# Transcript

**Anand**: [00:00] We hear about, on the one hand, AI being scary. 30,000 people using auto-generated AI in one day in one company. On the other hand, there are lawyers who are fined because they are using AI, and AI has come up with cases that don't even exist. What kind of a stupid lawyer comes up with a case that does not even exist? And they are calling this hallucinations, and we are afraid of that.

**Anand**: [00:29] Most execs that I speak with, when I ask them about AI, they say, "Oh, we have the most advanced AI used in our organization. We are at the forefront in terms of policies, in terms of using it." Then I ask them what they use. "No, I don't use AI." You're not part of the team? There seems to be a bit of a dichotomy, sometimes even within the same person. And that's okay. It's new technology. How old is it?

**Anand**: [00:55] What I want to do in this session is only one thing. **Have you used AI daily in a new case that you have not tried earlier? That's our sole objective today.** And in order to do that, you will be opening ChatGPT, you will be opening Claude on your phones, on your laptops. Don't bother too much listening to me. What you do with your phones and laptops is more important. This is one of those weird sessions where I'm literally telling you, don't listen to me. Open your laptops, use WhatsApp. In fact, you will be using WhatsApp in the session. And that's another story.

**Anand**: [01:33] The way we'll be running this is, Anand and several of the people at the institute will be helping you in case you run into trouble as we go through these exercises. What I'll be doing here is sharing how I use AI. And the short answer is, **I use AI like an intern. Like a chotu [intern/errand boy]. Sometimes I use AI like a plumber.** And I will tell you what I mean. Like a secretary, like a banker, all kinds of things.

**Anand**: [02:00] And you may think, wait, hold on, a plumber? This was in fact my first major use of AI. I was at Seoul, Korea, where I was stuck in a toilet where the water was just not going through. And clearly something needs to be done about this, but I couldn't find any place where I could press, pull, whatever to open it. Okay, how many of you know how to open this? Just raise your hands. Few of you know how to use it. Exactly, you press it. I had no idea.

**Anand**: [02:31] After 10 minutes, I took a photo, this photo, sent it to ChatGPT and asked it, "How do I open it?" and it said, "Press it." Saved. So the water drains finally, thank heavens. But in the same toilet, the same night, I was stuck a second time because I was trying to figure out where the flush was. And this looked like some flush thing, so I pressed it, and a siren started ringing. Apparently it's an emergency alarm.

**Anand**: [03:00] After a while, it turned off, but the light was still on and there were other flashing lights happening, so I called up the reception, and they were speaking in Korean. I don't know a word of Korean. I tried explaining, look I pressed the emergency thing, whatever it was, nothing wrong. And he was saying something along the lines of, "Water? You want water?"

**Anand**: [03:22] So then, turn on ChatGPT. ChatGPT has a voice mode. So I kept the phone like this and said, "Translate everything that I say into Korean." Kept the phone here and said, "Look, I just pressed the emergency button, but there is no emergency." And it said "[Korean translation]". And he said something "[Korean response]". They were speaking to each other for a minute or so. And then finally he said, "Ah, okay, no problem, no problem."

**Anand**: [03:47] Small little things. But what I am finding is that **voice is a really powerful API. Why bother typing when we can talk to it?** The best part is that in terms of quality of models, ChatGPT does a better job of transcription than Claude or Gemini, so I use it extensively. Gemini speaks best in terms of the intonation and so on, it understands more languages. So there are pros and cons of using these. Claude at the moment is shaky on voice. But it can understand Hindi fairly well, it can understand Tamil fairly well, it can understand Telugu fairly well, it can understand a reasonable number of languages fairly well, including mixes.

**Anand**: [04:28] So here is going to be our first exercise. Let's each of us open ChatGPT on our phones. I do suggest you use your phone, but for whatever reason if you can't use it, then you can use your laptop, that's fine. And spend a few seconds just talking to it, asking for anything. It could be literally—so I'm going to say, "Aaj shaam ko soch raha tha ki Mussoorie mein kahin jaun, lekin pata nahi ki kaun sa jagah accha rahega. Main toh pehle se Mussoorie aaya nahi hoon. Thoda recommend kar dijiye." [Translation: "This evening I was thinking of going somewhere in Mussoorie, but I don't know which place would be good. I haven't been to Mussoorie before. Please recommend a bit."] And that is it.

**Anand**: [05:07] I'm gonna turn off the thinking mode, I'll come to those modes in a while. But you will notice that the translation is probably good. I'm not a great judge of language. But here's the thing, it will probably respond in Hindi, which I don't read anywhere near well. So I'll then have to ask it to translate and figure out the answer. Small little practical things like this. So let's take a one-minute pause. Please speak to your phones. Say anything. Get a feel for it. We'll go ahead.

**Anand**: [05:46] Once you're done, I will request you share your chat on WhatsApp. So please speak something shareable. And we'll go to the WhatsApp group and tell you how you might be able to share it.

**Debjani**: [06:00] [inaudible]

**Anand**: [06:07] Anand, you might want to just tell them how to click the voice button. The button will look like this.

**Audience**: [06:15] [inaudible]

**Audience**: [06:21] [inaudible]

**Anand**: [06:33] It will work on unpaid also, but you will need the paid version on your phone soon. A bit for the other exercises. So use the paid version. Speak to your laptop, it's okay.

**Audience**: [06:48] [inaudible]

**Anand**: [06:49] Yeah, if you use the unpaid version it may not work as well as the paid version.

**Audience**: [06:53] [inaudible]

**Anand**: [06:54] Speak to the laptop. Same button.

**Audience**: [07:05] [inaudible]

**Audience**: [07:11] [inaudible]

**Audience**: [07:44] [inaudible]

**Anand**: [07:45] Yes, that too. And we'll get into a whole bunch of things exactly. But yes, it will work on the unpaid one, but please try on the paid one. There are other features that we have an advantage on.

**Audience**: [07:55] [inaudible]

**Anand**: [08:11] Once you have some messages, here's what I suggest. Click on the share link. That will appear on the top right. And copy the link. Send it to WhatsApp. There is a social group for us. Please send your chat on the WhatsApp.

**Anand**: [08:32] **A big chunk of my discussions with people are: somebody asks me a question, I don't know the answer. I ask ChatGPT, and I send them the answer.** Just click, and I'll look forward to seeing your messages on this group.

**Anand**: [08:51] On the laptop, there is a share button on the top right. Once you click on the share button, there's a copy link. Then you can send that to WhatsApp Web. And if anyone is having trouble with WhatsApp Web, again, please reach out, there are enough people to help you.

**Audience**: [09:07] I'm having trouble with the language.

**Anand**: [09:09] Ah, try English then. On the free version, you may want to just try in English.

**Audience**: [09:14] I am typing... [inaudible]

**Anand**: [09:16] Version is a new one, but it may have misinterpreted the voice. Gemini is better at this obviously.

**Audience**: [09:30] [inaudible]

**Anand**: [09:31] Yes. And we are starting to see some messages come in, that's great. From Ramesh, from Manish, from Robert. Good, three messages. Okay, just keep sending your messages in.

**Anand**: [09:45] **This is one of the things that I play around with, which is using voice. And I think it's one of the most powerful unlocks.** The second major unlock that I find, that gives me a lot of joy, is when I use photos. I was in Singapore. A large number of restaurants serve non-veg and don't necessarily label them, and I'm a vegetarian. So I go to the restaurant, take a photo of the menu, and ask, "**Here is the menu, which of these should I have if I want vegetarian dishes?**" Practical and simple use case.

**Anand**: [10:22] Or I go to a bookshop, take a picture of all the books, and then go home and ask it to convert all the books into proper text and give me recommendations on which ones to read. All it takes is clicking on this plus button, and you will be able to attach a photo from your machine or directly from the camera. Especially on your phone, you'll be able to click on the plus at the bottom left and attach a photo.

**Anand**: [10:52] You can do all kinds of things. Like, for instance, I took a picture of my palm and said, "You should know palmistry, you've read all the books on palmistry. Use palmistry to analyze my hand." And it did. At the end of it, I also asked it, "Look, tell me how right you are, how right you are not." So here is its summary. It says, "A strong-minded, idea-rich, independent, imaginative, self-directed person whose life is organized less around stability"—I don't know if that's a compliment or not—"and more around meaningful mental engagement and chosen paths."

**Anand**: [11:26] And now it starts praising itself. It says, "What's genuinely impressive is that this is also part of the best matches that are visible both in our chats and its public record." So it's saying, "Oh, I'm totally right," and all that. But it won't trust any palmist to try to turn it into exact forecasts. So it's also saying, "Look, don't trust me too much, I'm not going to predict the future or any such thing, but at least I'll tell you what the palmistry books are saying based on your hand."

**Anand**: [11:51] Interesting, but not particularly useful.

**Audience**: [11:54] [inaudible]

**Anand**: [11:57] It's read all the books on palmistry. Including the pictures.

**Audience**: [12:02] [inaudible]

**Anand**: [12:04] Exactly. And based on that, it's running its description. And it also says, for instance, "Your palm looks slightly rectangular rather than square, your fingers are long and slim," that has all of these interpretations. So it has vision. Put another way, **large language models are not just language models. They are vision models as well. They can see.**

**Audience**: [12:28] [inaudible]

**Anand**: [12:29] Absolutely. And that includes pictures of what we have on the screen. So for instance, one of the things that I was doing was taking our schedule. This schedule, which has been distributed probably to all of us, I took a picture of, and said, "Research and find all errors in the schedule." Now what errors can there be in a schedule? We'll find out. We'll come to this in a few minutes.

**Anand**: [13:00] But here is my request. You can take a picture of anything. A document that's lying around you, of anything that's on your screen, go to any website, take a picture of that screen, or any object around us. Ask it anything. Doesn't matter. Even ask it if you can tell me something unusual or interesting about this object that I might not know about. Or, "Here is a screen, convert that screen into text and send it to me as an email." Anything.

**Anand**: [13:31] Exercise number two: Click on the plus button at the bottom left of your WhatsApp screen—wherever that is, my WhatsApp is gone, I say not WhatsApp, ChatGPT sorry, or Claude for that matter, whichever you're comfortable with. When you click on the plus, you will be able to add photos and files or camera or whatever. Select camera, take a picture of something, and ask it to explain what that is. Let's just see if it can see. Give it a shot.

**Audience**: [14:06] Any image?

**Anand**: [14:07] Any image. And if you have something on your phone already, yeah, send that.

**Anand**: [14:28] And once you've done that, please share the conversation. The image will not be shared. So on the WhatsApp group, please share what you've been able to interpret from the image. Keep in mind, your image stays private. That never gets shared. Only the conversation gets shared.

**Audience**: [14:46] So Anand, one excellent example of using an image to understand is understanding medical reports. So I have a senior mother, is 80 plus, so she has to constantly get her tested and half the time I can't understand. So I click a picture and I just upload it and say, "Help explain it to me in simple terms." And it does it so beautifully. So that's a great example of using image.

**Anand**: [15:13] **That is increasingly becoming a very powerful use case, partly of images, partly of medical reports even, that sometimes come to us as PDFs and we send it. What we are finding is that in terms of diagnosis, Claude, ChatGPT, etc. are better than 70% of the doctors.** So, totally.

**Audience**: [15:43] [inaudible]

**Anand**: [15:48] Oh, okay. So that's a photo of a sunset. Nice. And from Robert, we have a bottle. An explanation about a bottle. So as you can see from what I see on my screen, the image does not appear in the shared conversation. The image only appears for you. And what we were doing was merely testing that it can see.

**Anand**: [16:11] The thing is, if you have the more advanced models—and I'll come to the advanced models and what I mean by advanced models in a bit—the quality to which they can see is shocking. Things that you would not be able to see, it can see. Things that obviously we don't know, it might know. So which means that for all practical purposes, it is like us being there on the field and exploring and trying to understand the world. Vision is the second major unlock.

**Anand**: [16:45] **The third major unlock that I found is in the quality of models.** Let me talk a bit about the history of how models have evolved. This is a picture of most of the models, LLM models that exist in the world. Now, the way to read this chart is, the x-axis is the cost of the model. On the right side are the more expensive models, left side are the cheaper models. How is the cost measured? The unit is per million tokens. A token is roughly three-fourths of a word. A million tokens is roughly all of the Harry Potters put together, or the entire King James Bible in one sequence. That's one million tokens.

**Anand**: [17:36] If I paid somebody to read the whole thing end-to-end, how much will it cost? So if I paid a model like Claude 2.1, which is an old model, then it would have cost 8 dollars. On the other hand, if I pay a model like Gemma 3, which is a somewhat newer, not very new model, it would cost 2 cents. Let's look at the difference. 8 dollars. In fact, GPT-4 would cost 30 dollars. 30 dollars, 2 cents. So 30 by 10... 3 dollars, it's cheaper than 10 times. 30 cents... it's cheaper than 100. 3 cents... cheaper than 1000. This is 1,500 times cheaper. **That's roughly the difference between spending 1.5 lakhs versus 100 rupees.**

**Anand**: [18:31] Imagine one person comes to me and says, "I will analyze this entire encyclopedia which is as large as Harry Potter for 1.5 lakhs." The next person comes and says, "It will cost 100 rupees." You'll say, "Boss, that fellow must be doing a better job. You will be doing a terrible job." Is that the case? That brings us to the quality of the models. And the y-axis here is the quality of the models.

**Anand**: [18:58] This has changed dramatically over time. And the scale that we are looking at is: somebody who has the intelligence of a high school student, high school graduate, college first-year student, college graduate, master's student, PhD candidate, tenured professor. And these models have evolved over time. In March '23, we had models that were about as smart as a high school student.

**Anand**: [19:22] Then gently the models evolved. Over time, we had a big leap with GPT-4, where it was as good as a college student. Junior college student. That evolved, and then in—the cost of these models fell a little. GPT-4o Mini came, and was about as smart as the same college student, but at a significantly lower cost. So the cost frontier is also constantly moving. Then o1-preview came, as good as a master's student. Then we had GPT-4.5 preview, which is as good as a PhD candidate. And around June '25, several months ago, we had Gemini 2.5 Pro, which was about as good as a tenured professor on average.

**Anand**: [20:04] And models are getting even better. **Today, o1 thinking is better than a tenured professor on average.** And the cost of this intelligence is falling roughly 10 times every year. Which means that, today what is this cost? About 5 dollars per million tokens. Pay 5 dollars, 800 rupees—400 rupees is it? 450 rupees. And you would get a tenured professor to read a document and analyze something as large as all the Harry Potters put together. 450 rupees. And next year that will become 45 rupees. The year after that, that will become 4.5 rupees.

**Audience**: [20:49] How is it calculated?

**Anand**: [20:51] The cost is the amount of... the number of words that it will read, and how much companies like Anthropic or OpenAI etc. will charge you. And if you take a subscription, they offer a lump sum and don't bother counting too much, as long as you don't use it too crazily. But you can also write a program to use it, to do it in bulk at scale. For that, they charge on a per-word basis. They call words as tokens. It is roughly three-fourths of a word. And therefore the volume of text or images is what matters.

**Anand**: [21:26] Now what this means is that depending on the model that you're using, you may be talking to a college junior student, or you may be talking to a professor. Obviously there's going to be a difference in the quality of output. And you want the higher intelligence. Especially when it costs a tiny amount. **There is no better investment today that I can think of than the $20 that one pays for ChatGPT or Claude. You effectively have a thousand professors sitting in your phone.** You have a thousand professors sitting in your laptop. Now imagination becomes the bottleneck. How do we use them?

**Anand**: [22:05] But let us make sure that we use the good ones. Not the high school students sitting in our pocket. How do we do that? To do this, the first thing that we need to do is choose the model. I'll explain how you do this on ChatGPT, then I'll explain how you do it on Claude. The process is the same on phone or laptop, please make the change on both.

**Anand**: [22:26] On ChatGPT, first you have to choose "Thinking". That is pretty much the only change. At the bottom, after you choose thinking, you may see either "Standard" or "Extended". If you choose "Extended", it will think for longer. If you choose "Standard", it will think for a shorter duration. Up to you. **I permanently leave it at Extended, which means that if I ask ChatGPT anything, it will take a good 10 minutes to come back with the answer, but I'm perfectly happy because I get one of the best possible, most diligent, sometimes un-understandable, but definitely correct answers.** So, select the drop-down under ChatGPT, choose "Thinking". That is the important step as far as ChatGPT goes.

**Anand**: [23:14] For Claude, it is slightly different. Claude has three models and a thinking option. Haiku is the fastest. And you can use it left, right, and center. But Claude is generally fast, there's no problem using some of the higher models. **Sonnet is probably the default that you should use. I'll repeat, when you are using Claude, stick to Sonnet. That is almost as good as the best and pretty fast.** It is very rare that you will have to go outside of Sonnet.

**Anand**: [23:48] But if you're solving research-level problems—and when I say research-level, I mean cutting-edge research that nobody has done—then you want to go for Opus. The trouble with Opus is that you can probably use it only twice or thrice a day. After that, it will say, "No, no, no, you cannot use anymore of it," and start limiting you. Fall back to Sonnet. And if you want very quick answers and are not too fussed about it, then Haiku. Frankly, I never go to anything other than Sonnet, and on ChatGPT I never go to anything other than Thinking with Extended Thinking. Barring rare exceptions.

**Anand**: [24:26] This is a one-time change. Once you have set this up, you are now unlocked to go for the higher levels of intelligence. So, let's go back then to one of the things that I was exploring. I said I took a photo of the schedule and asked it, "Are there any mistakes in the schedule?"

**Audience**: [24:46] Thinking is not coming.

**Anand**: [24:48] Okay, then you don't have... ah, if you don't have Thinking, you don't have the paid version. Could someone please help... they need to log into the paid versions.

**Audience**: [24:58] [inaudible]

---

**Anand**: [00:00] We'll pause for a few minutes. Let's make sure that everyone gets onto the option of thinking.

**Debjani**: [00:11] Anand as you pause, just a disclaimer for everyone. I think what you are suggesting here in terms of whether Claude or ChatGPT, it is for personal use. I think it's an important disclaimer since we are all [inaudible]. So for the workshop and also for personal use if they so choose to. But I think for official purposes the preference today that MeitY has come out with etc is Sarvam etc. And I think there will be a panel that happens, departments will know what to use. So it's good to just clarify that this is not something you're recommending for official use.

**Anand**: [01:20] So it looks like the majority of you have a Go subscription. I'll request, and please do request, that you switch to the Pro version. But now you're talking to a policy...

**Audience**: [01:31] [inaudible]

**Anand**: [01:35] Yes, exactly. For those of you who have the Plus version, it will show you thinking. For those of you who are on the Go version, it won't yet show you thinking. It's very worth upgrading to Plus. But Go also, by default, will give you the higher level of thinking for a few hours and then the next day again for a few hours, next day again for a few hours, etc.

**Audience**: [02:00] So it's showing only thinking... 10 minutes a day...

**Anand**: [02:04] If you are using only 10 minutes a day, then it will by default choose the smarter version. It's okay. But soon you will want to use the higher ones.

**Anand**: [02:26] Okay, let's gently move. If you have Pro or Plus, great. If not, that's still fine because for the workshop, even the Go version is good enough for you.

**Anand**: [02:42] So let's take this. I took a photo of our schedule for this week and asked it the following question.

**Anand**: [03:00] Extended thinking is not great. Which is not too bad. So what the thinking mode means is it will try whatever is the level of thinking it thinks is required. It will auto-adjust. Not a problem.

**Anand**: [04:44] Thinking does a much better job of dealing with the intricacies. It's smarter.

**Anand**: [05:10] So here's a question that's coming up from many people. What kind of questions should I ask? That is a good question. That is the question you should ask. Literally. What kind of questions can I ask? Now here's the thing. It is as smart as a professor. And you don't need to feel ashamed about talking to the machine. So you go to the professor and say, "I don't know how to use you. What sort of questions should I ask you?"

**Audience**: [05:41] But it might help to give some context.

**Anand**: [05:46] Exactly. Now obviously it will say there are some 10 things that I can do. But I don't know you well enough. So maybe it is generic at first. If you introduce yourself, this is what I do, this is what I am interested in, maybe it will give better questions. Over time it will learn about you. The more you use, the more it remembers you, the more it will give you answers along the lines of what you want. To the point where in a few weeks you'll be able to say, "Analyze me, what kind of a person am I?" And it gives a pretty good answer.

**Anand**: [06:26] Okay, we are going to go back to the session. So one thing that we can do that's very powerful is fact-checking. And what I mean by fact-checking is taking any document, you can upload a file, you can upload a photo, you can upload anything you want, or even just copy-paste things, and tell it, "Tell me all the errors in this." Supposing you have a policy document that somebody's given you. Is this going to be able to give you the errors in that document? Supposing somebody's announced something, what are all the mistakes that it has made?

**Anand**: [06:58] So let's take this schedule. After 7 minutes and 43 seconds, it says, blah blah blah, it is certain about a whole bunch of things. And eventually it says, one of the errors that it thinks it found is Dr. V. Anantha Nageswaran is the person who's delivering one of these talks, not Rajesh Nageswaran. Then when I looked at the schedule, I didn't find any Rajesh Nageswaran.

**Anand**: [07:31] So my next point was, double-check. You misread, and it misread two things. It said there is a Rajesh Agarwal who is misspelled, and that there is an Anantha Nageswaran who is not there in the original. So he said both of them are there. But then it also said something about a Vijay Raghavan. So Professor Vijay Raghavan, it says, do you put a space out here, or do you not put a space out here? It is saying that his official pages are using Vijay Raghavan without a space. So maybe that is a small spelling mistake, it's not really sure.

**Anand**: [08:08] Now, after seven minutes it has gone through all the speakers, it has analyzed the content, and it has managed to get to this point where it's saying a few things which are clearly wrong. But errors are very easy to spot and fix. Meaning, when it has made a mistake like this, so I told it, "Do another pass, find more errors, fact check." The result of that is it's saying, "Look, I have confirmed an error. Professor K Vijay Raghavan should be K Vijay Raghavan without the space. His official government and institutional pages consistently use Vijay Raghavan." And here's an important thing. It provides citations. Here is page one, here is page two, here is page three. You click on each of these, it will open that particular page, and you can verify. Which means that fact checking becomes a whole lot easier. So when Professor Vijay Raghavan's here, please let him know that his name has been misspelled with a space even though he prefers not to spell it with a space. Or change the schedule.

**Anand**: [09:07] Now, tiny little things like this. And we find that it can make mistakes. But **the beauty of this process is that mistake spotting is harmless. If it gets it right, good, we have a benefit. If it doesn't get it right, no harm done.** Anyway, we are going to do a double check to make sure, and we are not going to blindly correct something that it says.

**Anand**: [09:31] So in that context, let us take this Press Information Bureau release, which was at 9:37 AM this morning, about the Pradhan Mantri Mudra Yojana. I ran the same process. Copied this, put it into ChatGPT, and said, "Fact check this press release. Cross-check all claims against authoritative sources. Where do the claims align? Where do the claims diverge? Where is the methodology unclear?" Now, you might think, wait, how do you know to ask this particular question? Actually, I didn't know to ask this question. Earlier, I just used to say, "Fact check." After that, it used to tell me, "I will do three things. First, I will check where the claims are aligning, where they are diverging, where the methodology is unclear." So I just copy-pasted this from one of its responses from earlier. We'll talk about this. It's called meta-prompting. When you don't know how to prompt, how to ask a question, you ask it how to ask the question, it will tell you, and copy-paste the question, you ask it, it will give you a good answer. And meta-prompting works beautifully.

**Anand**: [10:36] But for now, it is just a fact-check where I copy the text. And here's what it finds. It's saying, "Here are the bunch of things where the release aligns with official sources." That is not interesting. "Where the release diverges. It is saying that the Yojana has disbursed 40 lakh crores through 57.79 lakh crores [crore accounts?], but in the same release, the year-wise table explicitly labels the year-wise sanction amount. And the sanctioned amount, not the disbursed amount, is close to this amount. So maybe the headline and the quote which is supposed to be sanctioned, they've misrepresented it as disbursed." Then it's saying something else, something else, something else. Now at this point, I don't have the patience to sit and read all of this.

**Anand**: [11:28] I do what Henry Kissinger does. Henry Kissinger had this habit. Whenever somebody would submit a report to him, he would ask the person, "Is this your best work?" "Take it, plus go back. Let me come back to you." They'll go double-check, triple-check, make sure they make a whole bunch of corrections. They'll go back to him. Same question. "Is this your best work?" They'll go back late nights, log, come back. Same question. "Is this your best work?" "Sir, yes. At this point, this really is my best." Then he will read it. Why waste time?

**Anand**: [12:11] I therefore did not read this at all. I just posted the same comment last time. "Double-check all the items you mentioned as diverging. Fact check, did you flag them correctly? Revise if required and tell me any additional mistakes that you find." So now it thinks for another four minutes. And it says yes. The first one I found is in fact correct, not only correct, stronger. This press release has a mistake. It says disbursed over 40 lakh crores through 57 crores, but it should actually have been sanctioned, and the table in that same press release says sanction. This is a mistake. Minor one probably, sanctioned, disbursed, okay, fine, wording difference.

**Anand**: [12:49] **But the cost of checking it is 10 minutes of its time. My time is not at all wasted.** Now if you know this, you can start asking your team, "Have you checked this through AI? Have you checked this through a better model?" And you will run a cross-check. Ask it to find all the errors. No matter what anybody sends, ask it what are the errors in this. Something will come up. Give it back to them. But it also says some of the things that I flagged are not actually correct. Earlier it says I treated up to 20 lakhs as a divergence, that's not quite right. I further checked and the official sources say that now loans can go up to 20 lakhs. So this is actually correct. Here's something else that does still look wrong. Here's something else that does not look so wrong and so on. So a second check gives you better results. And on additional checking, it's also finding a few small little things. But net-net, **the key thing that I want to flag off is fact-checking is a power tool with AI. Let it make mistakes. You can anyway check if it's made a mistake or not if it has spotted an error, it is a bonus for you.**

**Anand**: [14:04] So that's what we are going to do next as a small exercise.

**Audience**: [14:07] [inaudible] same documents?

**Anand**: [14:09] It can check on, it can search on the internet. You don't even have to tell it, it will automatically search on the internet.

**Audience**: [14:15] [inaudible] any additional document?

**Anand**: [14:16] Any additional document that it can find. So it's effectively a cross-check against all public information. So let's do this as an exercise. Take anything. It could be a public website. You can copy-paste that and ask it, fact-check this. Or it could be something from a document that you have received. Copy-paste, and ask it, fact-check this. It could be some news that you've read this morning. Copy-paste, ask it, fact-check this. The only thing that you need to do is say "fact-check". Just those two words are enough. Fact check, and then paste whatever you want to paste. It will do the thinking.

**Debjani**: [14:55] Just don't paste any sensitive or official document, please.

**Anand**: [14:58] Exactly. Whatever is public is always safer. And it will check against all its knowledge and all public records. Give it a shot. And once you've done that, please share that on the chat. Again, keeping in mind that you want only public material to be shared. It will be good to see on WhatsApp what it is that it's managed to find in hopefully official sources that is actually wrong.

**Audience**: [15:24] I have a question. In thinking mode, I asked the questions, it's still generating... like from five minutes ago, it's still generating a report. So, I mean it's good, it could help me, so how do I save it and then move on to the new assignment?

**Anand**: [15:40] Yes. When it's doing something, to create a new chat, on the top left, you will see a new chat on the laptop. On your phone also, somewhere near the top left, you will see a new chat. And they will all run in parallel. You can have 10 chats running simultaneously. Good. You have more minions working for you. It's a great idea to run a whole bunch of chats in parallel. Just ask question number one, then question number two, then question number three. Why waste our time? Let it do the work. When it's done, it'll send a notification on the laptop or even on the phone, or even here you'll be able to see a little icon that indicates that it's done. But the exercise we are trying now is fact-checking something.

**Audience**: [16:22] Sir, older won't get erased?

**Anand**: [16:24] No, it won't get erased. All the previous chats will be saved.

**Audience**: [16:27] Will it always be saved in the history?

**Anand**: [16:28] All previous chats will be saved in history. Unless you delete it, it will stay there. And you can delete whatever you want.

**Audience**: [16:38] Sir does it suggest the data or other data?

**Anand**: [16:43] Yes, it can pick up data from various sources. If there are Excel files that are published, CSV files that are published, it can even download and cross-check against that. And yes, it will tell you, here are a bunch of other places that we can verify. Correct.

**Audience**: [16:58] This is very interesting. I asked it a fact-check about a document on forest cover and all. So initially it is saying 71%, which is right. But at the same time it is suggesting to use 45% of green cover. And that is equally true.

**Anand**: [17:15] Interesting. So it's saying that in this particular document, here is a better reference to use than the original one. You know, that is a good point. It's not something that I was covering in the workshop, but it's not just about finding the error, it is also about correcting the error. And in some cases it has that information. That's a very powerful use. Very true. Please do share this on the chat. On WhatsApp. I am sure all of us would love to see how this works.

**Anand**: [17:50] Okay, in some cases, you may be sharing a conversation that others will not be able to see. Let me explain. In the link, only if you click on the share and share the link will others be able to see it. If you directly copy the link and share it, that link is only visible for you, not for others. And a way of verifying that is if the link has a /s or a /share, chatgpt.com/s or chatgpt.com/share, in those cases they are shared links, not otherwise.

**Anand**: [18:31] Okay, or sharing a photo is also fine. So this is doing an analysis of, it's not clear what it was analyzing. So this came from... Karunaji, if you could share the chat on the top right rather than the photo, that will give us a much richer conversation. And what I'm hoping we'll all be able to do is go through each other's conversations, see what others are able to use it for. What is coming out of that? Is that useful for us? We will get ideas. That's part of the purpose of this sharing.

**Audience**: [19:15] This screenshot, it is coming [inaudible]? That is why I put the picture.

**Anand**: [19:20] Ah, instead what you could do is go to that conversation and click on the share at the top right. When you click on this, you will get a link, share that link. That is better than sharing the screenshot.

**Anand**: [19:37] On the top right you will see the share. As you fact-check, please put those on WhatsApp. I'll wait.

**Audience**: [19:48] For fact-checking, do we have to post an image or a document?

**Anand**: [19:52] Documents are better because an image has less information to fact-check. But it can fact-check an image also. Documents are better.

**Audience**: [20:00] Suppose we are checking some drawing of a machine. So whether that drawing is correct or not.

**Anand**: [20:06] Yes, in that case, if you're fact-checking against a drawing, yes, it is better to show a picture. But in that case, you probably got the drawing in a document, maybe a PDF. I would upload that PDF, as long as it's not sensitive.

**Audience**: [20:19] Does it check the rules and bylaws of something?

**Anand**: [20:23] It does a pretty good job of checking for the, yes, against the bylaws.

**Anand**: [20:30] Correct. So if you take an architectural plan and upload it and ask it to find all the errors, it will definitely find some errors. But in different professions, it has different levels of intelligence. So this is probably a good time to tell you about how smart it is in different areas.

**Anand**: [20:53] OpenAI conducted a study. This was in August last year, therefore is close to nine months old. They took the various professions in the US. Each box represents one profession. The size of the box represents the total salary for that profession. The color represents whether AI is better or the experts in that profession are better. Green means AI has beaten humans. Red means humans are beating AI. Accountants and auditors, for instance. When they were given tasks like "you are the auditor of an audit engagement, you have to run an audit" and it asks for full details and several attachments were given to it. That's an example of a task. These tasks were written by experts, executed by experts, evaluated by experts, and all different experts. And also executed by AI. Then experts evaluated, is AI doing a better job? Are humans doing a better job? When writing software, 70% of the time one of the better models at that time, O3-Mini, was better than the best humans.

**Anand**: [22:04] So clearly we are already at a situation where software developers versus humans for these kinds of tasks, the software developer's coding work is done better by AI. Now that does not mean, as Adrian had mentioned earlier, that all tasks that a software developer does, AI does better. Software development, for a software developer, AI does better. Financial managers are still doing better. Industrial engineers are definitely better at many tasks. And civil engineering and such verifications, it is still not doing a great job.

**Anand**: [22:37] How good is it? Maybe about 30% today. Meaning it will find errors. Some of the errors that it catches are not right. Some of the things that it finds even an expert may not find. But it is almost free. So as an additional cross-check it is useful.

**Audience**: [22:57] One of the risks is that you might [inaudible]. Because it's good that the mistakes you can't find, it can find. That is a great thing. But over a period of time, you get used to like, if mistakes it points out, you can accept it as a truth.

**Anand**: [23:12] Very true.

**Audience**: [23:13] And you don't have always the time to keep waiting for the issues that you ultimately will doubt yourself in the end whether the original was correct or double check was correct.

**Anand**: [23:20] Correct. And this is not unusual for technology. AI is, at least LLMs in the popular world, are about three years old. And we find that they make several mistakes. GPS used to make several mistakes in the beginnings. I remember once it led me literally to the middle of a driveway in Manipal. How? I was giving auto driver instructions. He said, "Sir, there is nothing here. We cannot go this place." I said, "No, no, no, Google Maps is telling me go here, you follow the direction." I had to pay 150 rupees extra. But over time it slowly improved. My navigation skill worsened. I don't need to be as smart as the auto driver in terms of the navigation skills. They need to know. I don't.

**Anand**: [24:08] So with this, two things are happening. Number one, the technology is not perfect, far from it. Two, therefore humans need to be accountable still. You need to check. It doesn't matter whether AI is telling you, or your assistant is telling you, or your brain is telling you, ultimately if you are signing it, it is your responsibility. That accountability is something that we have, we need to have, nothing can take that away from us. Thirdly, AI is improving. So there will be many areas where many people will say, "See, for me, I don't really care if it gives me a wrong restaurant to eat at, worst case I spend one meal." Big deal. So I will not cross-check and verify that. I will not call up somebody who can recommend a good restaurant. To visit some place in Mussoorie in the evening, I will not necessarily double-check. So for some things I will blindly trust AI. Some things I will not. And even there...

---

**Anand**: [00:00] Even there, he may check a chart like what we saw with ChatGPT right now. Is it a software developer? Therefore, is it writing code? Then okay, I may trust it. But is it handling my accounts? I may not because it is not very good enough. Civil engineering, high risk, I may not. So we have a gradation of where to take its word blindly because either it is very good and/or we don't care that much about the consequence, it's easily reversible. Over time it will keep improving and when that happens we will change our map accordingly.

**Anand**: [00:41] That is interesting. Okay. We have few more fact checks. Seven more fact checks actually, that is great. Let's continue with the discussion onto the next topic after fact checking. **One of the possibly most powerful capabilities is its ability to read unstructured text.** Unstructured anything for that matter, but particularly text. Our WhatsApp groups for instance. I have a WhatsApp group of the IIM Bangalore batch that I am part of somewhere. Yeah, here we go. And this is a group that has lots and lots of chatter, I don't really read it regularly. I need some way of catching up with this group.

**Anand**: [01:30] Luckily, on the phone, WhatsApp has an export chat feature. It appears on the top right. It won't appear on the laptop, but it will appear on the phone. What I did was I clicked on export chat, sent it to ChatGPT, and asked it only one question: "Summarize the last three months of discussions." I've barely been on this group and I have no idea what's happening. Tell me what is happening. And it tells me over the last... I thought for a good two minutes. Very deep thinker, especially when I've said extended thinking. Over the last three months, there were three things that happened. A big congratulatory burst. Prachi, who's one of my classmates, had a dance milestone. Karan Bhagat, who's one of my classmates, he had a major achievement of some kind. And then there was an annual award for outstanding contribution of something, and then Prachi was again recognized for a teaching-related thing. Then there was the reunion that's happening, our 25th-year reunion is happening soon. A bunch of other notable threads.

**Anand**: [02:33] Useful. Instead of having to go through something close to a thousand messages, I get a sense of what's happening with my classmates. Now I can quickly chat, "Reunion, who's coming? Should I also come? Where should we book? What should we plan for?" Very simple. Then I said, why stop here? If I have all these messages, let's see what else I can dig into this and find out. I said, "Chart the message volume from inception." From the beginning to the end, give me a sense of how many messages are there. Did it go up? Did it go down? And here is the interesting thing. **This is a really powerful capability of AI. It can write programs.** Programming languages are languages. These are large language models, so they are very good at writing programs. But they also run programs. So it wrote a program, it ran the program, and it created a chart that this is the monthly message volume for my classmates. Now this is interesting. I'm trying to click on this, but it is a little slow, so... okay, something went crazy there. So I'm not going to try and reload it, you're just going to have to see the right side of my screen. Oh here it is, okay.

**Anand**: [03:48] So now there is a huge spike. Normal activity, sudden spike, and then overall activity has generally increased. Something has happened in this group, and I'm very curious as to what happened at that particular time. So that was pretty much my next question, which is, "There is this spike, tell me about that spike." No fancy prompting at all, just tell me what happened. Then it says, what happened. A member, P Chidambaram, my classmate, not *the* P Chidambaram, began posting a large number of long, off-topic messages and links, and many of those messages were deleted. And then people started debating what is the norm, free speech, whether he should stay in the group. A few members became irritated enough to leave the group temporarily, somebody removed him. This is like hardcore drama and I missed it!

**Anand**: [04:40] Now here is the thing, so much is happening to a group that I am interested in, and this happens to be the way I end up catching up. We don't have time for many things, but there are interesting things happening in areas that we are concerned about, and here is a way of catching up with this. So here is my next ask of you, which will come in a few minutes, so think about it in the meantime while we run through a few other things. Find a WhatsApp group that you have not been following that you are interested in, and export it directly to ChatGPT. You can do that. And ask it anything. "Tell me what's happening." But we'll do that in a few minutes. What I'm going to do is share a bit more of the kinds of things that you can do with this. Then I said convert this into a presentation. Tell me all the things that are happening and so on, but we'll come to that.

**Anand**: [05:30] Next. I took the list of attendees for this event. Effectively, all of you. And fed it as context. The theme that I'm coming to is **when we start giving it information, the things that it can do with it are pretty powerful.** WhatsApp chat happens to be information that it does not have, it is your information. But when you give *it* that information, it can do things that you might not expect. And this is another piece of information. It obviously does not know who are the attendees for this session. I have a piece of paper, this is not sensitive. I take a picture, I send it to ChatGPT, and ask, "Convert these three photos into a list." Basically, give me a list of all the attendees. Rather than on paper, I want it on my phone. So it does that. All of these, every single one. Very useful. Because then I can ask the next question, which is, "Can you convert this..." Then I said, "Make sure there are no mistakes." I usually do that, and it spent 10 minutes and did a fact check of itself and gave me the result. Then I said, "Give this to me as a CSV that I can upload to my Google contacts." And after eight minutes it gave me that. So all of you are now part of my contacts. Thank you for joining my phone. But that's it, right? You go to an event, take a photo of the attendees, tell ChatGPT, half an hour later they're all there. 20 contact cards, everybody gives you their business cards. Put all the business cards down there, take a photo, tell it put it into my Gmail or whatever contacts list that you have. It gets the job done. We are just giving it context in any shape or form, and when you give it that context, the information that it has is completely amplified.

**Anand**: [07:27] Let's try that with WhatsApp. Take any WhatsApp group that you're interested in. Put it into ChatGPT and ask it to analyze. Ask it any question you like. And this might be better if you just verbally share what you've learned rather than putting that in chat because this is absolutely non-public information. You'll be surprised what you find about groups that you're not tracking regularly.

**Audience 1**: [08:05] Sir, I'm just wondering, can this be done to, say, a list of our Aadhaar cards also? And then...

**Anand**: [08:16] Absolutely. Yes. Actually take any data, I'm just giving examples. The WhatsApp chats are fun, but really, this works on anything. Give it context that it doesn't have. Tell it to do something, it will do it. Sometimes I feel a little bad talking to it like an 'it'. It has the intelligence of a tenured professor. And I'm asking it to write down contact details and things like that. It doesn't mind it.

**Anand**: [09:00] Whenever anyone finds anything interesting about the WhatsApp conversation or anything else for that matter, please do share.

**Audience 2**: [09:15] Export chat kahan par hai? [English translation: Where is export chat?]

**Audience Member**: [09:17] Yahi toh pooch rahe hain. [English translation: That's what we are asking.]

**Anand**: [09:24] You have the three dots on the top right.

**Audience 3**: [10:00] I didn't find export chat.

**Anand**: [10:02] Didn't find export chat? Then you click on the group list. Go to a chat. In that chat, you will find it.

**Audience 4**: [11:21] Sir, just like human beings have muscle memory, can it recall any missing links... suppose we feed some document or a report, and we delete the content because we run out of capacity. But will it have a certain memory of the report that we uploaded once we delete it?

**Anand**: [11:47] If we delete... If we delete a chat, 99.9% of its memory is gone. There is always a 0.1% that it might have captured but it's very tiny.

**Anand**: [12:15] Please share, please share. We are looking forward. Do share it on the mic please. Please come here, use the mic.

**Audience 5**: [12:28] [inaudible]

**Anand**: [13:13] Okay, we have new shares coming in. Let's... okay, and from Nikam ji on Claude. Okay, that's an update, Punjab, Haryana, Uttarakhand group. And we know who are the active members, what are the dominant themes, blah blah blah. For a discussion like this, it is going to be very interesting to ask a slightly more powerful question, which is, "Narrate the most exciting thing like a short story." Ask it.

**Audience 6**: [14:00] What is the basic differentiation of Claude? What is the basic advantage here? Because we are used to ChatGPT, we have accounts there. What are the features we can try with Claude maybe?

**Anand**: [14:20] You can try Claude directly. I'm just showing you on ChatGPT because I... for no reason actually. Claude will do almost the same things, sometimes better, sometimes not as good, but on average about the same. So that's another thing to generally know about which model to use when. Rule number one is **always use the paid models. The best paid models, that is always better.** Between the top paid models, there are subtle differences. Here is the only rule of thumb that I use. **Claude has style. ChatGPT has rigor. If I want it correct, I go to ChatGPT. If I want it classy, I go to Claude. That's it.**

**Audience 7**: [15:08] Classy is better or rigor is better?

**Audience Member**: [15:13] Actually ChatGPT says that Claude has the best reasoning ability as compared to Gemini or compared to ChatGPT.

**Anand**: [15:20] That is true. Though reasoning comes in many forms. If I had to ask a policy question, I will go to Claude. Its level of thinking is fantastic. But if you said, "Would it have made a mistake?", I'll say yeah, 1% chance. If I want that down to a 0.1% chance, I would take it to ChatGPT and say, "You cross-verify." And that is a useful technique. Take one model's output, give it to the other model. "This guy said this. Find the errors." That guy found all of these errors, you correct it. Play them one against the other. Ultimately they are just two things sitting in your phone, and we are just doing copy-paste.

**Audience 8**: [16:01] Is that prompt engineering?

**Anand**: [16:03] That is prompt engineering. It is just about learning how we get the most out of these. So what you've been doing and we already have a bunch of results that are coming in that are fantastic, is you've started giving it context. Now, this particular field which actually used to be called prompt engineering is increasingly being called context engineering. And there is a reason why it's called context engineering. Prompts are what we type. Context includes what we upload. And what we upload is becoming very important. WhatsApp chats, fantastic source. Upload analyst report, fantastic source. And this is stuff that it does not have, it is beyond what we can tell it. So what you're doing now is context engineering. Do more of it, give it more context. It can process thousands of documents. Well, you can't upload thousands of documents, you can only upload 10, but you can zip them, and then it will process it. So...

**Audience 9**: [17:15] Will this data be shared? If you use the paid version, will it be available to [inaudible]?

**Anand**: [17:21] That is the next question. Will this be shared, who will it be shared with? If you are on the paid version, it will not be shared, it will not be trained on. If you are on the free version, it will be shared, not shared, but it will be trained on. It will not appear for anyone else, but they will take that information, use this as the basis for future training. Therefore nothing explicit from there will go out, but it will learn from this conversation.

**Audience 9**: [17:49] Not for others? Are our documents available to others?

**Anand**: [17:51] Absolutely not. If you are on the paid version, it won't even go to ChatGPT. If you are on the free version, it just won't go outside of ChatGPT.

**Audience 10**: [18:00] Let's assume on the health data, if we give our medical reports and stuff like that. The interpretation is much better than the doctors.

**Anand**: [18:11] Yes.

**Audience 10**: [18:12] The problem is that again the basic question arises. Whether we can trust it or trust a doctor.

**Anand**: [18:17] Correct. And what I usually do is the same as what I used to do with Google. I would ask the doctor, Google is telling me this. Doctor will get irritated, but they have to answer the question. So then, after a few months the doctor learns to search Google. So I have trained the doctor and I have gotten better medical advice. So it serves both our purposes.

**Audience Member**: [18:40] Ha ha, hum doctor ko padha rahe iske liye. [English translation: Ha ha, we are teaching the doctor because of this.]

**Anand**: [18:46] So... right. Next thing that I did. Taking all of you as my guinea pigs, is deep research. This is the topic that we're going to explore next. **Deep research is where AI is able to search across many, many sites, many, many times, and create a research report.** So, I gave it all of your details, and said what are their most significant contributions that they would be proud to share with their grandchildren someday. Share it like a story, one paragraph each. ELI12 is "Explain like I'm 12." This spent more than half an hour, and searched 671 times. Meaning each search result would have had 10-20 results, and it read through all of those, but even the number of searches is 671 searches. Imagine giving a researcher a task where they patiently go through, search, search, search, search, search, and synthesize all of that information. And that comes through. For each of you, we have stories that you would be proud to share with your grandchildren, which I will share of course on the WhatsApp group in a short while. And this is one shot. One prompt. That is the power of deep research.

**Anand**: [20:18] How does one do deep research? On ChatGPT, you just give it a complex question. And you will be trying this in a short while. You give it some complex question and on the plus, there are a whole bunch of things, one of those is deep research. And when you select that and give it a question that requires a lot of searches, then it does a good job, but it'll take a very long time. Claude also has a research option which is very similar, and that comes when you click on 'Research'. Claude calls it Research, ChatGPT calls it Deep Research, Gemini calls it Deep Research. I'm going to just call it deep research because two of them call it Deep Research.

**Anand**: [20:58] What are examples that are suited for deep research? Anything that requires a large number of searches. So, I'm going to share something that you could try out. Here's an example. Deep search every IAS officer who disagreed with a political executive, but was right. What happened? Try this out. We all learn stories about this. Public information, you have to sit and analyze. Too much work, if you actually have to search. Just curious if it tells me some interesting stories, and good. So I am going to start this deep research, as much as you are, but I would encourage all of you to give this a shot. Take this question and put it onto either Claude or ChatGPT, whatever. Claude will do a faster job, ChatGPT will do a deeper job, but by and large they are both pretty good. Make sure you select deep research or you select research depending on what you're using. I'm going to change the model to Haiku because I want on Claude at least a slightly faster answer. And let it run on both. Give it a shot. My aim here is mainly to make sure that you have tried out deep research at least once. Please feel free to change the question. Doesn't matter what question you use. But give some deep research a shot.

**Anand**: [22:27] Now, sometimes it will ask you a question. In this case it's asking me... or it may ask you to edit the plan. It is saying, IAS officers who opposed politicians. Compile a list, compile primary sources. So it is saying here is how I'm going to run the research. If you want to make any changes, I will give you one minute. You can click on edit and make the change. So far, out of the hundreds of deep researches that I have run, I have never changed any of these. The reason I'm asking *it* to do it is because I want to delegate the task. Maybe 0.1% of the time I know better than it how to do that research. Then I tell it, but so far that's so rare to be negligible. Claude can connect to tools. You would not have connected them to tools already. But it will then start planning, researching, whatever, and run the result. ChatGPT is going to take a good 20 minutes minimum. Claude might get this done in four or five minutes. Why is Claude faster? Partly because it runs things in parallel. Partly because ChatGPT goes in sequence, and also digs a little deeper. But Claude is a little smarter in the way it does things. Very, very hard to see which one is better. So I don't even bother thinking about it. I run it on both. It doubles my work for review, that is true. So I try and minimize that. I take Claude's output, give it to ChatGPT, take ChatGPT's output, give it to Claude. And say now review. I got these other responses, incorporate the best. And then finally I take it all into Claude which has better style and say write it in a nice way for me to read. That's usually my workflow.

**Audience 11**: [24:12] The earlier survey you gave us, the prompt like to put your name and then designation into the Claude and ChatGPT. What I found after applying on both, the normal Google search is much better and gives very wide results compared to both of them.

**Anand**: [24:31] Very likely. And the reason is because you are playing the role of researcher. See, what's happening is you know which of those results are relevant. You probably even know what that result exactly contains. So since the data...

**Audience 11**: [24:48] Plus in the data it will catch only after 2010 onwards. Whereas the Google goes much before that. Like from the time I passed 2000s onwards.

**Anand**: [24:57] Did you use the paid version?

**Audience 11**: [25:00] Yes.

---

**Anand**: [00:00] There is a huge difference between the quality of the research done by Deep Research versus the paid versions versus the free versions. There are two notches of differences. If I put exactly the same thing, put it onto Deep Research, now you will see the difference. Not only will it surface things... So let's take a concrete example. My mrudangam [percussion instrument] teacher when I was a teenager, used to live in Chennai, Mandaveli. I have lost touch with him for 40 years. I was trying to locate him. Google search fetched me very little. Most of it was irrelevant. Bing search irrelevant. Search on Claude, nothing. Search on ChatGPT told me a few details. Yeah, he was there, he was famous, blah, blah, blah, here are a few links. Search on Gemini. Now Gemini, which is by Google, managed to, I don't know from where, dig up all kinds of weird things and told me he is deceased. His son held a memory in memory of him there was a concert and this is the hall where it was held. He continues the tradition of performing at festivals. There is a fund for his family. They are still living in the same area. Rich details which I was shocked that it was able to find. Why is that? Because in that particular case I was using the pro model of Gemini and this was about three months ago. Things may have changed. You saw the pace at which models... actually you didn't see the pace at which models are evolving, maybe we cover that, but in terms of intelligence, they are every three months roughly growing up as much as a human is growing in four years, give or take, two years, sorry, give or take. So given that in the last three months they have advanced in their intelligence as well, so today the results may be different. So if you're saying I find the results better on Google, it is very true, and I suspect that that is for two reasons. Choice of model or platform. Second, the fact that you intimately know this and therefore are able to judge the results better.

**Audience**: [02:14] Anand, if I can just add to that, I think Google gives you an entire list without any sorting based on whatever you have asked. But when you are using AI, I think **the trick is how you ask the question**. If you just say search, don't use AI. It's the least effective prompt to say search for this. **You have to give it context. And when you give it context, that's when you see the difference in the search results.** And I think that's the most common mistake people make where in Google they will put search for this, in AI they will put search for this, and Google is way better in that case, right? And I'll give you an example of how I search for it. So when you're presenting, say you're presenting to your secretary or cabinet secretary or whatever, right, on a specific topic, before the presentation you could actually get into whatever models you are using and say that, on this particular topic, go through this person's speeches, tweets, whatever of the last so many months and form an opinion of what this person believes in, doesn't believe in, what are the fears, what are the questions. Google will never do that for you. Right? So this is the analysis. And imagine having that, and of course you get it verified, but **having that information in hand before you actually make a presentation to that person so you know exactly what the red flags are** or where is this person going to sort of... So I think that's the difference we need to call out, that **AI as plain search engine is a waste of money**.

**Anand**: [04:05] Exactly. And which is what it's doing right now, which is not just being a search engine, but researching a large number of people on a theme, for which even for us to construct... So it's saying so far it's counted 600 sources. For us to even conceive of the kinds of things that we want to search for is difficult. In short, the kinds of things that we would use Deep Research for are different from the kinds of things that we would use search for.

**Audience**: [04:30] I have been struggling with past few minutes. I tried to find all the judgments of the Supreme Court of India which explicitly contradicted each other on the same issue. Every time it's coming error loading app. I have the proper app...

**Audience**: [04:47] [inaudible] woh bhi ayega nai [That won't come either]

**Audience**: [04:52] Wo wahi toh lag raha he mujhe, yeh phone ne ban kar diya he. [That's what it seems to me, this phone has banned it.]

**Anand**: [04:56] Oh you can try it on the system if you want. Are you sure that it's a free version?

**Audience**: [05:03] No, no this is the LDS version, it's pro.

**Anand**: [05:07] Oh it's pro. I couldn't see...

**Audience**: [05:10] It has started working at all.

**Audience**: [05:12] Anand, I also gave the similar situation. I think so we'll have to keep on trying. I got the answer in the third attempt.

**Anand**: [05:19] Oh interesting. I did not expect that.

**Audience**: [05:22] I've tried it 10 times.

**Anand**: [05:24] Find all Supreme Court judgments of India that have contradicted each other. Keep in mind that I'm not looking for trivial contradictions. I want real substantial contradictions. Also, not just overrules. That's a different thing altogether. I want true contradictions. So make sure that you fact check yourself and not just give results that appear to be correct at first glance. Be diligent in your research and give me something that is truly mind-blowing.

**Anand**: [05:59] **Emotional prompting works.** Now, I can run this with extended thinking, which is not as powerful as deep research, but for this, my gut feel is might be enough. And instead of taking half an hour it will take only 10, 12 minutes hopefully.

**Audience**: [06:17] [inaudible] afraid of content.

**Anand**: [06:21] Contradictions in Supreme Court judgments is not contempt. And it is yeah, pretty nuanced about that. It has a good sense of not just what would be against the law, but what can it get away with also. For example, if I asked it to draw copyrighted content, depending on the kind of copyrighted content, it would actually agree or not. If it knows it can kind of get away with some copyrighted content. Now I'm not saying that as an embodied intelligence it is making that decision. In its training OpenAI has put in some broad guidelines, which used to be somewhat strict, then became a little more liberal, then became stricter, now it's slightly more mid-to-liberal side. It keeps changing. Gemini used to be ultra-strict after they published an image where there were black Nazi soldiers. They said, wait, how can this be? And there was a huge controversy around it. After that they made it ultra-strict. Then they made it ultra-liberal. Grok was always ultra-liberal. So in short, if you choose the right model, you can happily break the law.

**Audience**: [07:31] Sir, can I ask a question please? This may just be beyond the exercise part of it. I mean, I'll be grateful if there is some answer. I was just wondering that all these models, and we have a couple of models, handful of models right now. I also searched the models, LLM models as we were doing. And of course they range from expensive to very expensive models available and also with specific in terms of efficiency and when to use what. My answer is that because technology is iterative and it improves with every iteration, and that is how acceleration happens, and even Claude has these various versions, Haiku, or whatever, right. My question is that all of these AI models feed on the global commons of data, right? And we know that Amazon has its biggest cloud data because it really started early. For a person who is on the consumer side, I just want to know, from the competitiveness of, because it's a buy model, you buy a licensed version, and you'll try and buy the best version. So to remain competitive in the field, you'll need access to the best, to the most, to the best and to the most available global commons of data, sir. So will companies who hold the data not restrict access and grow their own system? Because this is what is happening in the travel field. MakeMyTrip today or Visas and MasterCard today have, because of one business of swiping cards collected at the back end huge amounts of data and today they are utilizing it as a second business revenue stream if I can say. So can you throw some light on how this is happening and how the future of this AI based on this competitiveness of data will emerge in the future so that we can keep a look out on how we do our tender to be able to, you know, work with these AI models in the future?

**Anand**: [09:41] Data has been an asset for the last 15 years. Now the asset has priced upwards. ChatGPT bought Reddit's data, bought Stack Overflow's data, wherever there are pools of data, they are buying data. So there is no market for it. And companies that have special data or are generating special data are holding it more than before so that they can sell it at a higher price. They are not building their own models because that requires a very different kind of expertise. Selling it is proving the easier option. There are two derivative concerns. One, will we run out of data? The answer is no for a couple of reasons. A, machines are constantly generating data. Humans are constantly generating data. Like every action that we are performing, we are generating data. Secondly, synthetic data is proving very effective. These models are almost as smart as humans. So when they talk to each other, they are creating new kinds of data. That is also proving very effective in building new models. So we are not at risk of running short of data. As a result, there has been a downward pricing of private data. Earlier companies thought the value of the data that they have will go shooting up. Then OpenAI, Anthropic, and all looked at it and said, look, if you're going to price it at this much, we can generate our own synthetic data, thank you. And started hiring armies of people to create similar data. So the price shot up, has stabilized, maybe gone down a little bit. We do not expect this to grow, but there is value in private data. No question about it.

**Anand**: [11:16] So, let's look at what else we can do with deep research. So one of the things that, okay, so not just deep research, but the output formats. Some of you have... who all have used NotebookLM? Could you just raise your hands? One, two, a few people. Got it. NotebookLM is from Google. You can just search and get to NotebookLM. It has several good features. It's a great education tool, but one of the most powerful uses of it is to create a podcast. And I thought I had created a podcast of all of your work. Looks like there's a... Okay, I'll skip this bit. Now what else can Deep Research do? One of the things that I did was asked it to deep research the AI policies of universities across the world. Do universities have an AI policy? If so, what does the policy say? How can we compare them? And after a good half an hour, it gave me a detailed result which I fed into Claude and it created a website. This website is a full-fledged report on how... KU Leuven, not sure which college this is, University of Helsinki, University of Melbourne, UNSW, Princeton, etc. All of these have very comprehensive policies on AI, how the faculty should use AI, how the students should use AI. And right at the bottom are Tokyo University 49%, meaning in terms of the things that the policy covers, very little coverage. Ashoka University, IIT Madras, SUTD. I was a bit surprised at this because I asked it to choose 25 universities by itself. It chose it. Out of these, I have lectured only in three universities, and those are the bottom three. So this is the point where I have to clarify correlation is not causation. I did not have anything to do with this. But we have a full policy matrix now of how each of these universities are dealing with different policy stances. Should AI use be declared? And if I look at this, almost without exception, they're all saying they should be declared. But University of Tokyo is saying we recommend it, we are not requiring it. Which is interesting. Or let's talk about something like privacy. Are they limiting what data can and cannot go to AI? Most of them are, but SMU has not addressed it at all. And again, University of Tokyo is saying not necessary. We are not too fussed about it, which is interesting. So where University of Tokyo is having a policy, they aren't forbidding it. In fact, they're saying look, use your discretion, which is in contrast with many others. This is the kind of thing we can create as a derived product from deep research. Take a number of things, like countries, like departments, like states, like universities. And take a number of aspects, like performance on different policies, like whether they have an AI policy on this aspect, and so on. And we can start creating a matrix out of it, because if I had to populate let's say 100 cells, a 10 by 10 for example, for a human that's a lot of work. That is exactly the kind of thing deep research is meant for. So we can say give it to me as a table like this, save it as Excel. It will save it as Excel. It knows how to create Excel sheets. You don't even have to do the work. And give me links in each of those to the proof.

**Anand**: [14:44] Similarly, I tried the following. Find and research the national AI policies of various countries. Pick four countries that India would like to compare itself with and give me the policy. And it's given me a long, detailed one which I have not yet visualized into any form or created into a matrix. But as an... and then I said, okay, extend this with some more details. But the theme is that **whenever you have a long list or whenever you have a large matrix, deep research is your go-to tool**.

**Anand**: [15:17] Your deep researches will probably not have completed by now, and that's okay. This does take time. Let me check the status of the deep research that I was running to see if any of them... Okay, so they're all still running and that is not unexpected. Oh, wait. Okay, here's a story on Indian bureaucrats. Claude, the lighter version, did get us a result. Have you heard of Ashok Khemka?

**Audience**: [15:43] Yeah.

**Anand**: [15:44] Okay. So yeah, 57 transfers and a land deal worth thousands of crores, fair enough, 91 batch. There's Durga Shakti Nagpal, familiar? Okay, fine, so there is Sanjiv Chaturvedi who won Asia's Nobel Prize. There's TN Seshan who is fairly popular. U Sagayam who exposed a granite scam. Fodder scam which was exposed by Vinay Shankar Dubey. And so on, and clearly a longish list which will go on. But at one shot, for me as a layman, I get a sense of what happens when IAS officers go against their political executives. From where we can use this as raw material for the next question. What strategies work well? What strategies don't work well? Even for an understanding of how to execute political strategy, or how to market to a political executive, or how to sell an idea to different people, we can research them as Devjani mentioned. We can research the context, we can research what worked in the past. Feed all of this back to AI and say in this context, tell me what are my best strategies. Here's what I've tried in the past, here are the documents that I've shared. Here are transcripts of my previous calls, and we'll come to that in a short while. What can you do with this? Give it a shot.

**Anand**: [17:08] So I am now going to come to transcripts. I took a bunch of documents, specifically I took a recording of Devjani's talk a short while ago. What I did was uploaded the audio recording. How did I record it? I told Devjani, I'm just going to keep my phone here. I kept my phone there. During lunch, I copied this into the laptop, uploaded it. What you're seeing here is Gemini. Why Gemini? Because Gemini has a pretty good audio sense. Nothing wrong with ChatGPT or with Claude. Maybe one, two percent difference and I happen to have a free account that's a powerful free account. So I said let me use Gemini, and it gave me a full-fledged transcript. Now what can I do with such a transcript? Here's the thing. My next step was draw this transcript as a sketchnote. And I get, let me see if I can open this up. This picture. It drew Devjani as a man with a mustache and all that, but leaving that aside, it doesn't know what she looks like, it doesn't get it wrong. But **this is a nice sketch of the contents of her talk. Somebody can just go through this and read little bits and say, "Oh acha okay, unhone parin ayi agenti kiayi yeh kya he recipe... Oh, approval needed to be... nice anecdote that she mentioned, right?" This is going to be more relevant for somebody who was in the talk, they will know the context. But for somebody who is in the talk, it serves as a good reminder. "Oh yeah, she covered that. Oh yeah, that's... a one-page sketch which serves as a mind map of what are all the things that she spoke about." That is pretty useful.** But also for someone who has not heard the talk, it serves as an interesting summary of the kinds of topics that were covered.

**Anand**: [19:12] Secondly, I said, I want you to write a full-fledged article, write it like Malcolm Gladwell, write it like a New York Times article. And here is the article. It goes through, introduces the slides, it explains what she said, what are some of the numbers. It researched behind the scenes from the websites that she has shared, linking to those. So this is the NITI roadmap PDF on job creation which she referenced. And with all of these details, we have also at the end a set of takeaways. Timing is strategy. Workflow integration, not just ChatGPT. Biotech will be bigger than AI. End-to-end stack, not point solutions. Silos are an enemy. The themes that she was talking about, why this refers to... why these themes are important, etc. Let me send this to you. I'm going to put this in my... PB I did not publish it yet, but I'll send this to you and you can take a look at these. But **with a transcript the number of things that you can do with it is crazy. When you record more and more meetings, we are often on video calls. Those are anyway recorded. When we have those recordings and we transcribe those in bulk, we practically have an AI friendly memory of what the interactions have been about. Feed that as context and then ask it a question, "For such a person, how do I explain better? What mistakes did I make when I was explaining to them earlier? How do I convince them in a better fashion? Are there tactics that I can use to improve the conversion of the policies that I'm proposing? What are the questions that they will ask me next time? Here is the new document. Here are all the 10 transcripts of the conversations I've had with them. What are the five things that they will trip me up on?" Transcripts can be very powerful and are one of the best sources of information that you can easily start putting in.**

**Anand**: [21:21] And it does not just have to be transcripts of conversations you've been having. Let's take the Press Information Bureau. Yesterday there were a couple of videos that were released. The inter-ministerial briefing on recent developments in West Asia. Press conference by the Ministry of Housing and Urban Affairs. 34 minutes, 58 minutes. Who has time to listen to all of this? Instead, take it and get a transcript, step one. Step two, tell Claude, which has a better style, give it to me as a quick presentation. So here's what happened. Crisis response briefing, okay, what happened there? We are insulating the domestic economy. India has successfully... okay, let me see if I can get a bigger version of this. Download. Oh, it's creating PDF slides for download. Okay, which is convenient, but I'll still talk through it. So we seem to have done something on aviation, energy, and citizens. Several airspace restrictions have come in, good... Converting to PDF can be slow because it takes font by font and puts it in, but eventually gets it right. Okay, so we have made, okay, they canceled about 10,000 odd flights, and that has reduced the revenue streams of the airlines, and to make sure that the supply chain is okay, we are putting in some special cargo flights. Energy supply is okay, that's not a problem, and we are intervening on LPG and CBG, good to know. Diplomatic evacuations are progressing, and here are all the immediate actions that different sectors are doing. Good, I can go through this in about 4, 5 minutes max, instead of the 30, 40 seconds that I just did, and get a full sense of what the briefing is about. Useful. Not just this, I can share this with other people, and that can be personalized. Different people need to understand the briefing in different ways. All I have to do is tell somebody, "Create 10 such briefings based on the same video, but using their past context, what is it that they would be interested in," and share. The scale at which we can do this... So this was the next one, press conference on... by the Ministry of Housing and Urban Affairs. Again, another presentation. Here's what we're doing, here are all the details.

**Anand**: [23:55] It does not have to be just conversations. It can be something even more detailed than this, but we'll come to that in a few minutes. Let's take this as an exercise. A cabinet briefing by the Union Minister that happened a couple of weeks ago. So I put it into ChatGPT and said, give me a transcript. Not only did it give me a transcript, it also gave me from the internet the official slides and the briefing presentation, from which I can do whatever I want. But what did I do? I created a sketchnote among other things, same as before. Open this in a new tab. So it looks like there were four topics: IVFRT, immigration visa registration and tracking, modified UDAN, India's NDSC, total infrastructure progress. And these are the themes that it talks about. Good to get a quick sketch of what's happening in this. From thereon, we can start looking at what is the impact on my policy. Is that...?

---

**Anand**: [00:00] Is that something that we need to change? Is that something that already complies? If so, what is the specific nature of the impact? Let's give this a shot. What I'll request you to do is, I'm going to share this link which has the briefing audio and I'm going to put this in WhatsApp. Could you... so this is the cabinet briefing audio. Take this, put this into Claude or ChatGPT, and ask it to create any kind of presentation or any other format that you want. Entirely your choice on what you create out of it. **But do create and share.** This can take a good 7-8 minutes. So we'll let it run in the background, but initiate it. So next step, open WhatsApp, download this audio, put it into Claude or ChatGPT, ask it to create a presentation for anything. Maybe the impact on a policy that you are interested in. Maybe the impact on a specific state, maybe the impact for a specific person like a certain minister, or maybe what is the impact on a specific foreign country, a product, a company. Examples just based on the impact. Give it a shot.

**Audience**: [01:25] Sketch notes is on [inaudible]?

**Anand**: [01:27] Sketch notes is on [inaudible]? Yeah, I'm not going to cover that right now.

**Unsure**: [01:42] Any questions now?

**Anand**: [01:45] Getting tired, are we? Why don't we take one or two questions? There's only one thing that I wanted to show and then we'll dive into questions. No, actually, if I'm reading the mood right, 4:45, we are getting close to the fag end. Let's do this. Do this when you have time. **The most powerful thing that I was going to show you, I will not have time to show you.** So I'm going to skip it. Data analysis. I will tell you in a sentence what is possible. **Upload Excel files and give it a shot.** Specifically, I had sent a link earlier this morning. It was UDISE data. **Analyze it and predict what will happen to the girls' dropout ratio.** I'm not going to tell you how to do it. I'm not going to explain anything, because by now you've figured out that trick. **There is no trick.** If I'm asking you to do something, you ask ChatGPT or Claude to do the same thing. That's all. But you have to give it the context. Upload the data, and see what you find. The quality of prompt makes a difference. I will share that on the WhatsApp group. Part of what I will be doing today is sharing a summary of this discussion. Like you saw Devjani's discussion getting shared, but also some of the stuff that I did not cover today, and you can try that all by yourself. Before I get into the Q&A, there is one little thing that I would like to share.

**Anand**: [03:20] I asked Gemini, which has a new feature right now called create music. It can create videos, it can create deep research, etc. I said create a soulful vote of thanks, with patriotic Indian music playing in the background, naming each of you. And it did. Let us listen to it.

[03:45] [Music starts playing]

**Anand**: [04:10] It's playing from my machine. Is there a way to get it to play on HDMI? Let's try.

[04:19] [Music plays clearly]
"The morning sun rises over the secretariat corridors, illuminating the echoes of long nights spent in duty. To Ms. Vatsala Vasudeva, for steady hands in every storm. To Shri Shyamal Misra, for the silent strength of leadership. To Shri Amit Rathore, for the vision that breaks through the haze. To Shri V. Shashank Shekhar, for the quiet pursuit of excellence. To Shri Pankaj Kumar, for building bridges of progress. To Shri Robert L. Chongthu, for the pulse of the hills and the heart of the city. To Shri Sanjeev Hans, for the clarity of purpose in complex paths. To Shri Subodh Kumar Singh, for the resolve that never falters. To Shri Ashish Sharma, for the foundation laid with integrity. To Shri Ramesh Kumar Sudhanshu, for the spark that inspires the team. To Shri Kaling Tayeng, for the voice of the frontier in our council. To Shri Manish Thakur, for the wisdom that informs every decision. We stand as one, guided by the compass of our republic."

[06:07] "For the Republic we serve, for the people we honor. To Shri D.V.S. Kumar, for the dedication beyond the call. To Ms. Aradhana Patnaik, for the grace of leadership under fire. To Ms. Himani Pande, for the brilliance of thought and action. To Shri Rahul Sharma, for the steady hand on the rudder. To Shri Rajiv K. Mital, for the legacy of selfless work, for the grit that transforms challenges into victories. To Dr. Sira Karuna Raju, for the healing touch of administration. To Shri Anandrao Vishnu Patil, for the roots that hold firm in the soil. Shri Sudhir Kumar, for the light that guides through the complexity. To Shri Subodh Yadav, for the courage of conviction in every file. To Dr. Richa Bagla, for the wisdom that creates space for growth. To Ms. Indra Mallo, for the spirit of the Northeast in our hearts. To Shri Nilkanth S. Avhad, for the depth of knowledge that serves the many. And to Ms. Mugdha Sinha, for the innovation that shapes the future of our service. Thank you for the years of service. May the path forward be clear and the morning sun shine."

**Anand**: [07:21] Thank you. Really.

**Audience**: [07:23] We started late, so you can have this data analysis [inaudible] that is most important.

**Anand**: [07:35] Okay, let's do that, but please feel free to have questions coming in. Absolutely no issues.

**Audience**: [07:41] What about security?

**Anand**: [07:44] Security. Risk, risk, risk, risk. Ah, okay. The risk. As a simple rule of thumb, please don't put onto Claude or ChatGPT or any of these that have not been approved by MeitY, anything that is sensitive. Now, what constitutes sensitive, you know better than I do. Personal work? Absolutely no problem. Anything that is not sensitive based on your discretion, feel free.

**Audience**: [08:12] I am talking like personal information also.

**Anand**: [08:15] Ah, okay. For personal information, the rule of thumb that I use is, number one, there are some things that I don't want most people to see. Like my bank passwords, like... mostly just my bank passwords. But even that I have placed on Dropbox. So there are one or two providers that I will trust. Maybe Apple, maybe Dropbox, maybe Microsoft, etc. OpenAI is on that list for me. Google is already on that list for me. Anthropic is getting there, but short of that, I will not share this with anyone else. The next level is stuff that I'm a little uncomfortable to share. For instance, some office information which is, it's not well restricted, but I'm a little concerned about whether it should go outside. In that case, the rule of thumb I follow is, **if it's a paid version, I share, not if it's a free version.** For example, there are a whole series of free models out there. I don't have a Perplexity subscription. So I use Perplexity's free version. Or many of the Chinese models. Qwen, Kimi, MiniMax. MiniMax I have a paid subscription, but the others I don't. If it is a free version, I don't give it. Why? The rule is the same for all the free versions. Free version, they will train based on that data. Otherwise, we don't know. If it is public data, it does not matter. It anyway has it. Keep in mind that a lot of the data about us is public. So I'd also do one cross-check. Can you find this information about me or this particular information that I'm interested in? If it can anyway find it, then I don't have a problem sharing it. It already knows it.

**Audience**: [10:00] We can see it's very powerful tools, but how does it work in the sense, those algorithms, maybe they are redefining what is happening, what we see. So those algorithms, are they frequently updated? And who finally decides what kind of algorithms are there? And yet, it finally circumscribed by the human element only because it cannot go beyond the input which is given to it by humanity.

**Anand**: [10:29] It works very similarly to the human brain. There are neurons. There are neurons that are connected to each other, and when some neurons fire it sends electrical signals to other neurons. Over time, it has learned like a human. Humans see, hear, feel, touch, and it is the same way these are being trained. That years of learning cycle has been compressed into a few months, and then they copy. And that copying is the powerful part of the process. I have one brain. I can copy that a thousand times. Who decides that? The companies that are training on those models. Anthropic, Google, Qwen, Kimi, Alibaba, etc. Will this be outdated? It's constantly getting updated. AI itself is discovering new ways of training AI. Some of these changes are small. For instance, what DeepSeek found was, by... instead of... the earlier process was, we will tell AI, "If you get this question, this should be the answer." Now, to do that a human has to prepare a long set of questions and answers. Very slow process. DeepSeek found that instead of having a human prepare that, if they asked it a question and told another model, "Evaluate it based on this criteria," meaning I don't give the answer to each question, but I tell another model how to evaluate it. Then two things can happen. I can randomly generate thousands of questions, and the answer rubric is available, it can train on that, which means that the human effort is a lot less. Net impact: they created a model that was so cheap, so advanced, that Nvidia's stock price crashed by several billion dollars. That's an example of one kind of advance, and like this people are constantly coming up with better ways of doing, better ways of structuring the brain, and so on. And you can see the impact of this getting implemented. So the companies that are building these models control it. It's built very similar to the brains, and it keeps improving. Over time it may have a different structure than the brains. I was keeping an eye on the time...

**Anand**: [12:53] So one of the things that I was exploring was, can we take some random dataset? So I went to the NITI Aayog's NDAP website, which has lots of datasets. I went to the data catalog and randomly picked the National Food Security Act monthly allocation. It said I can download the table. When I click on it, it sent me an email and I downloaded from that email. And I got a reasonably large 300 kilobyte CSV file. Then I took a prompt. This prompt is the result of about 10 years of my experience. Not that it takes 10 years to learn this, it's a very smallish prompt, which is basically what I teach our team, our students, etc. on how to do data analysis. Not that Claude or ChatGPT or others don't know how to do data analysis, but as of now, **I am still a better analyst than they are**. A few months later, they may not be. So by encapsulating everything that I know, and I've also checked whether these are things that they need to be told or not, and I find that by telling them it actually improves things. So I copied this long prompt and put it into Claude and added only one thing. Analyze this NFSA data, uploading this CSV file. The rest of it is exactly the same. Hunt for stories that may, blah blah blah... The rest of the prompt, the full details is exactly the same. It did the analysis, and here are some of the findings.

**Anand**: [14:30] The data collection blind spot. So March NFSA data has an 85% transaction drop. Which means that suddenly in March we only have 1/6th of the original data. You say, okay, hold on, that kind of makes sense because in April they probably published incomplete March data. And it's telling you what exactly happened. In Feb, 28 states were reporting 141 million transactions. In March, only 18 states were reporting. Bihar, Haryana, blah blah blah, are missing. **When somebody gives you data, without even bothering to look at it, if you want to know if there are some problems, give it to an AI model and ask it to find out what's unusual. Do general data analysis.** It will flag off these anomalies. Keep in mind, **this is better than most data analysts**. There are very few people on my team who will be able to do this data analysis. I definitely cannot do this data analysis. It will take me too much time. Why does this matter and all it explains, and confidence level, it is 100% sure. And this is where the power comes in. What it does behind the scenes is, it is not using its language intelligence. **It is using its programming intelligence.** You saw earlier that it beats the majority of the programmers in the world. It wrote a program, it ran the program, it did the analysis, which is exactly what a data analyst does, and is showing you the result interpreted like a business consultant. Why would I hire a business consultant?

**Anand**: [15:57] Second. Goa and Telangana's hidden delivery models. These two states are allocating food grains, but they distribute zero metric tons and yet show a massive transaction volume. What are they doing? Zero distributed, and yet there are transactions. Maybe these states, and its understanding of the world is not as perfect as yours. The numbers are correct because it's based on the factual data. Interpretation, you will have to apply your judgment. Are they running a cash transfer implementation of NFSA? Not giving grain in kind, so the transaction confirms that people are actually accessing the benefits, just in a different form. Maybe, maybe not, you could verify. Maybe it's smart enough to have searched online and figured this out. And this is robust. It's saying that at least on the numbers I have not made a mistake. This is policy. Okay.

**Anand**: [16:47] West Bengal. India's largest ration card program, but it has an activation crisis. There are 51.7 million ration cards issued. But monthly activation is only 18%. This is the lowest among the large states, and therefore the number of inactive cards is about 43 million. What's happening there? Something worth digging into. Are people close to this field familiar with it? Probably. If you get transferred newly into this department, how long will it take for you to get familiar with it? Now the answer is five minutes. That is the kind of capability. Not just you, you could just randomly poke at anyone else, one of your friends and say hey, what is happening here? And they will ask, how do you know this?

**Anand**: [17:33] Next, and again, this is real. Aadhaar saturation. The correlation between Aadhaar authentication and card activation rate is practically zero. That is, most of these states have 99-100% Aadhaar enrollment, but the transaction volumes are all over the place. Some of them are the activation are a range of about 0.3% or 110%. Now how can you have an activation more than the number of Aadhaar rates? Okay, Aadhaar enrollment is possible. But that seems to be anomalous. Therefore, Aadhaar readiness does not guarantee operational effectiveness in the distribution of seeds. So what actually does make a difference? Worth checking. Who's leaving food on the table? See, most states are actually distributing only a fraction of the allocated food grains and it is not uniformly distributed. Goa and Telangana, we saw that as a cash transfer model. But Odisha, look at the percentage. 0.01%. Something is happening there. It's probably not underutilization in that obvious sense, but something is happening there. Rajasthan, 210%. How is that happening? At the very least, we have data reporting problems. And now we know where the data reporting problems are. We can dig deeper, have a conversation with them, tell them to fix it. And when this sort of a fix can be flagged off in just a few minutes, we have power.

**Anand**: [19:00] This can translate into presentations. For instance, we have... let me just randomly take... yeah, which one shall I take? Hmm, this is an interesting one. The transaction volume swings 18% predictably. Consistently, March has the lowest percentage, September has the highest percentage. Why? Following the harvest cycles and monsoon patterns. And therefore, the opportunity is, currently we are doing a flat allocation. That's wasting 25% of the buffer. What if we rebalance the allocation itself to be seasonal? So that it can lead to, and it's based on a simulation saying that there can be a 10-20% improvement in working capital recovery. Yet another thing that we can start asking for based on data. You run a simulation. If we rebalance, what is the level of rebalancing that we need to do for an improvement? Exactly what we would hire a consultant for. Not that we don't need a consultant, but they can start doing the validation on this, compare against policy, see what's working, what's not working. You have access to rich data. You can demand some very powerful data. **The fact that the models do not have this is your advantage. Because that means that you can get some insights that others cannot.** And to the extent that this is something that you can share with these models, or use Sarvam's models where they have that capability, the kinds of things that you can come up with is incredible.

**Anand**: [20:30] And that is the exercise that I am going to leave with you offline. I have shared on the chat earlier, and I'll share again, a link to UDISE data. It's reasonably large, 1 zip file. Download it. Upload the zip file to either of these two. And ask it two questions. What determines girl child dropout ratio? And how can I improve it? If I do this, predict how much it will improve by. Feel free to ask follow-up questions. In which state does this matter more? Where is the correlation more? Where is the correlation less? Any of these. And **trust me, of the various capabilities that I have been exploring, this takes the cake.** **With data and a language model combined with the power of computing, you can, on a morning walk, speak with it. Have a conversation. By the end of your morning walk, you will have a ready presentation, data-backed, that will tell you exactly what policy interventions will have what kind of an impact.** Give it a shot.

**Anand**: [21:38] If you have questions, please feel free to reach out. My contact details are out here. Of course you have my WhatsApp number, but my email ID and all the others are available as well. What we do as an organization is help with AI for various companies. And most of my focus is how do we enable people to use AI more. You will find that there are several instances where you can use AI more. Please reach out. If nothing else, we'll get on a call and chat about it. But in the meantime, happy to take any questions. I'm conscious that we are past 5 o'clock. Thank you.

**Host**: [22:27] AI as a tool of problem-solving, solutions, as well as the opportunities. Today we have a very interactive session with you, and I think so far we have the most enjoyed... to our session in AI. I... On behalf of the Phase V 17th Mid-Career Training Program and the Academy, thank you both of you for the most enlightening and interactive and enjoyable session. Thank you.