# Transcript

**Anand**: [00:12] Carlton said he's not able to join. JK, we have you, Jaideep. So, yeah, shall we get started, JK?

**JK**: [00:23] Yeah, yeah. I think we can start.

**Anand**: [00:26] Cool. So, I want to begin with the weekend experiment.

**JK**: [00:31] Hmm.

**Anand**: [00:32] Firstly, how familiar are you with the food court question that we had?

**JK**: [00:37] I am fairly I am fairly familiar with the flow and what is required. I think we discussed it last week also. I mean what has needed to be done, so I understand, and I had gone through the chat transcripts in the group, so I know it has been a fairly [laughter] contentious assignment.

**Anand**: [00:58] The food court, I thought, was far less so. I, over the weekend, ran a ran a training for a data science course, experienced professionals, gave them exactly the same assignment.

**JK**: [01:15] Hmm.

**Anand**: [01:16] Part of what I was curious about was would they cooperate more or defect more? And would they score more or score less? But what's your guess?

**JK**: [01:29] See, I think what is going to happen, the data science professionals are not going to cooperate more. I mean, I think the level of cooperation that TDS students achieve over six to seven weeks will be much, much more than and there See, there is a higher carrot over here, right? Marks. So, I am expecting students to be much more cooperative. Strategy-wise, I believe data science professionals may have a better strategy, and the convincing power, I'm hoping they will have better than our students, at least because that's they are from industry. So, I'm hoping they have that convincing power, at least to management they should have, right?

**Anand**: [02:07] Let's test all three hypotheses. I'm kind of sharing my screen. I have no idea what or how much is visible. Hold on. I think I need to Yeah. There's a blur. And we may need to do something to unframe it. No, I need to pin it for everyone.

**JK**: [02:30] Yeah.

**Anand**: [02:31] So, the whole thing becomes visible. Okay. And obviously, I'm experimenting with a whole bunch of things, so bear with me. So, what... Okay. It's probably not fully visible, right? On the right side, it's

**JK**: [02:50] Yeah, the the right side is getting trimmed out, yeah.

**Anand**: [02:52] Or in your case, left.

Okay, let's see if changing the aspect ratio does not help. Let me stop sharing and

share again,

and try our luck. If not, I will

Oh, okay.

Oh, okay, okay. I think I know the problem.

I think that is because it does this reframing thing based on where I am.

And sorry, I'm going to have to

**JK**: [03:44] be invisible for 5 minutes. [laughter]

**Anand**: [03:47] I will have to stall this. Let me try rejoining for a minute.

**JK**: [03:52] Yeah, yeah.

Hey. Hi, Prasanna. Hi, Mayank.

Hey, hi. Hi, students.

**Prasanna**: [04:05] Hi. Hi, JK.

**Mayank**: [04:06] Hello.

**JK**: [04:08] Yeah. So, for uh from While Anand is joining back, uh people who attended the session, how was it, I mean, the trading?

**Anand**: [04:29] I'm waiting for others to answer in case anyone answers now.

**JK**: [04:34] Yeah. So, would be good to know from students what was your experience, I mean, I believe some of you attended multiple sessions, so what was it uh in general? What was your learning experience from the session, the food corner uh food court uh the discount mapping exercise? Not discount mapping, discounting across a month.

Pardon me and Anand because we are munching something on the sides. We are supposed to have something at this point of time even though we are in two different time zones.

**Anand**: [05:15] Failing which I will not have anything for next 4 hours.

**JK**: [05:18] Yeah. [laughter]

**Anand**: [05:20] Anyway,

the coming to this one,

first column, on the chart on the left, is what the professionals scored.

And first round, second round, third round, our students are clearly scoring higher. Well, these are big differences, 88 versus 92.5, 95, and 94.

**JK**: [05:43] Hmm.

**Anand**: [05:44] And you could say that even when you take them playing the first time, there is a solid 4% point difference on an overall range that isn't particularly high. Percentage of 20% discounts,

yeah,

basically half of them put in a 20% discount. And it was reasonably clear that that completely kills the population.

So, okay. Your guess is okay. A, they will cooperate less, you said. Correct. B, they will probably score uh uh no you, aha. No, you your theory was they may score more because they may have better strategies.

**JK**: [06:23] Hmm.

**Anand**: [06:24] Uh they were even more clueless

than the the students. They said, "Oh, but, you know, we don't know how to run

a food court." Well, it's not as if students know how to run any business, let alone a food court. Some of them do, but majority don't. And it was a bit of an ego, well, not ego, a surprise for them.

**JK**: [06:48] [laughter]

**Anand**: [06:48] But your third guess was interesting.

Post-negotiation,

score dropped

**JK**: [06:55] Hmm.

**Anand**: [06:55] from, on average, 92

to 86.

**JK**: [07:01] Wow.

**Anand**: [07:02] So, of course, we probably need to overlay whether it dropped for students also,

but on the margin, it would have been better for them to stay silent rather than speak,

**JK**: [07:14] Yeah.

**Anand**: [07:15] which I thought was an interesting thing.

**JK**: [07:19] Yeah, but I I thought people will have better I mean, at least working professionals will have a better convincing power, and they will have a better understanding of the situations a little bit better.

**Anand**: [07:33] The conversations were not particularly rich.

**JK**: [07:40] Okay.

**Anand**: [07:41] I Maybe they are better at speaking than typing. I'm not sure.

It's a possibility.

At least, it's kind of asynchronous communication that was a.

That apart, the GA

three results, I think, are now fully available including the bonus marks. I'll share my screen again.

And

yeah.

Right. This is

the distribution for

Question 6. Let me just open the questions

as well.

scoring test pad.

Yeah. Question 6 was the 3D model of the face.

And

that's the bonus mark distribution. Half to seven 7.5 were the bonus marks. Nothing

unusual

about this distribution.

We've shared that.

The next question was about morphing faces. Again,

**JK**: [09:05] Hmm.

**Anand**: [09:05] similar, maybe a bit more right-skewed.

**JK**: [09:08] Right-skewed and yeah.

**Anand**: [09:10] So, which means that the

there were a reasonable number of good marks.

**JK**: [09:19] What were the N for each of these, Anand? I mean, that's also an interesting thing to look at.

**Anand**: [09:23] Uh-huh.

**JK**: [09:24] Was the N consistent?

**Anand**: [09:27] No.

Slightly different, but in the 600 to 700 range.

**JK**: [09:37] Okay. Okay, so, I mean, there's a consistent participation in all both the questions.

**Anand**: [09:44] Yeah. Though I mean Q8 seems to have had a lot higher. And question 8 was just uh can you figure out which ones are easier to morph and which ones are harder to morph? So, effectively, just an opinion opinion guess the audience opinion kind of a question than uh do something kind of a question. So, understandably, participation was higher.

**JK**: [10:09] Yeah.

**Anand**: [10:11] And on that question, it was uh quite heavily right-skewed.

**JK**: [10:16] Heavily right-skewed, yeah.

**Anand**: [10:18] Yeah. Again, understandably so.

So, those results would uh get I've shared the bonuses. We should be uh

getting them. So, just FYI, the mechanism for for this was uh Bradley-Terry. That is,

I'll give you

a

good one, an example.

It is somewhere below, I guess.

**Anand**: [10:53] Yep, here we are. Right.

**JK**: [10:54] Hmm.

**Anand**: [10:55] So, if we just directly took the score

versus if we took a reliability versus a percentile at the start with the HOS Bradley-Terry, the idea being that, suppose, and let's take a real situation.

For question 6,

the student got 13 wins out of 26. That is 13 people thought the result was better, 13 thought the result was worse.

Same for row number six four 548, for two different rows. But

in the latter case, the 13 that they won

were tougher

**JK**: [11:37] Hmm.

**Anand**: [11:37] opposition.

**JK**: [11:39] Okay.

**Anand**: [11:40] And the 13 that they won was weaker opposition. So, if you score higher against something that has been

rated higher,

**JK**: [11:51] Hmm.

**Anand**: [11:51] sort of like link propagation or Google's PageRank kind of a thing, is what Bradley-Terry seems to apply, which was then scaled

from 0 to 7.5 on

**JK**: [12:03] Hmm.

**Anand**: [12:03] whatever the scores are.

That, I guess, I wanted to explain. And the Yeah. Hey, so,

the

thing that I was curious about was, for these, which of these tend to have a right skew?

Firstly, is it close to normal, and how skewed it is?

And this one, the two where people had to create things,

whether it's morphing or a mesh,

**JK**: [12:35] Hmm.

**Anand**: [12:37] the effort clearly seems

to be required is something that adds up to some kind of a measurable skill that is easy to

**JK**: [12:50] Got it, got it. So So, uh correct me if I'm wrong. I'm I'm actually looking at an actual scenario. So, uh uh in a real-life setting, I mean, this uh especially with this uh digital twins and other kind of things that are coming up, uh this essentially means the effort that is required I mean, if I am a media company or a gaming company kind of organization, it essentially means that this kind of twinning uh effort is going to be significantly larger with AI, right? Uh I mean, can we do uh some kind of extrapolation from I mean, I know it's a weird extrapolation, but uh essentially, what I'm thinking of is imagine that I have a uh a media company that I'm So, I'll give you a perfect ex- example. So, how many of I mean how many of us thought that Jensen Huang coming on Office was a, I mean, it di- it was like a an amazing thing, right? As in as a content creator, and that's what significant fraction of people are these days. So, as content creators, does it mean that the effort required to put in uh that additional thing on media and other artifacts are going to be still harder, or is there a is there a business uh opportunity in that segment?

**Anand**: [14:17] What I'm seeing, the content creators are strongly getting into the act.

**JK**: [14:23] Hmm.

**Anand**: [14:24] The formats are expanding.

What's not clear is, is there more demand

because the formats are changing, and

well, there is more budget to feed more people,

or people are becoming more efficient and, therefore, a smaller group of people will be able to cater to a larger audience? The verdict on that is out.

**JK**: [14:47] Hmm.

**Anand**: [14:48] But the shift seems to be that a reasonable number of people who are doing the usual kind of content creation earlier,

they're uh losing out. One of our clients,

um

how can I say it?

Huh. Petroleum company,

and uh they said, "We pay our uh graphics agency a huge amount."

**JK**: [15:12] Hmm.

**Anand**: [15:13] For all petrol stations, we put out an ad. Now, ultimately, how do we create the ad? We say, "You take the ad for Hong Kong, take their festival, just replace all of the same things, and now put it for Indonesia."

**JK**: [15:25] Oh.

**Anand**: [15:26] Now, for that, I will have to give them a brief. They will have to take all of those uh graphics created differently, etc. And that's it. Now, the cost for that versus if AI can do it, and my team can do it, why don't we use it?

So, their jobs are at risk.

And someone who can do this with AI, therefore, can cater to a larger audience. That's one part of the story.

The second part of the story is equally interesting.

So, we said, "Look, here is the proof of concept. Now, shall we empower your team to do it?" They said, "No."

"Why?"

"See, supposing there is a copyright claim, some copyright-related issue, or some cultural sensitivity,

if it is a graphic design agency, I can say,

'Here is the clause in the contract. They are to blame. Insurance is covered. Their insurance is covered. All of these are covered.'

My team does it, I don't have any of this infrastructure in place.

I need an external neck to catch.

But I can use this to show them and negotiate the price,"

**JK**: [16:35] Hmm.

**Anand**: [16:36] which makes No, no, it's a It's a valid, I mean So, it's the same thing, na? Any issue comes, you need somebody to point your fingers to. That's that's universal, I think. [laughter]

**Anand**: [16:47] Correct.

**JK**: [16:48] So uh okay, so I think uh one more thing is this um the GPT uh No. Uh Claude uh releasing Sable 5.2, uh Ethan Mollick had uh put out a uh blog with a video creation of Claude.

So, he had asked for uh I mean, so he had given a phrase, and he asked Sable to create a uh one-shot video based on that. Uh the video was I mean, it it was the parallel words's worlds video, and it was really nice. And there was another one which they created uh Okay, I forget. I mean, it was one he created parallel worlds, and another he gave a one Uh I think uh it was it was an I mean, you'll have to go and visit Mollick's page and actually see it. So, the point is all these models are becoming very powerful in multimedia, in general, and uh what is happening is this um general templating works. I mean, everyone will create a kind of thing. So, uh one of the argument, I think it was uh Simon's uh uh the I mean, Sinek, right? Sinek and somebody else was having a conversation, and they were saying is, "What differentiates humans in this is the aesthetic sense that we have." So, for example, I mean, exactly what you said. I mean, a Hong Kong cultural festival, the interpretation of a human would be highly contextual to their experience uh of it uh, whereas the one created by AI will be like based on a lot of rules, templates, and generic things, which will lack that uh finesse that a true artist can have, which, again, the point that you said is the same. I mean, the normal guys would all fade away, and the uh I mean, the people who have that niche niche craft, they are going to survive the AI uh Yeah, I mean, the AI boom or whatever. [laughter]

**Anand**: [19:01] I particularly experienced that with the Did Did we go through the picture stories results last week?

**JK**: [19:09] Yes. Uh the I mean you We compared this the two images and what were uh I mean, what were uh being I mean, which is the worse and which is That one only, right?

**Anand**: [19:20] Uh yeah. But did we go through the results as such?

**JK**: [19:24] No, no, no. We didn't go through the results. We just uh looked at a few of them.

**Anand**: [19:27] Okay, I'm if I'm Maybe we already uh saw this. In which case, I should just uh not uh bother. Let's see. Group 1.

**JK**: [19:42] Oh, we went through this during our um Wednesday meeting, not the Not Monday meeting, is it?

**JK**: [19:47] Yeah, not on the Monday meeting.

**Anand**: [19:48] Oh, okay. Okay, fine. I'll share the link here uh in case anyone wants to go, but these are the results.

And uh Swarnam has a comment: "AI-created content will have liability issues."

Well, not uh exactly. Uh what's interesting is most of the companies are willing to indemnify, that is Google says, "I will in- indemnify you

uh if somebody sues you saying this was based on their copyrighted content." They'll fight your case for you kind of thing or whatever, if you sign the enterprise clause.

So,

I wouldn't, in a blanket way, say that AI-created content has liability issues. But there is a

percentage of uh businesses for whom it can create a liability issue, because they may have contracted with their clients saying, "Oh, we will not use AI," or whatever. And uh it's a shrinking uh number because contracts are getting renegotiated because of facts.

In any case, uh let me take this one. I'll share the link.

Yeah, oh, I already shared it. Okay.

So, in terms of scores, the uh highest-scoring one was Bhagavad Gita, which I thought was pretty uh interesting.

And uh the second-highest- uh scoring story was this. Actually, I didn't fully understand um

Okay. Guy Rich businessman gives a bribe to a cop. Beggar gets uh shooed away, and is sad. And then politician gives people money, or people are giving money to the politician? Not sure.

**JK**: [21:48] No, I think uh I think it is the I mean, he gives more to the politician. I think that's getting

**Anand**: [21:53] Aha.

**JK**: [21:54] I mean, I was also looking at that trying to understand.

**Anand**: [21:57] Yeah.

It would be different.

Here is one where student narrated that he only gives to the politicians, whereas the poor doesn't, which he uses a satirical uh story. [laughter] I thought it was interesting that satire gets such a high score, or at least something that you interpret as a satire uh gets a high score, because uh a reasonable number of the Okay, skip some of these. Um

Now, let's actually look at the once the

patterns on this.

So, if I look at the uh so, on the right-side chart, let me make it bigger.

Oh, is it uh going outside uh? Yeah, okay. Is it No, it's probably just flooding out. Fine.

So, uh let me see. Yeah, okay. This might be a cleaner view. Let me see if I can

Aha. That is a good So, x-axis is uniqueness,

**JK**: [23:05] Hmm.

**Anand**: [23:05] and y-axis is whether people liked it or not.

So, we'll come to the outliers in a bit, but the least unique stories that people really liked,

**JK**: [23:19] Oh, wow. Instareels. [laughter]

**Anand**: [23:21] this is a total prototype. I mean, the sheer number of So, this is one.

To me, this looks like that it it is almost frame for frame the same story.

**JK**: [23:34] Yeah, yeah.

**Anand**: [23:34] Frame for frame, same story.

Uh

**JK**: [23:38] Hey, this is why mass cinemas work, Anand. I mean, [laughter]

**Anand**: [23:41] No, exactly, right? Exactly. Look at the preference. [laughter] I mean, this is so hard to beat. This is the other prototypical uh story. People totally love this.

So, yeah.

**JK**: [23:54] This is I mean, this is like Insta-influencer, I mean, motivational speaker kind of,

**Anand**: [24:00] It is stunning. It is stunning.

And, yeah, this is a second uh prototype. So,

yeah,

**JK**: [24:09] Hey, what would be interesting is if these people shared the prompts with each other. I mean, that's also another

**Anand**: [24:16] Uh possible, possible. In which case, it was a pretty good uh prompt on, at least, one criterion, which is preference. They've got the pulse of the audience, right? Uniqueness, of course, scores the lowest.

**JK**: [24:25] Hmm.

**Anand**: [24:26] And uh understanding is high, also. So, yeah, on two out of three counts, you score uh pretty well.

In contrast, so let's look at the other extreme. People uh It's really unique stories, but people do not like it. And there is a negative correlation here.

Gift given, but girl doesn't like that gift. Still says, "Thanks. I love it." Very uh Yeah, unique, I thought. Well, don't know, clearly it is unique. Another one,

---

**Anand**: [24:59] Another one, Armstrong, presumably, uses a mobile phone on the moon, and possibly hand drawn at that, which makes it even more interesting, but possibly AI-generated, also, carbon sketch, either way. Funny and interesting. This one was unique.

**JK**: [25:17] Oh, wow, it's

**Anand**: [25:20] Yeah, four bars for the networked world of AI, or however you want to put it. Again, very unique, but all three had surprisingly low preference rating. People liked it a lot less compared to the other. Clowns, this, it scored really high, and this, which again, scored really high.

This in the Bhagavad Gita, I can kind of understand, or rather, that it is high preference is reasonable, that it was so unique was Unique? Yeah. Not many people used this prompt is, I think, an indication of, okay, yeah, here's where the creativity of the prompting starts coming in. This looks like a proper original story, and it could be AI-generated, but the curation of it is certainly something that matters.

And, I was intrigued to see the other quadrant as well, which, yeah, is basically the The dog story? Yeah, the dog story, but now in a cartoon Correct. Not as well executed. Oh, people don't People don't love cartoons and yeah. Interestingly, right? Yeah. [laughter]

**JK**: [26:39] Oh, fine. Okay. They want CG and everything inside. I mean, actual humans, actual animals.

**Anand**: [26:48] Exactly, yeah. These were middle of the road. But you know, I should test that hypothesis, whether cartoons, or cartoonish things tend to score a little lower than realistic ones. But yeah, no, I, I personally thought of the various exercises, this is one that brought home the importance of partly, a certain degree of creativity to move to the Right side, yeah. right, as well as an understanding of the pulse of the audience, and to communicate with clarity. Both the understanding, as well as preference, which both of which are

**JK**: [27:27] No, so I think it is also I mean, **what we have to do is juxtapose this against with I mean, at least what is happening in Hollywood, or some of the major movie industries, as to how they source content, right?** I mean, look at Marvel franchise. I mean, it is I mean, they just came out with another endgame. [laughter] I thought Marvel was done. [laughter] So, yeah, I think See, one part of it is this I mean, why these the entire activities of the last I mean, the last set of assignments were interesting is because we were playing with multimedia a lot. And the part that I'm and I I mean, yeah, see, somebody who puts in a satire See, even in regular world, satire is so hard for people to get across the border. [laughter] So, I'm surprised that satire scored so high in the understanding and popularity. But yeah, I mean, it's also important I mean, it is a social satire, so where they empathize with with the boy.

I would want to see a reverse of it, where somebody is actually putting the satire from the businessman's perspective, and produces a different version of it where he loses out money, both to [laughter] politicians and cops, and he's trying to justify what happens. "I'm not getting anything out of it." So, something like that would be would be interesting to see.

**Anand**: [29:07] A few other observations, on the forecasting one, the prediction market. So, in case anyone forgot that question, it's: Can you predict AI Pipe's spend last week? Or might have been week before. The actual spend was $5.31. 620-odd students had gone with predictions. Initially, 123 chose the correct range. By the final prediction, that had almost doubled to 242, getting it in the right range.

It was interesting, one student changed their forecast, for instance, from $5 to $10, all the way up to $100 and $200.

**JK**: [29:56] Wow.

**Anand**: [29:57] And then started backing down when the spending was contradicting their assumptions. But, what I quite liked was students, which we stopped extrapolating trends and started using constraints. We've already spent $5.24. Okay. Uh-huh. Now, in order to cross $10 tomorrow, we need another $4.76. Is that even realistic or possible? It's a smart way of reasoning about it.

But the other interesting thing was students with better information, actually, sometimes got worse marks.

**JK**: [30:37] Oh.

**Anand**: [30:38] And that seems to be because early predictions were ending up scoring more, because they were holding the prediction while the price changed to the point where you could earn a larger profit, so to speak. So, effectively the marks were capturing both a certain degree of participation discipline, as well as forecasting ability, and some basic mistakes, also. Some students calculated $5 to $10 correctly in their reasoning, but they selected $0.05 to $0.10 in the interface.

**JK**: [31:21] Okay.

**Anand**: [31:22] Good reasoning, poor execution.

**JK**: [31:25] See, I think See, I I see this this particular thing repeating across multiple contexts. So, I see the same happening. So, in the numerical questions, I see people actually coming up with the same I mean, the actual numerical answer, but the difference in order will be around 10 or 100. It essentially says that, okay, they have applied all their models correctly, but the actual calculation, they made a mistake, because it is a pen-and-paper calculation that they do, and one place here and there, they will just go wrong by that many order. This is something that I has been I've constantly seen across courses, not just over here. And what is surprising is this was something where they could actually use an a calculator, right? I mean, it's not even They don't need to See, I can understand human mistakes, but this kind of modeling mistakes, when they put the input functions also of an order difference, I think that is that is worth looking at. I mean, how are they going off in terms of order is worth worth an inspection separately.

**Anand**: [32:32] My theory on this is that each of us have a certain kind of, what do you call it, blind spot, absent-mindedness, whatever. And some of them may be consequential, some of them may be less consequential. I cannot remember faces. For a big part of my life, it didn't make a difference. I cannot remember numbers. I swap digits, etc. It became less important with calculators. My short-term process memory is bad. That was important at an early stage, continues to be important, so my note-taking ability shot up like crazy, even at an early stage. Because, one thing I am comfortable about it is that when my memory starts fading, in my '80s, whatever, I'm at no significant disadvantage.

**JK**: [33:30] You have the notes, yeah.

**Anand**: [33:32] My note-taking ability anyway is I have the memory only of someone with that, not Alzheimer's, but pretty poor. [laughter] So, I'm guessing that we develop ways of compensating for what matters if we pay attention to it and are conscious of it.

Now, like I said last time, here, in a course like this, I don't think we are as much teaching students by force feeding, as we are creating an environment for, or an opportunity for students to pick and choose what they learn. There are 100 things one can learn from any arbitrary exam or course. Majority of them learn a set of things. Minority learn a different set of things, and it is up to the experience of each student. Because somebody who's already good at A, B, and C, there's no point giving them A, B, and C again.

**JK**: [34:27] Yeah.

**Anand**: [34:27] They'll get E, F, and G. So, which is why the menu of learning exists. So, from that perspective, what you said about what I'd call a silly mistake, $5, I replaced it with $0.05,

**JK**: [34:44] $0.05, yeah.

**Anand**: [34:46] maybe it matters, maybe it doesn't. B, it's also a function of the student being able to recognize it, figuring out whether it is important or not, and choosing to correct it. The last requires discipline. The second requires either foresight or knowing roughly what direction they want to go in. The first we are providing by saying you're scoring lower here, or higher here, so at least you'll recognize it. But, I'm open to, is there are ways of, how do we improve this structurally? At the very least, we can provide more realistic trick questions so people know where they will stumble, find their blind spots, and decide whether to improve or not. Something like this, you tell an agent to do it, it is less likely to make this kind of a mistake, and that may be part of the practice.

**JK**: [35:46] No, so I think, see, one aspect of it is I I think one analysis question like this in ROE, where you take one of the existing assignments that they have done, so they have a familiarity with the context. It could be picture story, for example. So, we give some hypothesis, some intuition, or I mean, would be good to create a question around the results of one of the earlier assignments. Can they do some analysis, and do something with the insights that are coming out from there? Because, I think, what is the differentiator? I mean, so, for example, for me, the differentiator would be, if I were I mean, if I were doing a market research on what is going around in the I mean, in terms of let's say, K-pop or K-dramas, so to speak. So, what should work in India? So, is is the is the I mean, K-dramas being consistently The consumption is large. So, does does a variation of that actually makes sense? Would be something that is interesting for that type of businesses, and generally, I mean, if you look at it, it's also another part of it. A lot of businesses in Kerala, so for example, most common businesses that you see along the road is restaurants, car washes, and possibly this citizen digital I mean, this citizen services, so to speak. So, you have you become an agent for several of the citizen services, and there is a one place where printout and all those things. I mean, these are like mushrooming kind of things, and after few days, these shops close by itself. So, a lot of expats are coming back and without a knowledge of the market I mean, the area and others, they are trying to get into these kind of businesses without doing that what is what is the unique selling point, right? For that, I mean, without doing that, they are trying to do it. So, if we can tweak it in some way where the I mean, to take borrow the other idea of understanding versus creativity, if they can push it to the upper quadrant, would be nice. I mean, if there is an assignment or a a question, a short-time-frame question, which actually looks at this, that would be I mean, I'm not sure. I mean, how much will you do in a 45-minute time period? I'm not sure, but still, this is an assignment that they are already familiar with. This analysis also we are putting it in advance, so what comes up in that would be a good thing to see. Just what I feel.

**Anand**: [38:23] It's interesting. What you mentioned, almost perfectly mirrors questions eight and nine in the project.

**JK**: [38:29] Mhm.

**Anand**: [38:29] So, earlier, they had shared facial expressions. Now, train a classifier for those facial expressions.

**JK**: [38:38] Correct.

**Anand**: [38:38] Same data set. But, with the constraint that you have to do all your triers and all of that, but you can do Oh, actually, okay, no. This, I think, is just submitting it. But, there were a few, where you could do this in the set of sessions, time sessions? Yeah.

**JK**: [39:00] Yeah, yeah, yeah. They're They're there.

**Anand**: [39:02] Yeah, these are the timed sessions. And, this is for the invisible ink, that is motion ink.

**JK**: [39:10] Hey, somebody has created a solver for this, right?

**Anand**: [39:13] For this one, if I'm not mistaken, the solver was not as effective as a viewer. There was an artifact that somebody had created, and it was very both surprised and impressed. Let me see if I have the artifact. It was a good one.

**JK**: [39:37] Yeah, yeah. I also saw that. I mean, I I think Yeah, the meta thing that we were discussing last week.

**Anand**: [39:44] So, for some reason, I'm not able to click the window. [whispering] Wait a second. I know it's not visible on my screen. This is such a pity. I was hoping to get to that artifact. Okay, I give up. It's not appearing easily enough for me. But, the accuracy of the solver was partly low because there wasn't enough data, but I think mostly low because this is actually a tough problem, and for an agent to solve. But interestingly, you can look at it and figure out, huh, okay, this is probably this letter or that letter. And I don't know why a model is not able to do that with its with its rather stunning vision capabilities in many areas.

**JK**: [41:08] Yeah. I mean, I saw

**Adwitiya**: [41:10] Sir, Fable performed quite well on this problem.

**Anand**: [41:14] Did it? Okay.

**Adwitiya**: [41:15] Yeah, I I created a solver. I Anthropic launched 5.2, right, Fable.

**Anand**: [41:20] Huh.

**Adwitiya**: [41:21] So, there was an access, a limited time promotional offer. So, I utilized that for this question. It it performed quite Like, of course, I can't say that with 100% certainty, because I don't have the answers.

**Anand**: [41:31] Okay.

**Adwitiya**: [41:32] But yeah, as far as I could see the solver I created, it was coinciding quite closely with what I was interpreting.

**Anand**: [41:39] Got it. Interesting. Now, and that's the nature of some of these. It's one thing to say, okay, choice of model and choice of prompt, and share that. It's another thing to build a deterministic solver that you can share the link to, and others just upload their file, and it works for them. The reason I'm saying that, is we're trying to see if each person has to do something, or

**Adwitiya**: [42:09] Sir, that's what I meant, actually. I It was a solver. It was independent of the data that I provided it with.

**Anand**: [42:17] Got you. However, the person who's using the solver would need Fable 5.2, right?

**Adwitiya**: [42:23] No, no, no, sir. It was a stand-alone model, just an HTML page encoded with the weights of the model, and no AI model was required for that.

**Anand**: [42:33] Interesting. Were you able to share it in the Google group, or Google, what do you call it?

**Adwitiya**: [42:39] Sir, are you a part of that, the TDS G-Space?

**Anand**: [42:43] I might be. I'll have to I think I am, yes.

**Adwitiya**: [42:47] Sir, so if you can just check the pinned messages. There were two of us. I mean, I was there and Shrikar was there, so two of us created solvers for the specific question.

**Anand**: [42:57] Let me just take a quick look. Okay, yeah. I can see the message, one Okay. I'm not sure on how to see something pinned, but Or if you can just paste on the chat here, the link.

**Adwitiya**: [43:22] I'll just paste the link. Just just once.

**Anand**: [43:42] I would really love to see how it does on some of my datasets.

**Adwitiya**: [43:54] Sir, can you access this?

**Anand**: [43:55] Yeah, I am opening the chat.

**Adwitiya**: [43:59] And sir, there is another one, a student called Shrikar Sharma created it. I'll also paste the link for that.

**Anand**: [44:06] Thank you. Okay, downloading the solver, and opening it. Give me a second. Okay, I will now share my screen. Great. Right, okay, now, let me pick a file. I've got my JSONL, and that seems to be solving it, and that Okay. So, it is three letters, it could have four letters, it could be five letters. Okay. I'm just somewhat doubtful that this is a V on the right side. But is it still solving?

**Adwitiya**: [45:08] Yes, sir. It's still solving.

**Anand**: [45:10] Okay, okay. Let's see if it

**Adwitiya**: [45:13] I mean, I I actually eliminated that dependency for Python on the end-user's machine, so it will take a few seconds more.

**Anand**: [45:20] Sure. Which is totally fine. But this one Okay, this stroke seems Is there a way we can Okay. Well, we'll have to see how well it actually is doing against the results. Based on a visual inspection, I think, for instance, out here, the F seems to be reasonable, but this E is a bit of a stretch.

**JK**: [46:00] Yes, yeah, exactly.

**Anand**: [46:01] V might be a stretch as well. H is definitely a stretch. Y is possible. This, your guess is as good as mine.

**JK**: [46:10] Yeah, exactly. I mean, there are a lot of things on to the left. Okay, P, C, D.

**Anand**: [46:15] Yeah, this

**JK**: [46:16] E, I'm not sure.

**Adwitiya**: [46:20] Sir, I've also pasted the link for the GitHub repo of the of Shrikar's solver, if you wish to check that, as well.

**Anand**: [46:27] Let me check. And Shrikar's repo, and this Okay. Uh-huh. That setup would be a Okay, we'll have to install. That's going to be more work than I'll do in this session, but nice, good. If both of these are deterministic, and both of these are, and any others are able to solve this in a way that does not require tokens, that's saying something. That's saying something. And this, again, is part of what we want people to learn, that is, what does it take to use an intelligence like Fable 5.2 to create a deterministic and distributable solver. And obviously, therefore, now I need to go start looking for problems that are in the cusp of what is possible, not so hard that it's no one can do it, not so easy that everyone can do it, which is where we get the normal

**JK**: [47:28] What would be I mean, because it was shared in that group, another thing to look at is how many of the other students are scoring at at a higher rate, right? I mean, see, these are solvers that are deterministic solvers that have been put up. So, how many of the other students And this we have been repeatedly seeing. One thing is we get an idea of what what kind of solvers are there, and there is a lot of students who are in that group, so we I mean, the basic thing is how many of in the how many of them in the group are actually making use of it? Because that TDS 2026 is not a this-term group. It is, I mean, it is a historic group which has seen [laughter] a lot of people from across the terms joining in.

Would be nice to see how, yeah, how that is I mean, are we seeing a network effect? And this we have been talking about multiple times. I mean, are there groups? Are there cohesions that we can actually track? And if there is one way or the other, it would be nice to see that coming from the entire TDS exercise.

I think we should also give an opportunity for students to give a citation from where they solved, at least in ROE, or in the other, P2 at least.

**Anand**: [48:53] That's a good idea. One way in which that emerges is, if there is a single URL, people copy-paste the URL, that effectively becomes the citation, with probably the first submission being the source. We've seen that kind of network happening.

So, I'm guessing there are three kinds of learnings here that emerge. One, I can't solve it, I don't know how to solve it, and I don't know how to find people.

**JK**: [49:23] Mhm.

**Anand**: [49:24] Solving it, the agent can do. Finding the people is what you need to learn. So, that is the find a community to be a part of, form a community to be a part of kind of a learning. Second, I have found a community, I don't know how to solve it, somebody solves it for me, I take it and use it. Do I still learn something?

**JK**: [49:47] Mhm.

**Anand**: [49:48] And the answer is absolutely. That it is solvable, it is solvable in a certain way, and it doesn't matter whether it's tomorrow, next year, or whenever. You can always just come back, point to a solution, saying it is done.

---

**Anand**: [00:00] **90% of what we really need is an exposure to, "Here is a kind of problem, it is solvable this way, rest the agent can take care."**

[50:10] The third kind of learning is for the person who creates the solver. Not everyone will create every solver. There'll probably be a few people that'll create a large number of these solvers, and for them, that'll probably be the path that they want to take, being the person who creates these sorts of products, potentially, however it works out. And that's a different kind of learning, and not only not applicable for everyone, it's probably a bad idea for everyone. **The vast majority of people need to learn how they can apply what people have created, not how they can themselves create.**

[50:48] Hopefully, which is kind of what I guess we were talking about, when you said that there are different kinds of learnings for different people at different points.

**JK**: [50:58] Yeah. And I think, I mean yeah, since we are almost coming to the end of the session, I think one of the bigger thing is also about what we are planning for ROE. I mean, you will see a very, what Anand has promised is, you will see a very different ROE. [laughter] And I'm hoping he has started working on it. [laughter]

**Anand**: [51:15] Yeah, I have a collection of questions. One of the themes I'm thinking of is either a text adventure game, or these days, it doesn't even have to be a text adventure, a real game, multi, multiple people going around, bumping into each other, negotiating, buying, selling, solving problems sometimes together, sometimes disparately.

[51:37] But because this is a game, it's not as if there's a fixed set or a fixed sequence of questions. This is something that people should know, and yeah, TAs, please let people know that: A, there is no order of questions. Earlier there was no, what should I say, identification. You could say question number one, question number two, etc. Now it's, "Okay, I'm in this room where there is a chest or whatever, and I opened the chest, and I found the problem." Except the same problem might appear for somebody else inside a knife.

[52:12] So, that is one complication, and it doesn't necessarily appear at the same time. I'm in one room, I see one problem. Somebody else is in a different room, they see a different problem, and how we wander around is entirely up to us. So, which means that coordination ends up being a little trickier, and people might want to plan for that because anyway is real-time. But what this means is that **if you put your agents at the game, you have to be very careful about figuring out what prioritization you want to give your agents.** Can it really help you score high marks?

[52:59] Now, unfortunately, you will not easily be able to run multiple agents in parallel because you have only one character. One person that goes into one room at a time, fight and try to control it, and let's go back and forth. So, at any point, you can be in only one place, and is that where you want to be? Do you want to disregard a particular problem, skip, move on to another because that might have higher probability, higher score, or do you want to persist on this? Have you been able to find somebody who has solved this particular problem, therefore there is a solver, can your agent pick it up while you are hopefully sitting dazed in the middle of the exam, trying to steer the agent?

[53:43] That sort of a skill will start Adwitiya, your hand is raised.

**Adwitiya**: [53:49] Yeah, sir. Sir, so this was just a suggestion. Since we are, you know, experimenting with the games and you are also interested in the data side of it, so right now, we have we are a big thing on anonymity, at least while we are interacting in the session, right? So, but it would it will be interesting to see how the students or the players respond if we, you know, switch that up with accountability.

[54:14] What I mean is if we do show your real name, at least it would because there are always two sorts of motivations if you look at the psychology of a person. One is personal gain or personal benefit, and the other one is how you are seen in your peer group. Okay. So, right now, one thing that I felt, at least in my sessions, was that there were a few students who weren't giving it their best, and that was maybe because they had things in mind like, "I have one more session. Why to waste my attempt here if my score has already gone down?" And that ended up hurting the entire team. So, if there had been an element of accountability, then perhaps they might have, you know, participated a bit.

**Anand**: [54:58] Completely agree with this could be an interesting experiment and possibly a learning. So, I have I did consider this, and just FYI, for the is for the food court game and similar games, the part of the learning was the other, which is: **can you make sure that that fourth person, even though anonymous, is someone you can bring back into the game? What does it take to be able to do that?**

[55:31] First time, okay. We don't really know everyone. But can we, within the span of a day, get the infrastructure required so that we are able to pull another person back in? And I wasn't sure if 1% will be able to do that or 10% will be able to do that, etc. And it seems closer to less than 10%. So, this is the kind of thing that over time, what I imagine a community like this will do is say, "Of course, everybody has to be on one register. Whoever is taking TDS, get them all to register."

[56:11] "Now, whoever is going to join this kind of a game, before that, start a game and then drop your you know, forcefully tell people to save what is your handle, put it in there." So now you can start tracing. It's possible for a group to be able to enforce that. A long shot, but the steps that one takes towards that is one direction.

[56:35] Another direction is four people get together and say, "Look, we will play the game at a really unusual time. Probably not many other people are going in. There's a chance that we might form the group." In this particular case, that dynamic wasn't possible. But there'll be other games where you would, just through synchronicity, be able to pick.

[57:02] So, completely agreed, Adwitiya, and just saying that: A, will anonymity reduce accountability is an experiment we might already have data for, and I should think about exploring that. Second, we may be able to conduct more such. From a learning perspective, for the food court, it was a different kind of learning, which I just thought I'd take the opportunity to share.

**Adwitiya**: [57:27] Understood, sir. Thank you.

**Anand**: [57:30] Cool. I think there are a few comments. Okay.

**JK**: [57:35] So, Preeti was mentioning about how they came into a GMeet and discussed instructions and others.

**Anand**: [57:41] Yeah. Exactly the kind of thing that is required. So, **a good ROE will require a higher level of collaboration**, and that's what is happening.

**JK**: [57:53] And an intelligent It's not a I mean, intelligent higher order, not just in a cognitive level, even in a collaboration level, right?

**Anand**: [58:01] Yeah. And at trade-offs as well. I'm thinking we discussed this "call a friend" kind of an equivalent, right? So, I'm thinking, in some cases, you have to choose between, let's say, do you want five minutes extra versus do you want one mark? Do you want to be able to swap questions? Options like that from a game dynamic perspective where I'm not sure if a human will choose better or an agent will choose better, or what combination will be required.

[58:27] But I suspect that some guidance from a human will be required. At least, I think this kind of a thing we will be able to solve better, and especially based on the WhatsApp chats or Google Chats, etc., to get a feel for, "Okay, I need to steer my agent this way."

[58:45] And if somebody can rig this up that the agent can monitor these conversations and use that to feed that in, that will take the system possibly to a slightly higher level if they, as a human, can keep up with the agent well enough to be able to steer it.

**JK**: [59:03] Yeah.

**Anand**: [59:04] We have fun for the 70 minutes.

**JK**: [59:06] Yeah. And a general I mean, a general input to all students, I mean, do share your AI sob stories as well. I mean, "I got so [laughter] so much I mean, I got into depression because of AI." These are stories that we would also want to hear because one part of it, while AI is making a lot of things easier, there is also the other side of it. So, how many of us are discovering it and what kind of stories are emerging, that is also good to hear.

[59:39] So for example, I have heard stories where freshers have said, "Okay, I can't do anything more because my tokens have been consumed and I'm not working today and going off." [laughter] Which, I think the manager was like, "Okay, [laughter] what am I seeing?"

[59:55] So, I mean so anything of these sorts if you And I think in ROE, I mean, generally, students who are using a limited-budget AI, they would feel these a little bit more. I mean, would be interesting to listen to all these because one part of it is also it is instructive to the course team to understand the different challenges that people are facing with AI. So, that would be, yeah, that would be really interesting.

[60:26] And there is one question from Preeti. Okay, she's trying to get more information about ROE. "Why negative marks?"

**Anand**: [60:32] "Why negative?" Oh, for that.

**JK**: [60:34] "Why negative?"

**Anand**: [60:36] What difference does it make whether it is +10, -1, or +9, 0, or +11, 0, whatever? It's largely a psychological thing. That, to one negative mark against a possibility of 10, then the ratios are similar. Yes. I was just curious.

[61:04] But no, that was partly it. Partly, the other thing was randomly people would say, "I'll just tell an agent to submit a question, a response," and I don't this one, I have to sit and read, and I wanted to make sure that I read good responses.

[61:18] Unfortunately, majority of the responses are rubbish. And that the agent had generated this isn't isn't what makes it rubbish. It's that the critique of the agent has been really poor. So far, I've received less than 10% of responses Sorry. There are only 65 that I would even consider reading. Out of those 65, only one out of 10 seems to be, "Huh, okay, I will give this positive marks," kind of a thing. So, in short, probably half a dozen people will get something between zero and 10. Rest, even in this case, will get a minus one.

[62:12] It's just minus one on a GA. It's a tiny component of a tiny component, and the amount of emotional impact that it has [laughter]

**JK**: [62:20] So, also yeah, yeah, let me let me just put that I mean, when you see 10 bonus marks, when you convert the 10 in a GA, I mean, see, everything will be rationalized based on the GA's actual weightage. So, the 10 that you see is actually a percentage of that GA that is going to come as bonus finally, and maximum bonus that one can get out of any course is +5.

[62:45] Okay? So, +5, and there could be exceptions. Every year, we look at the overall distribution and then decide on, and maximum that it can go is +10, not beyond that. And that also is capped to 100. I mean, you can only reach 100 in your T-score, not more.

[63:03] So, yeah. So, every time, so one of the important thing to look at is, what is the add-on and what is the drawback that I mean, what is the penalty that you are getting is worth checking for all data science students. [laughter] Yeah, I mean, Adwitiya is right. Everyone is of the opinion that any any extra mark gained is a gift. [laughter]

**Anand**: [63:24] We had a history teacher in class seven, Sheela Miss. Normally our quizzes were for 10 marks. She set a 100-mark quiz and one of the questions had was a 30-mark question. "Miss, 30-mark question? How can we write so much?" What difference does it make whether it's three out of 10 or 30 out of 100? It's exactly the same thing. "Oh, 30-mark question, we have to write 40 pages."

**JK**: [63:53] Yeah. And yeah, Preeti, you will not understand it right now, but let's come to two terms with another degree-level courses, and then look back on TDS and the marks that you get.

**Anand**: [64:06] Next time what I will do is might give -0.1 on a GA with a total score of 0.5. [laughter] "Oh, -0.1, what is the big deal?" [laughter] We should do that in the ROE. ROE is out of one mark. Everything is just 0.1, 0.05, 0.02.

**JK**: [64:25] Yeah.

**Anand**: [64:26] Okay, we should drop.

**JK**: [64:28] Yeah, yeah. Okay, anyway, so that is a secondary part of it, but generally from an operational thing, all the bonus marks that you see is on the basis of the the total value of that assignment, I mean, be it project or graded assignment. And what what we are looking at bonus marks when we put it out, we are looking at an actual, innovative effort that is done by the students. So, that's what Anand said. I mean, that is why certain criterias and considerations are put in for each of them.

**Anand**: [64:59] Thanks, everyone.

**JK**: [65:00] Yeah. Anand, and just a summary of this, just reminding you, I'll I'll just keep an eye on the Talks page.

**Anand**: [65:09] Yeah.

**Adwitiya**: [65:10] Sir, sir, just one request. You had a screen-shared a part of your website that had a very interesting summary. I think that's what sir just mentioned. Could you paste the link for that?

**Anand**: [65:20] It's already on the chat.

**Adwitiya**: [65:23] Okay. Thank you. Thank you, sir.

**JK**: [65:27] Yeah. And for all the, yeah, for all the students who have otherwise joined, a general check is after the session gets over, the next day, visit Anand's Talks page. Most of the sessions that he conducts, most of the discussions that we have, one or the other would be put in the Talks page in general as a summary.

**Adwitiya**: [65:53] Sir, I just scrolled through the chat. I believe I am not able to see the link. If you could, sir, or anyone if you could just paste it again, that would be really helpful.

**JK**: [66:03] Which one, the The the Talks page, sir. The one they were There are automated summaries of it along with the That's at talks.s-anand.net, s-anand.

**Adwitiya**: [66:17] Okay. It's a sub-domain thing. Okay, got it, got it.

**JK**: [66:19] Yeah, yeah. All his talks are available over there.

**Adwitiya**: [66:26] Great. Thank you, sir.

**JK**: [66:28] Yeah. Thanks, thanks, everyone. Jayadeep and others, I'll stop the recording and livestream then.

[66:50] [laughter]

## Chat

> 00:09:26.170,00:09:29.170
> PREETHI SIVARAMAKRISHNAN: feedback for chat window was that we had to scroll down for chatting & scroll up for adjusting the discounts... if i recollect properly.. a floating chat window would have been better I think
>
> 00:10:37.532,00:10:40.532
> Jayakrishnan Warriem: Got it.. But how was the discount negotiation experience?
>
> 00:11:07.389,00:11:10.389
> PREETHI SIVARAMAKRISHNAN: it got better with each iteration
>
> 00:12:28.691,00:12:31.691
> Jayakrishnan Warriem: Good to hear. Did the intuition increase with each iteration?
>
> 00:12:46.213,00:12:49.213
> PREETHI SIVARAMAKRISHNAN: Definitely..
>
> 00:18:20.206,00:18:23.206
> Swarnam Palaniswamy: AI created content will have liability issues. There are a few cases in court ( I am not sure if it is in India, but definitely in the US)
>
> 00:19:57.791,00:20:00.791
> Anand S: https://files.s-anand.net/pages/tds-2026-09-p1-q7-picture-stories/
>
> 00:33:37.623,00:33:40.623
> ADWITIYA CHAUHAN: Patterns you expect often end up overshadowing the patterns that exist in your memory
>
> 00:37:09.338,00:37:12.338
> ADWITIYA CHAUHAN: Just a suggestion: It might be interesting to inspect whether or not students perform better when we substitute anonymity with accountability in the assignments
>
> 00:43:52.505,00:43:55.505
> ADWITIYA CHAUHAN: https://chat.google.com/room/AAQAO77CWr8/YbzbOmdeMYI/YbzbOmdeMYI?cls=10
>
> 00:46:18.699,00:46:21.699
> ADWITIYA CHAUHAN: https://github.com/sreekarsarma55/temp-repo\
>
> 00:47:56.122,00:47:59.122
> PREETHI SIVARAMAKRISHNAN: https://claude.ai/artifact/Qs2N2ZyM91NFcA1Nuf34hp - this claude artifact that someone had shared was super helpful
>
> 00:55:02.529,00:55:05.529
> PREETHI SIVARAMAKRISHNAN: in the data chart question we had a gmeet with my team members for a clearer transfer of instructions & data
>
> 00:55:37.334,00:55:40.334
> PREETHI SIVARAMAKRISHNAN: since it was not mentioned that it is not allowed in the rules..
>
> 00:55:53.078,00:55:56.078
> Jayakrishnan Warriem: Which assignment was this part of?
>
> 00:56:45.982,00:56:48.982
> PREETHI SIVARAMAKRISHNAN: data poster - GA4 Q3
>
> 00:57:03.525,00:57:06.525
> PREETHI SIVARAMAKRISHNAN: time use charts sorry
>
> 01:00:17.530,01:00:20.530
> PREETHI SIVARAMAKRISHNAN: GA4 Q4 -- why -1 marking for inventing ROE question ?? Will it be implemented for real..??
>
> 01:02:09.339,01:02:12.339
> PREETHI SIVARAMAKRISHNAN: ok got it
>
> 01:03:13.732,01:03:16.732
> ADWITIYA CHAUHAN: Sir, in TDS we treat every mark gained as a gift given to us by the cosmos :)
>
> 01:03:40.492,01:03:43.492
> PREETHI SIVARAMAKRISHNAN: so true & sir is talking about -1 as if it is nothing...  :(((
>
> 01:04:53.384,01:04:56.384
> PREETHI SIVARAMAKRISHNAN: thank you Sir
>
> 01:05:59.382,01:06:02.382
> PREETHI SIVARAMAKRISHNAN: https://files.s-anand.net/pages/tds-2026-09-p1-q7-picture-stories/
>
> 01:06:09.150,01:06:12.150
> Swarnam Palaniswamy: https://files.s-anand.net/pages/tds-2026-09-p1-q7-picture-stories/
>
> 01:06:13.224,01:06:16.224
> Jaideep: https://talks.s-anand.net/
>
> 01:06:18.017,01:06:21.017
> Jayakrishnan Warriem: talks.s-anand.net
>
> 01:06:26.999,01:06:29.999
> PRASANNA S: https://talks.s-anand.net/
>
> 01:06:42.474,01:06:45.474
> PREETHI SIVARAMAKRISHNAN: thank you for the links
>
> 01:06:47.698,01:06:50.698
> PREETHI SIVARAMAKRISHNAN: thank you Sir

## Links

1. AIPL Session 3 Food Court Game Creation
    https://chatgpt.com/c/6af92045-6224-83ec-8caf-b9f275823c59
2. Picture Stories We Understand and Like | S Anand
    https://www.s-anand.net/blog/picture-stories-we-understand-and-like/
3. Picture Stories · TDS Project 1 · Results
    https://files.s-anand.net/pages/tds-2026-09-p1-q7-picture-stories/?tab=patterns
5. TDS 2026 Sep GA3
    https://exam.sanand.workers.dev/tds-2026-09-ga3#hq-3d-face-model
7. TDS 2026 Sep Project 1
    https://exam.sanand.workers.dev/tds-2026-09-p1#hq-face-expression-classifier
8. Invisible Ink · TDS P1
    https://tds-network-games.sanand.workers.dev/realtime/air-ink-classifier/
9. * GitHub - sreekarsarma55/temp-repo · GitHub
    https://github.com/sreekarsarma55/temp-repo/
