## Transcript **Speaker 1**: [00:00] ...get a deal, take a 15 minutes break, and then we'll do that. So I think without much of introduction and other things we will just straight away start set, right. Like so Anand, I think everybody knows about you. I have mentioned. So your profile is very, you know, familiar to the... to the people here. And it's our Alumni and Corporate Relations office. So mostly people who are doing... I think most of the the people who have been doing fundraising, sales and marketing activities. Then donor stewardship, which is essentially the reporting team which does all your impact analysis and other things. Sales operations team which does profiling of people and things like that. So I think these are basic things, right? So maybe we can get straight away started. Is there any introductory thing that you would like to? **Anand**: [00:42] No, let's dive in. The content is more important than introductions. **Anand**: [00:46] The screen. How do I... What is the best way to connect? **Speaker 3**: [00:50] Oh. Do you have a... **Speaker 1**: [00:54] You just stand side. **Unsure**: [01:03] [inaudible] **Speaker 1**: [01:15] So your NPTEL program is going on? **Anand**: [01:19] NPTEL program? **Speaker 1**: [01:21] You had a teaching course in NPTEL, right? **Anand**: [01:23] That was a one-time thing. **Speaker 1**: [01:25] AI tools and whatever that thing. **Anand**: [01:26] Correct, yeah. I mean, the course is still alive, but the intent was for it to be a one-time... **Speaker 3**: [01:32] In key presentation you want to have? **Speaker 1**: [01:34] No, no, no. We will cast it out. How to cast this... AC [inaudible]? Theriyuma? [Tamil translation: Do you know?] **Speaker 3**: [01:42] I will ask [inaudible]. **Speaker 1**: [01:46] About login... panninuma? [Tamil translation: Should we do?] Zoom open... Zoom link... I can send it to you. **Anand**: [01:53] We just login to Zoom. **Speaker 1**: [01:56] From this my laptop is connected to this... **Speaker 3**: [02:03] ...WhatsApp web or something... **Speaker 3**: [02:06] Can I send it to you by WhatsApp? **Anand**: [02:08] Yeah, absolutely fine. **Speaker 3**: [02:15] There is so much happening in this world, right? This is very difficult for us to kind of keep track. **Anand**: [02:22] Yeah, maybe we... Oh, internet connection is what I see. Sorry. No, **it is difficult to keep track, no question about that.** What I'm also finding is that that is true for everybody. Even those who are keeping track are saying there is so much to keep track of. But you know what, this is not a new problem. In that, 50 years ago, people would have said, "Oh, there is so much happening in medicine. I am having to learn only ENT. I can't learn about any of the other parts of the body." **Speaker 1**: [02:59] People online able to hear? **Unsure**: [03:03] Yes. Yes, perfectly fine. Yes, we can hear. **Speaker 3**: [03:07] Only thing is that... **Speaker 1**: [03:08] Now he's switching into his... **Anand**: [03:10] Yeah, you can't see my screen, but I'm not doing anything on my screen. I'm just talking. **Speaker 1**: [03:13] But your screen is also now visible to them. **Speaker 3**: [03:15] Not yet. He's got a logic first right. Oh. **Speaker 1**: [03:18] No, it's not yet there. Just hang on. Yeah. **Anand**: [03:22] Yeah, we'll keep talking and then we will deal with that. **Anand**: [03:24] The **difficulty is going to be there, and it has been there for a very long time.** So what I'm saying is we are just seeing a new version of the difficulty. It's okay. When we are young, it's a difficulty we grow up with. The next generation will grow up knowing how to deal with that difficulty. Just like we grew up knowing how to deal with the difficulty of dealing with hundreds of engineering. Our grandparents would have said, "What kind of specialization are you doing when we were there, we were studying all the subjects," that sort of a thing. **Anand**: [03:59] Now, **what seems to be different about AI compared to some of the other technologies is that it is very horizontal. About as powerful as the internet or mobile.** And you've been through both revolutions. So it's not like you're not familiar with that. **Speaker 1**: [04:25] So should we make even the audio here and remove the audio from there? **Anand**: [04:28] Oh yeah. I got that. So let's talk. **Speaker 1**: [04:31] So he has muted the desktop... you have to be... Ithu green pannanuma? [Tamil translation: Should we make this green?] His mic... **Anand**: [04:38] Am I still audible? **Speaker 1**: [04:40] No, no. Because you have to unmute your audio. **Speaker 4**: [04:43] Yeah, now you're audible, but you were not audible earlier. Just now I picked it. **Anand**: [04:48] Got it, fine. Cool. Let me share my screen then. And not that I'm specifically going to do anything on the screen at the moment. But we will see what we do. **Anand**: [04:58] So in light of that, part of what I wanted to share was what you can do. **The number one rule that I am suggesting is use ChatGPT or whatever, Gemini, Claude, whatever, 50 times a day.** **Anand**: [05:18] And you see, "But wait, I barely have things that I can ask it five times a day." Very good. **Then make this your sixth question: "What can I ask you?"** **Anand**: [05:28] **Second, probably even more important: use voice.** So there is just so much you can put into a voice request. **Anand**: [05:40] I am sitting with the IIT Madras team that's working on, among other things, funding and exploring corporate partnerships... Wait, this usually shows a little bar... Okay, no, let's look at... **Speaker 3**: [05:57] I think you are... you are on mute right? Let me read on mute... No... Oh yeah, no actually you may have a point. Ah, there you go... Let me check. We can disconnect. **Anand**: [06:09] No, no, I am unmuted on Zoom, so I am okay. **Speaker 3**: [06:12] Oh on Zoom, yeah. **Anand**: [06:13] So let's talk. I'm discussing with the IIT Madras team that's focused on, among other things, funding. Now, just search online and tell me what are some interesting use cases for them. **Anand**: [06:28] I don't need to prepare for this talk. It can do that for me. Now I could just send this. But what I've found is, why bother sending it to one? **Send it to three.** And asking the same question. If one of them does a better job, and I find that it does a better job ten times, then great, I will use that more. Especially when it does... when there is a pattern around it. **Anand**: [06:57] And **the other thing that I then do is log this pattern.** So for instance, here are my notes on different questions I asked. So I asked, "How does an innovator build accountability?" And I found that Sonnet did a better job than GPT-4 looking down on Gemini 3. **Speaker 1**: [07:13] So this is a Notepad file? **Anand**: [07:15] This is just my notes, yeah. Exactly. Just the fact that it's in black means nothing. **Speaker 1**: [07:19] But you're using it in a Visual Studio. **Anand**: [07:21] I'm using Visual Studio Code for my notes. You can use Excel, use whatever you want. **Speaker 3**: [07:24] You did a comparison. Oh, you did it when you can ask even the comparison, and then you look at that? **Anand**: [07:33] Quite possibly, but I don't trust the comparison yet, because I have not compared the comparisons and seen if it is enough aligned with my style. But I do that a lot. I send the comparisons to one and then tell it, "Look, take ideas from here and put it here. Take ideas from here and put it here." You figure it out. And then it will anyway get the best of the rest of it. **Speaker 1**: [07:54] Just one thing before that. So we can use a Microsoft Word or any of these things, right? Or do you use... Why I asked is because are you going to use this for any further... nothing, right? **Anand**: [08:07] Copy pasting. As long as you can copy paste, that is good, or upload it, it is good. I sometimes upload this. What I've done is, for instance, I've taken all of these comparisons and sent it to, and asked it, "Look, look at all the comparisons and then tell me what each of them is good for." So that way I get a sense of what they are good for. **Anand**: [08:24] Now let's take a look at any of these answers. Now... Yeah, it's asking which team am I referring to. Right. So it... Okay, and which kind of use cases? Fundraising. Great. **Anand**: [08:39] Now, here is the thing: **it is smart enough to do some research, find out about us, and give answers. Which is great.** So **the bottleneck to 50 conversations a day is not that we don't have ideas. It is that we are not trained on asking it for ideas when we don't have ideas.** If you force yourself, "I have to use it 50 times a day," ideas will come, because you have no choice. **Anand**: [09:05] And **if you use voice conversations, when you're out on a walk, you're going for lunch, have a quick conversation, ask it a question.** "I am walking by every day I go for lunch. Is there something interesting that I can do in this 5-minute walk that I go for lunch?" You see somebody, you just get a thought, dump the thought at it. And say, "I got this thought. I don't know what I can do interesting with this thought. You meet somebody, a complete random stranger. I met this stranger. I have no basis for any useful information, but just because I told you that I met this particular person, now can you think of some interesting idea that I can use for fundraising operations?" **Anand**: [09:47] In other words, **you don't need to know; it knows. You just have to keep asking it.** What a supervisor does, right? Ultimately, the guy is doing the work. You are simply saying, "Work harder, work harder." The Egyptians used to have that whip, slave driving. Now you have these machines that are doing the work. **All you have to do is crack the whip until your arms drop.** **Speaker 1**: [10:10] Don't just scan them also. Like, you know, I start asking a lot of things. In some days I felt that my lack of any better word, I thought I will take out the... History was getting anyway all over the place. As opposed to asking a few sharp questions and then further sharpening its responses. Why what you say is, okay, you can ask examples. And then it's all over the place. Or does it really matter? Like you know, if we start asking everything into ChatGPT, in its memory, there is a lot of... It will start getting to profile me in the right way? **Anand**: [10:40] Good point. Possible. So, and I would ask it: "Is that the case? If so, how can I fix it?" See, **this is exactly the habit that we need to build. We have a question, immediately ask.** You're asking me, is a valid one, you will get another opinion. But also ask it. Secondly, it's not just questions, it is any unresolved emotion. I am feeling a little uncomfortable, I don't know why. Ask it: "I'm feeling uncomfortable, I don't know why. You interview me so that I can find out why." **"You interview me" is a very powerful technique, because then you don't have to figure out what you want. It will help you figure out what you want.** **Anand**: [11:24] And to your point, yes, it can be a problem. You can clean up memories. And you can tell it what to ignore, what not to ignore. You can ask it to do an audit and say, "Which of these should I keep? Which of these should I delete? Can you automatically delete?" And it'll say, "No, I can't automatically delete." "I still want to automatically delete, can you write a program for me that will automatically delete?" It will write a program. Say, "But wait, I don't know how to run this program." It'll tell you how to run that program. And maybe you will still not be able to do it, that's okay. At whatever point you're not able to, stop, give up, move on. Because the number of things that you can do is infinite. **Anand**: [12:00] One of the things that I love doing is taking all the data that I have, and sometimes don't have, and dumping into it. Let's say we are having a conversation. So we had a conversation yesterday. My phone is set to auto-record calls. So which means I have all transcripts. Therefore, what I did was, while walking down here, I made it create this list of use cases. This was created while I was walking here. **Speaker 1**: [12:32] All these use cases? **Anand**: [12:33] Yes. **Speaker 1**: [12:34] And how, how did it do that? **Anand**: [12:35] This is the prompt that I gave it. Which is not opening, because I need to click harder. Yes. So I was dictating, saying—and I can dictate to ChatGPT very well, you saw that it does a fantastic job of dictation—"Blah, blah, blah, I want you to research this team, find out who they are." The transcript of the call with Kaviraj and Vasudha is in this particular file, because I have the transcripts. Now read, and I'm giving it instructions, "See, I have written a lot of blog posts on AI, I have lots of past transcripts, I have lots of extracts of AI, I document all the things that I have learned." In short, I have basically the equivalent of a Google Drive cloned on my machine. Based on this, do whatever you want, and share what tips and tricks I can share with them that'll have the highest impact for them, and put them into an ideas file. And then second prompt was convert that to a HTML web page. **Anand**: [13:28] In other words, I had this conversation with somebody. **Here is all the Google Drive of material about me. Use the combination and create something that I can deliver for them.** And the result is this. **Speaker 1**: [13:41] And you have... I am asking some, some basic questions here. So you have created a page on GitHub, is it? Do we have to like do it? So we have not started using GitHub. **Anand**: [13:48] Don't. Ask it to create it locally and then double click and open it. That's all. I am putting it here because I want to send it to you. Which I could have done by email, but I don't know all your email IDs. And I anyway want to publish it for the world to see. **Speaker 1**: [14:01] So essentially for that we have to go to GitHub and create an email, I mean an account or something like that and then... **Anand**: [14:08] No. Don't. **Speaker 1**: [14:12] And to do what you did, like making it open to, you know, large team without having to send emails? **Anand**: [14:18] No. You still don't need to. Because what we can do is we will use Gemini for this, but I will only dictate to ChatGPT because it has a better thing. Now you will be doing it for your organization, I will do it for my organization. And I will give this prompt: **Anand**: [14:35] Go through Google Drive and my emails and everything else in Google Workspace. I want you to create a crisp presentation for the Office of Institutional Advancement about what Straive can do in terms of services for them. This presentation should be to the point, introduce Straive a little bit, but the focus should primarily be on what kinds of AI use cases we can deliver that will benefit them. **Anand**: [15:03] Some random prompt. It is not very thought-through or anything. The next thing that I'm going to do is go to the Office Gemini where what I've done is somewhere in the settings, I have... Oh, what is all this? Okay, close it. Settings and help. Okay, this looks familiar. Fine. Connected apps. I have enabled this Google Workspace, which means that it can connect to all of my... **Speaker 3**: [15:30] Did you go to settings and help? **Anand**: [15:31] I went to settings and help. And then connected apps. And then it showed this thing. I can turn it off and on. And that's it. **Speaker 1**: [15:40] Gemini is now connected with Google? Your Google files. So you can access all your... **Anand**: [15:44] All your documents on Google Drive. Correct. **Anand**: [15:46] Then I'm going to paste the same chat. There are a whole bunch of things that you can do here on tools. But the important thing here is also to say... [inaudible]? Oh Canvas. If you say Canvas, then it will create a thing on the right side. **Anand**: [16:03] Now the next thing to do is put it on Pro. You will likely have Pro, it's smarter, always use the best model. What is the... then make spelling mistakes corrections. Our organization's name is S-T-R-A-I-V-E. I didn't mention at IITM. Any other mistakes? No, I don't think I made any other mistakes. And then let it run for five minutes. It can take five minutes and we will do other things in the meantime. **Speaker 1**: [16:30] Now, the ChatGPT window also allows to go through your Google Drive? **Anand**: [16:36] Yes. But Straive uses Google, not ChatGPT. So I can't connect my ChatGPT account to my Google Drive. I can connect my Google account. **Anand**: [16:46] Then your question is, but why did you go to ChatGPT in the first place? You could have just dictated to Gemini. Gemini does a terrible dictation. ChatGPT has excellent dictation. So **wherever I want to use voice, I only go to ChatGPT, copy from there and paste everywhere else.** Tomorrow they will improve. But today I know that this is the best that I can get. So ChatGPT is just for dictation. **Speaker 1**: [17:07] In spite of Google having Google voice and all those things? **Anand**: [17:11] Google actually has the best voice transcription. So if I upload an audio file to Google, I get a better result. And that's what I use for our transcription also. But dictation they have not enabled it for whatever reason. And ChatGPT has. **Speaker 1**: [17:26] And you go to your mobile and check auto-recording and then it will save... **Anand**: [17:30] Whatever that MP3 file or whatever. **Speaker 1**: [17:32] MP3 file which I then convert to markdown. I upload it to Gemini. And then say, give me the transcript of this file. **Speaker 1**: [17:41] So this is a basic question. Obviously your phone memory is only 512 GB or whatever it is. So these all this voice notes will take up a huge this thing no? **Anand**: [17:48] Tiny. Guess how long our call was. No no, it's the audio itself, but how how big a recording of a, a guess how big a call recording was... **Speaker 1**: [17:59] Half an hour or something like yesterday. **Anand**: [18:01] Half an hour call yesterday. What was the file size guess, roughly? **Speaker 1**: [18:03] No idea. **Speaker 1**: [18:05] Some... now that you have mentioned, this way I will start checking. KB and some format. **Anand**: [18:09] Yes, please. **Speaker 3**: [18:10] But an audio format tends to run into MBs actually usually. Usually general right? Let's check format... **Anand**: [18:14] You know let me find out. Kaviraj, Vasudha... dot Opus. The audio file. No, that is the transcript. That will be tiny. The audio file itself. Kaviraj... Vasudha... 550 kilobytes. Now one photo is 4 MB. Give or take. **Speaker 3**: [18:36] Do you connect your phone to the thing? **Anand**: [18:38] I just record it on the phone itself. And then I copy it from the phone to my machine, yeah. Correct. If I run out of space I can always copy it to my machine, but it's not like I am running out of space that often. To even worry about it. How many calls will I make in a day? **Speaker 3**: [18:52] Now whatever talk you are doing and recording. Can you convert that into a PPT also, Gemini and give us? **Because there is so much. It will really help us. Very clearly.** **Anand**: [19:00] Absolutely. And I do that very regularly. As a practice, what I do is... in my GitHub, some page doesn't matter, I just take a transcript of my recordings. So recently, for instance, I was talking at NIE. And I asked it to give me a full-fledged data story. And it created a very interesting data story on all the things that I said, with pictures that it's created on my behalf. Some of these are the slides that I had used. And at the end, it puts in a "What is the takeaway that people need to know?" These are the links. Yeah, here are the top takeaways that people need to get out of this. Or it could make it in slides as well. And these are the slides I had happened to have the slides beforehand. But in other cases it can create the slides. **Speaker 1**: [19:48] And even all those cartoons and all this were created by it? **Anand**: [19:51] Exactly, yes. **Speaker 1**: [19:53] Did you specify the format for it to create? **Anand**: [19:55] Yes. I said, "Create a sketch note based on this transfer." That's all. I know that "sketch note" is a format. Top transcript anyway I can give it, and it creates it. Now here is the thing: **the important thing increasingly is knowing the vocabulary.** That there is something called sketch note, that it looks interesting, that this is a way in which you can present it. So what are all the different styles, etc.? We have to build the vocabulary, and that comes now from digging and doing a different kind of research than we are used to. Now if, instead of researching how we do stuff, we have to research what words to ask for. That's a different approach. **Speaker 1**: [20:30] And where is this vocabulary? Where will I get all this from? **Anand**: [20:32] I didn't know, so I asked ChatGPT. So exactly. So, and I asked it, and because I don't want to ask it every time, I am storing that vocabulary. For instance, when it comes to images, I'm saying, "Look, sometimes I just create popular formats. But when it comes to comics, what are different styles of comics I can ask for?" And it gave me some 20, 30 odd styles of comics. It said you can do a Tintin style comic, you can do a... I don't know what this is, then you can do a Peanuts style comic. And these are the prompts that you can give for it. So now I know, okay, maybe if I want an RK Laxman style comic, I can give something along these lines. Maybe if I want something that is darker, I will use this style, etc. See, I have a variety of different formats. Amar Chitra Katha is also there. If I want an Amar Chitra Katha style comic I can give this format. But here is the thing, **this entire thing was based on the question: "What are different styles of comics?" And, "Give me the prompts for it, and also create the images in those formats so that I can use it as a ready reference."** **Speaker 1**: [21:36] Anand, and then if I really want to actually go to GitHub, create an account, and do this, and then I can save it... Go to the whatever Canvas thing and paste it here inside like so? **Anand**: [21:46] You could do that. Or you could tell it to go to GitHub and create an account for you. If you are using Claude Computer Use or Claude Desktop or any one of those desktop versions if you are using it, you can tell it to do the work for you. It will probably do a decent job. **Speaker 1**: [22:04] And it will create all this stuff. Now every time I go to ChatGPT, I can ask it to "refer to these styles on this URL and then choose whatever?" **Anand**: [22:11] Absolutely. And create it in that format. **Speaker 3**: [22:13] See, based on the audiences to what will cater better. Correct exactly. We do an informal comment, formal, whatever. It's how you think and how you want to use it. **Anand**: [22:20] Exactly. So this thing has created a presentation from Straive to you. And let's see what it has. It says "We are data and AI pioneers. We have an academic heritage." Yeah, actually we are an edtech company. And what can we do for you? There are three things: predictive analytics, NLP and Gen AI, intelligent documents, blah blah blah. And it's created a presentation around it. Donor profiling, very relevant use case. That's something an AI company can do for you. CSR matching, seems like a good use case. Okay, I am not sure why this picture didn't come up. But yeah. "Stop guessing whom and how to reach out to. Predictive models can do this." Smart. **Speaker 3**: [23:00] IDP. What was IDP? **Speaker 1**: [23:02] Individual development... **Speaker 3**: [23:04] Individual development program plan... Oh this is from your... **Anand**: [23:06] Could be. "Pull key data points from complex gift agreements." I can take gift agreements and pull data. "Transform unstructured memorandums." Okay, actually that's a useful in many case which will have a good impact. Blah blah blah. **Anand**: [23:29] The point is the presentation is ready. So **before you're talking to somebody, here's what I would do: I would take all of their history** which then is gone through Google Drive and found out the kinds of things that it can do, **put in their transcripts, put in the information of their past fundings, put in research from wherever they have given previous grants.** Look at who are similar people who are giving such grants. Where has this person been before? Put everything into one folder, and tell it: "Look at this folder. Look at the folder that has everything that we have, want, etc. Create a presentation suitable for this." **And then 5 minutes before the presentation, gently read through the presentation. Enough that you can talk about it.** Keep it short, because otherwise your head will explode initially, and present it to them, and send it to them directly so that they can also skim through it. **Speaker 3**: [24:21] And how will you send it? So the presentation will transport into Canva? **Anand**: [24:26] Ha. So you can export to Slides. Which it's, okay I have no idea what it's doing, but I'm guessing it's going to create Google Slides. **Speaker 3**: [24:32] Oh Google Slides. **Anand**: [24:33] Then export it. **Speaker 3**: [24:34] Oh okay. Oh and it switch over. **Anand**: [24:35] But will it go directly to Canva? I don't know. **90% of the time, what happens is what you want, it doesn't necessarily give. But what it gives can be fantastic.** So I treat it on two parallel tracks. I know what I want. AI can help me 10, 20, 30%. I don't care too much about what I want, as long as it's good, and AI can automatically do it almost entirely. So now you have a combination of the two things. Negligible effort, maybe fantastic. A little effort, maybe good. **Speaker 3**: [25:08] Now, why I am asking this question is because certain elements may not be what we require like what you just mentioned. There might be further edits that I might want to do, you know in Canva. There might be so many other elements which is available in Canva and I would like to enhance it even further. So that's why I'm asking if I can export this? **Anand**: [25:27] No. Most likely not. Maybe there's a 5% chance. **Speaker 1**: [25:30] ChatGPT has a Canva integration. **Anand**: [25:33] Yes, it does. Gemini, to my knowledge, does not. Claude, I think has an export. **Speaker 3**: [25:37] Oh so I can download it. I can download the slide and then I can go and upload it in Canva. **Anand**: [25:41] No, it will not integrate completely. Not say manually. You download the file, and then you log into it and then any of... just like any file when you upload it to ... it doesn't integrate completely in Canva. **Speaker 3**: [25:52] No, no. That's what, how will you create the PPT on Canva? You go log in, and then you do your... Right, but normally when you edit, we do it online only directly the PPT slides, right? So you download it as a Word Doc, if need be. Then you put that, and then that's the only option. But that integration is not there. **Speaker 1**: [26:08] Or upload it to ChatGPT again. Let's start building fine. And ChatGPT is integrated with Canva, let you start it. **Anand**: [26:14] Possibly. Or have ChatGPT generate it in the first place. Yeah. All of those are possibilities. **Speaker 1**: [26:20] You tell us one thing, like for making PPTs, which is the best? Is it Gemini or Claude or ChatGPT? **Anand**: [26:26] Claude tends to make the best presentations today. Followed by Gemini, followed by ChatGPT. But ChatGPT does the best analysis. Gemini does the best research. Claude has the best style. Gemini has the best information because of Google. ChatGPT has the best brains. Analysis, rigor. Tell it, it will get it done correctly, exactly as you said. If you say the wrong thing, it will do the wrong thing. **Speaker 1**: [26:57] What about this Excel and all the Microsoft packages? **Anand**: [26:59] If you want good analysis in Excel, ChatGPT. If you want good presentation in Excel, Claude. **Speaker 1**: [27:12] And Copilot is not figuring in any of these things, is it? **Anand**: [27:14] Copilot is... Think of it as ChatGPT about six months ago to a year ago. **Speaker 3**: [27:24] Perplexity. Standard then? **Anand**: [27:26] Perplexity... Good for research, maybe, but with Google, Perplexity does not necessarily... You don't need to use Perplexity. No harm. I don't use Perplexity because I have never found the need. **Speaker 1**: [27:41] Because we all got free accounts. We started using Perplexity also. **Anand**: [27:45] No harm. Familiar. No harm, no harm. Good for research, nothing wrong with it. **Speaker 3**: [27:51] Another question. This we can use day-to-day basis mostly. But if we want to create something more strategic and maybe going out of the Q&A session I can ask you that, if it is not here. **Anand**: [28:02] You can ask now. Go on. **Speaker 3**: [28:03] Yeah, so the struggle right now is you know a strategic way of looking at high net worth individuals. Right now there are many platforms where vendor is selling it to you at a cost and a huge cost right. And we have smart kids up here on the campus. If we have to use them and create a workflow in exactly what we want, like what the vendor is selling you out at a price. And if I need to create it and then what... and literally in my mind, look for donors bracket at 50 to have a 100 or 500 or 1000 crores. Also look at their philanthropic interest feeding. Also correlate that to my database of 60,000 alumni that I have. Also give me an alert that anytime these guys go IPO. Give me insights on these becoming big, suddenly a Forbes 30 or Forbes 40 or Forbes 50. So I, I have this workflow very clearly in my mind. Now, I really want to, you know... create something, don't want to pay money, and reduce the cost by ten, fifty percent. So what could you advise? Should we, the first three steps that I need to do to get here? **Anand**: [29:08] Let me talk about the different models you could use, worst to best. Worst is: get a large number of people to build this for you. Better is: get one person to build this for you. More people, more coordination, more communication, more chaos. One person, more controllability, because now there are only two heads, yours and theirs. Even better: use only one head. You build it. Even better: don't build it. Use an existing tool. Or a combination of existing tools. **Speaker 3**: [29:46] Not pay the cost. That's the whole idea. **Anand**: [29:48] You have Perplexity, you have Gemini, use existing tools. Put another way, I would say, I would start with: here is my workflow. Which of these parts... --- **Anand**: [00:00] Which of these parts of my workflow can be done by which of the tools that I already have access to? Do it manually once for a few people. Then I will know better what I want, how what can be done and what is missing. Then there are only a few gaps to be plugged. Still check with a few people and maybe if required build. **Anand**: [00:21] Second step is can I build it? Now I don't know how to code. At which point I will go to ChatGPT and say, "But I don't know how to code." And it will say, "No, no, no, in this day and era you don't need to know how to code. **People who don't know how to code can produce better programs than people who know how to code**, because they will stupidly try and interfere with the agent, the agent can do a good job." So, and this is true. Today, when I have to build an urgent immediate client demo, I go to interns. **Unsure**: [00:47] Interns? **Anand**: [00:48] Interns. I do not trust anyone with experience because they will mess it up. I tell an intern, "Here is what I want. I am giving you an audio recording. **Do not apply your brains, you do not have any.** Take that, give it to a coding agent. At the end, use your some validation muscle, you kind of have been working with me for the last few weeks, you know what I like, what I don't like, etc. Do one basic filter. At the end of it show me, if I like it I will take it, if not I am not even going to give you feedback, I am going to drop it. Finish." **Because cost is nothing. You're just dictating and reviewing. You can do that ten times.** **Unsure**: [01:26] But right, see, this is going back to the data, right? I'm going to look at the data. I want validated data. Right now when I look at a data of a person, I am getting different net worth, you know, of an... of a well established [guy?]. And I want a trusted source. And the reason why these platforms sell it at a cost is because it, you know, they, they... it's a verified data and they say they update it. So how... so how does that happen? **Anand**: [01:51] Absolutely. If you know what you want and it has to be exactly what you want, AI can only give you a 10% boost. Then you pay. We say, "Look, I can still get the outcome that I want. **It doesn't matter whether their net worth is 5 billion or 50 billion. I'm only going to ask for a 1 million grant. And between those, who cares?**" Then you don't have a problem. If you say, "No, I need to get into a conversation where I have to mention to them given that your net worth is 5 billion dollars", then I can't get it wrong. Then you have a problem. **So can you change your workflow? So that you can take advantage of all of the free things that you're getting, is another question.** **Anand**: [02:30] **And this is where experience comes in and becomes a problem. All of these years you've been doing something and gotten the process to such a good way that you are unable to take advantage of that incredible capability, exactly like my senior engineers.** And I would say, "No, interns, you come in. Don't use your brains. Do the funding work, CSR work, etc., by using LLMs. You will be ultra efficient and you will not have any of the biases and it's free. So it's a leg up anyway." But on that note, my apologies. I need to jump into another call. Take a 15-minute break and then we can resume. **Unsure**: [03:09] Sure. --- **Kaviraj**: [00:34] If I have to write an AI agent or whatever, right, to carry out a specific task, how will I go about doing it? **Anand**: [00:44] The easiest way these days is to use an online... okay... **Kaviraj**: [00:55] Specifically, like, I want to crawl all the LinkedIn profiles of IIT Madras alumni and come up with a list of people who are vice presidents and, you know, and blah blah blah. All of those things and do that right. So, perhaps it's already existing but it requires customization because this is a business rule that I am adding that IIT Madras alumni and those kinds of things, right. So how will I go about doing it? **Anand**: [01:15] For that particular use case, to crawl, I would just go to Gemini deep research, let me try and move these things somewhere, I would yeah select Gemini, not Gemini deep search, probably use Claude or ChatGPT deep research. **Kaviraj**: [01:27] Could we do a little bit of a demo? **Anand**: [01:28] Yeah, let's do that. What exactly do you need? **Kaviraj**: [01:31] I want to get... one second, one second, I'll yeah... I would like to get the list of IIT Madras alumni based in Mumbai who have designations of vice presidents or above based on their information given in LinkedIn. **Anand**: [01:46] LinkedIn... if it will not be able to, that's another story... or other authoritative sources. Let's make it get... and add to deep research and it will take a good half an hour. But that's... **Kaviraj**: [02:07] Why won't it access LinkedIn public profiles? **Anand**: [02:10] LinkedIn has blocked it. **Team Member**: [02:11] Public profile also? **Anand**: [02:13] For ChatGPT obviously. **Kaviraj**: [02:16] For all generative AI, OpenAI, all of them it is blocked, is it? **Anand**: [02:22] No, but if it is public information available on the internet, it should be searchable right? I mean, not that if it's protected under that but even then to access that you make money. **There is a big difference between you think it should be versus LinkedIn thinks it should be.** **Kaviraj**: [02:42] Right, I agree, I agree with that. Yeah. But even public available information is not crawlable essentially? **Anand**: [02:47] **LinkedIn is not crawlable. Twitter is now not crawlable. Reddit is now not crawlable.** **Kaviraj**: [02:53] Which is the third one? **Anand**: [02:54] Reddit. Reddit is not crawlable. And you'll now see the pattern of now not crawlable, used to be until recently. **Kaviraj**: [03:03] Because they are restricting... **Anand**: [03:04] Yes, because they are saying that I can just go here, read what I need instead of going there. **Kaviraj**: [03:10] Yeah, yeah, that's spot on. **Team Member**: [03:13] The other instance, some people wouldn't have updated their LinkedIn. It would be like they would have left the job and it'll be old data. So I wanted to also look at other sources to give me up-to-date accurate info. That's a test. **Anand**: [03:28] That is a... **Team Member**: [03:29] That's again a prompt that I can give. **Anand**: [03:30] Yes, give it a shot. And keep in mind that there are two parts here. One, stuff that you will get, that's easy. Stuff that you will get that you want, that is not easy. Take what you get, leave what you don't get. And therefore one doesn't need agents specifically for this. Now you say no, I actually do want to use LinkedIn. Is it possible? Yes, it is possible. How can you do it? I won't tell you because it is not worth the effort. **These things are becoming easy so rapidly that what I tell you might take you a month to learn, and in two months it will be available anyway in a tool.** You wasted one month and how many like this will you learn? Don't you have other things to do? So just wait a few months, keep up to date with what's happening to the extent that you can, by asking it what is new that is useful for me every week. **Kaviraj**: [04:25] Why this is getting done. If I wanted to say that like now we have attendance data and all of that available on let's say Zoho attendance right. If I wanted to, I can give access because I'm the internal system. If I wanted to try an agent that will say look at attendance patterns of people on a weekly basis and come back to me with a list of people who are... **Anand**: [04:43] Send it to me now. **Kaviraj**: [04:44] Oh okay... are you hooked on the taxes? **Anand**: [04:48] I've been... what data? **Kaviraj**: [04:50] But how will I go about, why would we do that, like how will we go about writing a code for that, like essentially? **Anand**: [04:57] I'm doing a demonstration of how we can analyze attendance data for insights. And this is for the IIT Madras Office of Institutional Advancement. So think about the kind of realistic attendance data they would likely be collecting. This is being collected on Zoho. And give me a rich synthetic dataset with the kinds of hypotheses that you would normally find in such data. So step one, I don't have data. I still want to run the demo. I will ask for data. **Kaviraj**: [05:33] It's simulating data. **Anand**: [05:35] It is simulating data. So step one of the exercise which I want it is doing. Now the interesting thing is it is writing code. The fact that it is writing code I don't care about. But fine, it is writing code, good. Once it has written code, the next prompt that I have, which I will keep ready, and we will run in a short while and even if I don't run you will get the idea, is: I have uploaded the attendance data from the IIT Madras Office of Institutional Advancement for the last few... well actually I don't know how long it is for. So you take a look at it. And what I'd like you to do is analyze it. Now I'm not even sure what to analyze it for. So if I were leading this team, then put yourself in my shoes. What kinds of useful insights might I want to get from this? That is step one. Then step two, using that think about what are the best ways of analyzing this data accordingly, and run those analyses. Step three, interpret those analyses. Some of your hypotheses may be right, some may be wrong, etc. But the point is not what you have done, the point is what I can do based on what you tell me. So tell me the most actionable, surprising, non-obvious and useful insights that you can get from this. Begin with a crisp executive summary of one to three bullet points of what I should do. Then go into all the details. Is the prompt that I would give... **Team Member**: [06:58] Wow. The prompt itself is very clear on what I can get. **Anand**: [07:04] Yeah, but you see this is not coming as much from a deep thought as frustration. **I've seen very, very dumb analyses that people give. And this is basically look, all of the stuff that I've been telling you 20 people for the last 20 years on what not to give me, is what I am dumping in here.** Dump your emotions into it. See, intelligence it has. Emotion is what you bring. Now this is generating some data, blah blah blah, and where is the download? Campaign... okay sometimes it gives me so much... ah okay, so here is the thing, I will eventually download it, fine. Now this has downloaded the synthetic data. And let me take this synthetic data as if this were the real data. And it will do the analysis. Again it will sit and write code for itself. Two, how can I have an agent analyze it? ChatGPT is an agent. Gemini is an agent. Claude is an agent. Now why am I using ChatGPT? Because it does better analysis in my opinion than... **Kaviraj**: [08:11] This synthetic data will be in a proper format and all of that, right? Like after you click clear then... **Anand**: [08:15] Yeah, let's take a look. You can tell me if this is... let's sort by modified date, the other way, and just open the synthetic data. So yeah this has... oh this has taken student... oh it's taken alumni data and not your attendance data... **Team Member**: [08:42] Office of Insti... **Anand**: [08:44] Okay, some secondary data with city they're locked in from, all of that, is it? It's saying every user is alumni and this is the alumni ID and you have invited them to events perhaps, and how many events have they attended, how many have they not attended, how have they checked in, etc. Which is also useful information. That's a different kind of analysis. **Kaviraj**: [09:05] But essentially, okay, so two things here. How do we create, is there some way of creating synthetic data relevant... is there some... or we just talk and it will create synthetic data? **Anand**: [09:14] **Talk and create synthetic data. What you have to do however is do it once and then look at it and say "Oh this is rubbish because..." and then start again.** "Give me a synthetic data, same prompt, and don't make these kinds of mistakes." Then it comes back with some stuff and then you say "Oh, but this is rubbish." Over three or four times you'll develop an intuition of the kinds of mistakes that it makes, and this is exactly what happens when you're dealing with a fresher who comes in into the organization anyway. And you train it, train it, train it. Except that the training happens to be in the form of your notebook where you will say "Here are five things I copy-pasted, and next time I will copy-paste this entire set of what to do, what not to do." **Kaviraj**: [09:51] Then we go back to the original code that you wrote. If I want to actually see the logic of what, let's say for example, you created a code in the previous screen, right? Ah, now... **Anand**: [10:01] It's Greek to me. Give to me whatever. **Kaviraj**: [10:04] So if I have to understand what... so can I say logically tell me what this code does and all of that, it will create easily it will create that? **Anand**: [10:10] Logically explain your code in three to five simple sentences for a layman. And done. That's it. **Team Member**: [10:23] All I can give that I can have, right? I can ask the PPT on explain what all I need to be looking at as an analysis. Then I can go and check if the dashboard has covered everything. If it isn't then I go back and say am I missing out something. Then ask it to create the code. **Anand**: [10:37] Absolutely. **The more... it all depends on how you want to think it...** **Team Member**: [10:41] Exactly. **Anand**: [10:42] Exactly. What you're saying is I will dump my process into it. And that is a fantastic step. But keep in mind that its process can sometimes be better than ours. So we can learn from it also. And these are not mutually exclusive. So I'll have one chat where I tell it... **Team Member**: [10:54] Same process, I'm saying the output what I want. **Anand**: [10:56] Absolutely. Same there. So I will have one window where I will say this is what I want and I know what I want based on my experience. But I'll also have another window in which I will ask it, surprise me, I have several years of experience doing this. What can you tell me that even I can learn from? **Team Member**: [11:09] That you open up a new thing. For me, yeah. **Kaviraj**: [11:11] I think Manoj will ask a question... JSON, I don't know whatever right, like you know this prompt, should it be in... Is Manoj on the call, you can also ask directly which I've forwarded to you Anand. **Anand**: [11:24] Oh wait, yes, I have Manoj's question. **Manoj**: [11:28] Yeah, hi. I saw your post on this JSON prompting. So when we ask a generic question in plain text format normally the result that comes is not very accurate. So if we do it in a JSON prompting also I think the ambiguity is not there, on the results. So just wanted to know how do we actually go about doing this JSON prompting? **Anand**: [11:57] JSON prompting is a way of writing the prompt in a structured way in a specific format. **The structured way is important. The specific format is not important.** So the answer to the question how do I write it in JSON format, it doesn't matter if it's JSON format or YAML format or just plain structured Word document. The structure is important. How do you write it in a structured way? You create section headings. Under each section you put in the details of that particular section. What should those sections be? That nobody is specifying. It's just saying instead of one big prompt where you're rambling, you might get better results by taking the prompt and converting it into something more structured. Now, I am rambling because I don't know how to convert it into a structured form, so how can I convert it to a structured form? Tell it. Take... I'll take this particular prompt, which was rambling enough a prompt, and let me ramble again to it and say "I am going to give you a prompt. Now this is a verbal prompt which I kind of rambled on. What I'd like you to do is first search online for this thing called the JSON prompting technique which apparently is about how to write structured prompts. After you research that, convert my prompt into a JSON format." And that is the first part. Then here is my prompt. So it will take that prompt, convert it to the JSON format, and you can give it. Now like I said, the fact that you have, you're giving it in JSON is not the important part. The point is structure. Instead of JSON format I could have also said instead... let's do this. **Anand**: [13:47] "I am going to give you a prompt. I am not an expert on prompting. So I want you to research best practices on prompting, and convert my prompt into the best possible prompt that I can give using the best principles of prompting." This is called meta-prompting. And we are using prompts, we got it to create prompts. Now this... the first one has after thinking for seven seconds still not giving me an answer. Fine, let it come up with the answer. This one is searching. And you can see that ChatGPT does a decent amount of thinking which makes it slow but also makes it a little more reliable for certain kinds of use cases. And okay yeah, the analysis is going on. So you'll see that it's writing code after code after code to do the analysis on this. And what it comes up with will be usually A, correct. I have not seen it make basic mistakes or even advanced mistakes. B, complicated, so you'll almost need a translator to explain it. Yeah, so I usually take ChatGPT's output give it to Claude or Gemini and say boss explain it to me in English. And C, sometimes actionable. Not always actionable. But that's okay. If I ask for 50 things, it gives me five things that are actionable, my job becomes to filter. First few times I'll be very happy. Wow, this is such a fantastic tool. After that my calibration goes up. I'll say, why am I having to review 50 things when I'm getting only five out of it? I'll get greedy, and then start looking for how do I get stuff that is even more efficient, and keep progressing like that. But the kickstarting is basically when you're sure, not even sure what to ask for, ask it. **Kaviraj**: [15:32] What are the versions you using? Eighteen thousand rupees per month for that ChatGPT or that 20 dollars? **Anand**: [15:37] Twenty dollars. For all three. All three twenty dollars. Yeah well, two of them. Gemini is free for us. **Kaviraj**: [15:43] One way to improve the quality of response as you said is ask more questions and get our memory and all of that. Apart from that is there anything that can be done on the settings and other things which can improve the quality of the responses? **Team Member**: [15:53] Along the same lines, the meta-prompts that you gave, can that be set as a standard for all my prompts going forward? **Anand**: [16:03] Got it. So firstly, in the settings go through all the settings and enable all the settings, that's generally a decent rule of thumb. Can you permanently apply it? There are a couple of ways of doing that. One is in the user preferences. And different applications have different locations for user preferences. Not able to find it on ChatGPT but usually out here. So if I click on this plus and go to personalization on ChatGPT, it lets you give custom instructions. So in my case my custom instruction which ChatGPT doesn't do a great job of following is write in simple conversational language. And even after telling it write in simple conversational language... **Kaviraj**: [16:48] It outputs everything? **Anand**: [16:50] Yeah, no, it does this so much. So I have to tell it write simply, write simply, write simply, and even after that it is very complicated. But here is where you might put it. Will it always do it? Well, it hears you and it will decide to do it probably, ChatGPT does it better than most. Instead, what I also do is I copy-paste the prompt that I need, and I have shortcuts to insert the prompt that I need whenever I need. So a small prompt library always helps. **Anand**: [17:21] If you wanted to absolutely do it more importantly than anything else, this used to... no, this is no longer relevant. Earlier I would have said you can put it in capitals and all that if you want it to really listen, but these days that does not help, it can hurt. So put it in the custom instructions, or put it, better yet, put it in the prompt directly. **Anand**: [17:41] Now, what has it done? The JSON prompting, it's saying blah blah blah, I've done my research, and for the same thing, rich synthetic attendance dataset, supporting analysis, blah. So it has created, taken the same thing and told me how exactly to prompt it using JSON. Then... maybe I can review it and figure out how I might want to change it. Or the same thing it is giving in a structured form. Between that and this, there is practically no difference. **Kaviraj**: [18:11] And this entire thing can be copied and pasted as a parameter is it? **Anand**: [18:14] Exactly. Yes. And for me, editing this is so much easier. So I just say, find the best practices on prompting and give me a better prompt, then I can figure out if this is what I want or this is not what I want. **Kaviraj**: [18:27] I think all these variables and all that you can actually edit in that is it? **Anand**: [18:30] Yes. Yeah, exactly. You just copy it and paste it. Or add additional columns and things like that. Exactly. **Kaviraj**: [18:38] And when you go to these custom GPTs available like if you go to GPTs like Khan Academy and all these things, like I think five of them can be added or things like that, right. So... **Anand**: [18:48] I have stayed away from custom... ah, this below cloud EX... Oh right. And you can create your own as well, and choose from these. These are effectively prompts and tools that people have created. I have not used them, which is not to say they're bad, but I'm not hearing enough buzz about this. So to my knowledge, this is two years out of date. **Kaviraj**: [19:13] Canva I thought has got, so I had added Canva and all of that, yeah, Canva as a top right. **Anand**: [19:20] And you can in the settings... where are the connectors, data connectors, does anyone know where the connectors are? Apps, yes. So out here you can add Canva potentially, yeah, as a connector if you are, if you can log in. I don't have a Canva account. **Kaviraj**: [19:35] So these are the places where ChatGPT itself has got organizational arrangements with them to kind of crawl the things? **Anand**: [19:40] Exactly. So you can take stuff from there, you can put stuff into that and yes. **Kaviraj**: [19:44] See initially it asked for a login and it asked me from Canva login, generate create something. **Anand**: [19:50] It just connects. **Anand**: [19:52] So, the analysis of fundraising. So what it has done is analyze that synthetic data and says, here are your recommendations. Split the engagement events from fundraising events and manage them differently. Attendance is similar, but donation yield is wildly different. Which, yeah, might... it's a synthetic dataset, but it is a reasonable outcome. Spend less time optimizing channels, and more time improving audience selection. So it's analyzed saying, okay, doesn't matter what channel, the results are the same. Choice of audience is what matters. Useful. Fix RSVP forecasting and event instrumentation. A number of people who say "maybe" or "no" still attended. Oh, okay, so your RSVP data is not a reliable attendance forecast, which means that we need some mechanism where somebody walks in at the last minute I should still be able to do research about them, target them, and so on. I don't know, Canva is a search, something that can automatically identify them, or even just as somebody quickly going in at the reception, noting down who it is, adding it to a queue, and it does a research so that somebody can look at it... **Kaviraj**: [20:54] No, it has based on the synthetic data it consistently created it in a compelling way. **Anand**: [20:58] Exactly. Correct. But what you're seeing is number one, even with synthetic data, the advice that it's giving is sensible. **Kaviraj**: [21:06] Yeah yeah. This is a very good analysis, use cases actually. **Anand**: [21:09] Correct. And with real data, it will be sensible and correct. Which is... **Team Member**: [21:14] Somewhere all these thoughts would have come into my mind at least when I read... **Anand**: [21:19] Exactly. So this is where I find asking it "just think of random thoughts, teach me something based on that", works quite well. So well that I built an ideator tool. **Anand**: [21:30] This ideator tool takes two inputs... one, and how did I build it, I went to Claude and said build something, and I want to save it, and I saved it on GitHub, that's all. This takes two inputs. Basically from my notes. I just organize my notes into a variety of topics. So let's say I'll take something related to LLMs and something related to things I learned. So the internet is forking into human internet versus agent web. Something I noted recently. I keep forgetting QR codes need a white border for them. Absolutely unrelated. But these are two random thoughts that I had. So I am going to now ask Claude to combine these two ideas into, what shall I combine it into? A research topic, or no, a fundraising technique. Why not? Completely random. This will have nothing to do with fundraising. Right? But that is the point. So now the prompt is, and I'll trigger it and then... [inaudible] Okay, this caution, malicious, yeah okay that's okay, run this. The prompt basically says, you are a radical concept synthesizer. **Kaviraj**: [22:40] How it's custom text... **Anand**: [22:42] I'm finding is that we need adjectives more than nouns these days. **Anand**: [22:51] And generate big, useful, blah blah blah concepts aligned with a fundraising technique, and you have to combine these two concepts. And here's how you have to think about it. Generate five ideas, rate them on novelty, usefulness, feasibility, here's how I want you to structure the output. Keep it interesting and simple. And it's generated five ideas, and it's saying a human-verified match campaign. So the internet is splitting into agent web and human web. Physical QR codes become a last hard to fake proof of humanity checkpoint. So what I should build is... okay corporate match programs are getting gamed. Bots can click links, fill, simulate donations, etc. So sponsors are scared of matching AI astroturfed campaigns. They want proof that humans care. Okay. So print a QR code at your event. Mail in postcards on coffee sleeves, whatever, that give them a badge, a signal. When the human scans, they land on a donation page that generates a one-time presence token, tied to that physical QR code. It basically knows this person at this spot etc. And then you can go to your corporate sponsors and say that look I will deliver 500 verified human donors, not random clicks etc. These are people who have physically been there. **Kaviraj**: [24:14] Now, this is understood. Now that other query that I had asked Mumbai Vice Presidents and all of that has it given a result here? **Anand**: [24:21] I'll come to that in a few minutes. Now here is the thing, because we are able to generate ideas of so many kinds, **what I'm finding is, and it is practically free, the way to use AI is often not by asking the question "how can AI help me in my job?" The question to ask perhaps is "how can AI help me find out what my job is going to be when I don't even know what it is going to become?"** It's a very different way of thinking about it. It can tell us stuff that we don't even know to ask for. **Anand**: [24:56] Now to your question, which is something that we have asked for. Oh, okay, it's ok, no it's still running, it's done 151 searches but currently it's looking for Naveen Tahilyani. **Kaviraj**: [25:10] Who is in this BP list? Naveen is 95 batch... Naveen is prudential VP exactly. So Varun Srinivas... **Anand**: [25:21] So Varun Srinivasa Sridharan from Apax looks like a solid candidate. Devika Ghate from MSCI. **Kaviraj**: [25:29] Now, I'll direct you to one more question Anand. How did you actually... what has been your learning journey on this like... I mean you have built analytics, all these things right. So what has been your, when you started, there was a day zero, how did you actually go about doing this whole thing at least actually practically if you can aspect to learn. **Anand**: [25:51] I, when ChatGPT was released, the first time I heard about GPT was when somebody took a comic generation tool and got a GPT version 1.5 to generate comics out of it automatically. Which looked like a very cool thing, but I didn't really understand it. ChatGPT was released in November 2022. I tried it, I didn't get what the big deal was. April '23 or June '23 I had to give a talk for a client on what AI was. So I very quickly, last minute, one hour or one day before the presentation, had to at least use it so that I know what it is. I bought the twenty-dollar subscription, started playing around with it. At which point, this is smart. Then I searched on Google... at that point searching on ChatGPT was not a thing. For "how are people using it?" Okay, that looks interesting let me try it. This looks interesting let me try it. It's like, there are at least half a dozen things that are useful for me, and out of those, half of it it totally messes up, half of it it does a good job of. But that is fantastic, that is practically free. Since then, I've been just trying to see how I can make my life more interesting. That is all. And because every attempt I make at it gives me such a big kick. Useful, maybe. Interesting, always. That I just wake up thinking of what can I do today with it that I haven't tried before. It's like a new... something is creating a new movie for me every day that I can watch and it is exciting, releasing a new episode of Game of Thrones if you will, every single day. It is that exciting which is what keeps me going. Now this is not true for someone who's trying to use AI to get a job done. For them it is a tool. For me it is my hobby. Which allows me to go a little further. Now therefore shouldn't be a hobby for you? Not necessarily. On the other hand, this is such a vast area that you'll find some niche where it is interesting. **Anand**: [28:00] For me for instance, my daughter was turning 20. The night before that I had a series of calls, I couldn't even go out and buy a cake. What am I to do? Until then, my practice was clear. Anything I don't know or even don't even know that I don't know, I ask AI. So I went to Claude and said look, it is ten o'clock in the night. Tomorrow morning for her birthday, what can I do? And... **Kaviraj**: [28:29] For asking to plan for the... **Anand**: [28:31] Yeah, exactly. And I said look, these are... I have basic household things, I just looked around I see I can see bread, peanut butter and all, I see what all can they solve? What can I do? And it gave me such a brilliant idea. It said, take the bread, toast it until it's slightly crisp. Put peanut butter inside it and a little bit of jam. Coat the outside with Nutella. It actually looks like chocolate frosting. Put a little bit of peanut butter mixed slightly with water, not peanut butter, normal butter, you know, normal butter, melt it a little bit, make a slightly pale mixture, run it out along the top. It will look like, you know, that nuts will come out there. Do a little bit of sprinkling. And then using peanut butter which has slightly different color, or butter, you'll have normal butter, melt it a little bit and write happy birthday Dhia on top of it. I said but wait, I don't know how to write that well. It said very simple, take a plastic milk packet or something, cut a small hole at the bottom and then you will be able to squeeze it in just that right shape. Then it said, you have these what candles are they? **Kaviraj**: [29:32] Spot candles. **Anand**: [29:34] Ah, that sort of thing, those small little ones, tealight candles. You have enough of those, arrange them in a decorative pattern, take 20 of those, three along the diagonal and two on the straight, effectively creating a star-like decoration. Or it gave me several other patterns for decoration. It was stunning the effect that I could create with that, and it took one hour. Which is the amount of time that it would have taken for me to go order a cake... --- **Kaviraj**: [00:00] ... to go order a cake and get it. Now after this kind of an experience, I will worship any tool that gives me this. **Unsure**: [00:12] Amazing, and fascinated with your usage of AI. Also in the beginning of this training session you were saying make at least 50 conversations with a tool. I am just curious to know the environmental factors, or how it affects the environment. **Anand**: [00:27] Yeah, I've calculated it. It is roughly the equivalent of, each chat is roughly the equivalent of three drops of water when taking a bath. **Unsure**: [00:38] So the hype that is getting created is not as much as what we see, is it? **Anand**: [00:42] I don't know how much hype is getting created, but 50 is basically about one third of a mug of water. So if I want to offset it, I will take a bath with one third mug less. **Unsure**: [00:54] That's your contribution, but others can contribute more if they don't take a bath at all. **Anand**: [01:00] Yeah, don't take a bath at all. That's okay. When you're talking quantitatively about this ChatGPT, and if somebody is saying no, no, I actually, I am using automated things and that consumes far more, yeah, contribute in a different way. See, there is a lot of consumption because of that. The consumption is happening because it is useful. Electricity consumes a lot of energy, but the alternative of burning coal is worse. So AI is generating much more value than the cost, that is fine. That is also not a reason not to reduce the energy usage. So do that, but don't do that by using AI less. **There it is delivering actual value. Use AI even more and figure out ways of reducing inefficient energy consumption, is my submission.** **Unsure**: [01:52] Is this the idea behind calling yourself the first LLM psychologist? **Anand**: [01:55] Oh, I was delivering a talk at MDI Gurgaon, and I needed some gimmick for the talk. So, and this was before I had the sense to ask AI for what kind of gimmicks can I do. So I manually was thinking of gimmicks, sadly enough, and found one tweet by Andrej Karpathy, who is an AI researcher, from November '23 or something like that, where he says the way in which we are prompting LLMs is now becoming almost like a psychology. And there will probably be a need for LLM psychologists in the future who will need to prompt. I said great. I went on stage. I called our Head of HR and said, do you have any problem if I start calling myself LLM Psychologist as my official designation? She said do what you want Anand. So I went on stage, opened LinkedIn on stage and said, change my designation and said, you are now talking to the world's first LLM psychologist. Branding. **Kaviraj**: [02:54] So one more thing on that, like this quality of the results that comes, so you see hallucinations right, actually many of these things when you ask it becomes... I asked once like giving some examples, demonstrations, and it said Toyota did this and then I realized that Toyota didn't do any of those things. It basically took a concept and enforced it on Toyota and said that Toyota did this. Right? So and it became very... if I'm giving a presentation it becomes very embarrassing. So how do I make sure that this is like not a scope for hallucination? **Anand**: [03:20] Very good. Yesterday I was at one IIM interview panel and these are people who have cleared CAT, very smart people. One guy who has about 37 months of experience in an analytics company, who is an analyst, Professor Subhagata Das... **Unsure**: [03:39] I think he taught me stats. **Anand**: [03:40] Ah, stats. He asked a question. US GDP, he showed a chart. You can see US GDP is 3% growth. Europe's GDP growth is 1.5%. What is the combined growth? He said 4.5%. We were like, is that so? Are you sure? Yes sir, 3%, 1.5%, combined GDP growth is 4.5%. We were stunned. US, much larger, Europe, much smaller. GDP growth, do you add them? Do you do weighted average? No, no, no, you add them. Very confident. **That is CAT-level human hallucination.** I compare that with AI hallucination today. And I ask myself, who would I trust? This guy who will probably get into a leading college, probably get into a leading consulting firm, who will probably give me advice in the form of a presentation, or this machine which is going to cost me for a thousand such presentations $20, and hallucinates less. Number one. Number two, how much am I going to trust that consultant? How much am I going to trust this? **I actually trust this less, which is good, which means I'll be more skeptical. I can do cross-checks.** How much is it going to cost to do the cross-check? All I have to do is ask it in Henry Kissinger's words, "Are you sure? Is this your best work?" Seriously, all you have to do is ask it again. Fact-check yourself. Do a good job. **Kaviraj**: [05:15] So I can bring that persona as this thing that all are source-checked and validated and fact-checked. **Anand**: [05:19] Exactly. And if I am not confident, I will say, "**I know you have three errors in this. Find them.**" Even if I don't know them. It will find, find, find, find, find. It can even hallucinate errors if you push it too hard. But it will do such a diligent job of finding the errors you will start feeling sorry for it. It will come up with, you know, "Here I used the phrase something, I could have used this phrase, it would have sounded better, and that is an error." **Unsure**: [05:53] I'm almost reminded of that Dobby in Harry Potter. **Anand**: [06:00] You slave-drive it, it will get there. But I can also ask another LLM to cross-check. I can say give me the citations. When it gives the citations it tends to hallucinate less because it is reading it. And then I can also verify. So with all of these, I have better tools to verify, I have faster tools to verify, so compared to a human output it is anyway better. **So the question often is what are we comparing against? Software's reliability or human's flexibility?** If you are comparing against human's flexibility, this is more reliable and slightly less flexible than a human today, but faster, cheaper, far more broad-based in its capability. If you are comparing with software, the flexibility is so much higher. And you can tell it to write software. It writes the code and runs the code. When it writes and runs the code, then it will not fail. So where you need analysis, you tell it, write the code and run it. It will not make that kind of a mistake. Fact-check. **Kaviraj**: [07:04] One of the benefits of the team, one last question on the current resources out there, right? So if we have to track some Twitter accounts or LinkedIn accounts which talk heavily about AI-related things which we can receive... **Unsure**: [07:16] Usage. **Kaviraj**: [07:17] Yeah, and on a regular basis we will be able to share or something... **Anand**: [07:21] I can... you will get easily overloaded, so I'll keep it simple. Follow Ethan Mollick and Simon Willison. **Unsure**: [07:28] Simon? **Anand**: [07:29] Willison. W-I-L-L-I-S-O-N. **Unsure**: [07:38] Can you type it... Ethan? **Anand**: [07:44] Yeah, E-T-H-A-N M-O-L-L-I-C-K. And even if you spell it wrong Google will get there. So Ethan Mollick is one, and Simon Willison is the other. **Unsure**: [08:00] Are they on Twitter or LinkedIn? Both of them? **Anand**: [08:04] Anything, they are on all kinds of channels. **Unsure**: [08:06] On both. **Anand**: [08:07] Yes, these are the two. **Unsure**: [08:10] Two good people. **Anand**: [08:12] Yes. **Kaviraj**: [08:16] Good. Okay, this is a... lot of inputs for us which you have given, right? And I get into this... As usual, like [inaudible] sit in like an Amazon pod and ingest it and [inaudible] to the team. **Anand**: [08:32] See, a lot of this, like I said, I'll reiterate. **It's just using it 50 times a day. The rest is detail.** And don't strain yourself. Talk to it. Ask it to tell you a story. I have to sort out what I have to make for dinner today, take a picture of your fridge, send it, say tell me what to make. I want it to be something different. It's stuff that you do on a daily basis. Just use it. **It is like an assistant that will follow you and do everything that you tell it. Use it.** All the best. **Unsure**: [09:10] Can we internalize all of this? This talk will be converted to... **Anand**: [09:14] Yes. **Unsure**: [09:15] How much he shared, and the PPT will be shared with the team. **Anand**: [09:20] Absolutely. And if you... you will find it on my talks page. Just search for 'S Anand Talks'. I am not even going to send the link. You will find it by tomorrow. **Unsure**: [09:30] Sure. Thank you.