# Transcript

**Sandeep**: [00:00] So she wants to do her series on health and aging and wellness.

**Debi**: [00:04] Sure.

**Sandeep**: [00:05] So she’ll curate four, five sessions, whatever, three, four, whatever we want to do.

**Debi**: [00:10] I think it's a good idea. Yeah, yeah, we should do that.

**Anand**: [00:13] Good afternoon.

**Debi**: [00:14] Hi Anand. Good afternoon.

**Sandeep**: [00:15] Hi Anand.

**Debi**: [00:17] So, Anand, I reached out to you, I think separately, and Sandeep, I also wanted to tell you that I think SINDA [Singapore Indian Development Association], I think Ravish called me up on behalf of SINDA to say that can Anand help the SINDA team do some work using Copilot. So I think my problem is that's why I reached out to Anand. Anand, listen, please be—feel free to say no, okay? Because there'll be a lot of these requests which are going to come your way. So I reached out and Anand probably said yes.

**Anand**: [00:53] No problem. Excellent.

**Debi**: [00:58] And I think we've done quite a bit of work with SINDA, Anand, and you know, which I think has been good for the community that they work with. So I think we are very happy as a team.

**Sandeep**: [01:13] Yeah, so Anand, this is something which can be low touch. You don't have to go overboard.

**Debi**: [01:21] Hi. I can see a VJ.

**Anand**: [01:30] Okay, more people will join.

**VJ**: [01:34] Hi, this is VJ here.

**Debi**: [01:36] Hi VJ. Thanks for logging in early.

**VJ**: [01:39] No, no, I just wanted to make sure I'm on time. But a fascinating course, so...

**Debi**: [01:44] I’m so glad. Thank you. Professor Anand is here.

**VJ**: [01:48] Yeah, I saw him. Hi, Professor Anand, how are you?

**Anand**: [01:52] Good, yeah. And you?

**Debi**: [01:55] So we’ll wait for the others, VJ. We'll start.

**VJ**: [01:58] Sure, sure. I’ll just go on mute. Thanks.

**Debi**: [02:01] Yeah, thank you.

**Sandeep**: [02:03] So Anand, we were thinking of doing this hybrid, but planning was a little shoddy on my part.

**Debi**: [02:14] So the last one we’ll do an in-person session, and that is something that I’ll tell people, that we will do an in-person session. We’ve got the venue identified. And I think, Anand, I had got the dates approved by you earlier, but let me just make sure that—just confirm back whether you’re fine, okay?

**Anand**: [02:35] Yeah, I think it was end of August, right? The last Saturday of August.

**Debi**: [02:39] Yeah. Hi Sonal.

**Sonal**: [02:42] Hi Debi, how are you? Hi Anand. Hello.

**Debi**: [02:51] Sonal, I'm very impressed. You’ve got a—yeah, you’ve got an AI background compared to others. Very impressive.

**Sonal**: [02:58] The thing is, you know, **this is what ChatGPT thought I looked like, which is not correct. But I’m not willing to correct it because it looks better than I do, so why not?**

**Debi**: [03:09] Sonal, no!

**Sonal**: [03:10] No, it did! It did! It actually picked up this image. I didn’t give it the image. So I don't know, I think it looked at my LinkedIn and my public profile or whatever and then just, yeah, decided this was me.

**Debi**: [03:28] Yeah, I think we'll do an analysis of this picture later on, Sonal. Hi Deepika. And hi Anju.

**Sandeep**: [03:43] Ah, Deepika came in, yeah. Dhananjay is there. Hi Dhananjay Patnaik.

**Dhananjay**: [03:51] Yes, hi Sandeep. Hi everyone.

**Debi**: [03:57] Everybody, we'll wait for a few minutes and then we’ll start off. Okay?

**Anand**: [04:02] Sonal, I’m using a note-taker also today, so we can compare, you know, how good Microsoft is versus the others.

**Sonal**: [04:07] Even better.

**Debi**: [04:20] Hi Deepika. Hey, hi! And I was telling Sandeep and—sorry, Anand—I think Deepika is somebody who has just joined. I think this is her first session. And she was asking me yesterday about what is it. I said, listen, I'm obviously biased because I'm an organizer, but Anand is a great guy, right? So you should definitely attend.

**Deepika**: [04:47] I’ve heard all good things. Yeah, looking forward. Thanks, thanks.

**Debi**: [04:52] Welcome, Deepika. And I think you should have got the recordings for the last two sessions, I think, Deepika, right? We’ve sent it out.

**Deepika**: [04:59] Yes, yes. I did receive them. Thank you.

**VJ**: [05:01] Hey Deepika, how are you doing? VJ here.

**Deepika**: [05:04] Oh my God! How are you, VJ?

**VJ**: [05:09] Yes.

**Deepika**: [05:10] So good to see you! And I can see Anju as well, Anju Dixit. Hi Anju. So many familiar faces. Sonal, hi!

**Sonal**: [05:22] Hi, hi Deepika. How are you?

**Deepika**: [05:32] VJ, we need to catch up separately. It's been a long time.

**VJ**: [05:36] Yeah, provided you come to Dubai.

**Deepika**: [05:39] Oh, that's where you’re joining from! I was just going to ask where are you joining from.

**Debi**: [05:46] VJ, you’re joining from Dubai?

**VJ**: [05:48] That's right. I used to be in Singapore for many years. I’m still on the group, but when I saw this course, I couldn't let go. It was too interesting, so I said let me join.

**Deepika**: [05:59] Yeah, I was a bit confused but then said no, you must come back.

**Debi**: [06:03] Well, thank you so much, VJ.

**VJ**: [06:05] No worries. No, so Deepika, you were saying, you know, what this course is about. It's **"what you wanted to know about AI but were afraid to ask."** So in four sessions, you’ll know about it.

**Deepika**: [06:16] Okay.

**VJ**: [06:17] And you’ll no longer be afraid after the four sessions.

**Deepika**: [06:20] No, yeah, after that it's—then you figure it out, right?

**Debi**: [06:29] Yeah, so if everybody’s fine, I’ll just wait till 2:35. Is that okay? We’ve got about 19 people, you know, 20 on the call. Is that okay with everyone? Then at 2:35 I’ll start sharp. Okay?

**Sandeep**: [06:41] Anand, summaries are very popular from what I—I have been told.

**Anand**: [06:46] Glad to hear that. We'll do more of those.

**Debi**: [06:51] I’ll just run and get a pen and paper and be back.

**VJ**: [07:01] Anand, while you're covering the session today, maybe sometime at the end, if you could tell us—especially with Claude—how do you ensure that, you know, you can reduce the use of tokens? Right? Even when I’m using Claude Pro, I seem to run out of tokens very, very quickly unless I ask it to strip it and, you know, work in a disciplined fashion and stuff like that. But if there’s any tips or tricks on that, that would be really useful.

**Anand**: [07:27] Most certainly. Though I’ll probably ask for tips on how you end up using that many tokens. I get worried that I’m not using anywhere near my $20 limit. So we should exchange notes.

**VJ**: [07:40] Sure, we will.

**Sonal**: [07:42] How do we know how many tokens we’ve used?

**VJ**: [07:45] So if you go to settings, and in the settings there’s an option called usage. In the usage you can go and click and you’ll be able to figure out how much you’ve used.

**Sonal**: [07:53] For ChatGPT as well?

**VJ**: [07:55] No, this is for Claude Pro. Yeah.

**Sonal**: [07:57] But does ChatGPT have the same capability?

**Anand**: [08:03] Yes, but you may not need to worry about it because you won't run out of it much.

**Sonal**: [08:12] Okay. Okay.

**Debi**: [08:14] I think in ChatGPT, if I'm not mistaken, Anand—I’m not sure, I don't know—I don't think we track usage. But if you really run out of it, I think in most systems you’ll get an alarm anyway, Sonal, right? Saying that, you know, you’ve run out of it.

**Sonal**: [08:28] Yeah. Not happened so far. Not happened so far. So not using enough, as Anand is saying.

**Debi**: [08:33] No, I think the problem is if you're going to do a lot of processing, which is what I think VJ must be alluding to, because I think that's where you land up using a lot.

**Anand**: [08:44] Right. Writing code takes a little more, and we'll be covering that today.

**Debi**: [08:49] Yeah, writing code has also taken a lot, yeah. Correct.

**Sonal**: [08:52] Yeah, I did ask ChatGPT a couple of—a few weeks ago, **"What is my carbon footprint of using ChatGPT based on the last 15 months?" It did a huge calculation and told me it’s one business class flight within Southeast Asia, which is something you would do in any case in a year. So don't worry about it.** It was very convincing.

**Debi**: [09:18] But Anand, there is—I guess after this, Sonal has to basically ask for references so that it can be—I mean, one can go back and see how verifiable it is. Did you try that?

**Sonal**: [09:33] No, the thing with ChatGPT is it's so verbose. Like it talks so much that you just run out of cognitive bandwidth, you know? You’re like, "Okay, okay, fine, okay." If you really needed to get into every source it's citing, it's just going to take up too much time. So I'm just—I mean, it's a low criticality item per se, so yeah, even if it's cheating just a little bit, it's fine.

**Sandeep**: [10:01] Yeah, Debi, we've got 32. I think we should start. It's 2:35.

**Debi**: [10:05] I’ll start, yeah. Okay. First of all, welcome everybody, and thanks everybody for joining in. I’m extremely happy that there are a few new people joining in today's session. I think this is all attributable to Anand, your star power, and the word of mouth that we have generated on these sessions. So thank you for that. I think a few house rules: During the entire session, I think it would be good that if you could put forward your questions in the chat section. I think Anand will pick it up during the session or if he can't cover it, you know, towards the end, I think Sandeep or I will pick it up, right, and will—what do you call—and will make sure that, you know, Anand covers them. Okay? And post the session, I think as always Anand sends across a very nice recording as well as a script of what has transpired and all of you will get it. If you have a problem receiving it, if you don't get it, just write to Sandeep or me in order to, you know, get that. Okay? So I think these are some of the house rules. I will pass it on to Sandeep. Sandeep will do a rapid fire today with Anand, you know, to start the session on an interesting note, and then we’ll ask Anand to perhaps do a few minutes recap of the previous two sessions and then get on with today’s—you know, what we plan for today. Okay? Over to you, Sandeep.

**Sandeep**: [11:30] Thank you, Debi. Welcome Anand. Thank you for doing this. And let's—I thought let's have some fun to start with. So let's have a quick rapid fire. Take you two minutes. So one-word answers for the first few. **ChatGPT or Claude?**

**Anand**: [11:44] **ChatGPT.**

**Sandeep**: [11:46] Type or Talk?

**Anand**: [11:47] Type.

**Sandeep**: [11:49] Oh, I thought you liked to talk. Agents or Tools?

**Anand**: [11:53] Agents.

**Sandeep**: [11:55] **AI Hype or Real?**

**Anand**: [11:56] **Real.**

**Sandeep**: [11:59] Okay. **Most overrated AI skill people waste time on?**

**Anand**: [12:03] **Prompt engineering.**

**Sandeep**: [12:06] Okay. The last time you used Google instead of AI, you mean Claude or ChatGPT?

**Anand**: [12:13] Today.

**Sandeep**: [12:15] Okay. **Who knows you better, your wife or AI?**

**Anand**: [12:19] **AI.**

**Sandeep**: [12:25] AI bubble or no?

**Anand**: [12:29] No.

**Sandeep**: [12:30] **One AI stock you would buy? This is not financial advice, so guys keep that...**

**Anand**: [12:37] **Nvidia, still.**

**Sandeep**: [12:41] **One job in the room AI will kill first?**

**Anand**: [12:47] **Personal financial advisor.**

**Sandeep**: [12:51] **Will you be replaceable by AI within the next three years?**

**Anand**: [12:54] **Yes, I already have been.**

**Sandeep**: [12:59] **Best use of AI you've stolen from someone else?**

**Anand**: [13:02] **Figuring out the—actually, somewhat obvious one, but connecting my emails to the entire context of my code, my documents, my notes, my blog, everything. And I hit inbox zero yesterday thanks to that.**

**Sandeep**: [13:38] Impressive. Thank you Anand, that's, you know—with that, let's hand it over to you. Maybe a quick summary of what we've done so far and take over.

**Anand**: [13:48] Before that, I have a counter question. Did you use AI for preparing these questions?

**Sandeep**: [13:52] Yes! I'll tell you how you prompted it.

**Anand**: [13:56] How did I prompt it?

**Sandeep**: [13:58] I'll have to scroll up because I've done a few rounds. So you know, I said, "Look, Anand's done this AI session, give me 10 amazing questions which will fox him or maybe also regale the participants." And it came up with a list. I said, "This is not interesting enough. I want shorter questions, one-line questions." Then a third prompt, then a fourth prompt, and then this is what it came to.

**Anand**: [14:24] But you do tell me the third prompt and fourth prompt were...

**Sandeep**: [14:26] The second one was, yes, then I said, "Yeah, talk about jobs and the one AI stock to buy." And then it said, "Yeah, then I said reorder them, you know, to start slowly and build it up." Yeah, that's it. So it started warming up, the money round, spicy finish, close.

**Anand**: [14:59] Very interesting. This is something we should probably come back to in this session and talk about it as well as VJ, what you mentioned about token usage, Sonal on carbon footprint, and Debi, what you mentioned about verification. Let's take these as themes, but first as you suggested, I'll very briefly recap what we covered. In the first session, we looked at how to give AI better context, broadly what's—Debi, you want to become the presenter maybe?

**Debi**: [15:33] I’ll try sharing my screen. Right now I'm just talking anyway. But just checking if I can share my screen. Is it visible?

**Anand**: [15:42] Yes. I’ll stop sharing for now and just talk for a minute or two. So we in the first session covered what is broadly called context engineering and how we tell AI what we need it to know. And that's comprised usually of four things: **Prompts** (which is what we type in), **Files** (which by and large we upload), **Memory** (which is allowing it to access past conversations and other stuff that's growing these days), and **Tools** (which are the connectors or things that appear in the settings that either let it automatically fetch stuff or automatically do things). **The basic idea was that AI gives much better answers when it knows more about you, your work, the problem, etc.** And we ran a few exercises, including creating an AI team's background, which—thanks, Sonal, for having that on the screen. And the most practical habit that we ended with was simply keep a prompt library and reuse good prompts. Some of the broader habits were asking AI to interview you when you don't yourself know the problem, running a post-mortem after important pieces of work to see how we might improve either the prompts or the files, etc., and learning from each other, such as the way I was trying to learn from Sandeep's prompt and see how we can improve. In the second session, this was more on individual prompts to complete workflows. We took the first workshop itself as an example, took the recording, transcribed that, created a comic, created an infographic, converted that into an HTML story. We also spent a fair time on a fairly critical piece of the workflow, which is checking whether AI is right, asking for sources, checking for facts that we already know to see if it is getting those, finding counter examples, getting several other models to cross-check one another. And how in different experiments we find that, for instance, a model that might have as much as a 14% error rate, when you have two models double-check or even five models act as a committee and agree, that error rate of 14% can fall down to 0.7%. We also looked at skills and schedules. A skill is roughly like a pre-created prompt that you can store that gets automatically called, and a schedule is something that lets AI do something useful on a daily, weekly basis without waiting for us to remember. So the first session was broadly about what AI knows, the second was about how AI can be used to work for us repeatedly. Today we’re going to be building on both of these and looking at analysis, how AI can examine data and content, find patterns and help us reach conclusions. But I'm quite curious actually and would love to hear from you: Has anyone been trying any schedules in the last few weeks? Yes, Sonal?

**Sonal**: [19:11] Yeah, I’ve been trying schedules at work already where I use Copilot, and also been trying schedules at home with ChatGPT. Yeah, but I'd like to move from it—move it from just prompting me to actually doing the thing that I'm asking it to prompt me to do. So I need to figure out how to do that. Yeah.

**Anand**: [19:35] Got you. But what are the more useful schedules that you find?

**Sonal**: [19:40] **At work, it’s about "Give me a summary of emails from the last one week or from yesterday that I need to respond to or where my inputs are required." In my personal life, it is about, you know, prompt me about doing a—you know, building my network. You know, what is it that I need to do once a week to expand my network or deepen my network? As in my professional network. Yeah.**

**Anand**: [20:11] Nice. Any others have been trying schedules? You’re welcome to put it in the chat window as well.

**Sandeep**: [20:17] Yeah, just using it more for work. Summarizing, you know, preparing for the day to start with, and getting summaries of meetings and that, and also I've been using it to schedule a LinkedIn posting schedule for a few things. It's not there yet, but yeah, trying to work that.

**Anand**: [20:49] Nice.

**Sonal**: [20:50] My prompt is also linked to LinkedIn, so it's supposed to prompt me based on my LinkedIn thing, but it can't, so I need to figure out how to get it to log in to LinkedIn etc.

**Anand**: [21:01] Got you. Others, any schedules? Okay, VJ says, "I get Claude to run..." sorry, I kind of came in... **"So I just set up a market sweep schedule, so a couple of them. One does a weekly, one does a daily. It goes, picks up different markets in the different prompts, insights, and it’s kind of I learned over the week to kind of keep re-tuning the prompt to give insights, pick one or two lines in terms of facts, and then pick three or four lines in terms of what people are doing in the market or expecting the market to perform. So it's kind of been playing around with it, but it gets—I get a nice summary."** Going back to your earlier thing, I can very clearly see you said that financial advisors will disappear, Sandeep and Debi may know, in the market there's lots of sales people—they will all disappear, you know, because I'm getting some very nice curated insights on markets through these schedules.

**Anand**: [22:06] Fascinating, VJ. No, that was very helpful. And we have a few more on the chat. VJ runs Claude to get a briefing on markets, currency, stocks, US and Asia. Rags gets ChatGPT to do a Saturday morning scan of stocks and markets and give ideas for trades this week, which again is on a similar theme, which is very nice. And **Bharath’s using it for a simple one to do an internship job search weekly for his daughter, which again is a very practical and real one.** Nice. And Shankar’s negotiated his position on an investment without a lawyer, which is very interesting, Shankar. But I'm curious, was that a schedule or just a use of AI?

**Shankar**: [22:57] Oh, it was a—I just did track. I just did track.

**Anand**: [23:01] Okay, okay, got it. And you fed it what context?

**Shankar**: [23:04] I was basically looking for an investment, a venture investment, and I had some clauses which were open-ended. So I just ran it and let it benchmark against two or three funds that I knew. It came back and gave me the entire negotiating positions, which I then—so I’ve been on it for the last two weeks and **I’ve got reasonably what I want without a lawyer.**

**Anand**: [23:33] Nice, nice. And hearing this kind of contract negotiation a couple of more times this week, so looks like the cursory legal analysis is another profession that might be...

**Shankar**: [23:49] It's not a cursory legal analysis. **In fact, in this case, it has beaten a lawyer whom I also benchmarked against. So finally I had to tell the lawyer that, "Look, Claude beat you."** So it was a two-week of engagement.

**Anand**: [24:05] Interesting. Has anyone been filing taxes with or verifying taxes with Claude or ChatGPT or something? If so, please do share.

**Sandeep**: [24:17] I also use the same thing, Anand. I negotiate a lot of NDAs and master service agreements, and I'm running them all through different—you know, Claude for the most part. And you'll be surprised what kind of, you know—people don't read these documents before they send them out. So **I just signed an NDA where people said, "This company's based in Bangalore, I'm based in Singapore," and he says it's—the applicable law is UK.** I said, "Guys, why is this UK law?" They had no idea.

**Anand**: [24:51] Interesting. And Sandeep, what I’m finding also, based on partly what you said, is that quite often it's an AI at this end and an AI at the other end that are, you know, one's generating something and the other's verifying something. And obviously, not that either is doing a bad job, it ultimately then boils down to the person prompting and what context, what questions they’re providing, right? And yeah, Arvind’s mentioned tax advice. And it's okay, interesting. **Tax advice from Claude was gibberish, which is a good counterpoint to have.** So tax advice is probably not in the—out of a job for a while. Though, Bharath, your experience is that it did do at least a preparatory work for this. Fascinating.

**VJ**: [25:42] The other one, the other one I used—and I think somebody's referred to it—is **basically 26AS, AIS, and TIS to try and reconcile it and then build the dividend schedule, you know, across those five days for each of the stocks that you own so that, you know, you can feed it in.** So yeah, it did a good amount of reconciliation on those and quite easily. There was some tweaking we had to do, but it was good.

**Anand**: [26:08] And VJ, that actually then brings us to the topic that I wanted to touch upon in this session. Reconciliation is where we're effectively passing it a data set or, in Bharath’s your case, a bank statement, and we're asking it to summarize. And these are effectively forms of data analysis. The interesting thing is that we've got a few examples that touch upon some of the common kinds of analyses that we do. So you mentioned for instance calculating the carbon footprint of your chats. That's effectively having it create the data. Or Krishna, in your case, it’s 401(k) and having it analyze across two different sheets to find out which one's right, which one's wrong. **The potential for analysis, however, is often a lot larger than many of us might think of**, and a part of what I wanted to do in this session was open our minds to what else is possible. And the way we’ll do that is I'll start with a small—couple of small examples of what happened in the course of this month, for instance. Let's start with—before I share my screen—GIPCL is Gujarat Industries Power Company Limited. They have a clause that says that when they generate solar energy, they have to make a forecast saying, "This is the amount of energy that we will put into the grid," and are penalized for not matching the forecast, plus-minus there is a penalty. They sometimes cannot control this because the wind can be very strong in certain seasons, at which point the solar panels have to be stored. The question then is, "How can we predict on the day whether we need to store it at particular times?" Because they get the wind schedules in the morning. The company that makes the software to stow at specific times can't be told beforehand before these arrive, so they have to do this manually. Here’s what they did. This was a one-hour session followed by them exploring and trying out by themselves. **They had an Excel sheet that created the forecast. They uploaded it into ChatGPT. They talked for two, three minutes explaining the problem. ChatGPT churned for about half an hour, one hour—it was the Pro version. They tested it. The next day, effectively the DSM penalty that they have to pay had reduced by—for that particular day—163,000 Rupees, covering well more than their GPT Plus subscription.** Because ChatGPT had simply given them a new Excel sheet and said, "I’ve gone through all of the past data that you've given us, and here are a few corrections to your model in Excel. And if you apply this, you will get a better forecast." And they did, and they’ve been continuing to use this on a regular basis. The takeaway is simple. **Sometimes, you can just give it an Excel sheet with a model, the financial model or any model, and ask it to improve the model.**

---

**Anand**: [00:00] ...financial model or any model, and ask it to improve it, or even build it from scratch. And some of that is what we're going to be doing today. And it can—let's take the other side. The Times of India said, "We have a property called **Statoyistics** (statistics with a 'TOI' in the middle), and this is a property they said they have to shut down because it's taking too much effort to create little cards that have interesting snippets." We said, "Okay, why don't we ask AI to do this?" Same process. Go to ChatGPT, ask it to search for some public data, create some interesting cards. And this is what ended up coming out. Let me share this.

**Anand**: [00:46] Over a few iterations, it took about a day to get to here. And an example of one of these cards is out here. This and many of the others that you saw were actually published on The Times of India because they looked at it and it was working fine. It was coming up with interesting stories—for instance, stories related to people preferring to stay away from hospitals, why there is a big difference between government hospital insurance, or why insurance won't help you irrespective of the type of hospital. All of this on the medical side, on the education side, public data stories where it was scanning the data, it was figuring out the analysis, it was figuring out which of those were interesting, and it was validating it, it was creating the visuals, and going through the entire process end-to-end. This is from PLFS [Periodic Labour Force Survey] data and it goes on. Many of these you may have seen actually on The Times of India.

**Anand**: [01:54] Question is: how does one go about doing it? Like with most things in AI, the answer is simply: "Oh, the process is very simple. We just tell one of these models to do it, it'll do it." We just didn't know it was possible, or there may have been one or two nuances about it that we may not have known. Let's go into that. But before that, a question from [the other] Anand on the wind model: "**With ChatGPT, we’re burning too many tokens to come to the output. In the old world, making a machine learning program to do an iteration could have given the same output.**" Entirely possible and that is in fact exactly what they did, except it was not a machine learning model; it was an Excel model.

**Anand**: [02:39] GIPCL has an Excel sheet. They said, "Anand, we want to hire a data scientist to do this job. How much would they cost?" I said, "Look, how about you spend 100 dollars first and hire a data scientist as an agent who came and fixed their Excel sheet?" Now they no longer need to pay even that 100 dollars. Frankly, I asked them after that, "Do you have any other problems to solve?" They thought for some time and said, "No."

**Anand**: [03:09] So, yes, I completely agree. **I don't think we should be using AI as a forecaster for every single thing, just as we probably don't want to use it as a—well, for anything that a program can do. We don't need to.** Also, every time we ask it to do something, it will likely give an answer like a human. Some things humans are better at. But if I wanted to multiply a large set of numbers, I would rather have it write a program that does it and keep running the program, which is again part of the workflow that we’re going to try.

**Anand**: [03:49] Let's take one data set and I’m going to share a link on this screen as well as in the chat. Well, let’s just go to [data.gov.sg](https://data.gov.sg). There's some fairly good, interesting data sets out here. You can just go to data.gov.sg itself and browse any of these data sets. Take a look at it across any of these agencies, across any of the different types of data. Is there any kind of data that captures your interest? Because one of the things that we'll be doing in this session is each of you will be coming up with your own data stories from any one of these.

**Anand**: [04:40] I happen to have done a quick search and I just asked it, "Look, what are some of the more popular data sets by downloads?" And it said, "**Acra information on corporate entities seems to be the most popular**, more stuff on Acra, then there's COVID weekly stats, and there's resale flat prices." And this looked somewhat interesting to me. If we have resale prices of flats—and I was looking at it a short while ago—there is a CSV file, in fact, that seems to have the prices of different kinds of flats (one-room, two-room, three-room, etc.) and how the prices have changed. And it also has in which town, what type of flat, what block, etc. Bunch of stuff. Is there something interesting that we could do with this?

**Anand**: [05:51] Now, this is a "solution in search of a problem." I mean, we have a data set, we're trying to do something with it. Not many people have that sort of a situation. It's usually the opposite: "I have a problem and I might want to find a data set that would work." We’ll cover both of these things.

**Anand**: [06:17] Here are two pieces of parallel analysis that I'm going to initiate, and then you can start as well in a few minutes. One is: **let's pick a problem that any of you are currently exploring and we'll see if there's any public data that might help with that.** Could you please just put in the chat window any problems that you are exploring that might require some kind of public data, or that might benefit from some kind of public data analysis? I'll repeat the question: **Type in the chat window any problems you're exploring for which public data might help**, and we will search for some things.

**Debi**: [07:11] Do you just want a problem statement, Anand?

**Anand**: [07:13] Yeah, just a problem statement.

**Anand**: [07:37] Okay, Sandeep’s comment is **electric car penetration in Asian markets**. And Sonal’s: **best use of credit card reward points**. Higher education seats in Singapore by department. Okay. Correlation between Singapore Dollar and INR and inward remittances in India. Okay. Let's give it a couple more. Okay, or in fact, while this appears, let me explain what I'm going to do.

**Anand**: [08:08] **My hypothesis still is that all activity can be delegated to agents.** So, while you are crafting these, here's my prompt to ChatGPT, which is: "I'm going to paste a series of tasks that I'm interested in researching. Which of these would best be supported by available public data? I want you to search online for the most relevant public data sets that could help with these tasks, and then rank order the tasks and supporting data sets that would have highest impact." Let's do that.

**Anand**: [09:01] This is not a very common—or this is probably closest to a consultant use case. Meaning, we reach out to a consultant and say, "Is there a data set for any one of these things that you can then do an analysis for?" So, let's copy the typos here and run this.

**Anand**: [09:24] There were a few questions earlier on token usage and how to manage it. Let me share what is working currently. I'm using ChatGPT, not Claude. And in ChatGPT, there is a "ChatGPT Work" and "Chat." In Claude, you would have Work, Chat, and Code. Now, **almost all of these are metered.** So what I mean by metered is if you consume too many tokens, then it will stop you from doing the work. But there is one exception, which is ChatGPT's "Chat" at the moment. Of course, this is also true for Gemini, which only has the chat and that works the same way. But what this means is that **you're not likely, even on the 20-dollar plan, to run out of credits if you just stick to ChatGPT Chat.** I don't know when this is going to change, but we have a reasonable window of opportunity. And that is what I'm leveraging here, saying, "Do my research."

**Anand**: [11:09] The good part is it is as extensive or as capable as most of the others—in fact, probably more in this context. In today's session, I'm almost exclusively going to be on ChatGPT because for data, for research, for analysis, **when the rigor of execution matters, it's hard to beat the OpenAI class of models. When it comes to taste, when it comes to strategy, Claude is hard to beat.** If I didn't know what I wanted, I would go to Claude. It would correct me and say, "Look, I'm going to do what I think you really need, not necessarily what you said." **ChatGPT will make sure that it does what you said to the letter without mistakes**, much more so than Claude. As of now—again, many of these things change.

**Anand**: [12:02] It's quick, and lastly, the model that I'm on, **GPT-4o**, that is probably what most of you are using by default. And I've set the intelligence to high, which—a month ago with GPT-4-Turbo, if a task took five minutes, it would probably take ten minutes today. It definitely thinks a lot longer. That's not a measured number, just saying it certainly seems to take a lot longer than it used to.

**Anand**: [12:35] And it's saying **the COE prices, there's excellent public support data**, there's high impact. **Rooftop solar growth**, there's published data, and it's pointing us to these data sets. So, let’s take, for instance, COE. If you just search on Google for data.gov.sg for COE, we have a bunch of data sets and that seems to be pretty extensive. And instead of HDB data, maybe we could take this and analyze it. Rooftop solar growth, there seems to be strong data and that is coming from [ema.gov.sg](https://ema.gov.sg) where you can download Excel sheets and a PDF report which has some stuff. Electric car penetration again seems to be reasonably strong. The weaker ones: best use of credit card for reward points seems to be reasonably weak and fragmented, not like we could do much by way of impact here.

**Anand**: [13:41] In other words, if we take the spectrum or the chain of what people do in the data analysis cycle, one of the first is: "**I have a problem, I need data. Where can I get it?**" Finding the data is just one prompt away. Use it. In this particular case, we weren't even saying, "Here is the data that I want." We're saying, "Here is a problem. Is there a data set that might help me?" and let it figure out what constitutes help. And then we can tweak it by saying, "No, no, that's not quite what we want." The other revelation is that **there is a surprisingly large amount of public data available, and you just need to know that it's there so that you can start searching for it.**

**Anand**: [14:28] The second is being able to actually work with this or such data sets. Now, there are a couple of things here. One, we have a problem in mind and we say, "Now, get me the answer to this question." The other is sometimes, "Oh, here is a rich source of data" or "I have a broad problem that I need to solve. How might I go about using this data to solve that problem?" The former is "I know what I need to do." The latter is "We don't know what I need to do." And again, AI can help structure those. We'll take a shot at both of these.

**Anand**: [15:04] Now, what we'll do in a few minutes is flip it over and actually, yeah, let's do that now. **Let's pick the HDB data set for now**, for no specific reason. Let me just confirm that if I pass this to you, you will get the same data set. Yeah, it gives you a CSV that is 22.3 megabytes. So, **Step 1: Please download this data set and upload it to ChatGPT.** So click on "download CSV" and you will get a CSV file. Put it into—well, and if you don't have a Claude subscription, please use Claude, or absolute worst case, Gemini or Co-pilot or whatever else. But for today's exercise, **if you have ChatGPT, especially the paid subscription, I would nudge you towards that.** And just drag and drop it into this file. For various reasons, I'm not going to drag the CSV file; I'm instead going to drag a zipped version of the same CSV file. And then go ahead and ask it any question, absolutely any question related to resale flat prices.

**Anand**: [16:30] And that question could also be: "**I have no idea what questions you can answer from this data. What interesting questions can you answer from this data?**" is a perfectly valid question. And let's see what you find, because from what you find, there are a few inferences that we can make. So, what I'm going to do is "Amuse me." Submitted.

**Anand**: [17:11] In the meantime, there are a few questions from Sandeep Lhari out there: "**Any value of college education in five years from now? If yes, what kind?**" I will want to take that next session, Sandeep. That’s certainly what I’ve planned for then, because after you've seen what it can do with code—the bypass-coding part, which is our fourth session—this question will probably be even more strongly on your mind. But the short answer is I think yes, but a very different kind. We'll talk about it.

**Anand**: [17:53] Comment from Debi: "**As a 60-year-old woman with no pre-existing health conditions, what coverage of health insurance should I buy?**" Okay, this is part of the data set search. Yeah, it would have been a great one. Let me just add it to the search. What about for Sonal?

**Anand**: [18:19] And [the other] Anand's question is: "**When you look at Singapore and also other global data sets, how many of them can be accessed in a form that GPT models can use without API issues or walls or a format issue of data availability?**" Which brings me to exactly the next point that I wanted to share.

**Anand**: [18:38] For someone as lazy as me, downloading it and then uploading it to ChatGPT—which is what I'm asking you to do—is a lot of work. **Why not ask ChatGPT to do that directly?** Short answer is earlier it was not possible, and now I'm stingy. Let's break that up. Let's ask ChatGPT to download something; we'll also ask Claude to download something.

**Anand**: [19:11] Now, let's start with Claude. Oh, let's try this out. That's good news. But we'll come back later. So, VJ, you had asked earlier how I manage token cost. I use haiku-3.5 unless I know that the task can't be done by something like haiku-3.5. The task that I'm going to give it is: "Download and very cursorily analyze Singapore HDB data from data.gov.sg." My intent here is to verify whether Claude can access the site or not. Haiku won't make such a basic mistake.

**Anand**: [20:03] And I'm going to ask ChatGPT the exact same question. My theory—and this is based on what I had kind of tested last probably about three weeks ago—was Claude is able to do it; ChatGPT is not. So, on the right side, Claude should be able to download the data set and it's right now doing a search. Both of them should be able to do a search, but ChatGPT might not be able to actually download it. And we’ll see if that in fact happens. Yeah, I think that is what is happening.

**Anand**: [20:41] So, it's downloaded the number from data.gov.sg. The data covers 2008 to 2021. That is—I'm sure we have more recent data than that, but still. Let me just do one more check. Is this the most recent data? What's the link you downloaded it from? And let's check.

**Anand**: [21:07] And Sonal, to your earlier question on where we can check the limits of usage. Again, against your name at the bottom left, you’ll see settings, as VJ had mentioned earlier, and there is a usage tab. That will show you how much of your current session you've used. So, this is saying **4% of my session, and in about almost five hours it is going to get reset.** And there are a bunch of other limits that it tells me about. The same thing is also available in ChatGPT where I can go to settings and there is a usage tab. But this is not as much for this. This says "Codex Work Workspace Agents" and "ChatGPT for Excel." It doesn't include chat conversations. Meaning, **what we put on chat isn't metered in the same way.** I think it's metered by the number of conversations and you can't have ten chats open and just blindly type in chats left, right, and center. But they don't publish it and it's not a limit that you would run out of very easily.

**Anand**: [22:23] In any case, it's saying "Here is the data set that it went through." Now, ChatGPT is still taking... Oh, okay, okay, sorry. This was a false test. I must apologize. **I have given ChatGPT access to my computer and it is downloading it to my local.** Let me put it this way: if you tried the same thing, run this chat on ChatGPT, it will probably say, "I'm not able to do that. You can download it and give it to me." So for what it's worth, here is the prompt for you to try.

**Anand**: [23:03] In my case, it's saying, "I've downloaded it to your machine," and the reason it's downloaded it to my machine is because I've given it access. **You could do the same. You could run ChatGPT on your desktop or you could run Claude on your desktop and there won't be any problems.** If you don't want it on your desktop, then that's fine. Now I have a problem here. I will not be able to run ChatGPT or Claude on my desktop because I'm on Linux and they've decided not to release the Linux clients. But almost without exception, all of you will be able to. And Anand, your question: "How can they be accessed in a form that GPT models can use?" **Short answer is by running ChatGPT Desktop or Claude Desktop.**

**Sandeep**: [24:02] Just one point, right, actually Anand, Claude also has a Chrome extension. So sometimes what happens is, you know, if you use Claude on another browser, it won't be able to give you what you want, right? So it's always better to use Chrome and also download the Chrome extension because it will prompt you. And it's better Chrome is always on when you use—at least I use only Claude. So it's better you keep Chrome on the desktop when you actually do the work on Claude because it's got its extension on Chrome. Not on, unfortunately, Microsoft and Anthropic don't talk to each other. Right? So for instance, **there is no LinkedIn connector. That's a disadvantage on Claude. Yeah, so you need to always download the archive and then upload it manually. And it doesn't have access to Outlook files, Outlook mails.** Whereas [Copilot] has got native connectivity into all Google [and Microsoft] products. So there are these tradeoffs.

**Anand**: [25:04] Exactly. And which is what some of us saw last session with the Chrome connector. And we will be looking at how the desktop tools are able to go beyond that. But let's take a look at—just checking if my screen's still shared? No, it's not. Let me share again.

**Anand**: [25:29] Okay. Anand ran a chat on this flat and okay, let's see. Can you look at URA... okay, I know this is a little small. I'll see if I can... that's about the largest I can zoom it to. **Trend of unit flat prices in Geylang per square foot and the answer seems to be it's clearly increasing and the slope is what—from 750 to 2000 in about 16 years**, give or take. And yeah, so there are three parts here that I’m taking away. One: **the ability to obviously process the data**, reading a CSV file, a spreadsheet, etc. Second: **doing an analysis as simple or complex as a log-linear trend**, which is what it's saying it's done. And third: **being able to plot this.**

**Anand**: [26:34] I had run an "amuse me" question and it's running. Uh, anyone who has anything that they've found, please do share. But let me review what... okay, download this. Okay, and let me just open the larger version of the same thing. Right.

**Anand**: [27:04] So, "The government data set enters a comedy club" and it brought in **receipts, remaining lease, and unexpected strong attachment to the number eight.** Okay, this part has certainly intrigued, if not amused, me. So, 1.7 million, this price is **1.728 million because even records prefer tidy arithmetic.** Does anyone know the significance of 1728? 172... possibly in the context of Ramanujan.

**VJ**: [27:49] Yeah, exactly, VJ.

**Sonal**: [27:51] 12 cubed, yeah.

**Anand**: [27:52] Yeah, there's... but I'm not sure if that is the exact significance that it's talking about here. Let's see. It will, I'm sure, come up with some interesting stuff. "**Negotiation Theater: Almost 80% of transactions end in tens of thousands.**" And yeah, fair enough. "**Okay, about 4700 transactions end exactly in 8888.**" Anyone knows why that is? I have no clue.

**Sandeep**: [28:25] That's the Chinese...

**Sonal**: [28:27] **It's the lucky—the Chinese lucky number. It's for wealth.**

**Anand**: [28:30] Wealth, yeah.

**VJ**: [28:32] They’ll pay an arm and a leg for 8-8-8-8 and five or whatever. Except four. Four is death. Eight is...

**Anand**: [28:40] Interesting, okay. And there is a **three-room terrace in Block 53 which at this size—one of the rooms—okay, well apparently about 300-odd square meters is huge in Jalan Membina.** Okay. And this is even larger. Four... okay, oh no, sorry, **highest price probably. And 39 times... okay, monthly pace of multi-dollar multi-million dollar transactions.** Oh, okay. **46 of all the 2017 multi-million dollar transactions are in 2017 versus 1055 in just the first seven months [of this year].** Okay, so the multi-million club is getting... okay, and let's see. Let's pick a serious... we have some interesting stats about each one of these. Okay.

**Sonal**: [29:46] Is this your personality which gets them to choose these as the outputs through the memory they have of what you like?

**Anand**: [29:54] Possibly, might be. I have been feeding a fair bit of my personality into ChatGPT. And the interesting thing is... [End of chunk]

---

**Anand**: [00:00] ...into ChatGPT. And one of the ways in which we could explore this is by running it. Well, firstly, we could take a look at what it was looking at, but **these days ChatGPT doesn't seem to be sharing its thoughts as well as it used to.** So what I could do is—and this is something that you'd find useful to do as well—click on the incognito or "Temporary Chat" button on the top right. And when you do this—oh, okay, sorry. No, that is not... I'm not sure this will work.

**Pragati**: [00:46] Anand, is there a way to put "skill" on ChatGPT? There is no "skills" on ChatGPT as far as I know.

**Anand**: [00:54] Depending on the price or edition that you have, the Enterprise version has skills that will appear on the left. If you don't have it on the left, then you're not on that version. I just copy-paste skills. But I thought that maybe a temporary chat would not use memory—that may not be the case. What I could do is turn off memory. If I disabled memory and reference previous recording transcripts, whatever, my guess is it would provide a less personalized answer, but frankly, I'm not convinced of it. **The most unpersonalized answer would be just creating a separate account**, opening a free account, and trying it on that and seeing what comes out of it. Let me share this chat for what it's worth, and I would love to see what some of the others have found. One second, we'll come to that.

**Anand**: [02:11] Comment from Sandeep: "**Chinese models are more creative; they can access your browser search on tabs, LinkedIn, YouTube, etc.**" True enough. Sanjay: "**I tried your prompt on Claude; it says no network access from my sandbox.**" Yes, and ChatGPT's chat makes sure that it won't access a link, I think, even if you explicitly give it a link. It just says, "Download it and give it to me."

**Anand**: [02:41] And here is Sonal's output. Let's take a look at this. "**Three rooms is apparently a philosophical concept.**" Yeah, it found the same 366 square meter huge HDB terrace. And the 8-8-8-8 trigger. And okay, "**Four transactions were registered exactly at $1 under $1 million.**" That is an interesting one. Million-dollar households are making news. That's reasonable. Most expensive flat was also covered. Cheapest transaction was $140,000. Okay. And yes, interesting stuff.

**Anand**: [03:44] We'll probably cover one or two more. Here is one from Shaibal. Oh, a superstition test. "**Four sounds like death in several Chinese dialects. Eight sounds like prosperity.**" These are the last digits of the block number. Okay, eight clearly sells at a premium. Nine is unlucky in Chinese, which no dialect associates with being ominous. Okay. Yeah, although I am fairly sure there is something behind the nine. I mean, just a $1 discount means the person is very sensitive to price, maybe? But this is certainly an interesting one.

**Anand**: [04:32] And we have from Michael an interesting comic story. We analyzed the HDB, yeah, and okay, oh, this is interesting: **Price versus Interest Rate. HDB resale prices kept growing even when interest rates rose, so that's not a one-to-one relationship.** This... let's come back to this in a short while because the correlation is clearly negative. And two sources of capacity. Okay, this is fairly deep. Michael, what was your prompt?

**Michael**: [05:07] I just chatted with it a bit and said, you know, "**Figure out what the hell is going on with this real estate market; prices seem to be up.**"

**Anand**: [05:15] Okay, but what did you pick up?

**Michael**: [05:18] And what did I ask? Two or three things I just said. So it came back with the first thing. I said, "**I don't know, fine. Prices went up ridiculously post-COVID. Nothing was happening before, no difference in size.**" So I asked, "**Why is that happening? Did interest rates make a difference?**" It said, "**No, interest rates didn't make a difference. It's only supply.**" I said, "**Okay, then what's happening with supply?**" And then they came up with this. And then I said, "**Make this into a nice comic which I can use to impress AI experts.**"

**Anand**: [05:54] Interesting, very interesting. And this was with ChatGPT.com or ChatGPT Desktop?

**Michael**: [06:02] ChatGPT.com. God, meaning... sorry, no, ChatGPT.com.

**Anand**: [06:09] Got it. Meaning if you had downloaded something and run it, then it's ChatGPT Desktop or the app, so to speak. And this may be a good time—does it add value versus the thing? Is there a difference in the app? I did not know this. It does. So yes, and in fact, let's do that next.

**Anand**: [06:33] So here's my request: **Could everyone now go to—if you have ChatGPT, then ChatGPT, or if you have Claude, download the Claude app—but just search for ChatGPT app download.** I'm sure that ChatGPT.com/download will let you download for Mac or Windows the app. And yes, it does make for data analysis a huge difference. For those of you who are using Claude, Claude download is probably the equivalent link. If you aren't or haven't been downloading these apps, please do.

**Michael**: [07:09] I'm getting three apps: ChatGPT, ChatGPT Classic, ChatGPT Beta. Do they make a difference or are they all the same?

**Anand**: [07:19] I don't know. Okay, you downloaded it for Windows presumably?

**Michael**: [07:27] On Windows, they're giving an option of three different ones.

**Anand**: [07:31] Could you share your screen? I'd love to see this.

**Michael**: [07:35] It's from the Microsoft store. It offers me ChatGPT, ChatGPT Classic, and ChatGPT Beta.

**Anand**: [07:42] I suspect it may be none of these and it's not available on the Microsoft Store. You may want to go to ChatGPT.com/download.

**Krishna**: [07:51] ChatGPT Classic is the previous version. They ran a new update and it's actually been giving me a lot of issues on Mac in the last two weeks. Like, now I can't delete conversations. But it is supposed to be an upgrade; they're just working through the kinks. So I wish I hadn't upgraded. I think ChatGPT Classic worked better on the Mac at least.

**Anand**: [08:16] Interesting, got you. Yeah, I have heard that the Microsoft Store isn't as up-to-date as their online version. But do give it a shot. I have also learned that Claude has now released a Claude Desktop for Linux, and that's helpful. I'm going to download it in a short while. But does anyone have the ChatGPT Desktop—or downloaded from the ChatGPT Desktop—installed on their system? If so, would love for you to share your screen, please. I can't.

**Pragati**: [08:54] Sorry, Anand, can you repeat that?

**Anand**: [08:57] If you have ChatGPT Desktop installed on your system—ideally from ChatGPT.com itself—would you be able to share your screen?

**Pragati**: [09:04] I just use it in the browser. You mean download it as an app on the desktop?

**Anand**: [09:09] Yes, as an app on the desktop, yes.

**Pragati**: [09:12] Anand, I have. I can share my screen if need be. I have ChatGPT downloaded.

**Anand**: [09:17] Thanks, Pragati. Yes, please, if you could.

**Pragati**: [09:21] Sure.

**Anand**: [09:25] Great. And yes, could you—you've shared the window or the screen, Pragati?

**Pragati**: [09:30] I've shared the window.

**Anand**: [09:32] Okay, that's good enough. And yeah, as you can see, this is not the browser; this is the downloaded version of the app. Now, the reason this matters is **you can give ChatGPT access to your computer, and that lets it do a couple of things that are more powerful than what it can do remotely.** The first is that you don't have to upload all kinds of stuff. It can find and upload stuff even if you don't necessarily even know it existed. And this it does for a lot of my archives. I just say, "**Here's my folder; this contains my last 15-20 years of emails. Search and take what you need from the attachments.**"

**Anand**: [10:25] The second advantage is that it can pull stuff from the internet through your computer as long as you give it access. There are two tabs on the top: Chat and Work. **Chat is where you usually do your regular kinds of stuff. Work is slightly more powerful; it writes code and gets stuff done, but is metered.** We are not going to run out of credits on Work that easily; it has fairly high and fairly liberal limits. So what we're going to do now is work on the Work tab of ChatGPT for a while. Now, this Work tab is also available online, so you don't necessarily need to have downloaded the app to do what we're doing in the Work site. But what I'm about to mention next would ideally be done on your local machine because we will ask it to download all kinds of data sets and play around with it.

**Anand**: [11:35] So let's take a task. Actually, Pragati, is there any data analysis or data-related thing that you are interested in exploring?

**Pragati**: [11:46] Yeah, I just posted like, maybe I was trying to plan my National Day weekend and if there is like **the cheapest flight that I can get from now to, let's say, 9th of August. And if there could be a schedule... I don't know if it's a data analysis or it's more of a schedule that it keeps scanning and telling me that today is the cheapest flight available to Bangkok, for example. Book it.** How can I sort of do that?

**Anand**: [12:12] Great, that sounds like an interesting problem. I have a theory though, that it might be able to solve that even if we just put it into ChatGPT Chat. So let me make it a little more complex. And what we'll do is let me do something. I'll type out a chat.

**Pragati**: [12:36] Should I give you access to this page if this is helpful?

**Anand**: [12:39] No, no, don't bother. I'm just going to share my screen and we will type out a prompt that you can just directly run. I would like to explore patterns of flight prices out of Singapore. Search online for the most appropriate data sets and download what you consider as most relevant. Analyze these to see if there are any interesting insights on when prices go up or down. And I'm going to give another link which is "data analysis skill" and we're going to use this for analysis. Let me paste that in the chat window.

**Anand**: [13:42] Now, let me stop sharing. Back to you to share screens. Could you please paste this prompt in?

**Pragati**: [13:49] Yes, I'll just share my screen and then paste this prompt. Should I use "Work", Anand?

**Anand**: [13:56] Yes, please. There are a couple of things that as it runs I'll highlight. But no, hold on, hold on. "Soul Light"... yeah, you can stop it; it's worth stopping it. Yeah, select that and set it to the last but one, please. Yeah, "High". And just tell it to continue.

**Anand**: [14:41] The reason I stopped it was the prompt that I had shared is something that will require a little more thinking. So let it run. Now, **one of the reasons this chat won't work on ChatGPT online chat is it doesn't have a container. That is, it doesn't have a place where it can go download stuff.** Actually, it does have a container—it can't access external links. But ChatGPT Work can, and this particular prompt any of you can try. No issues. You can just put it as long as you put it in Work. Let me share my screen and show you how I might do that.

**Anand**: [15:27] I'm going to take the exact same prompt and put it into ChatGPT, but not here—in the Work section. This is just the browser, ChatGPT.com. And I happen to be choosing, in this case, "Effort" as "High". Now what's the difference between these? Roughly how much time it will spend on the problem and how much money it will spend on the problem. There are limits beyond which you cannot go, but I suspect in the early days you won't be using it too much. There's also an option to run faster at one and a half times the speed and at a higher usage. I don't use this quite often. There are projects that you can organize this under, there are a whole bunch of plugins and all of that; we're not going to go into that today. But this will, unlike the earlier attempt, succeed because it will create its own machine that has better access to the internet and you can control what access it takes.

**Anand**: [16:35] What we're doing here is taking the analysis to the next stage. The first time we said, "Find me data sets that can solve my problem." It did. Now we're saying, "Download the data set by yourself." Last time we downloaded it and gave it to it. Now we're saying, "Download the data set yourself and play around with it." So much is fine. What we haven't done yet, however, is told it to do slightly more advanced analysis. What Michael did, for instance, at least was a correlation, saying between interest rates versus flat prices is there a relationship? Is there a more sophisticated analysis that we could do? Let's explore that. Pragati, could you share your screen again, please?

**Pragati**: [17:28] Yes. Anand, quick question: the "skills MD file" that you put in here, that is a predefined data analysis skills, is it? And then that's what you're asking here to use?

**Anand**: [17:39] Correct. And I'm going to come to that in a short while, but yes, that in a nutshell is exactly right. So let's open, while this is continuing, could you open a new work session? Probably find it at the top left if you scroll up. Yeah, "New Chat" is fine.

**Pragati**: [17:59] So should I just say "New Chat"?

**Anand**: [18:01] Yeah, yeah, this will continue in the background. And yeah, you're on Work. Let's ask a question: "**What impacts HDB resale price the most? Download, model, and verify.**" And if you could also paste the last line from teams of my previous chat, which is telling it to use the data analysis skill. That is sufficient; you can press enter.

**Anand**: [19:06] What it's doing now is very similar to earlier, but the only nudge that I'm giving it is this modeling business. **The realm of modeling used to be at one point we would use Excel to build models, more advanced Excel macros to build models. We'd get somebody to write Python code to build models.** They would then start building not just Python-based models but neural net-based models, deep learning models that would learn automatically. And then it got slightly beyond what many of us could be able to understand, certainly do. That is no longer the case because we tell it to write a Python script, we tell it to build a neural network; it will do all of that even if we don't tell it—it often is able to do it. We just need to know enough to get out of the way. And we will see some of that hopefully happening here. If not, the next prompt can simply be: "**Use a more powerful machine learning model to improve the accuracy of the forecasts.**"

**Anand**: [20:18] We've got a set of things running, so let them run in parallel. And once we get any of the results, we'll start looking at it. But in the meantime, I'll share my screen and take you through this prompt—this skill that we partly covered last time, but will go through in some more detail now. The link that I shared is for a somewhat detailed data analysis process. So far, I haven't found ChatGPT or Claude doing analysis quite the way I want it to. I'm not convinced that that is necessarily the best way of doing it, but it produces results of the kind that I find useful. So I've synthesized effectively about 15 years of what I've been learning and teaching on data analysis into this prompt, and I may have shared this last time as well.

**Anand**: [21:26] Here's what... and this is structured as a skill. This is something that you can upload into Claude. How can you do that? **In Claude, on the left side, there is a "Customize" section. And in Customize, you have skills, connectors, plugins. And out here in skills, you can add a skill by writing the skill instructions.** For instance, you can create a skill called "Market Analysis" and use "Web" for analyzing the market. And you may put in things like—well, you can tell it the kinds of things that it usually misses, or the way in which you want the analysis done. For instance, you may say, "**Prefer Singapore-related analysis**," which of course you could mention in your prompt also. Or you could say, "**When researching stocks, prefer prices from previous close rather than live prices; I don't do intraday trading**"—possibly a relevant thing. Or you may have your own methodology for how you do your analysis; put all of that in and this then stays in Claude's memory.

**Anand**: [22:50] **ChatGPT does not have an equivalent of that for most editions, so you may just store this and copy-paste it whenever you want.** That is in fact what I do. So on ChatGPT, I just have a shortcut—a keyboard shortcut—where I can just say "Data Analysis Skill" and whichever browser I'm in, it will paste that skill. But that's simply a fancy copy-paste. It helps to have things that you do on a regular basis stored somewhere. And in this case, what the skill does is tells it: first, find out who the audience is or take a guess on it; understand the data—here are some of the ways in which I find it useful to understand the data; look for what is interesting; and verify whether that in fact survives the typical logical fallacies; pick the most important stuff and narrate it in an interesting way. That's pretty much it. Without it, it will likely still get to interesting results; this will probably top it up. It's the equivalent of you then looking at it and saying, "Oh, but can you do X also? Can you do Y differently?" and so on, sort of like how Michael had prompted it to get to a couple of more interesting analyses. It may go that far at one shot, which is why I use a skill like this.

**Anand**: [24:23] **These skills are depreciating assets, not worth building unless you have a strong need.** What does that mean? Today the model is not able to do some of these. A year down the line, these will be doing analyses better than I can. So what's the point? Right now I need it, I do this a lot, so it was worth me investing time doing a data analysis skill. But the lifetime I expect is probably only one, max a year and a half from now. **Some skills will have more value, but just keep in mind that skill writing is not a skill you want to invest in. It's knowing what you want that is the more important skill to invest in.**

**Anand**: [25:14] Okay, Pragati, has yours come up with any results so far?

**Pragati**: [25:18] Yeah, there are a couple of results. One on the flight—let me just share my screen back. So like this one, it shows... but then I don't know why is it still showing in this also that it's still working on this one. But it's showing me the flight prices insights, but then there is another thread which opened up. So this one has some information and it's given me a Singapore flight analysis MD file.

**Anand**: [25:56] Hmm, but this is, I can see, clearly useful. Right? "**Short-haul Asia: start checking 4 to 6 weeks ahead**"—that's longer than what I had thought it would be. "**Long-haul is 8 to 12 weeks to track. And then within 14 days, waiting becomes speculative.**" That is interesting. Okay. "**Tuesday, Wednesday travels are cheaper than Friday, Saturday travel.**" Friday I might guess, but Saturday was not something I would have guessed. And yeah, use track tracking rather than repeatedly checking manually.

**Anand**: [26:30] **This again is a classic example where using a schedule can help us find and track whether the price is rising, falling, whether we should book in advance.** But there's nothing that stopping us from, as part of the schedule, asking it to do a data analysis on a daily basis to say: "**Based on today's data, where do you think the flight prices might go?**" Or as part of some of your existing workflows for many of you where you're already looking at market analysis, asking the question, for instance... yeah, where you mentioned doing a market scan of stock markets—maybe not just do a market scan, but also a technical analysis of that and tell me which of these might work.

**Anand**: [27:43] Let's pay a little bit of attention... sorry, Pragati, could you scroll down a bit on the same chat? Just scroll down where it said "linear"... okay, sorry, no, maybe could you go back to the other one where it said "Continue now"?

**Pragati**: [27:57] I think this used the light version and that's why it finished off. But the one where I used the "Soul High", it's still working actually.

**Anand**: [28:05] Yes. And let's—we'll click on a few things as part of this. Could you click on maybe where it says "Fetched Data Set"? The one two items below that. Or okay, the one below that, just below that. Yeah, are you able to click on it?

**Pragati**: [28:27] No.

**Anand**: [28:29] Oh, that has changed. Hmm, **I was hoping I could show you how it writes and runs code.** Let me see if my version—oh, okay. Hmm. We are on a $20 plan, so I don't know whether this allows us that. No, I think what ChatGPT Work does is hides the details, and that is such a pity because watching it code could have been so much more intuitive.

**Anand**: [29:30] Let me just check if... no, it's not showing it on mine either. Let me share my screen and the result is not very different. Okay. So here it's done something very similar for the same prompt. **It's fetched a bunch of files from GitHub, it's inspected a few APIs, it's read...** [End of recording]

---

**Anand**: [00:00] ...its AI, it's read, and so on. But the thing that it's doing behind the scenes, and the icon to watch here for that is where there's a little terminal-like icon. So this...

**Debi**: [00:14] Anand, your screen is not visible.

**Anand**: [00:16] Thank you. I'm sharing my screen now. So of these icons, the globe icon is where it's searching for something, but **the terminal icon is where it's running code on the fly. And because it's able to run code on the fly—write and execute code, which we'll cover more of in the next session—but in this case, it's analysis code, it opens up the entire possibility of what kind of analysis can be done.** I had assumed that we'd be able to see the workings here, but since we have not, let me do something slightly different. I will take the HDB data set that we had uploaded here and—yes, we are able to see this, right? Fine. Ask it a slightly different question.

**Anand**: [01:13] "I want you to predict the resale price based on this data. I want you to build the most accurate model possible. If you need to build a machine learning model that is advanced, you may feel free to do so. In fact, what I'd like you to do is test out multiple models. Make an initial guess on which models you think will have the highest chance of predicting accurately and test the three best of those. Benchmark them on how accurate they are. I want you to think about what is a good measure of accuracy—define that. And hold out some of the data set so that you test the models or build the models on some initial, let's say, 80% of the data and then verify if they are accurate on the remaining 20%. Use this to report to me which model has the highest accuracy and how much it is. Explain this in intuitive terms so that I can understand what it means if I were making a resale transaction. In other words, what would it mean if I got this wrong? How far off could you be for different kinds of flats, etc., more as a layman rather than as a data scientist." And let it run.

**Anand**: [02:32] But this time we are going to watch. I'm running this in Chat. Chat has the advantage that it lets us watch the code as it's writing. Work is slightly better in terms of coding, but only slightly, and unfortunately doesn't let us watch. I'm just going to take a quick check to see if there are any questions. Okay, Sonal has one output. Let's see that while this runs.

**Anand**: [03:13] Okay, strongest finding is waiting for a fare reduction is an asymmetric bet. Any idea what this means, Sonal?

**Sonal**: [03:31] No idea. It probably means that it's not equal on whether it's going to go down or go up. Not sure.

**Anand**: [03:42] Quite likely, right? 78% became more expensive, only 9% became cheaper. And for practical purposes, it's saying wait, or take, start six to eight weeks earlier. 21 days before is your decision point and avoid reaching the final 14 days without booking. Changing the travel month can matter more than the timing of the—ah, that is reasonable, but is a good insight. And there isn't enough evidence for universal cheapest weekday to book. Okay, that negative is also pretty useful. And pattern tracker that it built on Excel.

**Anand**: [04:32] **A lot of the work that I'm doing has shifted to having it write code and do analysis because it's a relatively easy thing to differentiate on these days. A lot of people are using AI on search and coming up with stuff. If I'm able to give it data and show something that is data-backed, it increases the credibility a certain amount and also increases the reliability of the work a certain amount, which is what makes this part particularly powerful.** Now let's spend a minute on what it's doing here. Big, let me make it bigger. You'll notice that it's writing some Python code. We don't need to necessarily understand the code, but it describes what it's doing and that is fairly helpful.

**Anand**: [05:23] So right now it's beginning by inspecting the data set—useful. It's considering multiple features for the model, like transaction date, location details, and cleaning methods, including CatBoost, LightGBM, XGBoost. These are terms we absolutely should ignore; doesn't matter what they are. But it sometimes is helpful to take a look at these, search online to get a feel for how advanced a technique this is or how simple a technique this is, etc. And XGBoost is a pretty good and advanced technique. **It produces some fairly accurate and high-quality results. In any case, it is likely to be more well-versed with forecasting than we are, and forecasting techniques, certainly much better able to code than we are.**

**Anand**: [06:27] So given that ability, what we have is a data scientist that we have hired. Exactly what GIPC had when they were forecasting the store schedules. It's written the code, it's running the code. We'll give it some time.

**Anand**: [06:53] Has the flight pattern analysis one here? Yeah, again, it's... September is the sweet spot, December is the danger zone, Chinese New Year is the peak. Reasonable patterns of insight. We'll let this one finish—not this one, let this one finish. In the meantime, I'm going to take a couple of questions that were raised earlier in today's discussion.

**Anand**: [07:31] Oh, before that, this was from Shaibal. Okay, "Singapore has the most predictable airfare calendars in the world." That is interesting. "Chinese holiday drags 4 to 5 percent price bump." You know, this is actually an interesting pattern of analysis. **One of the things I was exploring was: is it possible to find out which cities have predictable rainfall?** And here's something... this is something created by ChatGPT analyzing in which cities can the time of day tell us whether it's going to rain or not.

**Anand**: [08:24] Turns out that **in Caracas, there are very specific times at which it rains. You carry an umbrella between 12 noon to 6:00 PM; outside of that, you don't need to carry an umbrella. This captures 62% of the rain across practically every month.** The chance that you will need an umbrella outside of this window is less than one-eighth of the chance that you will need it in that window. Abidjan in Ivory Coast has equally predictable weather. Quito's is mostly predictable. Addis Ababa predictable in some months and so on. And Singapore depends—December fairly predictable, May it's a lot less bunched up at a specific time. And Mumbai is hopeless. There is absolutely no specific time window where it will rain; it just could, especially June, July, it could rain literally any time; there is 100% chance of you getting wet.

**Anand**: [09:40] And is the kind of analysis that we can do by just saying, "Download the weather data." In this particular case, the weather data happens to come from Open-Meteo, and there is easy-to-access APIs that Claude Code or Claude Co-work or, yeah, any other Claude, or ChatGPT Work can access directly and download, making it much easier for you to answer questions saying, for instance, "**Does weather make a difference to flights?**"

**Anand**: [10:11] **Possibly the single most powerful capability that you now have because you can do the analysis is not single data set analysis, but multi-data set correlations.** Do interest rates impact flat prices? Does rain impact flight schedules? Does construction schedule get impacted by who's in power or concrete prices? And going one step meta beyond that and doing what we did in the original prompt, which is asking the question, "What will make a difference to a certain kind of result?"

**Anand**: [10:52] Question from Michael: "**Are the GPTs in ChatGPT versions of skills made by third parties or do they really add value?**" Let me open ChatGPT. So the question for most people's benefit is: in ChatGPT, there is a GPT section. This was launched over a couple of years ago, and these are effectively mini prompts and some tool calls that people can make. Do they add value? **To the best of my knowledge, most of them do not. I have not used them in a long time.** I have not met people who have used them and told me anything good about these. ChatGPT doesn't seem to be paying any attention to this. There have been no new features released on this. There's certainly been no marketing around it. I'm just waiting for the time when these are going to get retired. ChatGPT has other plans for marketplaces; GPTs are, as you can see, kind of getting hidden.

**Anand**: [12:07] But "Sights" is a different beast altogether. That's a new one; that's a very powerful one. We'll cover that next session. Question from Debi: "**For building and refining forecast models, do you have a preference for Claude, Gemini, or ChatGPT?**" **For any analysis, I prefer ChatGPT. It makes far fewer mistakes.** It writes good back-end code, same as... Claude does a good job on back-end code also, nothing against it, but fewer mistakes.

**Anand**: [12:39] And VJ says: "Connectors. Recently found Claude Pro can connect to IBKR and place orders which need to be approved by human." Yes, the set of connectors that we have—and in ChatGPT, that's in plugins; in Claude, that's connectors—these are pretty powerful. I've been using my Gmail and Dropbox connectors to answer emails based on my notes, for example, which I put into Inbox Zero. And Figma I'm hearing a lot of good things about, especially the Claude Figma connector. For those who are doing any kind of financial work, Claude seems to be the place to be; they are placing extensive focus on financial services in general. That's a good ecosystem to be in.

**Shaibal**: [13:31] Anand, one quick question. So **when we use Co-work, right, in Claude, or Work in GPT, like from a security perspective, how do we make sure that it's not messing up stuff on your local drive, deleting stuff, and copying things which you don't need? How do you make sure of that?** Because just now experimenting this thing, I gave it a specific folder, but then it went and created the file in some other folder. So...

**Anand**: [14:05] A good point. And let me share a recent interview by Simon Willison with some of the Claude founders. This is the one. Simon Willison recently interviewed the Anthropic team on how they are using Claude and how they are making sure that it doesn't make a mistake. This is the answer that I'm now taking as state-of-the-art.

**Anand**: [15:08] What they're saying is—yeah, so many bad things can happen—so how do I make sure that I run Claude? And the way Simon Willison does it is he just gives it full permission but is feeling a little worried about it. "**YOLO mode**" is where you're running it saying, "I give you full access; do whatever you want." And Claude said, "**Why not use auto mode?**" And if you have Claude Desktop, see if you're able to find auto mode. I don't have Claude Desktop, so I can't locate it on the UI; I know how to access it in Claude code.

**Anand**: [15:49] But there is a thing called auto mode. And what Simon said was, "Yeah, maybe starting three weeks ago, I started using auto mode." **Within Anthropic, almost every single person uses auto mode. They've done some extensive testing, and what they're seeing is by and large we've pretty much mitigated every attack. Not 100%, but enough that the majority of the Anthropic team is comfortable just using Claude in auto mode and getting it to do stuff.**

**Anand**: [16:23] There are a few things that it cannot prevent. If we give it instructions that are ambiguous, it can place data in the wrong folder. It can possibly even delete files. But the deletions and incorrect results are happening more because of ambiguous prompting, not because of security reasons. So if you said, "Anand, I'm worried about external security, what if the model messes something up?" The answer seems to be more than 99% of people are not worried about it, starting maybe a few months ago—and when I say people, I mean people in these AI companies. If you're saying, "Anand, I'm worried about me messing stuff up," that's a fair point. Keep a folder in which it can work, and outside of it don't give it access; that Claude Desktop supports and your blast radius is contained.

**Shaibal**: [17:27] So Anand, one quick follow-up on this. Given that we are asking Claude or ChatGPT to go download files and bring it down and analyze on our computers, I mean, **there could be like a malicious file which it could potentially download and then run something on my computer. Is that a possibility and how do we make sure that whatever it's downloading is potentially scanned in a folder before it can execute it on my computer?**

**Anand**: [18:01] What the Anthropic team is saying is they in auto mode run it through Claude Sonnet, any of the commands that they are asking the application to run, and verifying if that's something that is in line with your original instruction and allowed to do. To that extent, you have the intelligence of a model like Sonnet 3.5 doing the scan on your behalf. And I would say it is, unless you are a cybersecurity expert, it is likely to do a better job than you are—significantly better. And of course, because it is going to be doing it 100 times in an hour, whereas we will probably be doing it only once a day, that increased security might only translate to comparable safety.

**Anand**: [18:52] But in short, it seems to be not too much worse in terms of security than the Anthropic team is willing to trust themselves to. Now, your question is probably at the next level, which is: **how does it make sure that it is not downloading a file from a source that has a virus?** Let us take specific examples. One of the sources by which a virus could come through is if it installs an infected package. Let us say it is downloading from MPX or NPM, which is a JavaScript package registry, or PyPI, which has Python libraries. What if it downloads a compromised source?

**Anand**: [19:34] There are registries that list which versions of which sources or which libraries are affected in some way and therefore should not be downloaded. **The models are designed to check those, the harnesses, the Claude code is designed to check those, and those sorts of issues are prevented. What if it tries to download some random package? It has generally been instructed not to download random packages unless there are reputed sources that tell it that it can go ahead and download.** It doesn't generally download. So the kinds of human practices that we follow for good security and hygiene, it has been trained on as well.

**Anand**: [20:16] Is it still possible that it would make a mistake? Yes, it is. The cost of that error has steadily been going down. A year ago we used to ask, "What if I ask it a question and it gives me a hallucinated case?" A lawyer did that and actually went to court and found that the judge found that such a case didn't exist and ended up getting fined. This happened two years ago. In the last one year, we have not heard of any such case because the models have gone way beyond that level of quality. I think we are now entering that stage where security is not a concern that will vanish, but is not a concern at the same level as it used to be, which is a naive concern. It's now far more sophisticated concern and I don't understand where the problem has shifted to.

**VJ**: [21:11] Got it. Thank you.

**Pragati**: [21:13] Yeah, Anand, so one, I have shared because the model finished its work, so I've shared both the links of the chats. **One of the questions that I had was on the skills piece, which is: you know, a lot of people these days are selling these skills online, and the work I am in, I'm in HR, so there are lots of L&D skills, talent skills that are available.** Now, is it how is it possible that we can create our own skills? Is it just a fancy prompt like or a more comprehensive prompt, or is there a way that we can build our own skills depending on our depth of knowledge or expertise that we have in our field? And is there a worth in buying some of these skills from experts when they're selling it across? So that was the question.

**Anand**: [22:04] **Learning how to create skills is not something that I would suggest investing in. Buying skills is not something that I would suggest unless it's priced really low, you need it immediately, and it's someone you're helping or whatever.** A skill is in many ways a fancy prompt. The main difference is that it can be auto-applied. Let me share my current list of skills and I'll come back to your output, which is pretty interesting.

**Anand**: [22:43] So this is my current list of skills. Different ones get used to different extents, and many of these I've updated within the last month. What I mean by updated is I'm constantly refining, iterating, changing little things here and there because these are the ones that I'm using most. One of the skills that... or let's take this one, "**Anand Objectives**." This is an ultra-niche skill. This literally tells Claude what my objectives are. And it is applying this to almost every scenario as an organization or as a division or whatever. And if you put in saying, "Yeah, these are what I want to do," then it doesn't matter what question you ask, it'll automatically pull that.

**Anand**: [23:32] Or let's take "**Expert Lens**." This is a skill that tells it to find out... first, write the task... let me start with "**Question Reframe**." Question Reframe is another fairly popular skill. This is based on my understanding that **I don't ask questions well. Sometimes I've thought them through, quite often I'm just not even sure what I'm asking or I'm asking in a hurry or whatever. It says my question is a draft. There's something that I need and there's something that I've put in words; the two need not be the same. You take a guess on what that might be, figure out if what you think is right, blah blah blah, and then come up with the answer.**

**Anand**: [24:23] Now the good part is—okay, let's take... recently there were some mathematical theorems that were disproven, conjectures that were disproven by Claude, ChatGPT, etc. There I explicitly told it, "Look, use three of these skills and any other relevant skills to get the job done." I find this convenient because when I remember that I have that skill, it will pull it up and use it. But the good part is it also loaded my objectives, it also loaded my memorable explanation skill without me explicitly asking for it because it thought those would be relevant. And **this is where skills come in handy. If you think that there are a set of prompts which should be used along with other prompts because they make other prompts more effective, it belongs in a skill.**

**Anand**: [25:13] In my case, I find that there are probably about 20 of these mental model type skills that make sense or personalized things like "write like me," "work towards my objectives," etc., that make sense.

**Sonal**: [25:30] This is only in Claude, not ChatGPT?

**Anand**: [25:33] This is only Claude. **In ChatGPT, you would have to manually copy-paste the skill unless you're on the Enterprise ChatGPT edition.** Now there are some, however, a lot of people confuse skills with prompts, and that includes me. An example of something that I confused recently was my "Talk Preparation" skill, my "Talk Event Scan"—let's take Talk Event Scan. I attend talks and events, etc. Initially I had crafted this as a skill, which basically said, "Think about the kinds of talks that I want to attend, events that I want to attend, etc., figure out where I should look for these," etc. Then I realized, hold on, this is not something that I want as a skill; this is something that I want to run as a weekly schedule. Why would I move it back?

**Anand**: [26:31] Another example was "Email Reply." At first I had put this as a skill, then I realized if I want an email reply, I will ask for an email reply. Why would I want to have it automatically create an email reply in the middle of... well, maybe it makes sense, I don't know. For me it didn't make sense. **So the test is: if I want some prompt to get automatically injected in the right context and that happens at least 5% of the time for my prompts, create it as a skill.**

**Anand**: [27:01] And lastly, how do you go about creating a skill? Quickest way is go to Claude and tell it, "**Create a skill for something. Here's everything that I know about it**," and dictate. And that's because Claude has a skill creator skill, and they invented skills, and the skill creator is a pretty decent one. This is a default skill. It will give you a reasonable starting point, and anyway Claude is slightly more easier to read than ChatGPT's output, so you'll be able to review it and see if it makes sense.

**Pragati**: [27:42] Got it. Super helpful. **Just on ChatGPT, we'll have to manually paste every time we have to use the skill, right? Like in the case of Claude, it automatically picked the ones that you had not explicitly called out, but in ChatGPT, we will have to put it that use this.**

**Anand**: [27:58] Yes. You can kind of work around it; you can say, for instance, you could give it a prompt something along the lines of "go to my past chat named something skill and use it" or even more succinctly as "read X skill chat." But you'll have to explicitly do that. It may pick it up automatically.

**Sonal**: [28:28] Anand, in plugins on ChatGPT, I just saw that there is an option for skills on—at the top. So which means that you could store a skill on ChatGPT as well, right?

**Anand**: [28:40] I did not know this.

**Sonal**: [28:42] I just discovered this as I was looking at plugins and then I realized there's skills there.

**Anand**: [28:47] Yeah, no, this is certainly more recent than... well, must have happened this week. So yes, absolutely. And let me put in a few skills here. I'm going to copy one of my favorite skills from Claude and we'll test it. But this is brilliant. Thank you, Sonal. That's very helpful. Edit, copy the Expert Lens skill. By the way, all of my skills are available on docs somewhere. Skills, yeah, I will just put a link to those skills for what it's worth. But keep in mind that what works for me will not necessarily work for you or even be relevant. Some of them are generic.

**Anand**: [29:48] So if I save these and ask it to apply it—well, let me pick some realistic question that I have in my to-do list. [End of recording]

---

**Anand**: [00:01] Sorry, just looking for some task. Yeah, here is... going to say no. I'll go back and create a new chat. For whatever it's worth, I'll say use the expert—no, actually I'm not going to. **As an expert, comment on this**. I will also remove access to my local server just in case it picks up a connector from there and runs it. What I'm curious to see is whether there's any visual indication that it's picking up the... okay. Did you use any skills?

**Anand**: [01:14] I'll add that for this purpose, and we'll get it a little faster. Okay. Could you use this skill? Okay, this does reframe the task. So **it looks like it's still in the early days. Even though I said "as an expert I want you to answer," it didn't realize from the context that it should have picked it up. But I'm sure it's not going to be long. And yes, thanks Sonal, this will simplify things for a lot of people.**

**Sonal**: [01:57] So Anand, I noticed that you did not put your data analysis skill in your skills library in Claude. Was there a reason that you just always want to pull it from somewhere else and not necessarily across all your chats?

**Anand**: [02:13] Yes. **What I use Claude chat for is thinking.**

**Sonal**: [02:22] And not for data analysis? And that's the reason why you don't have it in the skill section?

**Debi**: [02:33] I think Anand has frozen. Yeah. He's frozen. He'll come back, he'll come back hopefully.

**Debi**: [02:54] We'll do a quick time check at this point. It's about 4:30 PM. So I think once Anand gets back on, we'll check with him what else he wants to cover. But if anybody has any questions, can you start putting them down on the chat? We'll try and see what we can cover, please.

**Sonal**: [03:22] I think Debi, the one that you've asked about—is there a connector for LinkedIn—is something I'm interested in as well. So if we can ask Anand about that, that would be great.

**Sandeep**: [03:34] Yeah, Claude doesn't have that. I've looked at LinkedIn connectors. I don't know, GPT may or may not have it. **One workaround that has worked for me was the Manus model. It can take control of your browser and go to LinkedIn and do whatever you want.**

**Sonal**: [03:54] What model? Sorry, Sandeep?

**Sandeep**: [03:56] Manus. M-A-N-U-S. Manus AI.

**Debi**: [03:59] Yeah, M-A-N-U-S. Manus.ai.

**Sonal**: [04:02] Yeah, but I'll refuse that.

**Sandeep**: [04:04] Yeah, so I've done searches in LinkedIn and posting using Manus, but I haven't found any other way to do it.

**Anand**: [04:14] Say found a way to do what, Sandeep? Sorry, Anand you're back. Sorry, we lost you for a bit.

**Anand**: [04:19] Yeah, no, my system crashed. I just rebooted.

**Sandeep**: [04:22] Too much work! You make it work too hard. Too many skills!

**Debi**: [04:28] I think I had put down a question and Sonal had asked as well, right? Are there connectors to LinkedIn? And I think the team was just discussing that we haven't found anything which is in either through Claude or through ChatGPT or any of the others. So I think Sandeep said that he had used Manus AI, right? And I think that's managed to, you know, go into LinkedIn and crawl through it. Right, Sandeep? That's effectively what it's done.

**Sandeep**: [05:01] Basically it takes your browser—you can log into your browser, I mean you can log into LinkedIn, give it access to the browser—and it can do whatever you want. It crawls through it.

**Debi**: [05:13] It crawls through it. Yeah. Anand, I also was trying to do a time check right now. It's about 4:30 PM, right?

**Anand**: [05:19] We'll wrap up?

**Debi**: [05:21] Yeah, so a couple of things. One is that I've just asked people that if they have any questions to put down their questions in the chat mode. Second is that if there's anything else that you wanted to cover right now on this whole data analysis part, I just wanted to check on that.

**Anand**: [05:43] I'll just wrap up on this. So the key thing that I wanted to mention was because it's able to do machine learning on the fly (or deep learning for that matter), we can leverage that power. So for instance, in the earlier prompt which took a good half an hour or more to run, **it's tested CatBoost, XGBoost, and LightGBM—three models. And of these, the model benchmark is CatBoost; on average, it makes the fewest misses in terms of the error that it makes on the flat price.**

**Anand**: [06:22] And as a result, what I can expect is a typical error of just under $20,000 based on what it's predicted. Now, the way in which it's predicting, we don't know, we don't care. We just know that this is the kind of error that it tends to make and that this model is better than the others. **And the usage is now: Option 1—tell it "now that you've built the model, give me the price for a particular flat." And if there's any new data, rebuild the model accordingly. Or you can ask it to download or give me a program that I can download and run that will allow me the same kind of predictions.**

**Anand**: [07:08] **In short, use the models—use ChatGPT and Claude—not just for simple analysis of the kind that we're used to, but also for the sophisticated kind of data analysis that you would give to a data scientist.**

**Debi**: [07:26] Just put that page back again, Anand, what you were just showing. I just want to read that one. Okay. So the average miss is nothing but MAE (Mean Absolute Error), right? Basically, a percentage error, right? 4.11 for all of them. Okay.

**Anand**: [07:47] Whereas the typical, I'm guessing, is the median.

**Debi**: [07:49] Is the median. Yeah. And then the "comparable median baseline," which is the last line that they mentioned over there, which is 173,150. This is just a comparable of the two numbers, I guess. Right?

**Anand**: [08:06] I'm guessing that if you said "let's assume that flat price is a constant" or some such simple baseline, that's how bad it would be. A calibrated version of CatBoost is much better. "One-under-untouched comparison." Okay, I don't know what "untouched" means.

**Sandeep**: [08:26] So Anand, just to recap: we formed the data set, we did the analysis, and then we built the predictive model. So that's the three steps, I would say.

**Anand**: [08:38] Exactly. And **we can do this across any number of data sets and therefore the combinatorial capability is much higher.**

**Sandeep**: [08:48] And will these data sets... I mean, to improve the predictability of the model, as more data comes in, we can automate that as well, right?

**Anand**: [09:00] That can be on a schedule.

**Sandeep**: [09:01] It can be on a schedule. Because like in the case of HDB, there's a regular reporting schedule, right? Every certain date of every month they give the new data or something like that.

**Anand**: [09:12] Exactly. At which point the doing it becomes easier and figuring out what to do becomes tougher. I mean, there are so many possibilities—what else can I do? And **I find it fairly useful to... in fact, I have only one schedule that I run on a regular basis, and that is my data set scan. This is just looking for interesting new data sets on a daily basis.** And the most recent one that it found is NISAR L-band radar, which can identify fresh land deformation, flood, forest loss, crop moisture anomalies. And this is based on satellite data—public data.

**Anand**: [10:15] Fairly large, but I don't see why AI should not have a problem analyzing it. 64 GB of operational data from Norwegian schools on electricity and weather. So how much energy can public buildings save by changing control schedules? So effectively, **my scan is: search, find public data, find out what can be done. At some point, I will just copy this entire thing, put it into a new prompt, and say, "now do the analysis as well."**

**Sandeep**: [10:50] So you're not restricting this to any type of data? This is just open-ended? Anything which...

**Anand**: [10:55] No, I've got my preferences. The broad idea is it should be of relevance, journalistic.

**Debi**: [11:06] And this is on a schedule, right for you, Anand?

**Anand**: [11:09] Daily.

**Debi**: [11:10] On a daily schedule. Wow. Interesting. Yeah. I think Pragati has a question.

**Pragati**: [11:19] Yeah, Anand, just a quick one. Understanding: is there an architecture that we can use to build let's say a RAG (Retrieval-Augmented Generation) and combine skills in ChatGPT without like let's say using Codex or any of the Claude code and just create a simple project markdown file and then combine both? Is that possible?

**Anand**: [11:41] Without using Codex? Yes. But if it's for your use. If you want to build a software and give it to someone else, which of the two?

**Pragati**: [11:51] I want to get stuff done. Not giving it to somebody.

**Anand**: [11:54] Okay, then it is possible. I'll show you.

**Pragati**: [11:58] Like, I'll give you a use case also that I have in mind is: let's say I would be using... for example, **if I'm coaching somebody and then I need a... you know, I have a lot of transcripts that I could sort of build in my RAG and then give that data set. And then the skills I define in a certain way to always use the transcripts, but then it is able to also give me a certain way of coaching, etc. So it becomes my coaching tool and I'm able to create that without coding it as such.**

**Anand**: [12:41] Without entirely coding it... well, maybe, kind of. Let me show you how I'm doing it right now. So on ChatGPT, I have created a plugin called Local MCP. And what this does is it lets it connect to my machine. On my machine, I am running it by... let me show you what it looks like. Because I've written this program myself, I just say "run this particular MCP server," which has access to all of my data. And this is the instruction that I've given it: which is look, you can write code under my machine, here's where you'll find the skills, here's where you'll find all my transcripts, here's where my notes are, here's where my emails are, my talks, my data stories, my everything else.

**Anand**: [13:51] And then I can ask it practically any question by prefixing it by "@LocalMCP." So let's keep it as a fast one and ask it the question: **"What's your sense of the most interesting... okay, let's look at public record, right? What's your guess on the most mind-blowing discovery that I would have made in the last one week based on all my emails and transcripts and whatever else?"** And have it run. Now this connects—hopefully—it will do the equivalent of a RAG on the fly and connect to my machine, search by writing programs, get to the answer.

**Anand**: [14:48] Yeah, so here it's for instance saying "find me all the skills that are on Anand's machine and find all the activity logs in the last seven days, relevant notes about recently model-fied stuff." And saying hey, there are 18 meetings in 6 days, narrowing it to where I've said I changed my mind of something. And the ways in which it will do that will also be quite interesting. It will search for different kinds of words, try out different combinations. **So it doesn't need to do a RAG; it's able to do it by just running a search. And I can monitor what searches it's doing on my machine.** So it's searching: "surprise," "realize," "discover," "mind-blow," "change my mind," "didn't know." A reasonably good collection of words that would cover the way in which I might have captured or said this.

**Anand**: [15:53] And yeah, going beyond that, it seems to have found some of those results. Let's see if... okay, now it's looking in my calendar for the last eight days, see if there's anything interesting there. It's looking at all the stuff that I browsed over the last eight days. So a lot of this is about how much context we're able to give it as well. But **as long as we're able to give it that context and connect it somehow—in this case, I'm using a custom-made MCP connector—it would work. Alternative: just put it on Dropbox or Google Drive or whatever and connect it, that's it.**

**Debi**: [16:32] Okay. Perfect. Yeah. So Anand, I think the next session will be on whiteboarding, correct? So what would you suggest for the entire audience? How should they prepare? Should they come with a project? What would you suggest they do?

**Anand**: [16:48] Yeah, for the next session, I will send out an email at least a week in advance, but **we'll need Claude Code or Codex on people's desktops.**

**Debi**: [17:00] Okay. And apart from that, should they come with any ideas or any thoughts on...

**Anand**: [17:06] That would be good, but I will in fact go a step further and see if people can have white-boarded something before they come into the session and we do this offline. So that the session itself... the whiteboarding part is easy. There isn't much to cover there, unlike this session where people needed to know... the whiteboarding is just "tell it to do it." But **what fails is what we need to discuss.**

**Debi**: [17:32] Correct. So maybe when you're drafting the email, right, you could express this: that you know, let people think of a project, come out with something, right, and you know, we could perhaps get that presented and have a discussion around it. I think that'll be great.

**Anand**: [17:48] Absolutely. Great.

**Debi**: [17:51] Fantastic. Thank you so much, Anand. We've gone on for two hours and fifteen minutes.

**Anand**: [17:58] It's been longer and longer! It gets better and better. And Sandeep, do send the chat messages as well if you could.

**Sandeep**: [18:03] Yeah, I will. Yeah, there's some problem in downloading the chat, but I'll do it. I'll figure it out.

**Anand**: [18:11] Thank you.

**Debi**: [18:13] Thank you everyone for joining. Really appreciate it. Thank you.

**Anand**: [18:18] Thank you. Bye.

**Sonal**: [18:20] Thank you. Bye.

**Pragati**: [18:21] Bye. Thank you.