WEBVTT

NOTE Auto-generated meeting captions embedded in the recording; speaker tags normalised. Expect transcription errors.

00:00.000 --> 00:04.000
<v Jaideep>Hello
<v Anand>Hey, how are

00:08.000 --> 00:12.000
<v Anand>you? I know Cton and JK

00:12.000 --> 00:16.000
<v Anand>are not going to be so everything

00:16.000 --> 00:20.000
<v Jaideep>Surprising attendance is very

00:20.000 --> 00:24.000
<v Jaideep>low.
<v Anand>else. Yes, people are busy or happy for Diwali whatever good

00:24.000 --> 00:28.000
<v Anand>for them. We'll just get

00:28.000 --> 00:32.000
<v Anand>started. join at any point.

00:32.000 --> 00:36.000
<v Anand>Uh did you want to uh discuss the next

00:36.000 --> 00:40.000
<v Anand>GA offline or you want to discuss now? Either way is

00:40.000 --> 00:44.000
<v Jaideep>I mean uh it's fine uh we can discuss the GA but uh we

00:44.000 --> 00:48.000
<v Jaideep>want to do it anyway. So uh I mean we we don't want to

00:48.000 --> 00:52.000
<v Jaideep>obviously uh show the private report or of course it's your

00:52.000 --> 00:56.000
<v Jaideep>choice at the end.
<v Anand>okay. I'll leave it to you. Uh

00:56.000 --> 01:00.000
<v Anand>so then let's do this. we do it offline but I just want to flag

01:00.000 --> 01:04.000
<v Anand>off that I have a few questions that I wanted to introduce in

01:04.000 --> 01:08.000
<v Anand>J4 not because uh

01:08.000 --> 01:12.000
<v Anand>I'm partly capability uh

01:12.000 --> 01:16.000
<v Anand>questions for the platform itself to see if we can do uh

01:16.000 --> 01:20.000
<v Jaideep>But thank
<v Anand>a few things and

01:20.000 --> 01:24.000
<v Anand>uh partly just few followup kinds of things and and I wanted to

01:24.000 --> 01:28.000
<v Anand>introduce and one of the things that I'm curious about for instance

01:28.000 --> 01:32.000
<v Jaideep>you.
<v Anand>is does it help to

01:32.000 --> 01:36.000
<v Anand>have

01:40.000 --> 01:44.000
<v Anand>uh oh okay there's background noise that you're no that's okay

01:44.000 --> 01:48.000
<v Anand>um like for instance should we ask students what would

01:48.000 --> 01:52.000
<v Anand>uh they prefer uh option A versus option

01:52.000 --> 01:56.000
<v Anand>B in the ROE stuff like that I mean

01:56.000 --> 02:00.000
<v Anand>basically it's between the devil in the deep blue sea but if at

02:00.000 --> 02:04.000
<v Jaideep>Hey,
<v Anand>least students choose a I get uh

02:04.000 --> 02:08.000
<v Anand>input and uh B we can also steer the course in a certain direction so

02:08.000 --> 02:12.000
<v Anand>that's sort of a question I will probably want to add so I'm thinking

02:12.000 --> 02:16.000
<v Anand>I will create the scaffolding for it meaning one or two questions

02:16.000 --> 02:20.000
<v Anand>which you can then probably add

02:20.000 --> 02:24.000
<v Jaideep>uh yes uh one one more thing uh with

02:24.000 --> 02:28.000
<v Jaideep>regards to the questions I think it's uh at this point getting too hard to make them

02:28.000 --> 02:32.000
<v Jaideep>astra proof or opus proof
<v Anand>Yeah. Right.

02:32.000 --> 02:36.000
<v Anand>Yeah. What are your thoughts? How do we teach when the agent

02:36.000 --> 02:40.000
<v Jaideep>Yeah. Yeah. Yeah. I
<v Anand>can solve the problem? What do we

02:40.000 --> 02:44.000
<v Jaideep>mean, I think it's a problem not for us. I mean,
<v Anand>teach?

02:44.000 --> 02:48.000
<v Jaideep>even people above our grade like who are building

02:48.000 --> 02:52.000
<v Jaideep>state-of-the-art stuff, they're also running out of
<v Anand>Uh

02:52.000 --> 02:56.000
<v Jaideep>stuff.
<v Anand>I agree with 90% of your comment. The only thing I probably disagree with

02:56.000 --> 03:00.000
<v Anand>is above our uh grade because you are

03:00.000 --> 03:04.000
<v Anand>building state-of-the-art stuff. There aren't many people who are doing the stuff that's

03:04.000 --> 03:08.000
<v Jaideep>I mean uh let's say Andre
<v Anand>beyond it.

03:08.000 --> 03:12.000
<v Anand>Ah yeah. So like the few hundred

03:12.000 --> 03:16.000
<v Jaideep>Karpati. Yeah. I mean I don't know when
<v Anand>at most like

03:16.000 --> 03:20.000
<v Jaideep>Simon will last time handcoded
<v Anand>that. Exactly right.

03:20.000 --> 03:24.000
<v Jaideep>stuff.
<v Anand>And Patrick of Stripe, he

03:24.000 --> 03:28.000
<v Anand>tweeted recently saying he's thinking of liquid software that

03:28.000 --> 03:32.000
<v Anand>is a little button in the software where people give feedback,

03:32.000 --> 03:36.000
<v Anand>the software rewrites itself, creates their own

03:36.000 --> 03:40.000
<v Anand>tweaked version which potentially others can

03:40.000 --> 03:44.000
<v Jaideep>Yeah. So, uh have you taken a look at Omari a Linux
<v Anand>fork.

03:44.000 --> 03:48.000
<v Jaideep>distribution by DHH uh the creator of Ruby on Rails? So uh

03:48.000 --> 03:52.000
<v Jaideep>their idea is uh they have built a I

03:52.000 --> 03:56.000
<v Jaideep>mean something like a polished desktop environment for Linux and then

03:56.000 --> 04:00.000
<v Jaideep>they're uh integrating agent so hardly to the point they're saying uh if

04:00.000 --> 04:04.000
<v Jaideep>uh you find a bug uh in our operating system go and ask for it
<v Anand>No. Fix

04:04.000 --> 04:08.000
<v Jaideep>will uh rebuild
<v Anand>it. Love it. Okay. Uh,

04:08.000 --> 04:12.000
<v Jaideep>it.
<v Anand>sorry. What did you say was the

04:12.000 --> 04:16.000
<v Jaideep>Uh you can type uh
<v Adwitiya>Sir, it's called Oh, yeah. Sir, DHS, the creator

04:16.000 --> 04:20.000
<v Jaideep>DHH
<v Adwitiya>of Ruby on Rails, the same guy. He he's created

04:20.000 --> 04:24.000
<v Adwitiya>a Linux uh distribution. It's not technically a distribution. It's just

04:24.000 --> 04:28.000
<v Adwitiya>a bunch of, you know, uh customization scripts on top of

04:28.000 --> 04:32.000
<v Adwitiya>arts. But yeah, he calls it Omari. So, the spelling is O M

04:32.000 --> 04:36.000
<v Anand>reference?
<v Adwitiya>A R C H

04:36.000 --> 04:40.000
<v Anand>Okay. Okay.
<v Adwitiya>Y. Yeah. So initially he marketed it as

04:40.000 --> 04:44.000
<v Adwitiya>an opinionated opinionated project but now ever since he has

04:44.000 --> 04:48.000
<v Adwitiya>got funding from SpaceX and Grock and all stuff like that. So now he

04:48.000 --> 04:52.000
<v Adwitiya>is marketing it as an uh agentic

04:52.000 --> 04:56.000
<v Anand>Huh.
<v Adwitiya>OS. So at the time of uh startup you get to

04:56.000 --> 05:00.000
<v Anand>Interesting.
<v Adwitiya>choose your agent and it's more like you're interacting with a

05:00.000 --> 05:04.000
<v Adwitiya>computer not in a traditional sense but it's as if you're speaking to the machine.

05:04.000 --> 05:08.000
<v Anand>So I have to check this
<v Adwitiya>So that's the basic idea behind it.

05:08.000 --> 05:12.000
<v Anand>out here and you see they drove it

05:12.000 --> 05:16.000
<v Anand>for a while. Okay. What's your experience with it? What
<v Adwitiya>Sir

05:16.000 --> 05:20.000
<v Adwitiya>I I it's it's of course it's an interesting project

05:20.000 --> 05:24.000
<v Adwitiya>but uh from the security side there are a certain

05:24.000 --> 05:28.000
<v Anand>was
<v Adwitiya>flaws because when once you give like an agent there

05:28.000 --> 05:32.000
<v Adwitiya>are either you are very precise in what you type and what you want the

05:32.000 --> 05:36.000
<v Adwitiya>agent to do but otherwise it the agent has full system

05:36.000 --> 05:40.000
<v Adwitiya>access. Okay. So if it messes up something then it gets

05:40.000 --> 05:44.000
<v Adwitiya>a little tricky to debug it or to get it to revert it to the same state

05:44.000 --> 05:48.000
<v Adwitiya>as before. Of course you can create system snapshots but we are not considering

05:48.000 --> 05:52.000
<v Adwitiya>external software for now. And on top of that it is built on top

05:52.000 --> 05:56.000
<v Adwitiya>of arch like I said. So it's not a uh distribution in the traditional

05:56.000 --> 06:00.000
<v Anand>Here we
<v Adwitiya>sense. Right. Right. It's more of a customization aspect of

06:00.000 --> 06:04.000
<v Adwitiya>it. So uh that uh brings along with

06:04.000 --> 06:08.000
<v Adwitiya>it the drawbacks of Arch Linux such as the AUR not

06:08.000 --> 06:12.000
<v Adwitiya>being traditionally vetted as in Fedora and

06:12.000 --> 06:16.000
<v Anand>go.
<v Adwitiya>Ubuntu and you know the packages and all they are this they are still coming

06:16.000 --> 06:20.000
<v Adwitiya>from AUR and repos like this which are traditionally a little less

06:20.000 --> 06:24.000
<v Adwitiya>secure than you know uh more mainstream

06:24.000 --> 06:28.000
<v Anand>Got you. Interesting. I will check that
<v Adwitiya>dros.

06:28.000 --> 06:32.000
<v Anand>out. Cool.

06:32.000 --> 06:36.000
<v Anand>Uh before we move on to uh other topics,

06:36.000 --> 06:40.000
<v Anand>I thought one thing that I'll mention is uh in

06:40.000 --> 06:44.000
<v Anand>GF4 are there AI pipe related questions?

06:44.000 --> 06:48.000
<v Anand>Okay, we needed to rob but okay

06:48.000 --> 06:52.000
<v Anand>I think we need Jerry for this particular point. I will come back to

06:52.000 --> 06:56.000
<v Anand>it because among other things um it
<v Agrim>Hi,

06:56.000 --> 07:00.000
<v Anand>will mess up a few of the existing questions
<v Agrim>breathing.

07:00.000 --> 07:04.000
<v Anand>that people have.

07:04.000 --> 07:08.000
<v Anand>So do we

07:08.000 --> 07:12.000
<v Anand>have the

07:12.000 --> 07:16.000
<v Anand>uh just

07:16.000 --> 07:20.000
<v Anand>checking okay? Oh wait a

07:24.000 --> 07:28.000
<v Agrim>Hello hello hello good evening
<v Anand>good six is going to be you and

07:28.000 --> 07:32.000
<v Agrim>an uh yeah I don't think we have a I pipe
<v Anand>me good.

07:32.000 --> 07:36.000
<v Agrim>question in G4 currently
<v Anand>Okay, I would

07:36.000 --> 07:40.000
<v Agrim>uh
<v Anand>like I will be introducing one and uh yeah if you

07:40.000 --> 07:44.000
<v Agrim>Mhm.
<v Anand>could introduce a few as well that will be good because

07:44.000 --> 07:48.000
<v Anand>among other things it will be interesting to see the impact of

07:48.000 --> 07:52.000
<v Anand>that on project loan question one which was about

07:52.000 --> 07:56.000
<v Agrim>Mhm.
<v Anand>predicting the AI pipe token

07:56.000 --> 08:00.000
<v Anand>usage among other things one of the things I'm going to do

08:00.000 --> 08:04.000
<v Anand>now is uh increase the limit and you should probably

08:04.000 --> 08:08.000
<v Anand>let people know. Let me share my screen

08:08.000 --> 08:12.000
<v Agrim>Okay.
<v Anand>and do

08:12.000 --> 08:16.000
<v Anand>this. Uh so the

08:16.000 --> 08:20.000
<v Anand>IDM uh limit is uh $1

08:20.000 --> 08:24.000
<v Anand>over any 7-day period. I will make it at least

08:24.000 --> 08:28.000
<v Anand>$2 uh over the 7-day

08:28.000 --> 08:32.000
<v Anand>period. and deploy

08:32.000 --> 08:36.000
<v Anand>this. So I'll revert it

08:36.000 --> 08:40.000
<v Anand>later of course after a few weeks. But what this

08:40.000 --> 08:44.000
<v Anand>means and again this is a note to you to

08:44.000 --> 08:48.000
<v Agrim>Mhm.
<v Anand>please let everyone know that as a

08:48.000 --> 08:52.000
<v Anand>result project on one's

08:52.000 --> 08:56.000
<v Anand>uh forecast most people would want

08:56.000 --> 09:00.000
<v Agrim>Yep.
<v Anand>to revise the is

09:00.000 --> 09:04.000
<v Anand>uh you're not a position to use

09:04.000 --> 09:08.000
<v Anand>the new information that you

09:08.000 --> 09:12.000
<v Agrim>Okay.

09:12.000 --> 09:16.000
<v Anand>have. That was the one thing. the

09:16.000 --> 09:20.000
<v Anand>uh things that I wanted to

09:20.000 --> 09:24.000
<v Anand>share. Uh

09:24.000 --> 09:28.000
<v Anand>increasingly the time

09:28.000 --> 09:32.000
<v Anand>that uh it takes for features and

09:32.000 --> 09:36.000
<v Anand>capabilities to change is so high. As a result, obviously we are

09:36.000 --> 09:40.000
<v Anand>also changing the course fairly rapidly and we are

09:40.000 --> 09:44.000
<v Anand>using agents for a significant chunk of it. It helps to

09:44.000 --> 09:48.000
<v Anand>get a sense of uh let's

09:48.000 --> 09:52.000
<v Anand>do I think the exam rep needs an agent MD

09:52.000 --> 09:56.000
<v Anand>of some I would even call it an agent MD but more like a guide

09:56.000 --> 10:00.000
<v Anand>book and I took a shot at

10:00.000 --> 10:04.000
<v Anand>having GBT craft that based on all of our transcripts as well

10:04.000 --> 10:08.000
<v Anand>as uh emails, course

10:08.000 --> 10:12.000
<v Anand>material, etc. And

10:12.000 --> 10:16.000
<v Anand>here are the principles I'm thinking of. What I

10:16.000 --> 10:20.000
<v Anand>wanted to do was along with you go through

10:20.000 --> 10:24.000
<v Agrim>Oh
<v Anand>those. Is my screen

10:24.000 --> 10:28.000
<v Anand>visible? Okay.

10:28.000 --> 10:32.000
<v Anand>So alternately as you can see this is

10:32.000 --> 10:36.000
<v Anand>the context that we would want to share agents and

10:36.000 --> 10:40.000
<v Agrim>yes.
<v Anand>let's iterate on it. have made a few edits here and

10:40.000 --> 10:44.000
<v Anand>there but one is course philosophy why do

10:44.000 --> 10:48.000
<v Anand>we think TDS is different or not why we

10:48.000 --> 10:52.000
<v Anand>think why we are designing uh

10:52.000 --> 10:56.000
<v Agrim>Mhm.
<v Anand>TDS TDS is uh

10:56.000 --> 11:00.000
<v Anand>non traditional course is probably the

11:00.000 --> 11:04.000
<v Agrim>Mhm.
<v Anand>real intent here so

11:04.000 --> 11:08.000
<v Anand>one is

11:08.000 --> 11:12.000
<v Anand>agents if it is teachable agents

11:12.000 --> 11:16.000
<v Anand>will learn it. So there is literally no point teaching what is

11:16.000 --> 11:20.000
<v Anand>teachable. It's a weird way of putting it but

11:20.000 --> 11:24.000
<v Anand>that's where we are and that is the case and

11:24.000 --> 11:28.000
<v Agrim>H.
<v Anand>I think J this reflects what you were also saying earlier. What

11:28.000 --> 11:32.000
<v Anand>do we teach? The answer is perhaps what is it that

11:32.000 --> 11:36.000
<v Anand>humans need that agents won't learn. So that's

11:36.000 --> 11:40.000
<v Anand>perhaps a core principle. The way if an agent can do

11:40.000 --> 11:44.000
<v Anand>it, tell the lady agent to do it to

11:44.000 --> 11:48.000
<v Anand>and one of the things that that is dropped on

11:48.000 --> 11:52.000
<v Anand>is making sure that we are able to complete the task. AJS are

11:52.000 --> 11:56.000
<v Anand>good as a tool on intelligence

11:56.000 --> 12:00.000
<v Anand>by providing them additional tools. other agents

12:00.000 --> 12:04.000
<v Anand>people as capabilities we can drive them to get the job done or you get

12:04.000 --> 12:08.000
<v Anand>the job done and verify it but that's ultimately what

12:08.000 --> 12:12.000
<v Anand>human accountability will matter use

12:12.000 --> 12:16.000
<v Anand>whatever tools and get it done that's part of the principle

12:16.000 --> 12:20.000
<v Anand>are the curriculum is a well established principle for us

12:20.000 --> 12:24.000
<v Anand>uh

12:24.000 --> 12:28.000
<v Anand>and the fact that content pages are optional

12:28.000 --> 12:32.000
<v Anand>is secondary thing. Speaking of which, my is

12:32.000 --> 12:36.000
<v Agrim>Uh, no. Mango is not here.
<v Anand>on. Okay. So, one

12:36.000 --> 12:40.000
<v Agrim>Yeah.
<v Anand>of the things I wanted to flag off

12:40.000 --> 12:44.000
<v Anand>and on one thing that I want to flag off is

12:44.000 --> 12:48.000
<v Anand>as part of the

12:48.000 --> 12:52.000
<v Anand>DA sessions etc.

12:52.000 --> 12:56.000
<v Anand>uh we may be teaching

12:56.000 --> 13:00.000
<v Anand>things that an agent can teach. I happened to

13:00.000 --> 13:04.000
<v Anand>be in I think yesterday's workshop very

13:04.000 --> 13:08.000
<v Anand>briefly uh and M was

13:08.000 --> 13:12.000
<v Anand>sharing

13:12.000 --> 13:16.000
<v Anand>how to create valid JSON. There must have been a

13:16.000 --> 13:20.000
<v Anand>reason for it. But I was thinking hold on why should we teach

13:20.000 --> 13:24.000
<v Anand>that? Instead what I would do is have an agent create an

13:24.000 --> 13:28.000
<v Anand>explainer so that students

13:28.000 --> 13:32.000
<v Anand>will a learn how they can create their

13:32.000 --> 13:36.000
<v Anand>own explainers for the topic and

13:36.000 --> 13:40.000
<v Anand>B that can be published in any case for anyone else to refer.

13:40.000 --> 13:44.000
<v Anand>Therefore automatically creating reusable

13:44.000 --> 13:48.000
<v Anand>material reusable by humans and agents. But uh

13:48.000 --> 13:52.000
<v Anand>Jim uh are you conducting any

13:52.000 --> 13:56.000
<v Anand>uh classes for

13:56.000 --> 14:00.000
<v Agrim>uh I don't think I have conducted I think Jadep has take one session I

14:00.000 --> 14:04.000
<v Agrim>took one session but that was only a doubt session yeah
<v Jaideep>Yeah.
<v Anand>students. Got

14:04.000 --> 14:08.000
<v Jaideep>Uh I I have taken one session where uh
<v Anand>you.

14:08.000 --> 14:12.000
<v Agrim>uh
<v Jaideep>mostly I mean I was going through webmc where

14:12.000 --> 14:16.000
<v Jaideep>uh I it's uh at the end of the session it's

14:16.000 --> 14:20.000
<v Jaideep>just me using code using codeex and uh explaining

14:20.000 --> 14:24.000
<v Jaideep>students how to use codex to build stuff and

14:24.000 --> 14:28.000
<v Jaideep>uh
<v Anand>go which I think is fair

14:28.000 --> 14:32.000
<v Anand>and

14:32.000 --> 14:36.000
<v Anand>the challenge target whatever that I will

14:36.000 --> 14:40.000
<v Anand>uh place in front of all of you

14:40.000 --> 14:44.000
<v Anand>is how can we make ourselves redundant in the

14:44.000 --> 14:48.000
<v Anand>TA sessions meaning what would it

14:48.000 --> 14:52.000
<v Anand>take for agents to be able

14:52.000 --> 14:56.000
<v Anand>to fulfill the need of the TH sessions in the first

14:56.000 --> 15:00.000
<v Anand>place? side.

15:00.000 --> 15:04.000
<v Anand>If that becomes possible then

15:04.000 --> 15:08.000
<v Jaideep>Yeah. Uh but uh the way I conducted the session

15:08.000 --> 15:12.000
<v Jaideep>uh one point might be helpful where uh I have deliberately chosen

15:12.000 --> 15:16.000
<v Jaideep>myself to uh not look at WebMCP at all cost before
<v Anand>yes.

15:16.000 --> 15:20.000
<v Jaideep>the session. So uh each and every resource that is found is uh ad

15:20.000 --> 15:24.000
<v Jaideep>hoc. So uh what it actually helping me was uh

15:24.000 --> 15:28.000
<v Jaideep>assuming uh I'm at the of course we are we are at the

15:28.000 --> 15:32.000
<v Jaideep>same level to that concept. So how would we approach it uh having

15:32.000 --> 15:36.000
<v Jaideep>these agents how will we actually take on things? So

15:36.000 --> 15:40.000
<v Jaideep>uh we don't uh end up in information poverty.

15:40.000 --> 15:44.000
<v Jaideep>So just uh throwing every just tossing everything

15:44.000 --> 15:48.000
<v Jaideep>or even uh micromanaging things uh to the point where it's better

15:48.000 --> 15:52.000
<v Jaideep>to do it by yourself. So uh how to find that balance and how we

15:52.000 --> 15:56.000
<v Jaideep>approach things.
<v Anand>Good point. Completely agree. And what I'm taking away

15:56.000 --> 16:00.000
<v Anand>therefore is under preparation is an advantage in such

16:00.000 --> 16:04.000
<v Anand>sessions because you're transferring uh broader

16:04.000 --> 16:08.000
<v Anand>skill which is how you deal with such a question.

16:08.000 --> 16:12.000
<v Anand>Agreed. Who are all of us are doing that? Uh Aditya

16:12.000 --> 16:16.000
<v Anand>has commented saying wrong guides per session. Why do you have

16:16.000 --> 16:20.000
<v Adwitiya>So like you were looking for a way so that we can

16:20.000 --> 16:24.000
<v Adwitiya>basically more efficiently cover a larger information pool in the TA

16:24.000 --> 16:28.000
<v Adwitiya>sessions right. So for that what we can do is uh like

16:28.000 --> 16:32.000
<v Adwitiya>in my case just hunting through a 2our video for just a

16:32.000 --> 16:36.000
<v Adwitiya>20 minute segment that wastes a lot of time as compared to you know

16:36.000 --> 16:40.000
<v Anand>Adia?
<v Adwitiya>just having a prompt handy and asking the LLM or agent to you know

16:40.000 --> 16:44.000
<v Adwitiya>explain a concept to you. So if we can create if one of

16:44.000 --> 16:48.000
<v Adwitiya>the TAs whoever is taking the session if they have a specific topic

16:48.000 --> 16:52.000
<v Adwitiya>in mind they can just craft out maximum of four or

16:52.000 --> 16:56.000
<v Adwitiya>five points uh prompts sorry and then provide those to the

16:56.000 --> 17:00.000
<v Adwitiya>students so that if you're interested in seeing the execution the

17:00.000 --> 17:04.000
<v Adwitiya>live learning part of it you can watch the video but if you do want to explore the

17:04.000 --> 17:08.000
<v Adwitiya>content or your known or you know have a guided learning sort of experience

17:08.000 --> 17:12.000
<v Adwitiya>via an LLN you can just utilize those prompts and chat Start with an

17:12.000 --> 17:16.000
<v Adwitiya>LLM like chat GBT or something and you'll essentially be grasping the same

17:16.000 --> 17:20.000
<v Anand>It's a good point. Uh agree with
<v Adwitiya>concepts.

17:20.000 --> 17:24.000
<v Anand>the principle probably tweak the

17:24.000 --> 17:28.000
<v Anand>execution. So firstly it is good

17:28.000 --> 17:32.000
<v Anand>that TA is coming underprepared. That way the prompts will

17:32.000 --> 17:36.000
<v Anand>naturally emerge. uh

17:36.000 --> 17:40.000
<v Anand>it's useful that people then can

17:40.000 --> 17:44.000
<v Anand>understand the session through a set of prompts that were used or jump

17:44.000 --> 17:48.000
<v Anand>to the right portion. However, I think there are two kinds of

17:48.000 --> 17:52.000
<v Anand>audiences. one that would

17:52.000 --> 17:56.000
<v Anand>benefit from the answer but two and

17:56.000 --> 18:00.000
<v Anand>I suspect there's maybe majority uh who would benefit from the

18:00.000 --> 18:04.000
<v Anand>process that is the reason they are there in that session is because

18:04.000 --> 18:08.000
<v Anand>it hasn't yet struck enough people that you can get the answer

18:08.000 --> 18:12.000
<v Anand>just by prompting if they knew that they wouldn't attend the

18:12.000 --> 18:16.000
<v Anand>session in the first place. They just be sitting and uh uh

18:16.000 --> 18:20.000
<v Anand>prompting. So the session is meant to transfer that

18:20.000 --> 18:24.000
<v Anand>somewhat uh less

18:24.000 --> 18:28.000
<v Anand>tangible skill. Seeing somebody doing it, they say, "Oh, okay. it

18:28.000 --> 18:32.000
<v Anand>really is possible. The sec the

18:32.000 --> 18:36.000
<v Anand>other part of it though uh abstracting out content from the TA sessions

18:36.000 --> 18:40.000
<v Anand>is a useful uh thing and I

18:40.000 --> 18:44.000
<v Anand>don't think it should be the TAs that do this. I think it should be the students

18:44.000 --> 18:48.000
<v Anand>that look for instance if you said here's

18:48.000 --> 18:52.000
<v Anand>an agent that uh okay let's take it simple here's a

18:52.000 --> 18:56.000
<v Anand>prompt that from these YouTube videos will extract all the prompts

18:56.000 --> 19:00.000
<v Anand>share it in what say this course and let uh

19:00.000 --> 19:04.000
<v Anand>people create somebody else says oh I've already done this and here

19:04.000 --> 19:08.000
<v Anand>is a site where anyone can just get a sense of all the

19:08.000 --> 19:12.000
<v Anand>prompts that is great um one of the things that

19:12.000 --> 19:16.000
<v Anand>I do. Let me share my

19:16.000 --> 19:20.000
<v Anand>screen. This could be useful for anyone who wants to. We we used to do

19:20.000 --> 19:24.000
<v Anand>this for the TS sessions but stop. But

19:24.000 --> 19:28.000
<v Anand>my uh talk stage which

19:28.000 --> 19:32.000
<v Anand>is effectively yeah put in the link here.

19:32.000 --> 19:36.000
<v Anand>Uh it takes the transcripts or videos

19:36.000 --> 19:40.000
<v Anand>etc. and then converts it into a

19:40.000 --> 19:44.000
<v Anand>session where we go through what was covered

19:44.000 --> 19:48.000
<v Anand>including jumping to a specific point. And the past

19:48.000 --> 19:52.000
<v Anand>uh two weeks ago the meeting that we had is

19:52.000 --> 19:56.000
<v Anand>exactly what this

19:56.000 --> 20:00.000
<v Anand>is. The process is reasonably straightforward. I would rather

20:00.000 --> 20:04.000
<v Anand>people uh not even just discover it but implement it themselves.

20:04.000 --> 20:08.000
<v Anand>The reason I'm suggesting that is we will have more variety and learn

20:08.000 --> 20:12.000
<v Anand>what can be done from

20:12.000 --> 20:16.000
<v Anand>that coming back and of course at any point if anyone has any thoughts

20:16.000 --> 20:20.000
<v Anand>suggestions.

20:20.000 --> 20:24.000
<v Jaideep>uh I don't think even prompts should help uh because

20:24.000 --> 20:28.000
<v Jaideep>uh increasingly they are becoming to the point where uh each and

20:28.000 --> 20:32.000
<v Jaideep>every agent whatever the user is using is becoming too much
<v Anand>Yeah.

20:32.000 --> 20:36.000
<v Jaideep>personalized in the scene. So uh there is a huge bias in

20:36.000 --> 20:40.000
<v Jaideep>uh the way they prompt and uh how they interact with LMS with the context

20:40.000 --> 20:44.000
<v Jaideep>they build up over time. So uh I think we need to abstract

20:44.000 --> 20:48.000
<v Jaideep>even the prompts away and find something

20:48.000 --> 20:52.000
<v Jaideep>different.
<v Anand>Agreed. And yeah, that is literally coming back to the how do we teach what can't

20:52.000 --> 20:56.000
<v Anand>be taught kind of the nice. Yes.
<v Adwitiya>And

20:56.000 --> 21:00.000
<v Anand>Yeah.
<v Adwitiya>sir, if we are, you know, looking for an interesting tangent to this

21:00.000 --> 21:04.000
<v Adwitiya>conversation we could say would be that uh these days one thing that I've

21:04.000 --> 21:08.000
<v Adwitiya>noticed is that the more advanced the model gets as

21:08.000 --> 21:12.000
<v Adwitiya>compared to telling it what you want or or how it needs to be done you need

21:12.000 --> 21:16.000
<v Adwitiya>to tell it where to stop such as if I'm using

21:16.000 --> 21:20.000
<v Adwitiya>Opus 5.5 what I've noticed that it's programmed to give you the

21:20.000 --> 21:24.000
<v Adwitiya>absolutely correct answer so that sounds like a good idea at first

21:24.000 --> 21:28.000
<v Adwitiya>but if you're interacting with it in a manner that's limited you have session limits

21:28.000 --> 21:32.000
<v Adwitiya>you weekly limits and all of that. Then what's not good is that if you've

21:32.000 --> 21:36.000
<v Adwitiya>got if you've given it a complex prompt then it'll go in into

21:36.000 --> 21:40.000
<v Adwitiya>quite a comprehensive loop of self- testing before it provides you the

21:40.000 --> 21:44.000
<v Adwitiya>answer. So you do need to specify

21:44.000 --> 21:48.000
<v Adwitiya>what the limits of uh your like what are you expecting and

21:48.000 --> 21:52.000
<v Adwitiya>where it does not need to test. For example, if I just tell it to write

21:52.000 --> 21:56.000
<v Adwitiya>a just for an example a plug-in for a text editor. So for

21:56.000 --> 22:00.000
<v Adwitiya>that I already have the environment set up. So I won't really prefer to burn

22:00.000 --> 22:04.000
<v Adwitiya>my tokens on a test. I would prefer to burn it in debugging

22:04.000 --> 22:08.000
<v Adwitiya>rather than testing. So that's just something that we could explore in a

22:08.000 --> 22:12.000
<v Anand>How would you frame the
<v Adwitiya>question or two.

22:12.000 --> 22:16.000
<v Adwitiya>Sir I would one of the ideas was to use AI

22:16.000 --> 22:20.000
<v Adwitiya>pipe. So uh since I have seen in the previous GS as well

22:20.000 --> 22:24.000
<v Adwitiya>that you are able to analyze the output that we obtain essentially

22:24.000 --> 22:28.000
<v Anand>question?
<v Adwitiya>intercept it. So in that case you can also you know

22:28.000 --> 22:32.000
<v Adwitiya>uh time the responses or if not time you can

22:32.000 --> 22:36.000
<v Adwitiya>get an idea from the language whether or not the LLM

22:36.000 --> 22:40.000
<v Adwitiya>has tested something extra than what we told it to. Like

22:40.000 --> 22:44.000
<v Adwitiya>for example, if you are writing a script that should check

22:44.000 --> 22:48.000
<v Adwitiya>the uh that is not expected to fail on Python

22:48.000 --> 22:52.000
<v Adwitiya>3 for example, then it shouldn't just go on checking for some other

22:52.000 --> 22:56.000
<v Anand>Okay.
<v Adwitiya>irrelevant tests as well. So if you could just intercept the response that

22:56.000 --> 23:00.000
<v Anand>you but see here is
<v Adwitiya>you get, you will get an idea.

23:00.000 --> 23:04.000
<v Anand>how the student would and is expected to

23:04.000 --> 23:08.000
<v Anand>solve the problem and they will point it at the GA or

23:08.000 --> 23:12.000
<v Anand>they'll take CHP cloud whatever point it at the GA page and say you solve

23:12.000 --> 23:16.000
<v Anand>it and the agent will look at

23:16.000 --> 23:20.000
<v Anand>instructions instructions say make sure that you tell

23:20.000 --> 23:24.000
<v Anand>AI pipe to stop quickly it will craft a prompt that will tell AI pipe to

23:24.000 --> 23:28.000
<v Adwitiya>Yeah. So, but

23:28.000 --> 23:32.000
<v Anand>within. No, it
<v Adwitiya>in that sense at least this would inculcate that feeling of like

23:32.000 --> 23:36.000
<v Anand>wouldn't because the student won't even read
<v Adwitiya>right now it won't inculcate a sense

23:36.000 --> 23:40.000
<v Anand>the question. Not won't even should not even answer if it's framed this

23:40.000 --> 23:44.000
<v Anand>way. And
<v Adwitiya>of Yeah. Right.

23:44.000 --> 23:48.000
<v Anand>if and that's not a bad thing

23:48.000 --> 23:52.000
<v Anand>because if an agent can read it and solve

23:52.000 --> 23:56.000
<v Anand>it, that particular thing is not what we want to teach.

23:56.000 --> 24:00.000
<v Anand>So let's now think a little

24:00.000 --> 24:04.000
<v Anand>We want students to learn how

24:04.000 --> 24:08.000
<v Anand>to specify in a way that they conserve

24:08.000 --> 24:12.000
<v Anand>tokens. Okay. Why? Because that's a problem today. But if I tell Charge

24:12.000 --> 24:16.000
<v Anand>AP create a prompt to do that, it'll probably do a good job. So, do

24:16.000 --> 24:20.000
<v Anand>I still need to teach is question number one. Question

24:20.000 --> 24:24.000
<v Anand>number two is maybe for some reason I still need to teach.

24:24.000 --> 24:28.000
<v Anand>If so, how do we teach it so that even when it goes to

24:28.000 --> 24:32.000
<v Anand>CHP, it is not

24:32.000 --> 24:36.000
<v Anand>uh solvable? Think about it for a few minutes. You get a sense of the question

24:36.000 --> 24:40.000
<v Anand>design uh problem that we face on a daily

24:40.000 --> 24:44.000
<v Anand>basis. And let

24:44.000 --> 24:48.000
<v Anand>me open this up. Therefore, for

24:48.000 --> 24:52.000
<v Anand>subsequent gas, roe and projects, it's an open invitation to

24:52.000 --> 24:56.000
<v Anand>everyone. Uh please share

24:56.000 --> 25:00.000
<v Anand>ideas. What are the form? Uh this course is fine. The

25:00.000 --> 25:04.000
<v Anand>email is fine. Just make sure that all the TAs are copied.

25:04.000 --> 25:08.000
<v Anand>uh ideas for questions that

25:08.000 --> 25:12.000
<v Anand>we can include that will still teach if

25:12.000 --> 25:16.000
<v Anand>that question is literally copy pasted into JP

25:16.000 --> 25:20.000
<v Anand>or whatever and if it still

25:20.000 --> 25:24.000
<v Anand>teaches something that is useful should

25:24.000 --> 25:28.000
<v Anand>absolutely add it. So this is basically an

25:28.000 --> 25:32.000
<v Anand>open invitation opportunity whatever to everyone. You can

25:32.000 --> 25:36.000
<v Anand>decide your questions and each other's

25:36.000 --> 25:40.000
<v Anand>questions. We certainly will

25:40.000 --> 25:44.000
<v Anand>desperate for more such questions. So please think about it

25:44.000 --> 25:48.000
<v Anand>and we can iterate and discuss. Keep in mind that this particular

25:48.000 --> 25:52.000
<v Anand>skill of crafting a skill that will still teach when

25:52.000 --> 25:56.000
<v Anand>an agent can solve it is also one of the skills that you want to learn. Think of it as

25:56.000 --> 26:00.000
<v Anand>bonus learning if

26:00.000 --> 26:04.000
<v Anand>you and

26:04.000 --> 26:08.000
<v Anand>let me share my screen. Go back to the guide

26:08.000 --> 26:12.000
<v Anand>book. That relates

26:12.000 --> 26:16.000
<v Anand>very much to the point where we

26:16.000 --> 26:20.000
<v Anand>yeah teach people to delegate as much as possible.

26:20.000 --> 26:24.000
<v Anand>Uh

26:24.000 --> 26:28.000
<v Anand>and yeah, these are the four things that I still think have

26:28.000 --> 26:32.000
<v Anand>solid value judgment in some way, shape or

26:32.000 --> 26:36.000
<v Anand>form. Uh risk management in the

26:36.000 --> 26:40.000
<v Anand>sense we don't even know what all ways in which an agent can go wrong and we

26:40.000 --> 26:44.000
<v Anand>still have to make sure that we drive

26:44.000 --> 26:48.000
<v Anand>towards getting the job done. How will we know why it's

26:48.000 --> 26:52.000
<v Anand>failed and where it's failed? And finally, taking

26:52.000 --> 26:56.000
<v Anand>accountability. Doesn't matter whether the agent goofed up or not. I am accountable.

26:56.000 --> 27:00.000
<v Anand>That's what stays valuable. Yes. Therefore, it

27:00.000 --> 27:04.000
<v Anand>may be spotting something that we don't

27:04.000 --> 27:08.000
<v Anand>even state and

27:08.000 --> 27:12.000
<v Anand>implementing not just what is written but what is intended.

27:12.000 --> 27:16.000
<v Anand>This uh is the remaining human skill.

27:16.000 --> 27:20.000
<v Anand>So if a question had something vague along the

27:20.000 --> 27:24.000
<v Anand>lines of uh a CIO wants a dashboard with

27:24.000 --> 27:28.000
<v Anand>this data for an agent to say this

27:28.000 --> 27:32.000
<v Anand>is what a CIO wants is one thing but

27:32.000 --> 27:36.000
<v Anand>for the human to say okay look in that case I will tell

27:36.000 --> 27:40.000
<v Anand>CH GPT to do the analysis because it does it better and tell cloud to

27:40.000 --> 27:44.000
<v Anand>create the design because it presents it better is the kind of thing

27:44.000 --> 27:48.000
<v Anand>that will lead to a better dashboard.

27:48.000 --> 27:52.000
<v Anand>And that breakup is

27:52.000 --> 27:56.000
<v Anand>something that is essential. Better yet going beyond and

27:56.000 --> 28:00.000
<v Anand>telling JPD look a CEO will do not have time to read

28:00.000 --> 28:04.000
<v Anand>the massive amount of analysis that you have done. And telling plot

28:04.000 --> 28:08.000
<v Anand>that guy will almost certainly won't read it on the mobile device you've optimized for

28:08.000 --> 28:12.000
<v Anand>desktop is the kind of unstated requirement uh

28:12.000 --> 28:16.000
<v Anand>that judgment will bring in. And yeah,

28:16.000 --> 28:20.000
<v Anand>we now have to start teaching people this because earlier you would say oh that

28:20.000 --> 28:24.000
<v Anand>requires a lot of experience to get there.

28:24.000 --> 28:28.000
<v Anand>Every company is saying boss if you don't have that experience I won't hire you

28:28.000 --> 28:32.000
<v Anand>literally the room of one of my uh classmates

28:32.000 --> 28:36.000
<v Anand>who saying who are you going to hire and who are you going to fire say I just don't

28:36.000 --> 28:40.000
<v Anand>need judgment anymore I need people who have judgment so I should

28:40.000 --> 28:44.000
<v Anand>only hire senior people so which means you will get hired only if

28:44.000 --> 28:48.000
<v Anand>you build that experience and you build that experience by using agents

28:48.000 --> 28:52.000
<v Anand>to build up that experience having it interrogate so

28:52.000 --> 28:56.000
<v Anand>if in a host like This we guide people to

28:56.000 --> 29:00.000
<v Anand>solve problems which normally would require a lot of

29:00.000 --> 29:04.000
<v Anand>experience forcing them in this environment to pick up that skilling

29:04.000 --> 29:08.000
<v Anand>giving the tool which is the

29:08.000 --> 29:12.000
<v Anand>agent uh skip some

29:12.000 --> 29:16.000
<v Anand>of these. Okay. Or maybe not

29:16.000 --> 29:20.000
<v Anand>this the uh sourcing of context.

29:20.000 --> 29:24.000
<v Anand>uh should get on the context uh

29:24.000 --> 29:28.000
<v Anand>is the tough part and therefore

29:28.000 --> 29:32.000
<v Anand>finding or getting that whether it's from people or systems whether you

29:32.000 --> 29:36.000
<v Anand>connect to new systems to get the data whether it's connecting to

29:36.000 --> 29:40.000
<v Anand>people that's what's going to start

29:40.000 --> 29:44.000
<v Anand>mattering and seeing if we can spot the errors in the

29:44.000 --> 29:48.000
<v Anand>AI services partly the failure diagnosis partly the

29:48.000 --> 29:52.000
<v Anand>risk management and stuff. Finally, if you want

29:52.000 --> 29:56.000
<v Anand>to learn something by yourself, then

29:56.000 --> 30:00.000
<v Anand>before you tell ask the agent, You guess,

30:00.000 --> 30:04.000
<v Anand>see if the agent disagrees, agrees and use that to build your

30:04.000 --> 30:08.000
<v Anand>judgment. Those are the core principles. But the

30:08.000 --> 30:12.000
<v Anand>question to everyone, are there

30:12.000 --> 30:16.000
<v Anand>any principles that you think are wrong,

30:16.000 --> 30:20.000
<v Anand>missing? What would you add elite

30:24.000 --> 30:28.000
<v Anand>change?

30:36.000 --> 30:40.000
<v Anand>I'll publish this

30:40.000 --> 30:44.000
<v Anand>uh along with a bunch of other stuff and

30:44.000 --> 30:48.000
<v Anand>invite comments.

31:08.000 --> 31:12.000
<v Anand>Uh

31:12.000 --> 31:16.000
<v Anand>Jesus.

31:24.000 --> 31:28.000
<v Anand>This

31:40.000 --> 31:44.000
<v Anand>just the right place

31:44.000 --> 31:48.000
<v Anand>to put Yeah, it's

31:52.000 --> 31:56.000
<v Anand>fine. Okay,

31:56.000 --> 32:00.000
<v Anand>this I've added

32:00.000 --> 32:04.000
<v Anand>topics/guidebook.md. Make this

32:04.000 --> 32:08.000
<v Anand>prominently linked from the

32:08.000 --> 32:12.000
<v Anand>homepage.

32:24.000 --> 32:28.000
<v Anand>Still not

32:40.000 --> 32:44.000
<v Anand>good.

32:44.000 --> 32:48.000
<v Anand>and I'll invite comments. So on Wednesday when we catch

32:48.000 --> 32:52.000
<v Anand>up instructors and TAS maybe we can have a

32:52.000 --> 32:56.000
<v Anand>chat about what we need to incorporate the

32:56.000 --> 33:00.000
<v Anand>guide the Inex. of the guide book

33:00.000 --> 33:04.000
<v Anand>is so that when you have agents create

33:04.000 --> 33:08.000
<v Anand>questions, they refer to this. And therefore, if in the process of creating questions,

33:08.000 --> 33:12.000
<v Anand>you have found some prompts

33:12.000 --> 33:16.000
<v Anand>that you had to give on a regular basis, this is the place

33:16.000 --> 33:20.000
<v Anand>to be. It will soon become an agent.mmd

33:20.000 --> 33:24.000
<v Anand>for the uh exam

33:24.000 --> 33:28.000
<v Anand>record. Any questions?

33:32.000 --> 33:36.000
<v Anand>You got

33:36.000 --> 33:40.000
<v Adwitiya>Sir I was just uh thinking about what the discussion that we had

33:40.000 --> 33:44.000
<v Adwitiya>about the question how would I put it technically. So

33:44.000 --> 33:48.000
<v Adwitiya>something that came to my mind was of course it's not a very

33:48.000 --> 33:52.000
<v Anand>nothing.
<v Adwitiya>ethical thing to do but as long as teaching is involved. So

33:52.000 --> 33:56.000
<v Adwitiya>in the the GA interface you have an option

33:56.000 --> 34:00.000
<v Adwitiya>to choose the AI right which AI we want to choose. At the

34:00.000 --> 34:04.000
<v Adwitiya>bottom of that it says uh an option of copy to clipboard.

34:04.000 --> 34:08.000
<v Adwitiya>So if I click that button, the content will get copied to my

34:08.000 --> 34:12.000
<v Adwitiya>clipboard. So we could modify it in a way if we if

34:12.000 --> 34:16.000
<v Adwitiya>you wish to force the students to actually read the

34:16.000 --> 34:20.000
<v Adwitiya>questions, you could modify it such that the stream is interrupted and

34:20.000 --> 34:24.000
<v Adwitiya>edited before it lands up in our clipboard.

34:24.000 --> 34:28.000
<v Anand>Yeah, I know what you mean. My did that in
<v Adwitiya>So what I mean is if we go ahead and paste.

34:28.000 --> 34:32.000
<v Anand>one of the I mean last ro I think and
<v Adwitiya>Yeah. Okay.

34:32.000 --> 34:36.000
<v Anand>yeah
<v Adwitiya>Okay. Okay. And one one more thing. So sir what we as a

34:36.000 --> 34:40.000
<v Adwitiya>humans we know that certain phrases they are quite

34:40.000 --> 34:44.000
<v Adwitiya>natural to us. Make sure that your answer is correct. So we'll check it twice

34:44.000 --> 34:48.000
<v Adwitiya>at most. Right? Not a big problem. But if it is a complicated problem

34:48.000 --> 34:52.000
<v Adwitiya>that involves you know floating point decimals for example like

34:52.000 --> 34:56.000
<v Anand>yeah
<v Adwitiya>2.499 499 multiple times and I ask an AI to make sure

34:56.000 --> 35:00.000
<v Adwitiya>that the answer is correct and if it is a complex model it

35:00.000 --> 35:04.000
<v Adwitiya>will definitely you know burn through a

35:04.000 --> 35:08.000
<v Adwitiya>quite a significant number of percentage of tokens before it tells me yes

35:08.000 --> 35:12.000
<v Adwitiya>that I'm fairly confident this answer is correct that's because they are being

35:12.000 --> 35:16.000
<v Adwitiya>trained to make sure that what the output you give is perfectly

35:16.000 --> 35:20.000
<v Adwitiya>fine and secondly when I would read the question I wouldn't even think

35:20.000 --> 35:24.000
<v Adwitiya>that this is the line that's causing the problem. The the same goes for a

35:24.000 --> 35:28.000
<v Adwitiya>case when you know there are multiple correct options in a wide range of

35:28.000 --> 35:32.000
<v Adwitiya>options and you're just asking for one.

35:32.000 --> 35:36.000
<v Anand>That's
<v Adwitiya>So,

35:36.000 --> 35:40.000
<v Anand>promising. It is

35:40.000 --> 35:44.000
<v Anand>not agent proof in that what say

35:44.000 --> 35:48.000
<v Anand>Astra would do is run a whole series of uh

35:48.000 --> 35:52.000
<v Anand>props and and

35:52.000 --> 35:56.000
<v Anand>see which one will still

35:56.000 --> 36:00.000
<v Anand>solve the problem. But it is potentially a good armor.

36:00.000 --> 36:04.000
<v Anand>question because it just might be too

36:04.000 --> 36:08.000
<v Anand>long for Astra to solve. If you can find

36:08.000 --> 36:12.000
<v Anand>a way of personalizing that question

36:12.000 --> 36:16.000
<v Anand>in a way that is nonparameterizing what do I mean? Each

36:16.000 --> 36:20.000
<v Anand>student gets a different question. B the questions

36:20.000 --> 36:24.000
<v Anand>differences are not just in changing a number or a phrase or something.

36:24.000 --> 36:28.000
<v Anand>So that if if that were possible, what I would do is

36:28.000 --> 36:32.000
<v Anand>have somebody create a solver for it and share with everyone

36:32.000 --> 36:36.000
<v Anand>that solver will have done the full analysis. Spend the first 30

36:36.000 --> 36:40.000
<v Anand>minutes of the RO getting to the result and then tell people here is the

36:40.000 --> 36:44.000
<v Anand>prompt. Replace copy paste this number and this text and blah blah blah from your

36:44.000 --> 36:48.000
<v Anand>question. Put it in here. That is your answer. it

36:48.000 --> 36:52.000
<v Anand>will typically I'm very

36:52.000 --> 36:56.000
<v Anand>happy with the fact that people would have delegated also very happy with the fact

36:56.000 --> 37:00.000
<v Anand>that one person would have used an agent

37:00.000 --> 37:04.000
<v Anand>effectively to solve the problem. What I would like

37:04.000 --> 37:08.000
<v Anand>is for more people to have to use their agents in a clever

37:08.000 --> 37:12.000
<v Anand>way to get to that same result. And therefore, if you could

37:12.000 --> 37:16.000
<v Anand>design this question which I see has promise in

37:16.000 --> 37:20.000
<v Anand>a way that would still reach a larger number of people not

37:20.000 --> 37:24.000
<v Anand>100%. 10% of people learning this is good but it has to

37:24.000 --> 37:28.000
<v Anand>be 10% not 1%. That will be

37:28.000 --> 37:32.000
<v Anand>powerful. Do think about it.
<v Adwitiya>That's okay. sniffing

37:32.000 --> 37:36.000
<v Anand>share desk for a few more things.
<v Adwitiya>it.

37:36.000 --> 37:40.000
<v Anand>So I was just playing around with the results uh

37:40.000 --> 37:44.000
<v Anand>of uh G325 the bonus amount

37:44.000 --> 37:48.000
<v Anand>question to see if we

37:48.000 --> 37:52.000
<v Anand>can

37:52.000 --> 37:56.000
<v Anand>yeah PS code has a dictation feature which is pretty decent actually

37:56.000 --> 38:00.000
<v Anand>it's somewhat code friendly

38:00.000 --> 38:04.000
<v Anand>so in just looking at what are the

38:04.000 --> 38:08.000
<v Anand>questions that people are making drawings that are hard for AI to

38:08.000 --> 38:12.000
<v Anand>guess. So this one AI totally failed

38:12.000 --> 38:16.000
<v Anand>to guess but all students

38:16.000 --> 38:20.000
<v Anand>humans guessed it right. Donkey Luna thought it was a

38:20.000 --> 38:24.000
<v Anand>sheep or a cow. GM thought it was a horse or a sheep

38:24.000 --> 38:28.000
<v Anand>etc. It's a pretty decent guess. So

38:28.000 --> 38:32.000
<v Anand>yeah, plus five.

38:32.000 --> 38:36.000
<v Anand>We figure out the scoring for it. This was

38:36.000 --> 38:40.000
<v Anand>surprising. Uh, a ladder drawn this way ended up

38:40.000 --> 38:44.000
<v Anand>fooling most of the models. The

38:44.000 --> 38:48.000
<v Anand>other technique uh was

38:48.000 --> 38:52.000
<v Anand>ladybirds. Indian food GLM all the time but lady

38:52.000 --> 38:56.000
<v Anand>and bird is a clever combination where you're

38:56.000 --> 39:00.000
<v Anand>able to differentiate what you communicate But humor

39:00.000 --> 39:04.000
<v Anand>versus an agent and

39:04.000 --> 39:08.000
<v Anand>this yeah I don't know

39:08.000 --> 39:12.000
<v Anand>why they wouldn't think it was a pressure cooker and we consistently think of

39:12.000 --> 39:16.000
<v Anand>it as pressure cooker but it may be just

39:16.000 --> 39:20.000
<v Anand>linguistic and this one most models thought was a shuttle cock or a

39:20.000 --> 39:24.000
<v Anand>whizzel whereas it's according to us mostly a metal

39:24.000 --> 39:28.000
<v Anand>and these are all based on the same set of options uh that was given

39:28.000 --> 39:32.000
<v Anand>to the models of course randomized in different ways but options

39:32.000 --> 39:36.000
<v Anand>are more or less the

39:36.000 --> 39:40.000
<v Anand>same yeah mango versus pair not

39:40.000 --> 39:44.000
<v Anand>sure but one of the things uh I'm guessing at least in a few of

39:44.000 --> 39:48.000
<v Anand>these cases is people are getting an intuition for how

39:48.000 --> 39:52.000
<v Anand>to draw something that

39:52.000 --> 39:56.000
<v Anand>and that can differentiate between a human and an agent but in terms of

39:56.000 --> 40:00.000
<v Anand>the percentage difference It's uh relatively weak.

40:00.000 --> 40:04.000
<v Anand>The total number of people that could get something that a human could answer that

40:04.000 --> 40:08.000
<v Anand>agents couldn't. Let me see if I

40:08.000 --> 40:12.000
<v Anand>have the stats for

40:12.000 --> 40:16.000
<v Anand>that.

40:16.000 --> 40:20.000
<v Anand>Yeah. So, we were looking for something

40:20.000 --> 40:24.000
<v Anand>where the humans are always right and the agents are always wrong

40:24.000 --> 40:28.000
<v Anand>and only a could do that.

40:28.000 --> 40:32.000
<v Anand>Right. Overall metrics are not too bad.

40:32.000 --> 40:36.000
<v Anand>Uh that it's 112 where at

40:36.000 --> 40:40.000
<v Anand>least the the AI was uh only partly

40:40.000 --> 40:44.000
<v Anand>right while the humans were perfectly correct or humans were partly right and the

40:44.000 --> 40:48.000
<v Anand>AI was still wrong. It's okay slightly

40:48.000 --> 40:52.000
<v Anand>better fraction but obviously the AI capabilities are going to keep going up

40:52.000 --> 40:56.000
<v Anand>and learning different ways

40:56.000 --> 41:00.000
<v Anand>of identifying

41:00.000 --> 41:04.000
<v Anand>model gaps is something that

41:04.000 --> 41:08.000
<v Anand>I think we would like to prove. So as part

41:08.000 --> 41:12.000
<v Anand>of GA 5 and six can request to

41:12.000 --> 41:16.000
<v Anand>the uh TAS as you are crafting

41:16.000 --> 41:20.000
<v Anand>this uh please incorporate more

41:20.000 --> 41:24.000
<v Anand>mechanisms by which we can get AI to fail

41:24.000 --> 41:28.000
<v Anand>while humans cannot. We have the infrastructure to allow

41:28.000 --> 41:32.000
<v Anand>humans to test. So that will be pretty

41:32.000 --> 41:36.000
<v Anand>useful. Uh,

41:36.000 --> 41:40.000
<v Anand>sorry. Any questions or comments on

41:44.000 --> 41:48.000
<v Anand>that? Okay, move on. Another

41:48.000 --> 41:52.000
<v Anand>thing, uh, I'll share this link

41:52.000 --> 41:56.000
<v Anand>with let me wow.

41:56.000 --> 42:00.000
<v Anand>Uh, so what I did was

42:00.000 --> 42:04.000
<v Anand>catalog about 10,000 12,000 prompts that I've

42:04.000 --> 42:08.000
<v Anand>given coding agents as well as

42:08.000 --> 42:12.000
<v Anand>u non coding agents strategy cloud etc uh

42:12.000 --> 42:16.000
<v Anand>to extract the different techniques that at least I have

42:16.000 --> 42:20.000
<v Anand>been using for specifying requirements and verifying

42:20.000 --> 42:24.000
<v Anand>output which are the two broad

42:24.000 --> 42:28.000
<v Anand>uh agent era capabilities that people need

42:28.000 --> 42:32.000
<v Anand>and If you look at what at least

42:32.000 --> 42:36.000
<v Anand>for me would rank as number one

42:36.000 --> 42:40.000
<v Anand>is can we have it run the output

42:40.000 --> 42:44.000
<v Anand>effectively giving it an environment in which we can run the test cases and see

42:44.000 --> 42:48.000
<v Anand>if it can be deterministically validated. the environment could be

42:48.000 --> 42:52.000
<v Anand>a coding agent, it could be a chess playing agent, it could be a

42:52.000 --> 42:56.000
<v Anand>physical uh laboratory, it could be an API, whatever. But I think

42:56.000 --> 43:00.000
<v Anand>now we are at the stage where going beyond coding

43:00.000 --> 43:04.000
<v Anand>agents uh is important that the verification environment

43:04.000 --> 43:08.000
<v Anand>starts becoming

43:08.000 --> 43:12.000
<v Anand>broader and I leave it open to interpretation but therefore

43:12.000 --> 43:16.000
<v Anand>for GS and particularly project I would invite

43:16.000 --> 43:20.000
<v Anand>questions from the TAS but from anyone

43:20.000 --> 43:24.000
<v Anand>in general what kinds of environment

43:24.000 --> 43:28.000
<v Anand>uh engines this can be test. Now part of this is being

43:28.000 --> 43:32.000
<v Anand>incorporated in uh project one and I

43:32.000 --> 43:36.000
<v Anand>think uh G3 as

43:36.000 --> 43:40.000
<v Anand>well but let's see I think G3 might have

43:40.000 --> 43:44.000
<v Anand>had it was a meshy uh question in

43:44.000 --> 43:48.000
<v Anand>G3 it's technically not meshy it's just

43:48.000 --> 43:52.000
<v Anand>yeah deconstructing a famous phase

43:52.000 --> 43:56.000
<v Anand>and a lot of people have converged to a meshy plus charg GBTR

43:56.000 --> 44:00.000
<v Anand>cloud workflow as being reasonably good. Yes,

44:00.000 --> 44:04.000
<v Anand>that's not a bad uh uh solution. Um,

44:04.000 --> 44:08.000
<v Anand>however, it hasn't yet taken people to is there

44:08.000 --> 44:12.000
<v Anand>a verifiable environment that an agent can manipulate a

44:12.000 --> 44:16.000
<v Anand>3D model builder where it can iterate, improve, etc. with vision

44:16.000 --> 44:20.000
<v Anand>capability and that's the kind of next nudge that we would want to

44:20.000 --> 44:24.000
<v Anand>go towards. Uh the morphing question is similar

44:24.000 --> 44:28.000
<v Anand>and nobody

44:28.000 --> 44:32.000
<v Anand>very few submissions have uh any kind of decent

44:32.000 --> 44:36.000
<v Anand>morphing. Most are simply doing an animated transition between these. That's

44:36.000 --> 44:40.000
<v Anand>rubbish. What I want is the face to really I mean

44:40.000 --> 44:44.000
<v Anand>any frame in between should look like a real person's

44:44.000 --> 44:48.000
<v Anand>face and a real photo. That is good

44:48.000 --> 44:52.000
<v Anand>morphing and very few people have gotten that far.

44:52.000 --> 44:56.000
<v Anand>Again, is there an environment where we can do systematic morphing and

44:56.000 --> 45:00.000
<v Anand>FFM? It just does not do that. that. we

45:00.000 --> 45:04.000
<v Anand>can uh plug into an agent. So even just finding the

45:04.000 --> 45:08.000
<v Anand>existence of some of these verifiable environments becomes

45:08.000 --> 45:12.000
<v Anand>a useful teaching skill. Another one

45:12.000 --> 45:16.000
<v Anand>is how do we make sure that you stage something that is

45:16.000 --> 45:20.000
<v Anand>irreversible? Can we

45:20.000 --> 45:24.000
<v Anand>uh have a situation where students are

45:24.000 --> 45:28.000
<v Anand>forced to decide whether something is reversible,

45:28.000 --> 45:32.000
<v Anand>irreversible across a series of decisions

45:32.000 --> 45:36.000
<v Anand>and use that to decide whether they want to

45:36.000 --> 45:40.000
<v Anand>spend as you had said more tokens where it is

45:40.000 --> 45:44.000
<v Anand>irreversible, less tokens where it is reversible in a way that

45:44.000 --> 45:48.000
<v Anand>even if they gave the problem to GP tokens across or

45:48.000 --> 45:52.000
<v Anand>whatever it would not be able to in a

45:52.000 --> 45:56.000
<v Anand>generalized way solve the problem without spending a hell amount of

45:56.000 --> 46:00.000
<v Anand>tokens. So one way of doing it is to tell them

46:00.000 --> 46:04.000
<v Anand>look you're what you you can only give a prompt with

46:04.000 --> 46:08.000
<v Anand>instructions give the right instructions that will be executed by

46:08.000 --> 46:12.000
<v Anand>another agent. Yes codeex

46:12.000 --> 46:16.000
<v Anand>will figure out a good prompt for that even better. not sure how to design

46:16.000 --> 46:20.000
<v Anand>it but this knowing uh what's

46:20.000 --> 46:24.000
<v Anand>reversible what's irreversible is something that you will have to

46:24.000 --> 46:28.000
<v Anand>factor this could possibly be a good intern question and therefore

46:28.000 --> 46:32.000
<v Anand>personally if you're going to be designing the intern questions this is a

46:32.000 --> 46:36.000
<v Anand>classic set of tests saying what are these 10 things which ones of

46:36.000 --> 46:40.000
<v Anand>these would you commit and deploy or tell the agent to commit and

46:40.000 --> 46:44.000
<v Anand>deploy which ones of these would you have it just test and you verify for example

46:44.000 --> 46:48.000
<v Anand>A third one is

46:48.000 --> 46:52.000
<v Anand>how do we verify the verifier itself? Uh in a

46:52.000 --> 46:56.000
<v Anand>number of cases I've had situations where the test

46:56.000 --> 47:00.000
<v Anand>cases were wrong. Um where the

47:00.000 --> 47:04.000
<v Anand>judge prompt was wrong etc. and

47:04.000 --> 47:08.000
<v Anand>validating that becomes useful. I'm not going to go through the full list. Uh

47:08.000 --> 47:12.000
<v Anand>I just want to say that this is the order of

47:12.000 --> 47:16.000
<v Anand>um impact, frequency and

47:16.000 --> 47:20.000
<v Anand>uh

47:20.000 --> 47:24.000
<v Anand>effectiveness of techniques that

47:24.000 --> 47:28.000
<v Anand>I've been using. It's at least a starting point and

47:28.000 --> 47:32.000
<v Anand>I would suggest we create questions using these for verification

47:32.000 --> 47:36.000
<v Anand>as well as for

47:36.000 --> 47:40.000
<v Anand>specification. Any comments or questions from

47:48.000 --> 47:52.000
<v Anand>anyone?

48:00.000 --> 48:04.000
<v Anand>Okay, that actually covers the bulk of what

48:04.000 --> 48:08.000
<v Anand>I wanted to do.

48:08.000 --> 48:12.000
<v Anand>Uh, quick check to see if I missed

48:12.000 --> 48:16.000
<v Anand>anything.

48:20.000 --> 48:24.000
<v Anand>I think

48:24.000 --> 48:28.000
<v Anand>that covers. So primarily uh an

48:28.000 --> 48:32.000
<v Anand>open request to everyone to share Asian

48:32.000 --> 48:36.000
<v Anand>proof questions we will incorporate them.

48:36.000 --> 48:40.000
<v Anand>Karthik's uh question as well questions from G3 around

48:40.000 --> 48:44.000
<v Anand>these partly Karthik uh the idea for

48:44.000 --> 48:48.000
<v Anand>many of these is given the skills we know

48:48.000 --> 48:52.000
<v Anand>are relevant in the agent which of these are relevant but

48:52.000 --> 48:56.000
<v Anand>oh so the intent behind some of the GC questions

48:56.000 --> 49:00.000
<v Anand>was trickier than that let me share my

49:04.000 --> 49:08.000
<v Anand>Okay. Brackets

49:08.000 --> 49:12.000
<v Anand>open

49:16.000 --> 49:20.000
<v Anand>three. Let me go from the bottom.

49:20.000 --> 49:24.000
<v Anand>U which phases are easiest to

49:24.000 --> 49:28.000
<v Anand>mock is entirely based on the

49:28.000 --> 49:32.000
<v Anand>premise that look there is one thing

49:32.000 --> 49:36.000
<v Anand>you will find really useful in the future which is predicting another

49:36.000 --> 49:40.000
<v Anand>human's behavior. So here the point is

49:40.000 --> 49:44.000
<v Anand>not what you think is the ranking. Can you figure out if

49:44.000 --> 49:48.000
<v Anand>your ranking will match with your classmates's ranking? Now there are two ways of solving

49:48.000 --> 49:52.000
<v Anand>this. One, you have and build really good

49:52.000 --> 49:56.000
<v Anand>intuition which will teach you the psychology of people.

49:56.000 --> 50:00.000
<v Anand>Absolutely the most important during data science

50:00.000 --> 50:04.000
<v Anand>or it will teach you collaboration which is you somehow figure

50:04.000 --> 50:08.000
<v Anand>out the mechanism by which you and all the classmates rank in the same

50:08.000 --> 50:12.000
<v Anand>way cheat if you will. But

50:12.000 --> 50:16.000
<v Anand>uh that then teaches you collaboration which yeah okay I changed my

50:16.000 --> 50:20.000
<v Anand>rank that may be even more important than understanding the psychology. That

50:20.000 --> 50:24.000
<v Anand>was the intent behind question

50:24.000 --> 50:28.000
<v Anand>eight. Question seven is to teach you

50:28.000 --> 50:32.000
<v Anand>what uh morphing algorithms exist

50:32.000 --> 50:36.000
<v Anand>and figuring out a really good work obviously collaboratively

50:36.000 --> 50:40.000
<v Anand>and sadly very few people have picked

50:40.000 --> 50:44.000
<v Anand>up on the the right algorithms to do good morphing.

50:44.000 --> 50:48.000
<v Anand>There are some fantastic algorithms out there that are extremely easy to use

50:48.000 --> 50:52.000
<v Anand>with the right kind of research. This is a

50:52.000 --> 50:56.000
<v Anand>classic solver kind of a problem meaning it's all prompt

50:56.000 --> 51:00.000
<v Anand>kind of a problem. All people have uh to do is give

51:00.000 --> 51:04.000
<v Anand>one prompt to an agent and it will solve it. Luna will

51:04.000 --> 51:08.000
<v Anand>solve it with 10,000 tokens at

51:08.000 --> 51:12.000
<v Anand>most. Uh this again uh people are

51:12.000 --> 51:16.000
<v Anand>solving it with Mishi uh which is not a bad solution could be improved on

51:16.000 --> 51:20.000
<v Anand>but again is aimed at finding a good verifiable

51:20.000 --> 51:24.000
<v Anand>environment. The uh humans only uh question was similar to

51:24.000 --> 51:28.000
<v Anand>the last one. Is it possible for you to guess to create

51:28.000 --> 51:32.000
<v Anand>something that aligns with human judgment and misalign with agent

51:32.000 --> 51:36.000
<v Anand>judgment? Uh this would have been tough even with a soft so

51:36.000 --> 51:40.000
<v Anand>necessary learning. I had to convert it to a bonus mark question because

51:40.000 --> 51:44.000
<v Anand>the dates were uh too tight and a lot of people didn't see

51:44.000 --> 51:48.000
<v Anand>it. But it was also an opportunity to teach a very different thing all together which

51:48.000 --> 51:52.000
<v Anand>is in life anything changes at any

51:52.000 --> 51:56.000
<v Anand>point. An agent can sit and watch

51:56.000 --> 52:00.000
<v Anand>any change randomly. So in TDS

52:00.000 --> 52:04.000
<v Anand>now I'm making this announcement. The instructors and TAS are making the announcement that stuff

52:04.000 --> 52:08.000
<v Anand>will change. So build your agents

52:08.000 --> 52:12.000
<v Anand>collaboratively or individually to watch for changes. Better

52:12.000 --> 52:16.000
<v Anand>yet agents look go watch

52:16.000 --> 52:20.000
<v Anand>sites solve the questions and submit on my behalf if you can

52:20.000 --> 52:24.000
<v Anand>in a reversible or yeah preferably in a reversible way. Which brings

52:24.000 --> 52:28.000
<v Anand>us to why these checks are there. Uh 10 out of 10

52:28.000 --> 52:32.000
<v Anand>checks is very liberal but if your agent happily goes checking checking checking at

52:32.000 --> 52:36.000
<v Anand>some point it becomes irreversible. And that's where you need to have

52:36.000 --> 52:40.000
<v Anand>exercised the discretion to your agent saying everyday check. But if there

52:40.000 --> 52:44.000
<v Anand>are only a limited number of checks, leave some for my judgment at the end

52:44.000 --> 52:48.000
<v Anand>because you might mess up. So yeah, in a

52:48.000 --> 52:52.000
<v Anand>sense to your question, Karthik, what we discussed today are the kinds of

52:52.000 --> 52:56.000
<v Anand>things uh that uh we are trying to teach

52:56.000 --> 53:00.000
<v Anand>and making tokens. Yes, absolutely. As much as

53:00.000 --> 53:04.000
<v Anand>possible, burn it like crazy and learn from getting

53:04.000 --> 53:08.000
<v Anand>your fingers caught. Best way to go.

53:08.000 --> 53:12.000
<v Anand>Any questions from

53:20.000 --> 53:24.000
<v Kartikkumar>Uh so I don't know I joined a bit late uh

53:24.000 --> 53:28.000
<v Kartikkumar>so don't know whether this is the right forum to ask as well the PA

53:28.000 --> 53:32.000
<v Anand>anyone?
<v Kartikkumar>in PA right there were a couple of questions where there were issues and we

53:32.000 --> 53:36.000
<v Kartikkumar>submitted those issues but they were not resolved and uh I'm

53:36.000 --> 53:40.000
<v Anand>depending on the
<v Kartikkumar>a bit worried whether we lost marks on those or

53:40.000 --> 53:44.000
<v Anand>issue. What uh what
<v Kartikkumar>not. So for the uh the the

53:44.000 --> 53:48.000
<v Kartikkumar>one where we had to trade those services, right? Uh the deadline was extended.

53:48.000 --> 53:52.000
<v Kartikkumar>But even though a lot of us had uh recorded our services, it never showed up on

53:52.000 --> 53:56.000
<v Anand>issues uh correct? It will show after
<v Kartikkumar>the

53:56.000 --> 54:00.000
<v Anand>the data.
<v Kartikkumar>marketplace. But then how would we trade if it

54:00.000 --> 54:04.000
<v Kartikkumar>doesn't because uh I think the deadline for the person somebody to buy

54:04.000 --> 54:08.000
<v Anand>Okay, there will be round
<v Kartikkumar>it is already crossed if I'm not

54:08.000 --> 54:12.000
<v Anand>two. Yeah. uh there is no possibility of
<v Kartikkumar>wrong. Oh, okay. Okay. Okay.

54:12.000 --> 54:16.000
<v Anand>changing dates in the way the platform is structured. I'd have to redo the entire

54:16.000 --> 54:20.000
<v Anand>thing. So round two is meant exactly for that. So yes, this is

54:20.000 --> 54:24.000
<v Anand>uh part of the design and I will call it fixed or not fixed. But yes, there will be

54:24.000 --> 54:28.000
<v Anand>a round two and round two have slightly different
<v Kartikkumar>Okay. Okay. Okay. Thank

54:28.000 --> 54:32.000
<v Anand>rules. I'm still trying to figure out what the rules are. See right now uh the

54:32.000 --> 54:36.000
<v Anand>exchange mechanism was uh people

54:36.000 --> 54:40.000
<v Anand>saying get marks for exchanging stuff. So
<v Kartikkumar>you.

54:40.000 --> 54:44.000
<v Anand>somebody offer a service I don't care what the service is let's

54:44.000 --> 54:48.000
<v Anand>exchange uh and therefore get marks which is perfectly

54:48.000 --> 54:52.000
<v Anand>fine but what does it take to create an economy that actually does useful

54:52.000 --> 54:56.000
<v Anand>stuff and I'm trying to design around

54:56.000 --> 55:00.000
<v Anand>to so that the service actually produces something useful for

55:00.000 --> 55:04.000
<v Kartikkumar>Got it. Uh there was

55:04.000 --> 55:08.000
<v Anand>someone
<v Kartikkumar>another issue that I ran into which was related to the pictures. Uh you had

55:08.000 --> 55:12.000
<v Kartikkumar>to draw something and the AI had to recognize it,

55:12.000 --> 55:16.000
<v Kartikkumar>right? And then somebody else had to rank it saying that you know which ones are good

55:16.000 --> 55:20.000
<v Kartikkumar>and I think this is uh for the PA question seven uh

55:20.000 --> 55:24.000
<v Kartikkumar>picture stories. Yeah. So I the the competition

55:24.000 --> 55:28.000
<v Kartikkumar>for that has ended but uh we are not able to paste the the

55:28.000 --> 55:32.000
<v Anand>Yeah. Yes. Yes. citizen I
<v Kartikkumar>final token. I mean it still says the result is not final

55:32.000 --> 55:36.000
<v Anand>have to do the AI evaluation and people have to do the
<v Kartikkumar>yet.

55:36.000 --> 55:40.000
<v Anand>uh human part of the evaluation.
<v Kartikkumar>Okay.

55:40.000 --> 55:44.000
<v Kartikkumar>Okay. And we were expected to hand draw it, right? Like draw it hand or

55:44.000 --> 55:48.000
<v Anand>Yes, I think
<v Kartikkumar>like not generate the pictures from

55:48.000 --> 55:52.000
<v Anand>I had imposed that
<v Kartikkumar>AI.

55:52.000 --> 55:56.000
<v Kartikkumar>Okay. It says write your story

55:56.000 --> 56:00.000
<v Anand>restriction.
<v Kartikkumar>privately. I three lines of situation and upload one. Yeah. Draw it,

56:00.000 --> 56:04.000
<v Kartikkumar>photograph a drawing on a paper, make a collage and or use

56:04.000 --> 56:08.000
<v Kartikkumar>an AI tool. Okay. Because when I evaluated most of the people had used AI tools.

56:08.000 --> 56:12.000
<v Anand>Perfectly fine with that. Actually, the point is, can
<v Kartikkumar>So all right.

56:12.000 --> 56:16.000
<v Anand>you craft a good story that meets these uh

56:16.000 --> 56:20.000
<v Anand>criteria? Provide paper and pen

56:20.000 --> 56:24.000
<v Anand>versus uh
<v Kartikkumar>Okay. Yeah, easier to do it in an AI tool. It's much more

56:24.000 --> 56:28.000
<v Kartikkumar>challenging if you draw it. I mean, I drew it. So, but then I saw a lot of people

56:28.000 --> 56:32.000
<v Kartikkumar>had actually created it using uh AI, which is sort of

56:32.000 --> 56:36.000
<v Anand>Yeah,
<v Kartikkumar>easier to

56:36.000 --> 56:40.000
<v Anand>I didn't differentiate the method of drawing.
<v Kartikkumar>do.

56:40.000 --> 56:44.000
<v Anand>I didn't in this particular case have any point of view,
<v Kartikkumar>Okay.

56:44.000 --> 56:48.000
<v Anand>but it's something to think about. Let me think about whether

56:48.000 --> 56:52.000
<v Anand>No, I don't think drawing is a skill that I want to teach in tedious settings at

56:52.000 --> 56:56.000
<v Anand>least. It's a useful skill but

56:56.000 --> 57:00.000
<v Anand>drawing a picture story in itself I would re

57:00.000 --> 57:04.000
<v Anand>prioritize in this course but I will

57:04.000 --> 57:08.000
<v Anand>think if there are specific reasons why drawing comes hand drawing comes

57:08.000 --> 57:12.000
<v Jaideep>Uh I mean
<v Anand>Yes, it

57:12.000 --> 57:16.000
<v Jaideep>uh can we uh monitor the number of key

57:16.000 --> 57:20.000
<v Jaideep>strokes uh sorry the handstrokes that are being used for

57:20.000 --> 57:24.000
<v Jaideep>drawing and then uh we can evaluate how good the agent
<v Anand>is.

57:24.000 --> 57:28.000
<v Jaideep>is. Let the agents draw and uh with the least amount of key

57:28.000 --> 57:32.000
<v Jaideep>strokes possible.
<v Anand>It's an

57:32.000 --> 57:36.000
<v Jaideep>Yeah. Uh I mean how well they can direct

57:36.000 --> 57:40.000
<v Jaideep>their agents to get this work
<v Anand>optimiz but no that is the thing it is more

57:40.000 --> 57:44.000
<v Anand>agent intelligence given the question and the optimization

57:44.000 --> 57:48.000
<v Anand>criterion the agent will converge for solution

57:48.000 --> 57:52.000
<v Anand>I not entirely sure that will require human steering

57:52.000 --> 57:56.000
<v Anand>but again if you can craft a that requires that. Let's take a

57:56.000 --> 58:00.000
<v Jaideep>done. Yeah. Uh maybe we can uh do an uh internal

58:00.000 --> 58:04.000
<v Jaideep>level of that and then uh decide whether it's worth uh putting
<v Anand>look.

58:04.000 --> 58:08.000
<v Jaideep>out.
<v Anand>Yeah. Okay.

58:08.000 --> 58:12.000
<v Anand>Okay. Maybe time for one last question if anyone

58:12.000 --> 58:16.000
<v Adwitiya>Okay sir. So not a question just a suggestion once again on the same topic

58:16.000 --> 58:20.000
<v Anand>else.
<v Adwitiya>that we were discussing. So as long as the context is an ROE

58:20.000 --> 58:24.000
<v Adwitiya>and not a traditional GA. So the idea becomes

58:24.000 --> 58:28.000
<v Adwitiya>to teach something and also to I mean it's rather a crude way of putting it

58:28.000 --> 58:32.000
<v Adwitiya>but to get us to waste time so that yeah so that we are stuck

58:32.000 --> 58:36.000
<v Adwitiya>on one question whereas we could be scoring on another one. So that teaches how

58:36.000 --> 58:40.000
<v Anand>Priority is correct.
<v Adwitiya>to teaches us how to prioritize right correct. So

58:40.000 --> 58:44.000
<v Adwitiya>in this case what we could uh do is draft a question that

58:44.000 --> 58:48.000
<v Adwitiya>has an external file attached to it. For example uh there was one

58:48.000 --> 58:52.000
<v Adwitiya>question I'm just forgetting the GA but uh actually there were two three

58:52.000 --> 58:56.000
<v Adwitiya>questions where we had a separate file type a JSON or something which we had to download

58:56.000 --> 59:00.000
<v Adwitiya>and paste it along with the prompt to an agent. Now I doubt

59:00.000 --> 59:04.000
<v Adwitiya>whether even 1% of us would have you know actually opened the file and checked what

59:04.000 --> 59:08.000
<v Adwitiya>the JSON looked like or what the contents were because just downloading it

59:08.000 --> 59:12.000
<v Adwitiya>and pasting it becomes a more of a convenience

59:12.000 --> 59:16.000
<v Adwitiya>right. So now what if what we do is

59:16.000 --> 59:20.000
<v Anand>Agreed. No inter but uh
<v Adwitiya>give

59:20.000 --> 59:24.000
<v Anand>I align with the theme of your question and yes

59:24.000 --> 59:28.000
<v Anand>this is the kind of question that makes it somewhat agent pros.

59:28.000 --> 59:32.000
<v Anand>It is however susceptible to collaboration. one

59:32.000 --> 59:36.000
<v Anand>student gets it and puts it on the WhatsApp group and

59:36.000 --> 59:40.000
<v Anand>says guys add this you know telling

59:40.000 --> 59:44.000
<v Anand>it don't uh try too hard on question

59:44.000 --> 59:48.000
<v Anand>it uh becomes yeah less

59:48.000 --> 59:52.000
<v Anand>annoyable now having said that though it is possible to revise this in

59:52.000 --> 59:56.000
<v Anand>a way where even with that some amount of discretion is

59:56.000 --> 01:00:00.000
<v Anand>required if buried within it somewhere in a slightly hard

01:00:00.000 --> 01:00:04.000
<v Anand>for an agent to fashion. there is

01:00:04.000 --> 01:00:08.000
<v Anand>uh there is a data set or if the

01:00:08.000 --> 01:00:12.000
<v Anand>uh information were hidden in a strange let's say an

01:00:12.000 --> 01:00:16.000
<v Anand>audio
<v Adwitiya>But sir, we don't have to hide the information.

01:00:16.000 --> 01:00:20.000
<v Anand>encoding that is my information
<v Adwitiya>That was my see these days white text and all

01:00:20.000 --> 01:00:24.000
<v Adwitiya>is completely like AI is at an advanced stage that simply changing the

01:00:24.000 --> 01:00:28.000
<v Anand>correct but what if it was
<v Adwitiya>color of the text won't actually serve the purpose anymore. But

01:00:28.000 --> 01:00:32.000
<v Anand>not text if
<v Adwitiya>now okay

01:00:32.000 --> 01:00:36.000
<v Anand>it's in
<v Adwitiya>so that is a great idea but I was

01:00:36.000 --> 01:00:40.000
<v Adwitiya>thinking along the lines of that even in human cases if we interrupt the

01:00:40.000 --> 01:00:44.000
<v Adwitiya>stream of work and change a step that was carried on in the earlier

01:00:44.000 --> 01:00:48.000
<v Adwitiya>tasks. What I mean is let me say that there's a very complex

01:00:48.000 --> 01:00:52.000
<v Adwitiya>problem at the start of the statement I tell ni in the PDF for the attached

01:00:52.000 --> 01:00:56.000
<v Adwitiya>file is the part of the problem. Okay. So in the file at the start we tell

01:00:56.000 --> 01:01:00.000
<v Adwitiya>it x is equals to 10 y is equals to 5 and then you perform these

01:01:00.000 --> 01:01:04.000
<v Adwitiya>steps. There's a complex calculation and all of that but at the end we go edit

01:01:04.000 --> 01:01:08.000
<v Adwitiya>one x was equal to 5.5. So that would take a longer

01:01:08.000 --> 01:01:12.000
<v Adwitiya>time because it'll have to recalculate everything. And on top of that, a student who

01:01:12.000 --> 01:01:16.000
<v Adwitiya>actually reads the question would just modify the prompt and get an answer in

01:01:16.000 --> 01:01:20.000
<v Anand>Good idea. Uh what you're saying
<v Adwitiya>less

01:01:20.000 --> 01:01:24.000
<v Anand>is edits are

01:01:24.000 --> 01:01:28.000
<v Anand>tougher. Uh that is a
<v Adwitiya>time for the AI.

01:01:28.000 --> 01:01:32.000
<v Anand>good premise and So

01:01:32.000 --> 01:01:36.000
<v Anand>I will take that in. Okay. Sorry, I need to run but I will take that

01:01:36.000 --> 01:01:40.000
<v Anand>in. Thank you.