## Transcript

**Ashish**: [00:00] It’s a pleasure.

**Simran**: [00:03] Yup, Shivani.

**Ashish**: [00:13] Okay, we can go ahead. The recording has also started. Yeah.

**Anand**: [00:16] Fairly cozy group, so let's dive in. Very briefly, by way of introduction, I'm Anand. I co-founded a data science company some years ago and recently have sold it. Now, most of my work is researching LLMs. I've designated myself an "LLM psychologist," which roughly means I figure out how LLMs think.

**Anand**: [00:46] Let's go to the more important thing. As we go through this, this is a workshop, so in a workshop, _you_ will be doing the work. Quick question: How many of you have "vibe coded" before? Or, let's put it the other way. How many of you have _not_ had any LLM write any program? If you have never gotten any LLM to write a program, can you type "never" in the chat window? Just do a quick check to see if you are able to access the chat window in the first place.

**Anand**: [01:29] Just say "Yes" for "Yes, I have vibe coded," or "No" for "No, I have not written any code." Either way is fine.

**Anand**: [01:40] I can see your message, Ashish.

**Ashish**: [01:42] They can speak as well. I don't know if chat is enabled for them.

**Anand**: [01:46] Oh, I see. Is it possible to enable chat?

**Simran**: [01:49] Yeah, chat is also enabled. They can write.

**Anand**: [01:52] I see one response, but it says "Anonymous Attendee." Nope, got you. That's still helpful. Good. Others, just put in a message in the chat window saying "Yes" for you have vibe coded, or "No" for you have not vibe coded.

**Anand**: [02:11] Okay, that's another "No." It's interesting that the Q&A comes as anonymous attendees, but it's okay.

**Ashish**: [02:26] Now? I mean, I think...

**Anand**: [02:31] Sorry, by "now," Ashish, you meant what?

**Ashish**: [02:35] Can people type in their messages now? I think it was not enabled.

**Anand**: [02:40] I see.

**Ashish**: [02:45] Okay, chat just got enabled. Got you. Yeah, Nikhil just joined.

**Anand**: [02:50] Okay, so Nikhil, I guess the "Yes" is you have vibe coded. And Abhishek, okay, you're clear you haven't vibe coded. Pari, you're saying "Yes," you have vibe coded. Okay, good. There were a couple of "No's." If you could just put in your responses here again in the chat window, I'll know who is who. Or just unmute yourself.

**Anand**: [03:20] Got you, Surbhi [phonetic], that was a "No." Okay, we have four responses. Okay, it's easy enough for me to just manually check. And Shambhavi, I know you've just joined, but Shambhavi, have you vibe coded before? Meaning, written code using LLMs before? You can either put a Yes or No in the chat window or you can just speak up. Annie, same question to you. And Diya as well, if you could share whether you've vibe coded or not.

**Anand**: [04:12] Got you, Diya. Okay, and Shambhavi, you've used it for assistance. Fair enough. Annie? Maybe we could... yeah, you have. To limited extent. Got it.

**Anand**: [04:30] So I think we have a 50-50 mix. In that case, what I'm thinking of doing is, rather than treat this as an introduction to vibe coding—since at least some of you have been vibe coding—great. I'd like to ask for a volunteer who will share their screens and let's work together on this. And yeah, I know how this works. If I ask for a volunteer, it will take you a long time to figure out who will volunteer. So, Pari, you mentioned that you have vibe coded. Could you share your screen and let's do some vibe coding together? And could you also unmute yourself while you're at it, please?

**Ashish**: [05:14] Pari, I've promoted you to panelist. Just one second.

**Anand**: [05:27] And while we are at it, Dharmendra, I saw that you just joined in. Could you... okay, there's no mic against you, but...

**Ashish**: [05:35] Can you promote him to panelist? Pari? I'm not able to. Yeah, done now, I guess.

**Anand**: [05:46] Dharmendra, could you just put on the chat window a message indicating whether you have vibe coded in the past? Yes or no will suffice. And Pari, we're waiting for you to share screens and unmute yourself, please.

**Shivani**: [05:59] They don't want to be a panelist.

**Ashish**: [06:05] We need some other volunteer.

**Shivani**: [06:08] Why don't you pick? I'm not sure. We can... Nikhil can do it.

**Ashish**: [06:14] Hi, Nikhil, can you do it?

**Nikhil**: [06:17] Sorry, I didn't catch that.

**Anand**: [06:22] Could you share your screen?

**Nikhil**: [06:24] Sure. What do you want me to share?

**Anand**: [06:26] Just the screen. Then we'll...

**Ashish**: [06:30] You're the volunteer.

**Anand**: [06:34] You're the _Bakra_ [sacrificial lamb/scapegoat].

**Shivani**: [06:36] Volunteer is really too... It's like the forced volunteer.

**Anand**: [06:45] Then it's no longer volunteering, right? Actually, that's probably the right adjective for all of you. What? One day before your holidays, is it?

**Ashish**: [06:59] I think the exams are still on.

**Anand**: [07:02] Oh, I see. Okay. Let's just give Nikhil a minute to share screens. Okay, Nikhil's rejoined.

**Shivani**: [07:34] Nikhil, you can share your screen. You have the rights to share the screen.

**Anand**: [07:54] Perfect. And Gemini is on. Great. And please unmute yourself, Nikhil. We'll just talk. And others, please just feel free to unmute yourself at any point, unless of course you're...

**Nikhil**: [08:07] Yes. Can you hear me now?

**Anand**: [08:08] Yes, loud and clear. Cool. So let's do something, Nikhil. Let's vibe code something. What are you covering currently in any of your courses?

**Nikhil**: [08:24] So, no, I'm not a student actually. I'm a staff at Ashoka and I manage a lab. So I haven't vibe coded much. I've tried a little bit here and there, but not much.

**Anand**: [08:38] Got you. Fair enough.

**Ashish**: [08:39] He is an ex-Ashokan, though.

**Anand**: [08:42] Okay, fair enough. But I am curious. Should we maybe have a student then share and vibe code? I'll leave it to you, Ashish, whatever you say.

**Ashish**: [08:55] I mean, it's okay. Nikhil could also do it. He's quite connected with the students. Fair enough. If any student wants to do it, they can as well.

**Ashish**: [09:11] Nikhil, we can begin with you. I mean, if there is anyone else, we can perhaps ask them later.

**Anand**: [09:18] Let's do that. So in that case, Nikhil, currently what are some of the topics that the students are studying as part of their curriculum?

**Nikhil**: [09:30] Do you mean within CS or across Ashoka?

**Anand**: [09:34] Across.

**Nikhil**: [09:37] There are different domains of studies: humanities, and then political science, even arts, and there are STEM departments as well. So people from all across the domain are there.

**Anand**: [09:53] Let's pick something on the humanities side. I'm guessing... Indian Civilizations is probably part of the curriculum, right?

**Nikhil**: [10:04] Yes, yes. I think that's one of the foundational courses that UG students take.

**Ashish**: [10:09] Which course? Indian?

**Nikhil**: [10:12] Indian Civilization.

**Ashish**: [10:13] Yeah, yeah, yeah. That I think everyone studies.

**Anand**: [10:16] Cool. Now, let's vibe code. How about if we built an application that will do a better job than any textbook of teaching us Indian Civilization? How about if you ask Gemini to **build an interactive scrollable timeline of the Mughal period?**

**Anand**: [10:51] And don't press enter yet. But two changes. Under "Tools," could you select the dropdown... "Canvas"? And just select the dropdown once again. This is for anyone who wants to follow along. What Nikhil has done is gone to gemini.google.com, and all of you have access to Gemini, I'm guessing. And the things that you can create here—you probably tried all of this—but Canvas in particular lets you create documents, web pages, applications, and a whole bunch of things like that.

**Anand**: [11:31] The second change, Nikhil, that you might want to try is there's a dropdown next to "Fast" a bit to the right side. Could you change that to "Pro"? And you generally get better results with Pro. And now let's run this.

**Anand**: [11:46] So really, once this builds an application, there's a pretty decent chance it will work. You'll see that there's nothing much that one needs to learn about vibe coding. You tell the system to do it, and it does it. So it sort of begs the question: why are we even having a workshop on vibe coding? Why are people talking about vibe coding? What the heck is it? What's the big deal about it?

**Anand**: [12:13] I personally think _there is no big deal_. **The point is simply that we haven't gotten used to it.** A lot of us assume that programming takes effort—it does—and therefore most people _can't_ do it. The latter part is no longer true. And therefore, the kinds of things that we can build, we are not familiar with. The kinds of things we can ask for, we are not familiar enough with. So this is a session where we'll try and see what it is that we can do using our imagination.

**Anand**: [12:50] So what it's doing right now is writing code. A year ago, it would have written code that mostly would have worked, but maybe 10% of the time wouldn't. These days that 10% has become, I don't know, half a percent or so. And yeah, just play around with it, Nikhil. Scroll down, let's see what it teaches.

**Anand**: [13:12] Okay. Babur begins, Humayun came in... Okay, this is fine, but I would say it's interesting in that if we hadn't known that we could do this, this is a useful exercise. But what change would you suggest to this, Nikhil, for it to start becoming useful?

**Nikhil**: [13:38] Hmm. I think this seems good in itself, but maybe... I mean, I have done a little bit of vibe coding and I feel like it kind of depends on the user as well, like what kind of information they want to get. So maybe if they want to do it in a certain way or if they want to read the information in a certain way, they can maybe try those kind of iterations.

**Anand**: [14:08] Like?

**Nikhil**: [14:11] Hmm. Maybe a map of India of how, you know, the Mughal period expanded or contracted, and we can see across the timeline.

**Anand**: [14:23] Fantastic. Let's try that. Go ahead. How would you prompt for it?

**Nikhil**: [14:31] I would write something like... "Can you include a map of India and a scrollable timeline which shows how the domain..." No, go ahead, go ahead. It'll figure it out better than we can explain. "...how the domain of the Mughals expanded or contracted over the years."

**Anand**: [15:27] Which is cool. And let it run. Now, what I've usually found is that typing helps us bring a certain amount of clarity, but is also a little bit of a pain because then I have to think about what I wanted to say. So why don't we try something different? Could you open a new tab and go to ChatGPT? Doesn't matter if you have an account or not.

**Anand**: [15:52] And on the right side there's a little mic button. After you click on that, why don't you just tell it what you just typed, approximately? And then yeah, just click on... not that... yeah, that one. The dictate button. Yes, go ahead.

**Nikhil**: [16:11] "Can you build me a scrollable app which shows the domain and expansion and contraction of the Mughal period and also the entire Mughal history?"

**Anand**: [16:26] Perfect. That was easier. And you could copy it and paste it back. No need to press submit here. We just need a good transcription engine, and I find that ChatGPT has one of the best transcription engines that's available for free. So I usually just dictate into ChatGPT, paste it wherever I want, including my notes.

**Anand**: [16:45] **What I find though that happens—and I'll encourage you to try this the next time—is when we are talking, we can ramble. When we ramble, it doesn't hurt the prompt any. These things are smart enough to figure out what we want.** So why don't we do this? Let's take a quick look back at the Gemini window to see if something actually was... I mean, something is happening, going on or not. Okay, it's adding to the context, blah blah blah.

**Anand**: [17:14] Let's try something else as a new exercise. Could we maybe take a person and create a visual biography on a map? That is, their history charted on a map. Very similar to what you said, but it's a different kind of an application focused on an individual rather than a period. Could you just _ramble_ into ChatGPT on what you might want that application to look like?

**Nikhil**: [17:48] Oh, okay, sorry. I thought the whole point was you not typing.

**Anand**: [17:52] No, no, I want you to dictate. Okay. Yeah.

**Ashish**: [17:55] Arnold Schwarzenegger would be... he's well documented.

**Anand**: [17:58] Okay.

**Ashish**: [18:00] Or if it has to be Mughal, then Aurangzeb or something would be good.

**Nikhil**: [18:06] I was thinking of going with Einstein.

**Anand**: [18:08] Yeah?

**Ashish**: [18:09] Yeah, Einstein would be cool. Yeah.

**Nikhil**: [18:11] Okay.

**Ashish**: [18:13] Anand, can you make me the co-host? Then I can promote other people to talk and all that. Because you became the host.

**Anand**: [18:22] Oh, oh. I don't even know... okay. Ashish More... Make Co-host. Okay.

**Ashish**: [18:30] Yeah, yeah. Thank you.

**Anand**: [18:32] Thanks.

**Nikhil**: [18:33] Anand, quick question. Would you want me to ask ChatGPT to build a website? Or is it a picture are we expecting? What is it?

**Anand**: [18:45] No, I'm suggesting that you just use ChatGPT as a dictation engine, nothing more. We'll give the prompt back to Gemini where you have a Pro account clearly, and ask _it_ to build. Yeah, why not a website? But what else would you have in mind rather than a website?

**Nikhil**: [19:02] I mean, website would be good because just one picture generation would not be... would not test the capabilities of Gemini, I would say.

**Anand**: [19:12] That is a good point. So yeah, go ahead. But that gives me an idea and we'll do something next. But yeah, why don't you go ahead and ramble to it and paste into Gemini?

**Nikhil**: [19:21] Sure. Do I need to keep myself unmuted for this?

**Anand**: [19:37] Yes, please, so that we can hear you.

**Nikhil**: [19:40] Sure. "Can you go ahead and build me a website which shows the journey of Albert Einstein throughout his career? I want to see over the years how his... what his publications were like, what his research was like, the places he visited, spoke at, or the institutions he taught at or he attended. So I would want to look at this website and get a complete idea of what Einstein's entire life and, you know, achievements or story was about, that is there in the public domain. Is that good enough?"

**Anand**: [20:28] Yep. Go ahead. Copy, paste into a new Gemini tab and let's see what it comes up with.

**Nikhil**: [20:43] Tools... yeah, Canvas.

**Anand**: [20:46] And let it run. So let's see where the earlier one got to.

**Anand**: [20:53] Okay... oh. That's a pathetic map of India, but... keep scrolling.

**Anand**: [21:04] Okay, it has totally messed this up. And is not unusual. So, the problem that it has... okay, how would _you_ fix this problem?

**Nikhil**: [21:20] I would just simply try to explain in more detail what I wanted to do. So maybe something like... "I want a map of India and a control panel on..."

**Ashish**: [21:46] Map of maybe _present-day_ India and Pakistan?

**Nikhil**: [21:51] I think Gemini will get that automatically because we are talking about Mughal period?

**Anand**: [21:56] Hmm. And yeah, it kind of seems to have included that area.

**Ashish**: [22:12] And for those who joined a little bit later, we are building a timeline of the Mughal empire in one of the prompts, and the other prompt is going to be a website about Albert Einstein's life and so on. Just for some people who joined later.

**Anand**: [22:31] And not that we specifically set out to build this, we're just having fun.

**Nikhil**: [22:38] So maybe something like this and see what it gives?

**Anand**: [22:42] I would also add: make sure that I'm able to see the map while I scroll.

**Nikhil**: [22:49] Make sure... sorry?

**Anand**: [22:52] At the moment, when you scroll, the map vanishes near the top, right?

**Nikhil**: [22:56] Yes.

**Anand**: [22:57] You might want to tell it: "Look, I also want to see the map while we scroll."

**Anand**: [23:20] Cool. And let that run. What about the other one? Okay, that's still running. Oh, now it's done. Okay. Let's take a look.

**Ashish**: [23:33] This looks quite nice actually.

**Anand**: [23:35] Not bad. Though I was expecting a...

**Nikhil**: [23:41] It's all the same thing. Oh, okay, it's changing here.

**Anand**: [23:46] Yeah, yeah. Then... okay, and it filters and so on. Years, various publications... Okay, fine.

**Anand**: [23:56] Though when you said "journey," in my mind, I had jumped to journey on a _map_ rather than on... well, just a timeline. But yeah, which kind of begs a question: can it actually use something like Google Maps to do this?

**Anand**: [24:16] Sorry, your voice got cut. Oh, sorry. I was saying... but what I was hoping for is that it might use something like Google Maps or whatever to do this. But let me give that a shot. I'll share my screen in a few minutes. But before that, I wanted to share a few things that have emerged from this.

**Anand**: [24:34] First: **All it takes is an idea if we want to build something.**
**Anand**: [24:41] Second: **Once we build it, it may not work. All it takes is for us to tell it how to change, and it makes the change.** Which then may still be fine, may not be fine, that's another question. We'll come to that.
**Anand**: [24:56] These things happen slowly. So rather than just run one thing at a time, we may as well multitask. And you don't have to only multitask with multiple coding agents. I usually watch a movie while my coding agent runs, and then I come back and check. It usually helps to watch movies that you've already watched, that way you don't have to pay attention to either. Or pick a book, eat, whatever. But multitasking is a pretty decent idea.

**Anand**: [25:27] Now we tried maps and it didn't quite work, so... Sorry, somebody said something. Okay, maybe I missed that. So why don't you stop sharing screens for a bit, Nikhil? And I will share my screen and let's see if we can get it to draw a map.

**Anand**: [25:50] What I'm going to do first is go to ChatGPT and start dictating. Look, I'd like you to build an interactive map-based application—and when I say map, I mean Google Maps, OpenStreetMap, that sort of a thing. Use whatever library you need to build this. But what I want is to show the visual timeline of Einstein. I want to zoom in into where he was born and then, you know, show something about his birth, and then zoom into the next place he was and show something about what he did in that place, and so on. Effectively creating a full-fledged visual biography on a map.

**Anand**: [26:38] The reason I want to do this is, as part of my class assignment, I have to study Einstein and produce an essay about Einstein, and I'll probably get more marks if I create an innovative new format out of it. But also, the other thing that I'm curious about is if we start producing new and different formats, are there any kinds of innovation AI itself can bring in?

**Anand**: [27:09] So as I think about it, what I'd like you to do is introduce something novel. Something that maybe the world is mostly not familiar with or has never seen as a prominent visual feature. And just write down what it is. I may not even recognize it when I see it. So yeah, just give me that.

**Anand**: [27:32] Now, I was literally rambling. Midway through it, an idea popped up. I just shared it. It's okay. What is the big deal? And if it fails, it fails. So I'm going to ask Gemini Canvas Pro to do exactly the same thing. While we are at it, may as well ask ChatGPT also, except that I'm not going to put an extended thing. I'm just going to give it to the regular model, but also have it do the Canvas thing. And give it to Claude as well and see how that's going to work.

**Anand**: [28:07] Now, this is another one of those things that I've learned in vibe coding. It's free, cheap, I anyway am not having to do any of the work. So **what's the harm in having a dozen of these run in parallel? If I like one, that's good enough.** So don't bother asking one tool for an output. Ask _all_ the tools for an output and see what you get. That will probably give some ideas as well.

**Anand**: [28:40] Why don't you give me some things that might be interesting to vibe code and let's see where it goes. Any thoughts from anyone? And this is the point where I'll start my timer and see if you are able to beat NID's record of staying silent for two minutes and 30 seconds.

**Ashish**: [29:04] Something to do with stock market will be nice.

**Anand**: [29:06] You don't count, Ashish.

**Ashish**: [29:22] Nice. I think I'll have to pick people. Raghav?

**Anand**: [29:24] No, no! Don't ruin... you're destroying the experiment.

**Ashish**: [29:27] I'm destroying the experiment. Okay. Sorry.

**Anand**: [29:37] Okay. Bunch of stuff on the chat window. Okay. Did you mean in this prompt? Not necessarily, Diya. In any prompt. What might we vibe code?

**Anand**: [29:44] And Abhishek asks: "Can we vibe code an online course? Like a self-paced MOOC on behavioral science?" Why not? Let's vibe code it. I'm going to ask Claude... I'm going to ask ChatGPT to...

**Anand**: [29:58] Could you build me an online course? I'd like a self-paced online course on behavioral science. While you're at it, do some gamification. Make it interesting and engaging for me.

**Anand**: [30:15] And let's give it a shot. Which I'm going to paste back into Claude Sonnet and give it a shot.

**Anand**: [30:25] Now here's the thing. Part of what we need to think about when we vibe code is: why am I vibe coding? It's one thing to say I'm vibe coding so that I want to learn what we can do with vibe coding, and that's a great purpose. But otherwise, we are often stuck for ideas. And like in this case, you're having to think and say, "Okay, maybe I can use vibe coding for this," etc., right?

**Anand**: [30:50] **One of the things I've learned is: whenever we are asked a question, it is best to deflect that to an agent and ask _it_ the question.**

**Anand**: [31:04] So, let's ask it: "I'm talking to some students at Ashoka University and I'd like some examples of things that we can vibe code that will be interesting and useful to them. Many of them are familiar with vibe coding, so the point is not to _teach_ vibe coding. The point is actually to figure out what uses they can put vibe coding to that are interesting and innovative. Also, don't give me boring examples. Give me some fun, nice examples that are also non-obvious. We want something that can create a certain degree of surprise as well, because, I mean, what is the point of just sharing an obvious one? Okay. Yeah. Go get some good stuff."

**Anand**: [31:54] Let it transcribe that. I'll give that to Claude as well, which I think has a pretty nice creative sense. As well as to Gemini, which is not bad at all. And... for some reason, it's... Okay, reload. Claude has slowed down. Oh, is it because it thinks I'm talking to it too much?

**Anand**: [32:26] Okay. And... hmm. Here's one: "Your professor, but honest." You can feed a course syllabus and notes into a Claude-powered app that answers what will actually be on the exam. Okay. And "Which readings can I safely skip?" But I don't need to vibe code this. I can just feed it into Claude and give it.

**Anand**: [32:47] "Paste an essay..." Okay, now at this point, I'm not going to sit and read all of this. Hold on. For the first two, I don't need to vibe code. I can just pass it to the agent and it'll give me an answer. Give me stuff that _needs_ coding. And let it rerun.

**Anand**: [33:21] So you'll notice also that I didn't read the whole thing. The bottleneck is _my_ time. I don't want to sit and review its whole work and say, "Okay, fine, you did three out of five correct." Get it all right. Otherwise, don't waste my time.

**Anand**: [33:36] So it's saying: "Okay, we can create a campus mood map where students can drop anonymous emoji reactions tied to locations." Yeah, that is not a bad idea. And it builds a real-time emotional geography of the university. Just share a link and people can come there, come to your site and say... you know, automatically geocodes the location and put in an emoji on this. Which is not a bad idea at all.

**Anand**: [34:04] "Two people type their sides simultaneously in a split-screen interface, neither seeing the other's output until they submit. And then when we reveal the overlap, we show the gap." Okay. So we give two people a topic and say, "Write an essay about this" or "a paragraph about this," and then we start connecting. That's an interesting idea.

**Anand**: [34:25] "Personal inflation tracker. Running log where you enter the price of everything you actually buy." Okay, this is an interesting one. You know what? What I'd love to do is... I don't need a log of things that I buy. I _anyway_ have my bank statement. Most of my stuff is anyway bought through the debit card. So what I could do is just upload all of my bank statements. Add... where are my bank statements? I think in Dropbox... Yeah, here we are. Dropbox. Anand bank statements. Let's take my OCBC Singapore bank statements and I'll feed in... let's sort by date. So everything in 2025 and upload it.

**Anand**: [35:24] And say... ooh, okay. Maybe it doesn't accept... Oh, but hold on. I also have this as Excel files. Why am I bothering? That way I can give it my entire... give it to Claude. Yeah, why not? Where is my bank statement in Excel? Yeah. Okay, that was tax... Oh yeah, here's a CSV file.

**Anand**: [35:54] Now let's tell Claude: "I've uploaded my OCBC bank statement and what I'm interested in doing is tracking inflation of items. Are there things that I've bought—exactly the same thing that I've bought—whose price has changed over time? I'm looking for as many... in fact, a comprehensive list of everything whose price has changed over the duration of this bank statement. And tell me how much it has changed. In fact, give me a nice visual picture of how much it has changed."

**Anand**: [36:38] And take that and put it in here. Now, the thing about this is that data analysis is something it does really well. And my bank statement is data. So let it run. Now you would have noticed by now... hold on, we are giving it so many problems. Let's take a look at what it's created in the first place.

**Anand**: [36:56] So let's go back. We told ChatGPT to build this interactive application. Okay, here's Einstein's map application. Okay. So in Munich is where he lived, and he was sick in bed in Aarau. Okay. Chasing a sunbeam... This is...

**Anand**: [37:22] You know, the effects are... oh, okay! Now I see the effect that it produced. I'm not sure how clear it is for you, but do you notice the map sort of wave happening? That is nice. It's almost like the flashback effect that you see in movies. The entire picture changing. I didn't even know that such an effect was possible. So that's definitely an interesting way in which it's doing this. And also a "warp effect" is a pretty powerful one. Does it mention the details of this as to what the effect is called? No, it doesn't.

**Anand**: [38:05] But let's take a look at what ChatGPT created. It's usually not as—how should I put it?—visually suave as Gemini. And yeah, it literally failed. Fine, big deal. I asked multiple applications to build something. So that's the thing.

**Anand**: [38:22] Oh, this was Gemini by the way. Did we ask Claude also? Yes. Okay, and Claude also has an answer. "Begin the journey." Okay. And it's zooming in. Hmm. This is nice. Beautiful typography and all that. Let's go to the next. And... I need to click on "Next." That's the next place.

**Anand**: [38:46] Yeah. Oh, _acha_ [I see], okay. The nice thing that it's doing is all of these little flows showing roughly from _where_ Einstein moved. It's almost like showing the journey like a bunch of sparks that are flying in. And okay, this is called **"Cognitive Field Lines."** "Animated particle-based idea flow between locations, showing how Einstein's thinking evolved spatially."

**Anand**: [39:13] Now here's the thing. I asked it to code, but I also asked it to do something interesting. And now I've learned something new: that there is a "cognitive field line." And I'm absolutely going to use these cognitive field lines going forward because not only do I know what it looks like, I know what it's called as well, and that's powerful. And the alignment is perfect. So because it's a physics field visualization, so the concepts are kind of like electromagnetic waves. So yeah, there's a theme fit as well.

**Anand**: [39:48] And let me share this. So, and the good part is all it takes to share is... oh, okay, I have to download it and publish it somewhere. Fine, then I'm not going to share it right now. But there are ways you can create artifacts which you can just publish directly, and there are a whole bunch of sites that you can use for that.

**Anand**: [40:05] Diya has a comment: "An extension that helps with doom scrolling? Something that has an MVP reminding you of your screen time every time you pass the assignment time." Can we try that? Please try that, Diya! Just do that now. And in fact, why don't you share your screen? I'll stop sharing. I'll need to promote you to host or something? Promote to panelist. Okay. You are a panelist. Go ahead, unmute yourself, share your screen, let's do that.

**Diya**: [40:37] Hi, I'm actually on my phone so screen wouldn't make sense.

**Anand**: [40:40] Got you. Then well, once you get to your system, try exactly that. You'd be surprised how some of these will, at the very least, teach you new things, but also sometimes give you a sense of... get your job done.

**Anand**: [40:57] We had asked for an online course on human behavior. And okay, it seems to have created a course where the Module 1 is "Bias Lab" and I have 200 XP that I can earn from this. Okay, which is the gamification thing. Okay, so here's some content on bias related to Kahneman's work and so on. Okay, what is the main reason cognitive biases exist? Laziness, evolutionary shortcuts, poor education, random neurological events. Tempted to say D, but it's probably B. And I think that's pretty guessable. And gave me 30 XP. Okay, this is fun. Next lesson.

**Anand**: [41:40] And this is great because if I can get to the answer without reading it, that probably means that I've understood the concept or enough of it. So I'm just going to jump straight to the questions. "The manager believes their new strategy is working. What shows confirmation bias?" Okay, if he reviews both positive and negative, that's not confirmation bias. "Seeking only the metrics that show improvement." Yeah, that's easy enough. That has gotten me 40 XP more. And the very fact that I'm sitting and solving these problems in the middle of a vibe coding workshop indicates that I'm hooked. I want to get more XP and the gamification is obviously working. So obviously that I'll just try one more please.

**Anand**: [42:16] "High profile shark attack. People start avoiding beaches. This illustrates..." Availability heuristic, I think. Yes, got that right. Yay, I'm a speed learner. Awesome.

**Anand**: [42:27] Sycophancy is a problem, by the way. These models will happily praise you to death, and you'll have to explicitly ask for critique. But I have at least learned that I can pass the first module of a basic course on this. And again, you can just save this or... okay, I can publish this as an artifact. So let me do that. I'm going to publish the earlier one also, if I can, as an artifact. And I'm going to put that in the chat window. Both of these. Copy this link and... okay, this is only going to host and panelists. Everyone. Yeah. That's one link. And the second link as well. Which all of you should be able to access directly.

**Anand**: [43:14] Now, the point of this is not that the output is useful so you should access it. The point is for you to get a feel for the kind of stuff that coding agents are producing these days. And taking that as a starting point, there'll be an option in all of these where when you click on the customize button in the artifact, it will take you to a new chat or whatever where from _where_ it is, you can change it to something else. And you can say, "Okay, now instead of teaching me human behavior, teach me animal behavior," for all I know. Or "Teach me about the evolutionary history of cockroaches." Pick your topic and you have a new field.

**Anand**: [44:06] We... yeah, okay, these are examples of other things that we can code. Let's go to the price inflation tracker. I am curious. Oh, okay, it's created some fancy dashboard and all. It's saying that there are 30 items it found that I had repeated purchase of, I guess. 25 of these have gone up and 5 of them have gone down. Okay, interesting.

**Anand**: [44:27] "Job leads GMBH." I have no idea what this is. "Youfit"... oh, okay. That's my wife's fitness subscription which probably has moved from a monthly to an annual one. So technically it doesn't count. "Subway." Okay, that's probably me ordering for a group versus having an individual one. So this doesn't count either. Again, yeah. So in all of these cases—or at the majority of those that I see at a glance—it's just confused the units. What's the cost for one item versus the cost for multiple items? But even if it's identified _one_ that is the same, then that has given me value for money.

**Anand**: [45:12] Okay, that's for two people. 57, 56%... oh, this may be a real one. And yeah, this is probably the only real one that I can think of. The 1000 Delight price may have actually gone up from 180 to 280. But no, that's too short a period.

**Anand**: [45:37] So in short, this didn't work the way I expected because the data that I fed it is wrong. But there are two things that you should note.
**Anand**: [45:49] The first is that I gave it data, and **the fact that it used code to do the analysis as well is probably one of the most powerful things.** It's taken literally all of the data and started doing the analysis live implicitly. It's not hallucinating most of this. That is powerful.
**Anand**: [46:11] The second is the presentation. It started putting together a dashboard to present the information. That made it easy for me to review. That's another example of the kind of thing that you should learn to ask for. And how do you learn? By asking it to do crazy things. Like I discovered these cognitive field lines by asking it to do something crazy, and now I've built a vocabulary for it. That is powerful.

**Anand**: [46:35] Now, this **building the vocabulary is arguably one of the more useful things in the AI era.** Most skills are declining, like coding, because the system can do it. But _knowing what to ask for_ is a bit on the increase. So I generally build my own vocabulary. An example of that is art styles.

**Anand**: [46:56] So can I... and I had vibe coded this. When creating an image, are there different styles I can ask for? 2D animation is a style, 3D animation is a style, abstract representations is a style, anime is a style, balloon art is a style, diorama. Diorama is something that I had discovered _after_ I saw this. And this is nice. Or let's take something a little more obscure, like pure art styles. If I took a person in a cafe and asked it to draw it in different visual styles. One of the really nice styles that I discovered was this "Xerox photocopy aesthetic." That is, drawing as if an image has been photocopied, then photocopied, then photocopied. And it gives a nice gritty style.

**Anand**: [47:49] By knowing styles by name and also what the output looks like therefore, I can just copy this, put it into Nano Banana and say "Give me an assignment output, but it should have this distinctive style," not like the usual AI-generated stuff that everyone will say, "Oh, this is generated by AI, what's the big deal?" Even in that, this is kind of how we start bringing in a certain degree of taste.

**Anand**: [48:07] In short, where I'm going with this is: **vibe coding is the easy part. Knowing _what_ to vibe code, _how_ to vibe code, and the process of how do you put stuff in that will work—don't put in bank statements when the units are not there, is an example of something that I've learned what kind of data it needs as an input. And how do you make sure that the output is easy to review?** If you ask for it in a dashboard, that makes it easy for you to review. So it's mostly just knowing what to ask for again. That comes in handy.

**Anand**: [48:47] But the most important thing is really just practicing. If you practice, for no reason other than to build a skill, you would have learned how you can do certain things that others may not yet have picked up, and at least for some time you will have an advantage.

**Anand**: [49:01] My invitation to you therefore is: **Practice more vibe coding.** It's easy enough, and therefore everyone will find it easy. What is it that _you_ can push it to do that is probably a little tougher than someone else may know? And for this, leverage what is unique about you. Just introspect, or talk to ChatGPT or one of these, feed it what you know, ask it to interview you to find out what is unique about you. And leverage that in the process of scaling through software.

**Anand**: [49:36] All the best with that. And I'm happy to take any questions. You're welcome to put them in the chat or in the Q&A. I see one question... two questions that I missed.

**Anand**: [49:46] "Is it safe to upload bank statements to LLMs? Where should we be careful?"

**Anand**: [49:54] It is _not safe_ to upload bank statements to an LLM if you are worried about companies like OpenAI or Anthropic having access to your bank statements. My bank statements are already saved on Dropbox, are already saved on Google Drive, are already saved on Microsoft's OneDrive. They are... I receive my bank statements through email in any case, so they're already with Google. So at some level I have decided that I'm going to trust some companies. So in this particular case, it's simply a question of which companies am I trusting. And I've added Anthropic and OpenAI to that list recently. You decide whether you're comfortable. If not, if you're comfortable with Google, stick with Gemini or Microsoft's Copilot. And if you're comfortable with none of them, then use open source models and run the code locally. That works too.

**Anand**: [50:47] "How secure is the data that you put into these websites though? Is there any way to make it more secure?"

**Anand**: [50:52] It depends on _who_ you are trying to protect it from. If you're trying to protect it from people who might access it _on the way_ to the servers, then go for people who have... companies that have a strong level of security, go for stronger, more secure browsers. But these days, neither of these is a problem. Most browsers are pretty secure, most sites are pretty secure, if you've heard of it and it's a popular one. If you're saying, "But hold on, I don't want to put the data to the website because I'm not sure if the website itself is secure," then **there is no way of making it more secure.** You have no way of figuring out what is the level of security behind a website. You can only rely on reputation, audits, and things like that. So you can do your research and find out how secure you _think_ it is, but you can't make it _more_ secure. That's entirely their choice.

**Anand**: [51:44] Any other questions? "Can we get a recording of this session as I joined late?" Farheen? The session is being recorded, but I guess Ashish maybe you can comment on the sharing.

**Anand**: [51:57] Okay, Ashish is not on. But yeah, please reach out to Ashish and he might be able to provide you a copy.

**Anand**: [52:08] I know we're past the hour. Got you, Farheen. If there are any other questions, I'm happy to share my email... or my website will have all my contact details. So you are welcome to reach out. I'll put that in here. And all the best for your exams. Take care. Have fun vibe coding. Bye everyone.