things is: how do we prompt and how do we verify? AI does the middle stuff well. The stuff that we ask for and checking the stuff that we get out is where the quality of prompts becomes important.**

**Anand**: [05:41] So, as part of this exam—and I'll work with you on this, Prasanna—to share how you would be prompting under different scenarios. What does it mean for your preparation? Right now, if in an ROE or a graded assignment, I asked a question like, "What prompt will you give for this?" you will copy the question, you will give it to one of the models, it will give an answer—that's effectively meta-prompting. Good technique.

**Anand**: [05:58] My voice is echoing? Let me—is my voice echoing for others as well? Oh, Mayank is saying it's good. Okay, then Ishaan, it may be your system issue.

**Anand**: [06:08] So, you would be meta-prompting. Whereas in the end term, you would be relying on your own knowledge. So, **practice learning how to prompt, specifying problems well, and verifying solutions well.** These two will be part of the end term for sure, now that I think about it. And if this does not answer anybody's question, please feel free to edit and add on any follow-ups to your question.

**Anand**: [06:58] Another question is: "Is AI bubble real? Will agentic AI sustain?" My belief is yes. **Not only is it [AI] not a bubble, it is underestimated.** Usually, people underestimate long-term impact and overestimate short-term impact. Short term, maybe we have overestimated—I don't know—even that I don't think so. But supposing all progress in AI stopped today, the amount of potential that it has to impact the economy even with the current state is incredible. And not only is it becoming more capable, it is becoming more capable at an exponential rate, and we don't know where it will end.

**Anand**: [07:54] One theory is that it is growing on one axis, which is intelligence. And intelligence is overrated—that even if you have perfect intelligence, there is only a certain class of problems that we can solve. That may be true, but the sheer number of problems that we can solve given perfect intelligence is enormous. Short answer: I definitely don't think it is a bubble.

**Anand**: [08:21] Question: "This course is tipped by everyone as very tough. I realized it is as tough, at the same time it is preparing us for job interviews, more or less job-ready. Am I correct?" Absolutely. **TDS is basically how I would recruit and what I'm giving people out as tasks within Straive or amongst our clients.** So, it is not even preparing you for job interviews; it is preparing you for the jobs themselves.

**Anand**: [08:58] But keep in mind that the pace at which things are getting outdated—in a year's time, this material will be outdated. So, it is partly preparing you for what you would be needing to do, but hopefully also preparing you for _how_ to be prepared. When things are changing so rapidly, how do you even deal with that kind of a change?

**Anand**: [09:29] For instance, we removed the course content. We only have evaluations, and we're saying, "Solve it by hook or by crook. I don't care how you solve it, just get there." And the entire internet is there. Why are you relying on me for creating stuff? Which is how it is in the real world, right? We say, **"Get the job done. I don't care how you get the job done."** But in the process, you will hopefully also have learned how to go about figuring out what the task is: Why is the person asking me to do this? What am I supposed to learn in the process? How am I supposed to get it done in the easiest possible way? More importantly, how can I do it by spending the least amount of time so that I have time for other subjects, leisure, etc.? That's the underlying principle that I'm hoping that you will implicitly pick up.

**Anand**: [10:30] In that process, let me ask you a question, and I'd love for you to share your inputs both on the chat window as well as maybe unmute yourself and speak. **What do you think TDS is actually teaching you?** I'd love for you to share in one sentence, please. Just what do you think TDS is actually teaching you? You're welcome to unmute yourself and put it into the chat.

**Rajkumar**: [10:53] Hi Anand, Rajkumar here. So, I think from ground level, if you think, if you have practiced HackerRank, so for a given question we try to come up with an answer.

**Anand**: [11:07] One sentence, Rajkumar, please. And whoever wants to answer next, please just raise your hands. But yeah, Rajkumar.

**Rajkumar**: [11:13] Yeah, basically like **we want to cover all the scenarios that even AI would not think at the first go.**

**Anand**: [11:19] What AI would not think. Good point. Thank you, Rajkumar. Umapathi, what do you think TDS is teaching?

**Umapathi**: [11:27] Sir, it takes the students into another dimension.

**Anand**: [11:31] Okay, one more sentence explaining that.

**Umapathi**: [11:35] The place that is not contemplated by the student.

**Anand**: [11:39] That is very interesting—taking you out of the box in a way. Got you. Thanks, Umapathi. Angad.

**Angad**: [11:47] Yeah, sir. So, from my view, **TDS is actually teaching us how to think, how to solve the problems and work with the data, not only just using the AI tools and all.**

**Anand**: [12:00] Solve problems with data. Fair point. Shishir.

**Shishir**: [12:06] Asking better questions.

**Anand**: [12:11] How to ask better questions. Also a fair point. Thank you. Aditya.

**Aditya**: [12:12] Well, when somebody like you says it, it gives **permission to get things done and move on rather than endlessly diving into theory.** So, it's been useful to get permission from that point of view.

**Anand**: [12:25] Giving you permission to get stuff done. That's a perspective. Thanks, Aditya. Jai.

**Jai**: [12:28] Sir, I think it is to how to approach the problem efficiently and get the answer.

**Anand**: [12:37] Sorry, that wasn't clear. Could you repeat?

**Jai**: [12:41] Sir, I think it is to how to solve the problem efficiently and ask the right question.

**Anand**: [12:47] Solve problems efficiently. That is a fair point. Thank you. And I have a number of responses on the chat as well. Thank you for those—that will be helpful. Aayush?

**Ishaan**: [12:59] Sir, for me, TDS is a course which gives me a full opportunity to collaborate with my peers as well as—so that we could work on the edge cases and if there are bugs, so we can work on it. It is not about simply prompting and getting an answer. It's more or less working on every error which we face.

**Anand**: [13:17] Got you. Fair point. I got the gist of it. And yeah, just keeping it to one sentence. Thank you for that. Ishaan?

**Ishaan**: [13:25] I think TDS somehow gives me an opportunity to work in an environment where **AI is not rather than just your helper, but itself your co-pilot, also what we can say, or your partner which helps you to do your task faster.**

**Anand**: [13:48] Partnering with AI. Got you. No, that is helpful. Thank you. That is a useful set of responses, and I will go through the chat as well.

**Anand**: [14:15] Let me move on then to another question that seems to have popped up in the list of questions that you have submitted, which is: "Any broad career thoughts or musings on the future of technical careers? I'm a non-technical person from a traditional industry wanting to transition into tech or tech-adjacent PM roles."

**Anand**: [14:41] This is an interesting transition point. Tech roles themselves are changing. **The execution part is becoming easy. So, telling agents what to do and figuring out if that is correct is growing. Therefore, the role of someone who is non-tech in the traditional sense, maybe even tech-adjacent, is growing.** And this is exactly what a product manager does, which is telling a team what to do—and in this case, the agent becomes the team—and verifying if they've done their job right. So, in that sense, this role is growing.

**Anand**: [15:20] Somebody who understands the domain obviously has an advantage. But the other thing is, because these are so cheap—what I mean by so cheap is agents can execute things so rapidly—the number of people required who will be specifying what to do will be growing. So, my thought is the number of such roles will grow for what you might notionally call "non-technical people." But what exactly will that look like? I don't know—meaning what is the shape of those roles, I don't know.

**Anand**: [16:02] Let me move on to—hold on. I missed this—I'm losing track of the questions that I have answered. Yeah, there we are. Another question that came up is: "How to tackle complex tasks? And what are the most important student behaviors needed to tackle such tasks and how to work under pressure?" Very fair question.

**Anand**: [16:58] My opinion on this is: people can push themselves behaviorally 5%, 10%, 20% better—maybe even 100%—by sheer individual ability, willpower, etc. But how much ability and how much willpower do you have, and how many courses will you apply that to? **It helps to have something else that supports you.**

**Anand**: [17:31] Agents are one such. You give them a task; they get it done. Before this, programs—you pre-created programs to solve a HackerRank competition. That was an asset; that helped. Just having money helps as an asset because then you can say, "I don't have to worry about cooking; I can eat at a restaurant," and that is one problem solved. "I don't don't have to worry about washing or ironing; I can outsource that," that's another problem solved.

**Anand**: [18:04] Having friends helps; they can teach you things. You can leverage them contributing, therefore, to them, and that compounds. In other words, **any kind of asset that you can build is an advantage.** That is our—that is the environment that we live in. Now, what are assets? Your notes are assets; they keep compounding. Your relationships are assets; that keeps compounding. Literally how you organize the place you work—whether it is your physical environment, or your laptop, or your file system, or whatever—that's an asset that compounds. The way in which you organize your prompts, the kinds of scripts you have, the kinds of tools you have, etc.—that's an asset.

**Anand**: [18:58] The conversation that you may want to have with yourself is: **What are things that put you in a good position to tackle complex tasks without needing much by way of willpower or behavior change? That it happens automatically—that is the best situation.** For example, if you say, "I'm supposed to get up at 6:00 AM to do something." Will you get up by yourself? Maybe, maybe not. So, you set up an alarm; that's an environment—a change to the environment. Will you get up with an alarm? Maybe, maybe not. So, if not, then—well, in my case, I tell my mother to wake me up. That's another kind of a thing. Or you commit to your friends saying, "I will be there." One of them will call you or just come over physically shake you. That's a certain kind of a commitment. If that doesn't work, schedule a timer that will automatically, I don't know, turn on the AC, give you an electric shock at 6:00 AM—whatever. So, you're engineering systems, people, whatever, to get it to solve your problem without you relying on your limited abilities. Is how I think about solving complex problems.

**Anand**: [20:18] Which actually gets me to another question given that we're talking about how one solves complex problems. I'd love for you to tell me about the last or most recent TDS problem where you were properly stuck. What did you do? Open question to everyone. Again, please feel free to raise your hands and try answering in one sentence if you could. What was the last TDS problem that you were really stuck—actually even forget what was the last problem. In the last problem that you were stuck at, what did you do? Please feel free to put it into the chat window and/or raise your hand and I'll invite you one by one. Angad?

**Rajkumar**: [21:11] Yeah, Anand. So, basically, what happens—when it does not give an answer, we will ask it to re—go through the initial questions that were there in the each of the problem given, right? So, we ask AI to go through the questions itself so that it will step back and then think in different angles and perspectives so that it will get those nuances that is hidden in each problem.

**Anand**: [21:38] Got you, Rajkumar. Understood. Thank you. Though please wait for me to call you. Angad?

**Angad**: [21:45] Yeah, sir. So, last time in the GA5, if you remember, the question 11 is the toughest one. So, I got stuck for the two days almost. And at last, honestly, sir, if I say you, we actually put some system which read the link from the TAs that how they are doing. So, actually it—thinking by the brain very extraordinary ways.

**Anand**: [22:15] Got you. That is helpful. Thanks, Angad. Umapathi?

**Umapathi**: [22:21] Sir, if my memory goes correct, that is in GA6, question 7 or 8. I was stuck up really for a long time. I didn't realize that the GitHub repository I created—and I had to create a YAML file which was supposed to be in workflows. I didn't notice it, so it took a long time. But the issue was so small it took a long time for me to realize that.

**Anand**: [22:51] Got you. That is helpful. And we have a few comments. Himanshu says: "Ask friends, make solvers and guides." Varun says: "Spent $700 worth of tokens on Claude Code and shared the solution with others," which amortizes the cost, which is great. Devanarayanan says: "Build tools and use agents to get the human out of the loop," which is interesting. Any other perspectives from anyone? You're welcome to put it on the chat or answer verbally. One sentence would be great.

**Anand**: [23:26] Angad says, "I know that Jaydeep TA sleeps at"—I presumably—"at 3:00 AM." Okay, that is helpful—leveraging the TAs is a good strategy. Aditya's comment is: "Image coordinates or location in the ROE is something you couldn't get." Got you. And I'd love to hear from everyone what you did when you were stuck in each—what is the first thing that you did when you got stuck? Ishaan?

**Ishaan**: [23:55] For the coordinate—one, the question, I was really stuck on that question for around two hours and I was trying to get help from all the AI agents and all, but finally I just asked the Google AI mode—the Google Lens's new AI mode—and it gave me the answer in two seconds. So, that one really helped me in the image rectification of that question.

**Anand**: [24:19] Interesting. I did not realize that Google AI mode probably has excellent street view recognition. Got you. Gaurav?

**Gaurav**: [24:27] Yes, sir. We all got stuck at the GA5 question 11. But our TDS community will work on it, and our wisdom of crowd will solve that problem.

**Anand**: [24:37] Got you—reaching out to the community. Divyanshu?

**Divyanshu**: [24:41] Yes. So, for the picture one in ROE, I luckily got a picture with a landmark. So, that I put it in Google Lens and it directly gave me the landmark, then just lining up the road—that was easy. In GA6, I actually got a picture which was difficult, for even in ROE the people who got it, it was—

---

**Divyanshu**: [00:00] ...people who got it, it was quite difficult to get. And like if you put it in Google Lens, it tells its Vienna. But now pointing the coordinates, it took me an hour searching through the city just to get to that place. Though when I got there, I realized that I was quite stupid when I was looking through it. It could have been done in three to four minutes, but again you have to brute force some things and just be good with street view, how to look at, you know, signs, directions, sun, and everything else to figure out things.

**Devanarayanan**: [00:32] Sir, for the ROE question, I basically used the agent. I told it to use APIs like OpenStreetMap or whatever. And I gave it—this is the interesting part—I gave it a way so it can submit answers. So I take myself out of the loop and I can solve other ROE questions instead of like constantly copy-pasting into there and clicking the button. Which is a powerful technique. Got you.

**Anand**: [00:58] Thank you, folks. That was useful. Let me take some of the other questions. Here's one which says: "TDS looks good, but is it wrong to feel that TDS can be made even more challenging? GA5, 9, and 11—I guess that's GA5 questions 9 and 11—were quite interesting, but maybe in the future the difficulty could be increased a bit more."

**Anand**: [01:31] **The difficulty is a relative thing these days. What I mean by that is somebody who finds that an agent has solved the problem because maybe they've given the right prompt or have set it up the right way or whatever, finds that it's so easy. And somebody else who's figured out that they know the right people to reach out to finds that they have to take zero effort.** And the people who have not yet figured it out are struggling to get to this.

**Anand**: [01:54] For many years, TDS was entirely hackable, meaning all you have to do is figure out what is the API to send a request, what is—the entire set of answer scripts is there, you just put in the score that you want, it will directly save that score. In that sense, every question was hackable. The percentage that hacked it was less than 2%, 3% for almost a year. So I left it open. Now we're finding that that percentage has gone up to about 15-20%, so now it makes sense to make it a little harder.

**Anand**: [02:34] But here's the thing, right? If there is an open exposed endpoint, why are people not able to figure it out? There are many good reasons. They don't have enough friends; the community has not yet formed; they haven't thought about it. But these are also parts of things that we're teaching. When enough—and because the teaching is happening, not just from the TAs or from the content but socially—over time, and the content also gets built—people are building solvers, etc.—as a natural progression, the course becomes harder. It is also becoming harder because agents are becoming more capable anyway. So what was difficult earlier has become easier, so we have to keep pushing the boundaries up.

**Anand**: [03:26] So, the answer to the question is: **Can TDS be made even more challenging? Yes, TDS is continuously becoming more and more challenging.** However, at any given point in time, what we're trying to see is at least some people find it hard enough. And right now—there was a time when TDS was the easiest course. At that point, we had the luxury of making it harder so that we can improve the learning. Right now, TDS is not considered among the easy courses, so we probably won't make it more challenging than the natural pace of students' learning. But yes, short answer is it will continue to become gently tougher and tougher.

**Anand**: [04:12] Question: "What are my opinions on using AI tools because if basics are carried out by students using AI then the student will forego the manual effort? The major thing they learn is to use the tools to get the job done, which seems to be a generic tool." Yes, absolutely. That is in fact what I'm hoping to teach people, which is: **When a tool is available to solve a problem, the industry is not going to pay someone to do the same thing slower, costlier, and at lower quality. Today, many of the things that an agent can do, there is no point for a human to learn to do.**

**Anand**: [05:00] Now, the process can still be valuable. Today we do exercise. Why? I mean, is there anything as pointless as exercise? Lifting a weight—a machine can lift a weight better than you can, but you're doing this to build your muscle. Okay. Similarly, why are we learning mental multiplication when a calculator can do it? Because it trains our mind to think about things in certain ways. The rigor is important. Fair. Knowing roughly where numbers are so that we can spot patterns—that familiarity is important.

**Anand**: [05:43] So, the reason why we learn certain things is sometimes to get the job done, sometimes the process is important. **By allowing AI to do the job instead of us, we are learning how to get the job done. The process of doing it, we are learning a different process—not the process of doing stuff.** Why does TDS focus so much on just getting the job done? Because most of the courses are focusing on the process anyway. There aren't that many courses that are teaching you to get the job done. So I have the luxury of saying 80% anyway you're learning that. So in TDS, let me teach you that 20% which is not often taught. As simple as that. So it's not like you're short of the learning.

**Anand**: [06:33] But I also believe that there is a certain process skill involved in using AI. It is something fundamentally different. It is almost like management. See, **these agents are like people. They're smart enough. Now, management is a subject in itself where you get people to do stuff. In a sense, TDS is management for agents.** And that's the other skill that you're learning.

**Anand**: [06:56] Which I think is also the answer to someone else who asked: "I'm prompting LLMs, getting the answer, and pasting. What do I learn from TDS?" You're learning how to prompt, and paste it, and getting the job done. Now, if it were that easy, you'd be scoring full marks in TDS. If you're not scoring full marks, then there is something that you haven't learned. And if a reasonably large number of people are scoring full marks, we'll then move on to the next stage of learning.

**Anand**: [07:31] And if you yourself are, in fact, getting full marks in most of the subjects, then you are ahead of your peers by far because many people are not getting full marks. So recognize that, A: there is something to learn because others have not yet learned it. B: that this "something to learn," I'm hoping you'll figure out what it is by yourself—I'm not trying to make it explicit, partly because I don't know, partly because it is what I'm recruiting for, that is, your ability to figure it out yourself. Partly because if this you're able to figure out by yourself, you will know what to learn next when the agents are doing more. In other words, good question—figure it out, like many other things in TDS.

**Anand**: [08:15] Yet another... okay, but no, that actually is a good segue for one of my questions which is: **Can you give me an example outside TDS where you would now approach a problem differently because of TDS?** It begs the question, if what we're learning here is prompting, is it transferable elsewhere? So I'd love to hear, are there any places where because of what you have learned in TDS, you are now approaching a problem differently? Please feel free to share. One sentence would be ideal. Chat window and raising your hand, both are great. Ishaan?

**Ishaan**: [09:04] Sir, I think after the course of TDS, asking any question to an LLM or trying to learn a concept of a different subject via asking the LLM changes the way we prompt the LLM. So before the TDS course, I used to give a one-line prompt, but after using the TDS course I try to give a detailed prompt about how the LLM should act and behave and the process of thinking of LLM towards explaining me the concept or anything.

**Anand**: [09:39] Got you. That's helpful, Ishaan. Effectively the notion of more detailed prompts and how to prompt in itself. So which is a fair rebuttal to the earlier point which is: if what we're learning is prompting, what are we learning? The answer is we are learning prompting. Got it. Umapathi?

**Umapathi**: [09:55] Sir, TDS has helped the thought process change. So therefore the other subjects like System Commands, its bit—approach has changed and it has become easier now. And one more thing I'd like to add in this context, earlier probably I was checking in with AI with the full question on System Command for a code. Now I can prompt in three ways—I mean, splitting the question and learning the coding. So now I know the code. Whatever the code it gives, I know what it is and then I understand.

**Anand**: [10:37] Got you. That is helpful. Thank you. Angad?

**Angad**: [10:40] Yeah. So, for example, sir, while actually I'm working on a new project, like an MLP project, it actually helps me to like debugging the unfamiliar codes. Like also when I'm doing sideways, I'm doing bug bounty. So it will also help me to think very differently like how to down the problem and do some new stuffs and new experiments.

**Anand**: [11:07] Interesting that you mentioned bug bounties. That's something that we might be introducing in some of our projects going forward. Thank you, Angad. Devanarayanan?

**Devanarayanan**: [11:15] Good evening, sir. Actually, TDS has taught the attitude that nothing is actually tough. It can all be used with prompts. So earlier, like Telegram bot and all those, deploying seemed to be tough by reading the question statement. But now with the help of LLM, like in every subject, whether it be like Electronics or whatever I study, I can use LLM to build that stamina like nothing is tough, everything is doable, just that it requires rigor to prompt.

**Anand**: [11:40] Got you. Expanding the bounds of possibility, which is something that I'm pushing for. Thank you. Jai?

**Jai**: [11:51] Hello, sir. Am I audible?
**Anand**: [11:53] Yes.
**Jai**: [11:55] So, instead of just prompting, I try to plan the problem, how the AI should approach the problem first.

**Anand**: [12:02] Framing the problem as a transferable skill. Got you. Gaurav?

**Gaurav**: [12:05] By using the TDS knowledge, I crack the MLP project and get the rank two.

**Anand**: [12:13] Transference to other subjects like System Commands. Thank you. Ishaan?

**Ishaan**: [12:17] Sir, earlier we used like a lot of tokens of LLMs, but in the—as the course moved on from week one to week eight and nine, so I used the tokens efficiently and effectively. So TDS helped me in my personal project to use the tokens efficiently without exhausting them. And also the content available on the TDS portal like web scraping and all, it is like useful for everyday stuffs.

**Anand**: [12:47] Specific skills. Got you. Atul?

**Atul**: [12:50] Hello. Good evening, sir. So, I am an economics student and this is my part-time thing, TDS and BS degree. So as an economist, what we usually try to do is to interpret things or analyze things, right? So how Claude or TDS helps me is that human mind in itself is constrained with thought process or something like that. So you build a particular thought and then with that base thought from your mind, you try to level it up with this Claude as well as TDS. That is what how Claude and TDS is helping me.

**Anand**: [13:28] Cross-subject transference is an interesting one. Thank you. Rajkumar?

**Rajkumar**: [13:31] Yeah, there are two things. First thing, reverse engineering. So instead of giving question and then answer, it asks so many questions so that the AI itself will go back and think what could go wrong and then based on that, it will come up with a better answer. And second thing, from the initial weeks, like we were solving just through the ChatGPT UI. Then as the course progressed, like the questions were getting tougher and tougher, we cannot solve through chatbot. Now we started using—I purchased Cursor, so each question I actually create it as a separate project and then save it in my repository so that I can look out later onwards.

**Anand**: [14:11] Agentic techniques. That is powerful. And I see a few responses on the chat as well. Thank you for those; that is helpful.

**Anand**: [14:16] Let me take one of your questions. "How can we verify that an AI-generated answer is correct when we don't have enough expertise to check it ourselves?" **This actually is an age-old problem. How does a judge who knows nothing about engines decide on a patent case between two parties about engines? How does a regulator who knows nothing about the telecom industry pass laws related to regulation?** That actually seems to work reasonably well. How does an auditor who does not understand or have access to all of the information that a company does, come in and still verify and make sure that they're not violating any laws?

**Anand**: [15:02] In all of these cases, we have a variety of techniques and I think the easiest would be for me to share a recent talk that I had delivered at the DataHack Summit this weekend. The theme of the talk is "How do you manage something that is smarter than you?" I'm sharing the link on the chat and I will also share my screen and quickly tell you where it is. You can go to s-anand0.github.io/talks and okay, that's the link.

**Anand**: [15:48] There are a set of techniques which I will see if I—yeah, here we are. I will very briefly go through five fairly popular ways people are finding to verify. **One is a checklist.** Before you go and say, "Here are five things I expect in the answer. Is it working? Does it have it? Does it not?" Sometimes you may not know what the checklists are, but surgeons, pilots, divers etc., they use standard checklists that somebody has given them and you verify against that.

**Anand**: [16:18] **Another is receipts.** Every auditor or a journalist or whatever asks for proof. In our case, we ask for citations. Give me the original links, give me the actual files, show me the actual formulas. And this works very effectively too.

**Anand**: [16:35] **A third thing that we do is audits.** Which is: have someone else come in and inspect. Give the same logic, the input and the output, to another model and have it go through and find all the errors and create an audit list. See where it fails.

**Anand**: [16:51] **A fourth technique is voting.** Ask ten people and see what the majority opinion is. In certain cases that works well, certain problems that's not the solution.

**Anand**: [17:01] **The fifth is practically put an examination together.** That's what we're doing for coding, which is: the system writes—the agent writes a piece of code and then we give it to an interpreter, a compiler, whatever, and say, "Run it." If it passes the exam, great. And this is exactly what's happening with mathematics today, which is the agent says, "I have solved a theorem," and we pass it to Lean, which is a programming language that can verify mathematics, and we get the result.

**Anand**: [17:31] Can we do this for other things? Maybe. For instance, we have... what is the robotics verifier... I forget. For insurance contracts, for instance, there's something called Insurly (I-N-S-U-R-L-Y) by Cambridge, where you can take an insurance policy, convert it to a programming language of sorts, take a claim, convert it to the same language, and then verify if it meets all the clauses or there are exceptions. In short, and these are traditions from several professions in the past, and therefore how we verify need not be a new problem. It's something that we have solved in the past and we can learn from these professions as well. Check out the talk for more details.

**Anand**: [18:28] Question: "Why are we saving the last submission instead of the best submission? Since we can create the server only once and it is hosted on Cloudflare, it may be lost or reset causing all progress to be lost." Absolutely a valid question. Supposing I said we'll take the best response. Here is how I would approach the question. I would create 30 servers, tell it, "Create random answers." One of them is bound to be right. This is not a bad strategy. And I'm not trying to discourage that strategy either. In fact, now that whoever it is has mentioned it, I'm tempted to think, yeah, we'll have a few questions where we will also say, "Get to the answer somehow. Toss 100 answers at it; as long as one of them is right, we will take it." We will take the best for that particular question. I take that as a to-do.

**Anand**: [19:25] But we won't do this for all of them because we also want a certain amount of determinism. **See, the whole point of I'm hosting it in an environment where it might get reset—that is exactly the kind of real-life problem that an agent is not going to be able to solve for you.** And how—today agents can create so much code that creating the code is not the difficult part. It is: how do I get the code to be reliable enough in a deployment environment? So that when I'm running 10 services—agents will create 200 services for me—even 10 services if I can't run in one session, keep them all alive, then I haven't picked up the skill to be able to deal with the deluge of code and services that agents will be able to provide. That is precisely why we're saying all of them should be live at the time of checking. In other words, the thing that is difficult is exactly what we're trying to test if you're able to learn.

**Anand**: [20:25] Jai, if you had a question to post, please just comment on the form and I'll ping the link to the form again for anyone who needs it.

**Anand**: [20:41] Another question is: "In a fast-changing era of AI, lots of things are changing. If someone wanted to go into the future and figure out the direction of AI for at least five to ten years—what's going on now in the industry and taking direction?" I guess the question is: what do you do? I don't know, to whoever who asked this. I am trying to figure out as well and I'm in the same boat as you are. I'll tell you what I'm up to doing. Honestly, I'm waiting and watching. I'm seeing what is happening that is reasonably safe, what is happening that will completely be disrupted.

**Anand**: [21:26] What I'm realizing is I can't predict that far ahead, but there are certain things that I think I can reasonably say for sure because they don't change much. What are things that don't change much? The laws of physics don't change much. So if there is something that is constrained by physics, then I know that no matter how fast things change, technology changes, etc., that will be a reasonably safe space to work in.

**Anand**: [21:58] **If something is constrained by human behavior or by regulation, then that is a reasonably safe haven.** What does that mean? No matter what, drugs will have to go through a rigorous testing. People will not allow a drug to be released just like that. So I know that any kind of drug approval related areas will not change so fast thanks to AI. So I'm—if I want to spend my time learning something, that might be an area that is worth learning.

**Anand**: [22:39] Human behavior is not likely to change too much. Just because AI is going to make things fast, we're not going to start working two hours a day. We are working eight hours a day or however many hours we're working because that is the limit of what we can tolerate. So if AI can do more, we will do more. We won't do less. So human behavior therefore and how to deal with human systems is something that is worth learning. Also, therefore, these are businesses that one can invest in because you have something that is likely to be protected irrespective of technology. That is how I'm thinking about it.

**Anand**: [23:33] In other words, what I'm saying is I'm not even trying very hard to predict where AI will go, I'm trying to predict in areas where AI cannot go and say something along the lines of that therefore AI will not go there. That apart, I'm watching the leading indicators. What are people inside Anthropic saying? What are people inside OpenAI saying? They are able to see things three, six months at least ahead of we are. And if multiple people are saying the same things, I know that that is likely to happen. In short, AI I'm predicting three, six months ahead by following what the leaders are saying, and in safe areas I'm able to predict for longer.

**Anand**: [24:22] Which leads me to a question to you. **If AI becomes ten times better next term, what do you think I should stop teaching and what do you think should remain?** I'll repeat the question. Again, please feel free to put this in the chat. One line is sufficient, or raise your hand, or both. If AI becomes 10x better next term, what should I stop teaching? What should we continue? Any thoughts?

**Anand**: [24:41] Comment from Aayush, which is: "Manual marking should remain," which is a human evaluating it. Got you, Aayush. Jai?

**Jai**: [24:50] Sir, I think you should ask next term the student how should they approach the problem and then ask them to document it.

---

**Jai**: [00:04] If AI becomes ten times smarter, then you should ask students how they would approach the problem and ask them to document the problem and the solution.

**Anand**: [00:16] Document the approach. Got it. Rohit?

**Rohit**: [00:23] So AI becomes ten times smarter, which means it has so much intelligence that it can solve any task. So we can learn how to interact more comprehensively with humans, with masses. That should be done because AI will solve everything.

**Anand**: [00:43] **Interacting with humans rather than with AI.** Got you. Umapathi?

**Umapathi**: [00:51] One thing is that creating Fast API and GitHub URL repository and deploying it—that portion might become irrelevant. What would be more relevant is security. **The security concept will be more relevant as it becomes 10x; the likelihood of exposure will be more and therefore developing the security systems will be more important.**

**Anand**: [01:29] Security at scale is a fair point. Thank you. Uday?

**Uday**: [01:40] In my perspective, you are right in telling that physics basics will not change. In the same way, we should be able to strengthen our basics. Even though if AI is a great kind of thing, 10 times or 100 times, we will not be leaving our ABCDs. In that way, we should be more strong enough to beat out AI.

**Anand**: [02:06] The fundamentals. Got you. Dev?

**Dev**: [02:09] I think like in Project 1, there were questions like: the prompt which will prove the GPT wrong. So if it's 10x smarter, human should be able to build that intelligence to build—so prompting questions should remain and continue as it is.

**Anand**: [02:26] **Specification and verification is where I think you're going.** Got you. Rajkumar?

**Rajkumar**: [02:31] Yeah, I think instead of focusing on some common questions that can be easily solved by AI, we can go for some vague or innovative questions. It might not even have a better answer, but we can score them based on how well they come up with the answer.

**Anand**: [02:51] Effectively evaluating outcomes rather than... yeah. Fair point. Angad?

**Angad**: [03:00] So if AI becomes ten times better, I think we should stop teaching students to compete with the AI on execution. Instead, **I think they should train the students to define the right problems and working with the messy, real-world constraints and also challenge the AI output.** I think that would be better.

**Anand**: [03:26] Good point. Got you. Working alongside rather than against AI. Fair. Let me take up again... yeah. Thanks for those who put in the comments on the chat as well.

**Anand**: [03:38] Let me take up one more question that just popped up. "How do I overcome the TDS mindset creeping into other courses? The first few weeks of the Java assignment after TDS started, I had to constantly resist the urge to paste assignment into an agent and make it solve the assignment for me."

**Anand**: [04:02] This is like asking: how do I resist the urge to get onto social media? Or how do I resist the urge to just play video games, or any kind of behavior? It is no different. Ultimately, you will make choices. And each person makes a different set of choices, and I'm not necessarily saying that that is right or wrong.

**Anand**: [04:26] What I mean is, if you, for instance, just went to the agents fully and said, "You solve it for me," there are certain skills that you will not be picking up. But maybe it doesn't matter. Maybe that is not the area that you end up learning skills in. **Maybe you will spend the time that you save by doing that, learning something else.** What could that be? I don't know. Professional dancing? Who knows? And you switch. Ultimately, we make a set of choices. They may be conscious; they may be unconscious. It takes us in certain ways.

**Anand**: [05:07] So the answer to this question, "How do I overcome the TDS mindset applying elsewhere," is: just think about it. And the only thing that I will ask is: **whatever you're doing, try to do it intentionally.** Meaning, if the mindset comes in, just pause for a second saying, "I'm doing this. Do I want to do this?" And at that moment, if you say yes, do it. If you say no, don't do it.

**Anand**: [05:32] Secondly, write it down. Just say, "At this point, I thought about this, I decided to do X for whatever reason." Try it for a week or two. You'll find journaling to be one of the most powerful techniques. And you can always take that journal, give it to an LLM also, and say, "Here is how I'm deciding, how can I improve myself?" or whatever. But the crux of it is: **when you do something intentionally, at least you know where you're going and why.** If you write it down, you have a record of it. I'm not saying do X or Y, but just do it knowingly. That's all.

**Anand**: [06:13] Another question out here is: "TDS is hard, I loved it, but still traditional coding gets tough. So why is a fresher job hunting still so heavily DSA oriented?" I'm guessing DSA is Data Science and Analytics... [Correction from chat] Data Structures and Algorithms. Thanks, people, for pointing that out for the acronym of DSA.

**Anand**: [06:33] Short answer: the industry has not yet caught on. It takes a long time. How is the industry structured? There is an HR team who gets requirements from the business team, who gets requirements from the client project manager, who gets requirements from the client business stakeholder. If the client business person has not yet caught on to the need for someone with a very different skill profile, the entire chain has not yet caught on.

**Anand**: [07:05] Good part is, everybody is moving rapidly. So I expect the change to happen relatively rapidly compared to other industries. The amount of time that it took for Data Science as a role to catch on was three, four years. **I expect forward-deployed engineers to happen a little faster than that, probably a lot faster.**

**Anand**: [07:31] But the other side of it is, there really is a need for Data Structures and Algorithms because people are vibe-coding like crazy. **A big role that is emerging is "vibe-code fixer."** Somebody has created it, it's a mess—which is a good thing because somebody's finding it to be a useful mess. So somebody has to come and clean up the mess.

**Anand**: [07:56] So basically maintaining code, which people and agents have vomited out and nobody knows what to deal with, will be a growing thing. So I am saying there are two parts to it. **One, the demand for agentic use will grow a lot. But the demand for more structured programming will also grow.** The knowledge of Data Structures and Algorithms will truly be useful.

**Anand**: [08:24] We are almost out of time, so I will probably wrap up with one last question, but keep in mind that I will take all of the questions that you have and reply as an FAQ. "Can AI agents be used as judges or auditors to evaluate the work of human experts, especially in critical or high-stakes domains?"

**Anand**: [08:52] What I'm finding with high-stakes domains is: an easy entry point is actually verification. Meaning, we are doing some work. **If an AI agent comes in and says, "I think you are doing X right and Y wrong because of blah blah blah," it is so high-stakes that I will verify what it's saying as well.** Now it's giving me an extra pair of eyes. I wouldn't mind.

**Anand**: [09:20] **So if we use AI as an additional pair of eyes, it's both harmless and reasonably high impact because it might be able to literally look at things with a fresh perspective.** On the other hand, if I handover to it entirely, that may not be a safe thing to do given how risky it is, unless we watch it very carefully, try it out in a wide variety of scenarios, and even then we'll probably say, "Look, at least I will sample a few of its outputs and see if it's drifting or it's getting some things wrong."

**Anand**: [09:52] The benefit is that in high-stakes industries, you need experts. Experts are expensive. If you can reduce the load of experts by saying, "Look, 50% of the things that you check, this thing can give you... can prepare the stuff for you so that you just have to cursorily glance through it and approve or disapprove," that helps. So all the more reason to apply it in high-stakes industries.

**Anand**: [10:17] Put another way, I don't think the segmentation is: is this high-risk or low-risk industry? **It is: how we break the task. Is it generation or is it verification? Is it preparation of verification or is it approval of verification?** Is it verification of simple tasks? Is it verification of subtle tasks? That's the kind of tradeoff or the split that we may have to take.

**Anand**: [10:42] Now, I am conscious we are out of time. So let me say this: the form, by the way, will vanish in a short while. But what I will do is: for the next 21 days, till the end of the month, my email ID is available and I'll give you a specific email ID: `askai@s-anand.net`. I'm putting this in the chat window.

**Anand**: [11:14] **Any email that any of you send to this will be answered by my agent, which has all the context of TDS and my work and everything else, and you will get a response within a day.** It's a human-in-the-loop; I'm still very much looking at the inputs and outputs, but it saves me bandwidth. In short, `askai@s-anand.net` is an email ID you can send any future questions to and I should be able to get back. Sorry, I would not be able to take more questions or comments, but seriously, thank you so much for joining in. I found it helpful. Bye everyone.