WEBVTT
00:00:00.133 --> 00:00:06.266
Welcome to the 2025 podcast is our first, podcast in 2025.
00:00:06.266 --> 00:00:10.699
So myself Krithi and this is Chetan and this is Guru.
00:00:10.699 --> 00:00:28.365
Okay, so so we'll be starting off with the podcast, So, sir there's been a lot happening at Quoqo and especially, outside the legal space So basically the Qfin and Syn-Apt seem to be making waves, oh, so what's the perspective of yours on this?
00:00:28.365 --> 00:00:42.399
I think before we get into this, Guru I think in 2025, it started with a big bang correct ... more In terms of AI and events across AI and also, events at Quoqo Correct.
00:00:42.399 --> 00:00:48.865
And I think we should sort of like give viewers a perspective about what's happening in the world.
00:00:48.865 --> 00:01:01.765
I'm sure people have, heard of Deepseek and, what's going on in China and, how it is open source and is cheaper and can be so many things and things like that.
00:01:01.765 --> 00:01:09.965
So, I mean, where we left off with Code and Counsel in 2024, was we were we were not even at an open level.
00:01:09.965 --> 00:01:10.433
Correct.
00:01:10.433 --> 00:01:13.965
When it came to sort of OpenAI and things like that.
00:01:13.965 --> 00:01:28.165
But I mean, just to give, viewers a perspective in terms of large language models, in the size of these, if you go to Hugging face, Guru or if you, generally investigate, you will find a few large players, right?
00:01:28.165 --> 00:01:31.933
got obviously OpenAI, which is open or close sourced ...
00:01:31.933 --> 00:01:33.566
depends on whom you ask.
00:01:33.566 --> 00:01:36.965
If you ask Elon Musk ... he will say, I'm sure it's open source.
00:01:36.965 --> 00:01:40.299
But, anyway, there's, Claude by Anthropic.
00:01:40.299 --> 00:01:44.132
The 3.5 sonnet is is a super LLM ... right?
00:01:44.132 --> 00:01:55.799
And, then you've got Mistral, then you've got Microsoft's phi some math specific modules, there are many, many correct for all of these things.
00:01:55.799 --> 00:02:06.433
And then you have, the ability to run some of these locally, you know, using Ollama, whatever you need ... Qwen And I mean, you know, by the way, China also has its own models.
00:02:06.433 --> 00:02:07.566
You know, you've got Qwen ...
00:02:07.566 --> 00:02:13.233
you've got, Deepseek V2, you know, which came out, and the hints around some of these things as well.
00:02:13.233 --> 00:02:22.332
But, what's happening, I believe, is, you know, depending on how you encode, LLM can either be, text generation or reasoning LLM.
00:02:22.332 --> 00:02:27.432
And most of the OpenAI ones are reasoning LLM and things like these.
00:02:27.432 --> 00:02:41.432
So, and what Deepseek has come up is potentially, an O1 competitor and, you know, we, you know, I told you like, the day it got launched thing that I tested this and I thought that, you know, it may or may not be production ready yet.
00:02:41.432 --> 00:02:58.665
At least, you know, from where we, looking at things and we might need something much more, you know, and I thought that you know, and my first, inclination was, OpenAI still had a lot of edge over some of these things from a production perspective in terms of, our products and things like that.
00:02:58.665 --> 00:03:02.199
Why I'm giving this context is how things have changed.
00:03:02.199 --> 00:03:18.199
You know, you've suddenly gone from a paid model and deployed only on Azure and to, fully open source model where you can deploy on, on your servers provided you have the ability to sort of like deploy these.
00:03:18.199 --> 00:03:26.699
But, you know, most people don't also realize saying that if I want to I mean, you have different quantized versions of Deepseek as well.
00:03:26.699 --> 00:03:27.466
Okay.
00:03:27.466 --> 00:03:34.400
You may be able to run like a 1 - 1.2 Billion on a local comp, let's say, an eight core GPU.
00:03:34.400 --> 00:03:45.832
But you know, if you want the full length you know, if you want like the 64 billion or whatever that run it with or 72, you need some real, really powerful machines.
00:03:45.832 --> 00:03:46.165
Correct.
00:03:46.165 --> 00:03:55.233
So it's not something which, saying, is it production ready, is it something that you can sort of, like immediately adopt it?
00:03:55.233 --> 00:04:01.665
And, I would not necessarily sort of like immediately jump at some of these things. Correct.
00:04:01.665 --> 00:04:03.765
It really depends on the context.
00:04:03.765 --> 00:04:17.333
You know, then, you know, definitely take some time to understand how the models are performing, how it can be used for certain applications, etc., but at least Deepseek has been a ground breaking and it shatters all beliefs even the mindsets of people.
00:04:17.333 --> 00:04:20.100
All you need a lot of resources to train a model.
00:04:20.100 --> 00:04:29.132
Oh, it's not possible for anyone outside the Europe or the USA to develop models So if something that got shattered, going in the direction I think, you know.
00:04:29.132 --> 00:04:29.733
Yeah, yeah.
00:04:29.733 --> 00:04:45.165
So, but you know, this I was talking to, know, a VC recently, and I mentioned this thing, and I had read somewhere saying and it was a proper sort of like encapsulation of the moment thing, an LLM has become a commodity. Correct.
00:04:45.165 --> 00:04:56.432
And it's basically on top of what you build is what really matters and what other what you build is essentially like a wrapper, but you know, you don't need a VC to what I've had happen, right?
00:04:56.432 --> 00:04:59.033
You know, you don't need it requires an adoption.
00:04:59.033 --> 00:05:02.966
That's what I, I think that's what we've done for Quoqo for a long time.
00:05:02.966 --> 00:05:06.632
But, so does it change the mechanics overnight?
00:05:06.632 --> 00:05:10.699
I'm saying does it make it accessible or is it something that, that I can do it?
00:05:10.699 --> 00:05:19.966
before Deepseek maybe not at a level of a reasoning model, but, there are others as well which did the tasks pretty competently Correct.
00:05:19.966 --> 00:05:23.399
We have benchmarked, we've got our own QLLM which we did.
00:05:23.399 --> 00:05:30.165
We did not need, huge, quantized model to be, utilized we trained on some things.
00:05:30.165 --> 00:05:40.300
And we thought that whatever the current state of, requirements are you know, from a user in a production scenario perspective was more than sufficient with what we have now.
00:05:40.300 --> 00:05:58.466
Right now, if you if you're looking at a PhD, there were like comments just yesterday, OpenAI came up with something, very, very for a GPT pro user spending like $200 a month, you know, you've got like a PhD level, output window that, research reports and.
00:05:58.466 --> 00:06:04.065
Yeah, but then, you know, human ingenuity is something that you may not be able to replace.
00:06:04.065 --> 00:06:22.966
And I think Deepseek if you read that paper and you know the way that they've gone about sort of like, you know, adopting a non cuda, architecture to be able to sort of human sort of like develop that model, you know, gone above and sort of like use a very different way of looking at things.
00:06:22.966 --> 00:06:23.399
Correct?
00:06:23.399 --> 00:06:28.432
as you know, like if you're using GPUs, these are usually Nvidia ones.
00:06:28.432 --> 00:06:41.632
the reason why AMD is also been left behind is, you know, the world adopts Cuda architecture for, programing and, you don't find the same adoption in AMD or an Intel GPU kind of like a scenario, right.
00:06:41.632 --> 00:06:44.865
So now you have gone about and created something else.
00:06:44.865 --> 00:06:45.199
Correct.
00:06:45.199 --> 00:07:03.333
Which can work in parallel, you know, looks at Cuda as a separate, if it looks at a GPU as a separate, sort of like an entity from which it can tap in and tap out I mean, this is at a very simplistic level, and not that I also that I'm an expert at how this how they've come about doing this.
00:07:03.333 --> 00:07:16.466
But this involves sort of like, some, some element of deep research But how does this all sort of like add up and Kriti’s question correct, saying, got a lot of things that are coming up.
00:07:16.466 --> 00:07:19.766
And what is it, you know, you've got some new products also coming up.
00:07:19.766 --> 00:07:20.165
Correct.
00:07:20.165 --> 00:07:21.899
And what is it that, mentioned?
00:07:21.899 --> 00:07:37.466
Qfin and I Syn-apt So just to give users, an overview, if you go to our website www.quoqo.com then at that point in time you will sort of, you know, there are links to non legal tech applications.
00:07:37.466 --> 00:07:44.399
So Qfin is is a set of financial agents, almost 30 or 32 of them that we built using our expertise on.
00:07:44.399 --> 00:07:49.865
Q bot, that solves things like, financial consolidation of subsidiaries as an example.
00:07:49.865 --> 00:07:50.632
Okay.
00:07:50.632 --> 00:07:52.000
It's a huge problem.
00:07:52.000 --> 00:08:01.300
And also, you know, especially in large enterprises, it also works with audits sort of, with with auditing requirements in mind.
00:08:01.300 --> 00:08:13.899
So there are quite a few, you can also visit the sub-site qfin.guru So and then and then look at it Syn-apt is also on the agentic AI process.
00:08:13.899 --> 00:08:23.665
we can assist with some custom development of, agentic AI Very specific to a company as requirements as well, during development of Qbot what if you remember?
00:08:23.665 --> 00:08:33.765
Qbot is, specific to legal tech in terms of having a conversational interface for legal specific data sets, right.
00:08:33.765 --> 00:08:36.732
So you've got Qfin, which is also agentic AI Okay.
00:08:36.732 --> 00:08:41.799
Not exactly in the same lines of Qbot but Syn-apt, goes a little bit further.
00:08:41.799 --> 00:08:42.232
Correct.
00:08:42.232 --> 00:08:44.133
And uses different models.
00:08:44.133 --> 00:08:48.966
and different agents to be able to accomplish a number of different things as well.
00:08:48.966 --> 00:08:57.932
Qfin is advanced in the sense that, it contains a lot of technology and a lot of different LLMs to be able to accomplish what it does.
00:08:58.932 --> 00:09:08.033
When you're doing financial apps, as you know, Guru our agents, the, accuracy is a lot more important.
00:09:08.033 --> 00:09:08.899
Yeah. Important.
00:09:08.899 --> 00:09:09.200
Correct.
00:09:09.200 --> 00:09:17.299
You try and run a LLama I think .2 latest on and mathematical sort of like reasoning, you know, like, you know, it doesn't do well.
00:09:17.299 --> 00:09:17.799
Okay.
00:09:17.799 --> 00:09:26.066
So, and it's not I mean, you mentioned Llama, but and depends on, and we have our own thoughts around these.
00:09:26.066 --> 00:09:31.566
I don't want to make these because on these, proprietary and how we test some of these things as well.
00:09:31.566 --> 00:09:41.232
if you remember, in the early days of, GPT2 you could convince, the GPT that two equals one, you know, two equals zero.
00:09:41.232 --> 00:09:47.000
And you could just load it up with so many things that it's not like.
00:09:47.000 --> 00:10:01.033
I mean, I would be there, but, I wanted to give users the ability to try out Syn-apt and Qfin these are cutting edge, agent AI and, and fintech specific products.
00:10:01.033 --> 00:10:08.299
These are as cutting edge as probably Deepseek you know, when it comes to functional sort of like use cases.
00:10:08.299 --> 00:10:13.799
And we believe like we are the first ones in the world sort of like create all of these things as well.
00:10:13.799 --> 00:10:28.166
And by the way, you know, with, with Q fin, we can integrate into a number of, different architectures in terms of, also integrations, things like these, so that you can just plug and play.
00:10:28.166 --> 00:10:30.265
So this is our new thing.
00:10:30.265 --> 00:10:37.033
Same, Qfin and Syn-apt new things for Quoqo and developed.
00:10:37.033 --> 00:10:53.100
by utilizing the knowhow we do things on the legal tech side, I wish some of you And by the way Guru you want to speak about how enterprises are adoption these two things Yeah, definitely.
00:10:53.100 --> 00:11:11.100
You had demos with a few large companies you know, you know, like the day after, when they were looking at these use cases and they can, you know, decided evolution, all little, you know, all of them went to their own, yeah.
00:11:11.100 --> 00:11:16.200
And how they can, use something because most of these activities are really valuable to them.
00:11:16.200 --> 00:11:24.033
And, you know, the, the level of some of these, you know, artificial them will be very interesting, you know?
00:11:24.033 --> 00:11:24.466
Yeah.
00:11:24.466 --> 00:11:31.966
And, on that doesn't mean that it's probably going to be deployment soon as been. Okay.
00:11:31.966 --> 00:11:53.799
And, so, we had the opportunity to demonstrate a very large, in financial services company that did the work being done by about 24 analysts, you know, is essentially sort of like can be automated to an extent, with, lot more efficiency.
00:11:53.799 --> 00:12:07.865
And, you can and, and also it is quite a, I could execute tasks within, say, 20, 25 minute window, which is like, just like an eternity in computer processing times.
00:12:07.865 --> 00:12:08.700
Correct.
00:12:08.700 --> 00:12:16.033
Which, people would take like that twenty four people would take like a month to execute, essentially. Right.
00:12:16.033 --> 00:12:23.033
So it's reaching a phase saying, you know, it is possible to sort of like bring in a lot of efficiency.
00:12:23.033 --> 00:12:25.000
Cost savings. Right.
00:12:25.000 --> 00:12:27.865
Into the process.
00:12:27.865 --> 00:12:32.166
Not that, you know, that the human element is taken out of the mix, right?
00:12:32.166 --> 00:12:39.133
But the human element can be focused on better task of analyzing what comes out of it, rather than trying to dig for the information.
00:12:39.133 --> 00:12:56.700
You know, so I you yeah, I think the, the better way to do this is if you remember, Alta Vista which was the forerunner of, Google you know, if you try to run search in Alta Vista that used to be a big thing, like saying that you are able to sort of like look at stuff and search for it.
00:12:56.700 --> 00:13:01.332
I think that is the that is the use case in many enterprises right now.
00:13:01.332 --> 00:13:07.399
So the way they do things is at least about 35 to 40 years old. It has not been changed.
00:13:07.399 --> 00:13:10.732
And for a long time.
00:13:11.332 --> 00:13:29.633
And when you expose on these things, it's basically like it's like, it's a paradigm shift in thinking saying, oh, there are better ways to do things because as technology advances, reallocate resources, time and money, correct, for all of these things as well?
00:13:30.633 --> 00:13:46.966
So that is, that is how some of the agenitc AI systems are also sort of like scaling up, but sorry, it's a longwinded and sort of like, long monologue from my think it just also outlined how, how much we are using these applications in our day to day.
00:13:46.966 --> 00:13:48.432
activities.
00:13:48.432 --> 00:13:50.000
yeah, that's an important thing to highlight.
00:13:50.000 --> 00:13:59.500
Also, I think within Quoqo and Krithi you need to speak how you are using some of the, some of the agents that you have correct in your day to day work as well. Right?
00:13:59.500 --> 00:14:12.432
So for that, so like you mentioned, like the deekseek and o1 So I, I'm hearing many things from many people that, you know, they're comparing both of them like what is the main difference.
00:14:12.432 --> 00:14:25.299
between deepseek and openAi I think basically Deepseek is also trained on openAI LLMs.
00:14:26.566 --> 00:14:46.533
you know, responses, you know, they claim that only about 830 of them were used, from an API Correct, to the level that they use a different way in which, they use, reinforcement learning. Correct.
00:14:46.533 --> 00:14:55.633
They use a method called called Mixture of Experts it's like a developing smaller LLMs and integrating them with those bigger umbrella.
00:14:55.633 --> 00:15:04.633
And based on the question that you ask, it basically, you know, activates that particular agent, which, gives you the more pointed response.
00:15:04.633 --> 00:15:10.533
That's always the so would it be right to say Guru is like, you know, I remember root learning nowadays in schools.
00:15:10.533 --> 00:15:12.100
That is not encouraged.
00:15:12.100 --> 00:15:17.666
But let's say you are to root learn tables from 1 to 20, correct.
00:15:17.666 --> 00:15:22.399
That is, that is used to basically develop your first, language model.
00:15:22.399 --> 00:15:23.533
Correct.
00:15:23.533 --> 00:15:25.500
And then you know how to multiply, right?
00:15:25.500 --> 00:15:34.265
Then you teach a child or, computer to go to the 21st table or the 35th table.
00:15:34.265 --> 00:15:35.000
Correct. Right.
00:15:35.000 --> 00:15:40.700
And then you extended, you know, and then it figures out that you can do it, you know, for infinity.
00:15:40.700 --> 00:15:41.500
Correct as well.
00:15:41.500 --> 00:15:42.133
Correct.
00:15:42.133 --> 00:15:47.033
So and the way that you can learn the table is multiple ways you can do an addition.
00:15:47.033 --> 00:15:48.566
You can do multiplication.
00:15:48.566 --> 00:15:49.365
Correct.
00:15:49.365 --> 00:15:50.732
Can use squared methods.
00:15:50.732 --> 00:16:02.533
You know depending on so that these are so Deepseek basically explored one of the methods similar to this in which it's like a thinking Method too. Is that right Guru Is that right.
00:16:02.533 --> 00:16:06.166
Another way of looking at it is let us say, there is a closed room in which there are multiple experts.
00:16:06.166 --> 00:16:09.466
The finance guy and a marketing guy and there is a sales guy.
00:16:09.466 --> 00:16:13.700
And, you know, all this, you know, all these different types of people.
00:16:13.700 --> 00:16:15.033
Now you are asking question.
00:16:15.033 --> 00:16:21.765
standing outside the room and you will get the correct answers, because the person who is actually having expertise in that field is going to answer you.
00:16:21.765 --> 00:16:43.432
So that's basically how you are able to sort of like, yeah, but you know, earlier you do is instruction based, sort of like, you know, your you would have to change the context for the LLM or through prompting to sort of like put it in a different frame and make sure that and but now even the LLMs also sort of like you have the ability to have memory.
00:16:43.432 --> 00:16:55.000
So it's all and also one of the things is, you know, if you use, for instance, like 3.5, the early version of OpenAI ... not the turbo one correct, as an example from OpenAI. Correct.
00:16:55.000 --> 00:16:59.000
You know, it would quickly lose context, right, right.
00:16:59.000 --> 00:17:00.500
And then it forget.
00:17:00.500 --> 00:17:07.099
So, so all of those things are being tackled on a, much but on a much better way. Yeah.
00:17:07.099 --> 00:17:13.665
And have larger context windows that have better output, output token sizes.
00:17:13.665 --> 00:17:13.965
Okay.
00:17:13.965 --> 00:17:14.732
I will come to that.
00:17:14.732 --> 00:17:16.066
I have my own things on that.
00:17:16.066 --> 00:17:17.299
Yeah, yeah.
00:17:17.299 --> 00:17:23.032
It's because I think it's is it's very hard to implement in the local production.
00:17:23.032 --> 00:17:23.500
Yeah.
00:17:23.500 --> 00:17:26.400
I, I have my own views on this. Correct.
00:17:26.400 --> 00:17:27.865
Do you want to go first Guru?
00:17:27.865 --> 00:17:28.665
or should I.
00:17:28.665 --> 00:17:28.932
No.
00:17:28.932 --> 00:17:29.500
Go ahead, go ahead.
00:17:29.500 --> 00:17:30.232
Okay okay.
00:17:30.232 --> 00:17:33.665
So see when you deploy locally. Correct.
00:17:33.665 --> 00:17:37.532
You are assuming that you have servers.
00:17:37.532 --> 00:17:37.900
Correct.
00:17:37.900 --> 00:17:41.066
Which are capable of running an LLM.
00:17:41.066 --> 00:17:47.432
So let us say you remember the, you know, the logerithmic skale in school.
00:17:47.432 --> 00:17:47.799
Correct.
00:17:47.799 --> 00:17:50.566
So you have let's take the number two.
00:17:50.566 --> 00:17:55.965
You have to two squared, 2 cube, Two to the power of four to the power of eight and so on. Right.
00:17:55.965 --> 00:18:00.965
So if you take two squared, you know, you get a smaller number.
00:18:00.965 --> 00:18:04.865
If you take to sort of eight it's going to be large.
00:18:04.865 --> 00:18:15.965
If you take to the power of zero it becomes one The point being that these are all quantizations Correct.
00:18:15.965 --> 00:18:21.400
So the larger so you can take a very big model and take a sample of it.
00:18:21.400 --> 00:18:24.932
and you can run it on a smaller level computer.
00:18:24.932 --> 00:18:29.532
It only is like I only know two to the power of two. Correct.
00:18:29.532 --> 00:18:34.900
But if you take two to the power of 64 or whatever, I'm just simplifying this.
00:18:34.900 --> 00:18:43.432
You have a very large number to be able to compute If you do two to the power of 64 on your calculator app, on your computer will give you an error.
00:18:43.432 --> 00:18:44.799
It will give you an infinity error.
00:18:44.799 --> 00:18:45.165
Correct.
00:18:45.165 --> 00:18:55.099
So because beyond a certain point, the floating points it goes beyond you know, it goes beyond the possibility to create such a large number.
00:18:55.099 --> 00:18:56.000
Correct.
00:18:56.000 --> 00:19:03.465
But what happens is at that scale, you have a lot more information easily, which is not compressed.
00:19:03.465 --> 00:19:08.365
So for from the LLM’s perspective, you can access that information.
00:19:09.165 --> 00:19:20.432
there was a visualization from China about how an LLM then sort of like upgrades, different layers and how it's able to train and figure out what it is. Right.
00:19:20.432 --> 00:19:24.465
You know, in an uncompressed model, it has access to all of these.
00:19:24.465 --> 00:19:34.165
And then basically depends on the GPU's power to sort of like, chunk out and, be able to arrive at the computation that makes sense.
00:19:34.165 --> 00:19:34.532
Correct.
00:19:34.532 --> 00:19:40.865
So, it is like dealing with uncompressed files and dealing with compressed files.
00:19:40.865 --> 00:19:41.532
Yeah, yeah.
00:19:41.532 --> 00:19:46.900
And if you need to deal with a compressed file, it's going to take time for you to extract kind of open it.
00:19:46.900 --> 00:19:50.000
It may not have the same color accuracy.
00:19:50.799 --> 00:20:00.099
it may have removed a bunch of repeating colors as an example in a graphic and things like that.
00:20:00.099 --> 00:20:00.833
Right.
00:20:00.833 --> 00:20:05.133
But, you know, in an uncompressed format, correct.
00:20:06.799 --> 00:20:17.400
every pixel is a representation of an actual graphic Our video editors, as an example, color accuracy is very, very critical for them.
00:20:17.400 --> 00:20:20.165
So they have to deal with uncompressed formats.
00:20:20.165 --> 00:20:25.066
And when you actually view this podcast, it's a compressed version of whatever it is.
00:20:25.066 --> 00:20:26.665
It's not the original file.
00:20:26.665 --> 00:20:32.432
Similar to the LLMs, you will have differences in quality when you run all of these things as well.
00:20:32.432 --> 00:20:36.599
Now to back to your question, saying, is it easy to run all of these things?
00:20:36.599 --> 00:20:43.200
What I said is If you want to run the uncompressed version, it's going to take some very serious hardware.
00:20:45.865 --> 00:20:50.700
you need to have the proper servers, the GPUs, the data storage.
00:20:50.700 --> 00:20:55.465
And it's also a question of how fast do all of these can, do a token output.
00:20:55.465 --> 00:20:56.200
Correct.
00:20:56.200 --> 00:21:02.900
If you are in a production grade environment where people sort of, let's say you get a token output of 20 tokens.
00:21:02.900 --> 00:21:03.833
Correct.
00:21:03.833 --> 00:21:08.066
But your output window is, 6000 tokens.
00:21:08.066 --> 00:21:09.732
How long is it going to take?
00:21:09.732 --> 00:21:15.032
You figure 20 tokens per second to figure out something, which is $6,000?
00:21:15.032 --> 00:21:18.932
Each token and say is about 13, 14 letters, correct?
00:21:18.932 --> 00:21:21.000
You know, on an average.
00:21:21.000 --> 00:21:22.833
So it's going to take some time.
00:21:22.833 --> 00:21:23.066
Okay.
00:21:23.066 --> 00:21:28.532
If you use our products If you take Quoqo products, it's literally instantaneous.
00:21:28.532 --> 00:21:28.833
Correct.
00:21:28.833 --> 00:21:32.000
Because we have the infrastructure to run some of these things as well.
00:21:32.000 --> 00:21:41.333
So but the short answer is unless you have the hardware necessary to run all of these things, and the hardware keeps progressing.
00:21:41.333 --> 00:21:41.833
Correct.
00:21:41.833 --> 00:21:46.900
It's not that older hardware can't run or newer hardware is the only way forward.
00:21:46.900 --> 00:21:51.932
Really depends on what kind of application that you want and to whom are catering to.
00:21:51.932 --> 00:22:00.766
You know, for someone who is sort of, know, is able to wait correct, then you know, it is not mission critical, then you can wait.
00:22:00.766 --> 00:22:03.333
You know, maybe it'll land in your inbox one hour later.
00:22:03.333 --> 00:22:04.232
As an example.
00:22:04.232 --> 00:22:05.099
Right. All right.
00:22:05.099 --> 00:22:07.633
And, and then the results come over the next day or so.
00:22:07.633 --> 00:22:14.833
That may be a case where you may want it on this You know, on, older less powerful slower servers, things like this.
00:22:14.833 --> 00:22:30.299
But then if you want to run this, production grade environment and conforming to Gen Z standards of wanting everything instantaneously, right, then you need something very, very spontaneous and very, very powerful infrastructure.
00:22:30.299 --> 00:22:33.299
Yeah, yeah, yeah, I think that's very much so.
00:22:33.299 --> 00:22:34.865
And also to add to that.
00:22:34.865 --> 00:22:35.599
Okay.
00:22:38.133 --> 00:22:46.200
depending on what you run you know, if you run things locally, you're also limited by something called an output window size.
00:22:46.200 --> 00:22:46.500
Correct.
00:22:46.500 --> 00:22:48.833
It really depends once again on the infrastructure.
00:22:48.833 --> 00:22:50.400
Yeah, you can change that.
00:22:50.400 --> 00:22:54.232
But you know, more stuff running locally.
00:22:54.232 --> 00:23:01.665
You know, if you use open source stuff to be able to run and then locally limited to over 2000 tokens, whether 2000 tokens.
00:23:01.665 --> 00:23:04.200
takes for ever to generate also depending on the hardware.
00:23:04.200 --> 00:23:05.032
Correct.
00:23:05.032 --> 00:23:17.032
And if you want, if you want like, OpenAI research tool to connect to generate like 15 pages, it's gone to take time even for a OpenAI it will take time Easily.
00:23:17.032 --> 00:23:17.833
Some 15 minutes.
00:23:17.833 --> 00:23:21.165
Ten minutes even for them So, yeah.
00:23:21.165 --> 00:23:33.165
And if some of those things I don't think locally can run, you know, a need like an entire, Sever farm to be able to sort of like, be able to handle complex tasks.
00:23:33.165 --> 00:23:47.532
Also, you know, for shorter things, like you want a conversational bot as a fancy to run locally, I think it's great to have, a locally running LLM, but anything more than, you know, as an example.
00:23:47.532 --> 00:23:47.799
Correct.
00:23:47.799 --> 00:24:04.200
We have like Diligence project We have a product called Diligence, where it can review so many things together and basically, you know, come with a conclusion or a help you arrive I want to conclusion or another or an analysis on a transaction as an example.
00:24:04.200 --> 00:24:04.900
Correct.
00:24:04.900 --> 00:24:08.165
It may not be possible to run it with current output windows.
00:24:08.165 --> 00:24:12.700
You know, it's like it's like getting cut off mid-sentence.
00:24:12.700 --> 00:24:14.465
Okay.
00:24:14.465 --> 00:24:16.200
So you'll get a report.
00:24:16.200 --> 00:24:22.133
But you know, what's going to happen is the computer is going to wait for some time for this, output to come through.
00:24:22.133 --> 00:24:27.633
And, you know, the computer will wait for the time allotted and then say, okay, this is it.
00:24:27.633 --> 00:24:29.500
And then it has to complete the task.
00:24:29.500 --> 00:24:32.032
You are dealing with it at the end of the day of machine. Correct.
00:24:32.032 --> 00:24:34.732
The machine has like a start loop and an end loop.
00:24:34.732 --> 00:24:37.833
And this is essentially how it works in real life.
00:24:37.833 --> 00:24:42.566
But, that's what I think Guru. Guru is a lot more, completely agree with.
00:24:42.566 --> 00:24:43.465
Yeah, yeah.
00:24:43.465 --> 00:24:44.532
So a different use case.
00:24:44.532 --> 00:24:47.099
It really depends on the use cases that you are working with.
00:24:47.099 --> 00:25:01.799
Do you want the repository kind of a tool that and AI repository like Quoqo legacy where you can upload your documents and you expect, you know, documents to be available, maybe with, next day or know a couple of days later.
00:25:01.799 --> 00:25:08.365
So now we can now, use local models with, very high level quantization. Correct.
00:25:08.365 --> 00:25:17.932
But then if you want instantaneous results, Chatbot or also some of those other use cases, some of these things, may not be suited for the use case, basically.
00:25:17.932 --> 00:25:18.400
Okay.
00:25:18.400 --> 00:25:25.066
So maybe that is the short response on the higher the quantization level, the lower the quality it becomes.
00:25:30.965 --> 00:25:41.500
Speaking about the, infrastructures and, Quoqo has been doing some cool things, like micro data, you know, endorser centers like throught Quoqo what you're supposed to get around.
00:25:41.500 --> 00:25:44.400
It's like it's the way, like this.
00:25:44.400 --> 00:25:51.099
So I think, you know, when Guru I doing when we distribute as well.
00:25:51.099 --> 00:25:51.833
Okay.
00:25:51.833 --> 00:25:56.133
And I mean, how difficult you know, run some of these things.
00:25:56.133 --> 00:26:22.066
And you also, you know, you either have ambani scale of, you know, hundreds of data center that is yet to be built, versus, you know, you have public cloud like Microsoft or AWS and, Guru is an expert, at, and I always these guru saying he is always over buget by 5X Using the cloud.
00:26:22.066 --> 00:26:22.965
Right.
00:26:22.965 --> 00:26:31.232
And and you know, what is lets say your are teching a model you want to, Guru my trainng said the other day was 6MB.
00:26:31.232 --> 00:26:38.500
Correct But, you know, if I put it through a public cloud in India, it cost us almost ₹55,000.
00:26:38.500 --> 00:26:39.566
Correct. That's done.
00:26:39.566 --> 00:26:42.333
I mean, there were so it was cleaned.
00:26:42.333 --> 00:26:47.432
it was done something, you know, I wanted some special stuff on it or whatever the case is, right?
00:26:47.432 --> 00:26:48.333
It never got.
00:26:48.333 --> 00:26:55.833
And then after having burned through all that, it came back saying that there's an error in the file which it could not compute. Correct.
00:26:55.833 --> 00:27:08.465
So it's not very easy to solve because public cloud infrastructure, though they everyone thinks is very easy to train your model, clean your data set and collect the data set and all of these things.
00:27:08.465 --> 00:27:09.432
Right.
00:27:09.432 --> 00:27:12.400
And it's very, very expensive. Right.
00:27:12.400 --> 00:27:26.365
And, you know, for a infra company like Quoqo, and Quoqo is focused on legal tech, and, you know, you have different use cases and sometimes, you know, you may have data privacy restrictions on utilizing, data.
00:27:26.365 --> 00:27:32.000
And if you look at all the LLM right now, correct, that very focused on the US, as a law.
00:27:32.000 --> 00:27:38.865
Correct Because, know, if you want India specific stuff correctly are to create those data sets yourself and India.
00:27:38.865 --> 00:27:40.700
so unique that the legal correct.
00:27:40.700 --> 00:27:45.432
As an example, it uses the US templates it uses the UK templates.
00:27:45.432 --> 00:27:47.000
It uses India templates.
00:27:47.000 --> 00:27:48.965
All the common law stuff. Correct.
00:27:48.965 --> 00:28:05.133
But if you want to create India specific templates, there's, you know, it's only, and if you ask a human to do it, lawyer to do it today, let's say to draft and agreement, they have an inherent biases to it because they're only trained on documents outside of India.
00:28:05.133 --> 00:28:06.000
Okay.
00:28:06.000 --> 00:28:10.932
So the endeavor is to remove the biases in some of the LLms that we have, correct.
00:28:10.932 --> 00:28:15.633
An example of biases is using English to do this podcast.
00:28:15.633 --> 00:28:19.400
So, you know, we could have done it in any other Indian language, which we know.
00:28:19.400 --> 00:28:20.266
Correct.
00:28:20.266 --> 00:28:23.766
But so those kind of biases coming okay.
00:28:23.766 --> 00:28:31.266
So to remove that and clean all of these things take a lot of a lot of time, effort and energy.
00:28:31.266 --> 00:28:34.766
And it's not that, you know, you can simply have data put it together.
00:28:34.766 --> 00:28:40.000
and do a chat bot you know, we have seen those days where it gives you like, rubbish results.
00:28:40.000 --> 00:28:40.665
Isn't it Guru?
00:28:40.665 --> 00:28:41.665
Oh, yeah.
00:28:41.665 --> 00:28:44.299
And, when you do it, you know, there's a requirement.
00:28:44.299 --> 00:28:53.432
It's a we also have some patent technology, where it's not necessary to always train, but I won’t get, into very big details on that.
00:28:53.432 --> 00:28:58.865
It is, of course, you know, our internal IP for all of these things as well.
00:28:58.865 --> 00:29:11.432
But suffice it to say that, you know, if you're pulling through all of these different bits and pieces, you are going to consume a lot of power and want to consume a lot of computing data and, private cloud.
00:29:11.432 --> 00:29:11.965
Correct.
00:29:11.965 --> 00:29:16.932
I sorry, on public cloud, you will burn a lot of these things.
00:29:16.932 --> 00:29:21.700
And I don't know how many people are successful in getting you.
00:29:21.700 --> 00:29:26.032
You know, you will you will still get a Tesla T4 and never get h100.
00:29:26.032 --> 00:29:31.133
And why is it that you will not get an allocation and simply not to get an allocation?
00:29:31.133 --> 00:29:37.266
And also, if you're looking at India, correct, the concept is all about Services.
00:29:37.266 --> 00:29:40.465
It's never been about products and innovations.
00:29:40.465 --> 00:29:41.400
Correct.
00:29:41.400 --> 00:29:46.066
People underestimate, you know, it's like trying to get a chartered accountant. Correct.
00:29:46.066 --> 00:29:47.066
You will get one.
00:29:47.066 --> 00:29:51.799
Sorry for, you know, all due disclaimers on this.
00:29:51.799 --> 00:29:52.465
Correct.
00:29:52.465 --> 00:29:59.532
You will get somebody depending on how much you pay and you will get you know, it's across the spectrum.
00:29:59.532 --> 00:30:02.532
And this is true of lawyers as well and true.
00:30:02.532 --> 00:30:06.532
But India such a service is denominated economy. Correct.
00:30:06.532 --> 00:30:17.865
That the innovation piece and ability to actually do deep research, and do deep things is underestimated today. Right?
00:30:17.865 --> 00:30:23.566
So, you know, we've got a new product, on the hardware side called Qcore.
00:30:23.566 --> 00:30:28.665
I was, I think about two years back correct Guru about a year and a half back.
00:30:28.665 --> 00:30:33.032
We started experimenting with edge devices, perfected it. Correct.
00:30:33.032 --> 00:30:34.232
It was very difficult to get.
00:30:34.232 --> 00:30:45.932
We were probably the first to get, and we know that Nvidia what I'm saying that, you know, we were we are running, you know, actually LLMs on edge devices, which we had a very specific use case.
00:30:45.932 --> 00:30:48.032
I don't want to get into that.
00:30:48.032 --> 00:30:50.700
But it was possible to run tasks on.
00:30:50.700 --> 00:30:56.133
And at that point in time, you know, Nvidia's only company that made a GPU on an edge device, right?
00:30:56.133 --> 00:30:57.365
So available.
00:30:57.365 --> 00:30:59.432
So, it still does.
00:30:59.432 --> 00:31:04.900
I mean, you know, you you know, it's like many people that are running on Raspberry Pi, I can go ahead.
00:31:04.900 --> 00:31:09.900
Try, you know, see what you get like two tokens a second, you'll be happy.
00:31:09.900 --> 00:31:10.232
Correct.
00:31:10.232 --> 00:31:13.400
So but that it's an edge device.
00:31:13.400 --> 00:31:15.165
It is not a Raspberry Pi.
00:31:15.165 --> 00:31:15.500
Correct.
00:31:15.500 --> 00:31:23.799
So I an edge device needs to have a certain amount of output capacity to be able to run some of these things as well.
00:31:23.799 --> 00:31:26.432
And it needs to take into account concurrency.
00:31:26.432 --> 00:31:29.032
Network availability and so many other things as well.
00:31:29.032 --> 00:31:29.333
Okay.
00:31:29.333 --> 00:31:32.732
Sorry, I'm getting a little technical here, but that's how we started.
00:31:32.732 --> 00:31:33.432
Correct?
00:31:33.432 --> 00:31:35.566
Other damn thing would consume so much power.
00:31:35.566 --> 00:31:40.465
Also, it's not that an edge device, you know, is so energy efficient and things.
00:31:40.465 --> 00:31:47.833
So it's like running, in an old style, filament light bulb, but, 200W. Correct.
00:31:47.833 --> 00:31:48.833
So something like that.
00:31:48.833 --> 00:31:52.766
So and that can consume a lot of power very, very quickly.
00:31:52.766 --> 00:32:04.900
And the devices get hard because it's not necessary that you have cooling systems in, you know, run it in a, place in somewhere in India without adequate cooling and that thing with the smoke.
00:32:04.900 --> 00:32:05.266
Correct.
00:32:05.266 --> 00:32:06.299
So after some time.
00:32:06.299 --> 00:32:10.400
So we've gone through all of these, you know, I've learned a few things also, isn't it.
00:32:10.400 --> 00:32:12.700
So it's happened as well.
00:32:12.700 --> 00:32:16.566
And why I mentioned this is the Qcore is our data center.
00:32:16.566 --> 00:32:24.000
It's a micro data center, where you can run, your applications and inference tasks very, very quickly.
00:32:24.000 --> 00:32:24.665
Correct.
00:32:24.665 --> 00:32:26.900
Because it's made for application.
00:32:26.900 --> 00:32:33.865
And you know what the output levels look like and it can scale to a certain number of users and things like that.
00:32:33.865 --> 00:32:39.165
And it has a certain power capacity and tier 1 networks associated with it.
00:32:39.165 --> 00:32:47.232
It does not meant to become like a huge data center, because in India, people in the moment you go to a very huge data center, correct..
00:32:47.232 --> 00:32:53.000
And you know, the problem becomes one of cooling, of management, of power.
00:32:53.000 --> 00:33:05.465
And tier 4 for data centers, you need to, to have at least, two power sources in India unless you are in, in what you call as a Gadi bhaga (boundary area), which is a border area. Correct.
00:33:05.465 --> 00:33:10.965
You know, you will not have power supply from two discoms Correct.
00:33:10.965 --> 00:33:18.599
electricity board, most states in India have only one discoms So they will get the second discoms Right.
00:33:18.599 --> 00:33:20.432
And the second discom is the sun.
00:33:20.432 --> 00:33:20.865
Okay. Right.
00:33:20.865 --> 00:33:28.566
So that's also limited to approximately 12 hours a day and it's never gonna run on like a gigawatt capacity.
00:33:28.566 --> 00:33:30.333
And on that it's not gonna happen.
00:33:30.333 --> 00:33:39.833
So, you know, you have to do I think in most, places, you know, you can do like a tier 2 different tiers of data centers, no that It's better.
00:33:39.833 --> 00:33:41.333
What do you have?
00:33:41.333 --> 00:33:43.799
Tier 2 or 3 capability as well.
00:33:43.799 --> 00:33:44.599
Correct.
00:33:44.599 --> 00:33:50.032
And tier 3 is quite adequate, you know, to be able to run what you require.
00:33:50.032 --> 00:34:03.133
You know, you don't necessarily need to pay for, hundreds of acres of, you know, you know, to run like cutting, you know you need micro data centers, correct.
00:34:03.133 --> 00:34:09.766
In addition to it, I just also have to add that we want to charge customers in Rupees instead of Dollars .
00:34:09.766 --> 00:34:10.532
Exactly.
00:34:10.532 --> 00:34:11.300
Right. Correct.
00:34:11.300 --> 00:34:16.500
If you look at AWS, Asur or, any other global cloud.
00:34:16.500 --> 00:34:16.766
Yeah.
00:34:17.132 --> 00:34:22.333
So you pay them in dollars, but then that those charges have to be transferred to the customers.
00:34:22.333 --> 00:34:22.599
Yeah.
00:34:22.599 --> 00:34:25.266
That's something I think I think, you know, you're right.
00:34:25.266 --> 00:34:25.965
Well isn't it.
00:34:25.965 --> 00:34:27.333
It's a joke, right.
00:34:27.333 --> 00:34:29.599
People sitting in an next road in Bangalore.
00:34:29.599 --> 00:34:33.599
which charge in dollars to which we think is a joke.
00:34:33.599 --> 00:34:36.365
I mean, people buy it Also, okay, okay.
00:34:36.365 --> 00:34:39.932
So, yesterday I was, talking to someone in Chennai.
00:34:39.932 --> 00:34:40.565
Correct?
00:34:40.565 --> 00:34:44.199
They bloody sitting in Madras and right is charge in dollars.
00:34:44.199 --> 00:34:44.900
Correct.
00:34:44.900 --> 00:34:45.766
You know what? That.
00:34:45.766 --> 00:34:46.565
What's the fun?
00:34:46.565 --> 00:34:46.932
Correct.
00:34:46.932 --> 00:34:48.666
You have just adopted some.
00:34:48.666 --> 00:34:53.099
And if you look at the USD to INR it's indicating a 92.
00:34:53.099 --> 00:34:53.833
Yeah.
00:34:53.833 --> 00:34:56.432
Here it soon becomes unaffordable in India.
00:34:56.432 --> 00:34:57.833
and it's a service economy.
00:34:57.833 --> 00:35:02.333
How many people actually know guru that you need such huge data centers right.
00:35:02.333 --> 00:35:04.333
I know you think about it right.
00:35:04.333 --> 00:35:05.300
Doesn't make sense.
00:35:05.300 --> 00:35:05.632
Right?
00:35:05.632 --> 00:35:08.032
It's making way for the US based companies.
00:35:08.032 --> 00:35:15.465
to take over the indian company And also, you know, to large extent data center is real estate and power play.
00:35:15.465 --> 00:35:15.865
Correct.
00:35:15.865 --> 00:35:22.400
You need to have a real estate to be able to have scale, and you need to have power to be able to rent.
00:35:22.400 --> 00:35:22.565
Okay.
00:35:22.565 --> 00:35:28.699
Now how many one gigawatt stations, you know, per data centers are there in India, Guru at least to my knowledge.
00:35:28.699 --> 00:35:30.400
Zero. Right.
00:35:30.400 --> 00:35:34.500
You know, you top out, say 250 or 300 K.
00:35:34.500 --> 00:35:35.666
We are and things like that.
00:35:35.666 --> 00:35:41.199
And that's the max that you can, you know, I'm not maybe I'm you know, it's kind of stand corrected on these things.
00:35:41.199 --> 00:35:54.565
But to get a gigawatt of power correct requires like, you know, can you imagine the power, supply for some when you need to have, like, coal based nuclear power, nuclear power source?
00:35:54.565 --> 00:35:55.233
Okay.
00:35:55.233 --> 00:36:00.632
Fusion is something which the Chinese and the French have not yet perfected since I did come to India. Correct.
00:36:00.632 --> 00:36:02.699
So they have a lot of things to do?
00:36:04.032 --> 00:36:04.333
Yeah.
00:36:04.333 --> 00:36:21.632
Hopefully it's, you know, we have our own internal jokes where, you know, I feel that, you know, for, for product companies, especially in the AI space and India and want be research and I think you will need these micro data centers.
00:36:21.632 --> 00:36:27.365
And I think the government has an ambitious plan of procuring, thousands of GPUs.
00:36:27.365 --> 00:36:31.500
And, you know, anyway, Nvidia has annual event is coming up.
00:36:31.500 --> 00:36:35.666
So we we know that there's going to be another major technological leap.
00:36:35.666 --> 00:36:36.500
Correct.
00:36:36.500 --> 00:36:42.266
You know, I mean, India wants to subsidize GPUs, what,₹64 an hour? Correct.
00:36:42.266 --> 00:36:43.766
That is the thing.
00:36:43.766 --> 00:36:48.632
But I am asking some people saying, well, that may be an approach towards availability.
00:36:48.632 --> 00:36:49.032
Correct.
00:36:49.032 --> 00:36:50.166
And affordability.
00:36:50.166 --> 00:36:50.932
Right.
00:36:50.932 --> 00:36:55.565
And Who the people who are building these applications right.
00:36:55.565 --> 00:36:58.733
You know, people are living longer hours, right?
00:36:58.733 --> 00:37:00.500
Who is actually doing the deep tech guru?
00:37:00.500 --> 00:37:01.500
I mean, isn't it.
00:37:01.500 --> 00:37:07.900
You can read it as number of research, how research is happening in India for very, very less amount of research.
00:37:07.900 --> 00:37:10.865
Now, what will you do with a grasshopper scale?
00:37:10.865 --> 00:37:16.599
You know, GPU and one one rack is probably enough to run so many things.
00:37:16.599 --> 00:37:23.032
I read a report today that India's fastest supercomputer has, 800 GPUs. in It.
00:37:23.032 --> 00:37:26.833
Okay, you know that maybe older scale or whatever it is.
00:37:26.833 --> 00:37:45.132
We know we've been to enough stuff on, you If you go to a Taiwanese trade show or an Nvidia show, correct, for some of these things, you know, you will have some rack servers and fully contained ones that can probably run like a quarter of India, isn't it?
00:37:45.132 --> 00:37:55.733
You know, in terms of just a small rack, you know, it's no longer about like having a acres and a acres because the, the computing has reached such an extent.
00:37:55.733 --> 00:37:58.365
It's like one, it's like one, 2 or 3 nanometers.
00:37:58.365 --> 00:37:58.599
Right.
00:37:58.599 --> 00:38:07.532
Even the CPU sizes and GPU sizes and the ability to sort of like colitis and weld the technology has reached a stage where you can or GPU, right?
00:38:07.532 --> 00:38:17.532
I mean, the if you look at the, the grasshopper versions that I go to, they basically take like two dice and then Weld in the GPU.
00:38:17.532 --> 00:38:22.199
And, you know, they have the ability to set off like really it's not as simple as just taking these two.
00:38:22.199 --> 00:38:35.900
Obviously I'm oversimplifying, but if you really look at the technology and the ability to interconnect all of these things, the ability, you know, to be able to sort of, you know, have the bandwidth in the GPU to process so much.
00:38:35.900 --> 00:38:36.632
Correct.
00:38:36.632 --> 00:38:38.166
It is decreasing day by day.
00:38:38.166 --> 00:38:48.800
you know, there have been examples, people say I can run, LLM's correct, let's say so many tokens per second on interconnected, Apple Mac mini.
00:38:48.800 --> 00:38:49.965
Correct? Correct.
00:38:49.965 --> 00:38:57.833
And we are on the fourth lesson about M4 pros, but the M2 Pro has more bandwidth and more cores for the GPU, correct?
00:38:57.833 --> 00:38:58.932
Than the M4.
00:38:58.932 --> 00:38:59.766
Okay.
00:38:59.766 --> 00:39:02.565
You know, so you're going to be looking at that kind of throughput.
00:39:02.565 --> 00:39:06.699
It's not necessarily like the older generation is bad.
00:39:06.699 --> 00:39:07.865
And therefore cannot.
00:39:07.865 --> 00:39:14.465
It's more in terms of what is the use case, what does it require to run all of these things?
00:39:14.465 --> 00:39:15.199
Yes.
00:39:15.199 --> 00:39:19.800
And, do you require some huge things to be able to do this right?
00:39:19.800 --> 00:39:22.532
But what is the use case in India for this scope.
00:39:22.532 --> 00:39:22.965
Yeah.
00:39:22.965 --> 00:39:32.065
When we look at OLA, which probably was the first to sort of like AI turning, how many users are there stuff to do all these things.
00:39:32.065 --> 00:39:38.166
And, you know, the question is, you build the wrapper you dont need the kind of, yeah.
00:39:38.166 --> 00:39:39.400
You don't need need it.
00:39:39.400 --> 00:39:45.766
And India is in level of Creating wrapper is indian is not deep tech correct.
00:39:45.766 --> 00:39:47.666
And that's the point, right.
00:39:47.666 --> 00:40:02.800
And the other question was before you started recording this podcast, he was saying, you know, this happened that, Can’t innovate necessarily started people moving and etc..
00:40:02.800 --> 00:40:19.266
But, you know, if you look at someone like you running a company, correct, there's a skill level of people that you get sometimes and sorry, it's it's, you know, it is not, you know, people who say, oh yeah, jobs that are not enough jobs at the end.
00:40:19.266 --> 00:40:30.800
You know, you've been, you know, with all of these things, skill levels have not reached the level which all the knowledge that it was. Okay.
00:40:30.800 --> 00:40:35.865
And there's a gap for this, and this is a huge gap.
00:40:35.865 --> 00:40:36.532
Okay.
00:40:36.532 --> 00:40:42.199
Now, in the government, you know, the don't give it to people like you had them.
00:40:42.199 --> 00:40:50.632
It was a good idea for for a particular property people look at me like this that technology has been for 25 years Yes.
00:40:51.632 --> 00:40:52.065
Okay.
00:40:52.065 --> 00:40:56.132
Correct 20 years atleast maybe not 25.
00:40:56.132 --> 00:40:56.733
Right.
00:40:56.733 --> 00:41:02.266
So and they're saying this which I mean, this is a huge gap.
00:41:02.266 --> 00:41:08.666
And in one industry players and this is and only.
00:41:08.800 --> 00:41:36.432
No, no, no really guru, I mean, you know, it's like, you know, in this particular episode So they seem to have gathered the most for all of these things, but these are all things which I thought that, you know, if I'm doing this, I might as well say it or probably also wasn't saying, today, the, government calls out for all these people to bring to build like and huge GPU cluster, which is maybe going to get outdated by the time it’s going to surface they come through.
00:41:36.432 --> 00:41:47.500
And then you want to build in in India specifically, that's where the huge skill deficit, where people tend to figure, you know, when we build QLLM I put out on LinkedIn post for saying we all laughed at guru.
00:41:47.500 --> 00:41:50.766
We did not even have the resources, right, you know, to do it.
00:41:50.766 --> 00:41:57.965
But we would sort of like, use ingenious methods, putting so many things together to get things running correct.
00:41:57.965 --> 00:42:07.166
And that seems to be the real question mark, saying, you know, one, are you at a skill level where you can do all of these wonderful things?
00:42:07.166 --> 00:42:09.532
The answer is absolutely not correct.
00:42:09.532 --> 00:42:12.900
You are the skill level as a country, correct?
00:42:12.900 --> 00:42:13.733
To do services?
00:42:13.733 --> 00:42:15.266
Absolutely correct.
00:42:15.266 --> 00:42:17.199
Are you at a skill level to build products?
00:42:17.199 --> 00:42:18.532
Absolutely NOt correct.
00:42:18.532 --> 00:42:21.699
So I think that sums it up, isn't it Guru?
00:42:21.699 --> 00:42:25.965
You know, in fact, when we go on sort of like do stuff internationally, correct.
00:42:25.965 --> 00:42:28.632
know, with our innovation, sometimes people are surprising.
00:42:28.632 --> 00:42:29.432
Are you in India?
00:42:29.432 --> 00:42:31.900
Why not come here Why not move here?
00:42:31.900 --> 00:42:34.565
You know, all of those things are just sort of given out.
00:42:34.565 --> 00:42:41.065
But if you look at most of the companies that exist today, they're just copies of things that have come up.
00:42:41.065 --> 00:42:42.099
Yes.
00:42:42.099 --> 00:42:45.932
I think it's just copies of stuff that has come on. Now.
00:42:45.932 --> 00:42:46.833
Where is innovation?
00:42:46.833 --> 00:42:50.233
is the innovation in the UI because it looks nicer.
00:42:50.233 --> 00:42:58.833
Mine is, mine looks my app looks nicer than yours kind of stuff, which is not innovation And I know that right is the innovation in getting things.
00:42:58.833 --> 00:43:01.266
it's like, deliver on time.
00:43:01.266 --> 00:43:02.132
Correct.
00:43:02.132 --> 00:43:04.666
You've got a huge population, which is jobless, correct.
00:43:04.666 --> 00:43:25.132
You know, but you know, therefore utilization But then there is a real, you know, you know, where is a really a vishwarya moment where you can, you know, build, a Dam, or you can really think of, you know, maybe, you know, early 1900s by the British where the Maharaj of Mysore decided to sort of like give a grand opening.
00:43:25.132 --> 00:43:28.733
IASC maybe the scientists at ISRO. Correct.
00:43:28.733 --> 00:43:32.699
That really sort of like, innovation correct to get things done.
00:43:32.699 --> 00:43:35.000
Maybe not saying that it doesn't exist.
00:43:35.000 --> 00:43:36.766
Maybe in certain portions. Correct.
00:43:38.000 --> 00:43:47.132
where some of these things come up, but can we still do, like, booster, the rocket that comes back and hitch itself to, you know, the launch pad?
00:43:47.132 --> 00:43:49.000
No, no, we're not there yet.
00:43:49.000 --> 00:43:58.532
So that requires some amount of innovation, some amount of skill giving some amount of, some sort of dedication to actually take the country forward.
00:43:58.532 --> 00:43:59.666
Well, correct.
00:43:59.666 --> 00:44:08.500
Easiest thing is to leave the country, go outside, and try do something because, again, you don't need a lot of resources to be able to build.
00:44:08.500 --> 00:44:11.599
But you certainly at some point, I don't think there is ever.
00:44:11.599 --> 00:44:19.166
Have you ever faced a situation where we could not buy or do something Guru You know, there's always I think, the difference also is in the mindset.
00:44:19.166 --> 00:44:20.432
Yeah. okay.
00:44:20.432 --> 00:44:27.065
To build, a LLM we need only big companies like OpenAI or Microsoft or Google, etc.
00:44:27.065 --> 00:44:38.132
but then came the deepseek and shattered all the AI and also early attention made here, also saying, we don't you only need like a Mini LLM why should we need a proper LLM?
00:44:38.132 --> 00:44:39.599
Why should we do this?
00:44:39.599 --> 00:44:48.565
We are some interesting, ex company, like three lakh people doing services, you know, you know, the power of AI.
00:44:48.565 --> 00:44:49.800
I don't need to say more.
00:44:49.800 --> 00:44:50.666
Correct.
00:44:50.666 --> 00:44:58.565
But, you know, the early attention made to sort of like, you know, also hijacked the, on the headlines and things like these as well.
00:44:58.565 --> 00:44:58.932
Correct.
00:44:58.932 --> 00:45:17.733
It requires a certain amount of, also brute sort of, brute headedness, you know, to be able to sort of like saying, I'm doing this for the, for whatever it may be, may or may not turn out to be monetarily, you know, successful it may or may not, but it certainly gives you a, I think, guru.
00:45:17.733 --> 00:45:32.166
Can you also say that in the last year we worked on some so such critical technology that has given us so much, happiness that we actually can build this, you know, this this can only come from when you actually build something.
00:45:32.166 --> 00:45:32.865
Correct?
00:45:32.865 --> 00:45:40.699
When you invent something new, when your not just important stuff and trying sell it here, there not like an import export business.
00:45:40.699 --> 00:45:46.199
where import some and try and sell it here Yeah, it's not, not that it is a bad thing. Okay.
00:45:46.199 --> 00:45:52.800
You know, a joke, like very recently, you know, we asked for some, Exim bank support.
00:45:52.800 --> 00:45:53.532
Correct.
00:45:53.532 --> 00:46:01.965
name a particular banking institution and the state of affairs saying I need some support in this part of the globe because we are doing here.
00:46:01.965 --> 00:46:09.166
But, you know, it usually gets delayed because of, you know, the country's tax policies and things like that.
00:46:09.166 --> 00:46:14.900
Can you say, you know, the the answer we got back to Italy was, if you are a chappal manufacturer, we can help you.
00:46:14.900 --> 00:46:20.400
I'm like, you know, this is the place that this is where made in India currently stands Correct.
00:46:20.400 --> 00:46:24.900
And this is where, you know, the state of affairs is when it comes to innovation.
00:46:24.900 --> 00:46:26.965
And we don't say this enough. Correct.
00:46:26.965 --> 00:46:33.166
And even at Quoqo you know, we strive for you know, we like you know, please, innovate.
00:46:33.166 --> 00:46:38.065
And it's not that we don't have a lot of these things cut out or, you know, built correct.
00:46:38.065 --> 00:46:40.800
It's not that you need to use something which is over there.
00:46:40.800 --> 00:46:41.099
Correct.
00:46:41.099 --> 00:46:49.965
You create yourself, you create your own data centers, and you are able to sort of like be able to do that, but you need to have some critical mass for all these things as well.
00:46:49.965 --> 00:46:51.000
just say this.
00:46:51.000 --> 00:46:55.032
Sorry, that's my, that's my gross of the season also, isn't it?
00:46:55.032 --> 00:46:57.532
I'm sure a lot of founders also feel the same way.
00:46:57.532 --> 00:46:58.599
They want to change.
00:46:58.599 --> 00:47:09.733
What do you think? Guru What do you think Krithi LLM , when you were talking and LLM you see LLM the minimal resource, if that, you know, could realistic.
00:47:09.733 --> 00:47:14.766
More like how is it like more realistic using LLM the minimal resources.
00:47:14.766 --> 00:47:18.099
people are talking that LLM are the new when you was bringing with it.
00:47:18.099 --> 00:47:24.599
So you mean building LLM and everything and LLM minimal resources like how is it more realistic?
00:47:24.599 --> 00:47:40.632
So basically, you don't really need, very large number of resources to build a model that can cater to certain, know, specific now, use cases, let's say you are only dealing with legal documents, for example, to take or, talk about what?
00:47:40.632 --> 00:47:42.266
That's been our approach as well.
00:47:42.266 --> 00:47:43.599
Is that okay? Yeah.
00:47:43.599 --> 00:47:59.333
So you don't really need a model that conquer’s the world then that would just need to be able to understand your documents, try to go to, solve the problems if you are working on, so with respect to resources, you don't really need a great number of resources, but the process itself is quite grueling.
00:47:59.333 --> 00:48:04.800
You need to go through the data collected, probably because garbage in, garbage out, you have to clean it properly.
00:48:04.800 --> 00:48:09.333
You have to ensure that the document formatting is proper, etc.
00:48:09.333 --> 00:48:19.333
and then you need to, train it so we need some amount of GPU power to be able to train a model that does a decent, job giving you the response Yeah.
00:48:19.333 --> 00:48:33.000
So just to add there, you know, we've taken an approach where you have a specialized data sets, like actually, you know, and some of these, can be accessed as demo ones, like we have, all india Supreme Court cases.
00:48:33.000 --> 00:48:34.432
We have legislation.
00:48:34.432 --> 00:48:35.065
Correct.
00:48:35.065 --> 00:48:43.065
we have tax bots for member states like these that you can use to query very, very specific data sets. Correct.
00:48:43.065 --> 00:48:55.400
And some of these come together, you know, let us say, but, you know, to create a mini language model for a very specific task that you can potentially serve and apply on each devices and things like this.
00:48:55.400 --> 00:48:59.900
You know, your iPhone 15 is quite capable of running some of these things as well, right?
00:48:59.900 --> 00:49:01.500
It's got enough computational power.
00:49:01.500 --> 00:49:03.632
Not that you lack any of these things today.
00:49:03.632 --> 00:49:06.199
It's just the use case. Correct.
00:49:06.199 --> 00:49:18.032
And also when a lot of these things come up and you have a large enough data set and you also need to sort of train them, you will need a combination of real data, synthetic data.
00:49:18.032 --> 00:49:20.532
And there are synthetic elements on it suddenly.
00:49:20.532 --> 00:49:26.132
So it's also called as bull shit data and it's because it's like completely synthetic data. Right?
00:49:26.132 --> 00:49:28.965
You're just generating random numbers.
00:49:28.965 --> 00:49:30.565
Correct, for some of these things.
00:49:30.565 --> 00:49:30.833
Right.
00:49:30.833 --> 00:49:38.266
So and so when you reach that scale, when you have a language model where you want it to sort of train and give an example.
00:49:38.266 --> 00:49:38.733
Correct.
00:49:38.733 --> 00:49:48.599
You know, let us say you have in India, you have, backlog of millions of cases people avoid court correct for what, reason.
00:49:48.599 --> 00:49:56.032
Guru you go there until you file when you are 20 and you get out when you are 60 I'm saying this if you are lucky, correct.
00:49:56.032 --> 00:50:10.965
There are cases running for, tens of years with no, you know, it's because for lack of trained cadres lack of in sense that, the number of judges may not be sufficient for the population.
00:50:10.965 --> 00:50:11.865
Correct.
00:50:11.865 --> 00:50:16.365
And, you know, we follow a procedure which is not necessarily Indian.
00:50:16.365 --> 00:50:26.032
It's anglicized because we adopted a process when we were, you know, when we are colonized, trained, we were, you know, for a number for hundreds of years.
00:50:26.032 --> 00:50:26.333
Correct.
00:50:26.333 --> 00:50:42.766
So there's a process that is coming to this, which may not necessarily actually match the Indian ethos as well today, but let us say that you've got judgments or the thinking of 200 different judges over the last 200 years. Okay.
00:50:42.766 --> 00:50:47.800
And you want to resolve this issue of cases, correct?
00:50:47.800 --> 00:50:54.733
we can train the, LLM exactly sort of like analyze cases the way depending on the genres or other things.
00:50:54.733 --> 00:50:55.365
Right?
00:50:55.365 --> 00:50:59.000
I mean, I have to do this so that, you know, the resolution.
00:50:59.000 --> 00:51:20.365
Correct for a case can come in a format, where, you know, you may have a sitting judge or someone else who looks at it and is able to pass, you know, the process, the documentation, you know, if it is streamlined and you know, you can have a resolution in a matter of days versus 50 or 60 years, right?
00:51:20.365 --> 00:51:22.365
So that is my thinking on this.
00:51:22.365 --> 00:51:29.666
But, you know, that is the that is the ingenuity and that is a deep detail that is required for all of these things as well.
00:51:29.666 --> 00:51:33.333
I think, as Quoqo we are in a position to execute some of these things as well.
00:51:33.333 --> 00:51:37.065
But it also requires saying, is there a political will to do this?
00:51:37.065 --> 00:51:51.365
Is there technological skill and or and evaluated by LLM, you know, can I you know, at the end of the day, you need a resolution you know, it is not necessarily a mediation that you require for some of these things as well.
00:51:51.365 --> 00:51:54.000
that's my thought process. Longwinded.
00:51:54.000 --> 00:51:54.733
Yeah.
00:51:54.733 --> 00:52:00.333
I in addition, that there might be a few other, limitations to local models. Is it.
00:52:00.333 --> 00:52:07.900
I spoke to again the output kind of, output window how would you can actually get as an output it already?
00:52:07.900 --> 00:52:13.965
So if you are building smaller models, will be adding some outputs or it may not be able to do that.
00:52:13.965 --> 00:52:18.465
Some of the limitations that might be there for, you know, models that are. Yeah.
00:52:18.465 --> 00:52:25.132
So I'll give an example to take your judgments in a this and it's like 20, 25 pages long.
00:52:25.132 --> 00:52:33.199
You know, the reason being you have to, you know, a judge has to be justice has to be seen fair and not just sort of like be fair. Right.
00:52:33.199 --> 00:52:40.199
So there's and you go through law school, you know, so all of these things come up where you have to be seen to be fair and square, right?
00:52:40.199 --> 00:52:45.199
Like if you walk into if the police intervene in a conflict, they're not going to immediately take sides.
00:52:45.199 --> 00:52:53.932
I mean, notice this, because the police has to sort of, like, appear at least be fair they will never take side immediately.
00:52:53.932 --> 00:52:56.166
And it has a process of its own. Correct.
00:52:56.166 --> 00:52:57.333
So similarly on this.
00:52:57.333 --> 00:53:02.333
So one is when you are doing this data, it has to be clear of biases.
00:53:02.333 --> 00:53:06.065
India has a history where the judges have been biased.
00:53:06.065 --> 00:53:06.900
Okay.
00:53:06.900 --> 00:53:09.699
You know you have judgments either which way and across.
00:53:09.699 --> 00:53:22.132
You know for whatever it is, if you go down LLM through a graphic program and see and today we can map or if we go for which judge you know, what is it that you can expect as a result? Right.
00:53:22.132 --> 00:53:24.400
You know that technology is available.
00:53:24.400 --> 00:53:26.166
it does not require you.
00:53:26.166 --> 00:53:39.099
And that's how some of the funds there and litigation funds that invest, correct, will invest, use this technology to figure out whether, you know, there's stand a chance if they pick up, a piece of, litigation. Correct.
00:53:39.099 --> 00:53:41.865
And on it on its own, others on its course.
00:53:41.865 --> 00:53:54.000
So this technology is available, but the other thing that access to justice has to be cheap, you know, to solve major issues like justice and the backlog of cases.
00:53:54.000 --> 00:53:58.099
And at the end, you want something with an enforcement.
00:53:58.099 --> 00:53:59.000
Right.
00:53:59.000 --> 00:54:18.432
And if you don't want something like drag on for 50 years correct, and you you have to go to the court but even this has been the problem for the last 100 years now, you have really thing Deep technology can solve these things?
00:54:18.432 --> 00:54:22.599
So I think you have a commercial to dispute, you dont have to go with the court.
00:54:22.599 --> 00:54:32.500
And then I want to basically say, okay, you know, I wanted to be paided, you know, for all of these things, you know, that can be done very, very quickly.
00:54:32.500 --> 00:54:33.900
traffic fine.
00:54:33.900 --> 00:54:34.565
Right?
00:54:34.565 --> 00:54:58.266
your traffic fine board is very, very, very, great and, you know, people prefer to pay for to the government rather than more litigate, you know, traffic fine, including, unfair locking up my bike or car or whatever the case, when there no road and forced to park you cannot come and lock it up basic common sense so these are things that are required.
00:54:58.266 --> 00:55:02.099
Fundamentally, outer look I would think that.
00:55:02.099 --> 00:55:08.599
Ai in india is fantastic, Legal is fantastic, is’t it Guru? right?
00:55:08.599 --> 00:55:14.233
You know, you are able to solve so many things, so many inefficiencies.
00:55:14.233 --> 00:55:40.000
And one of the things that also Krithi you know, when you use this here, it applies, you know, AI it doesn't take a site we where on a demo recently not very long ago, there it basically said stuff on the financial side where I could have hesitated it's like, the most straightforward of things, you know, it's emotionless as well.
00:55:40.000 --> 00:55:44.233
But at the end of the day, it's also going to tell you exactly what it thinks.
00:55:44.233 --> 00:55:51.065
And it never sort of, I mean, that's, original thinking when we get there is another thing altogether correct.
00:55:51.065 --> 00:55:59.000
This is this can be used as an exploitative technology, or it can be as a transformational technology, like any piece of technology that you have. Right?
00:55:59.000 --> 00:56:11.065
At this point, the countries are going to exploit through technology, like dark stores dark pricing dark this stuff that you've got competition commission coming in and doing all of these things.
00:56:11.065 --> 00:56:14.432
It's about time to really think about it as transformative technology.
00:56:14.432 --> 00:56:16.132
And we are very specific to India.
00:56:16.132 --> 00:56:23.233
What this, and yes, I'm, I feel that I am entitled to speak on some of these things as a practitioner, as well. Correct.
00:56:23.233 --> 00:56:33.365
and these are things that require you to some sort of I would think that I would use this at every village level, taluk level and district level.
00:56:33.365 --> 00:56:34.166
Correct.
00:56:34.166 --> 00:56:38.465
You know, a lot of the cases can just be solved in a few days basically.
00:56:38.465 --> 00:56:38.833
Yeah.
00:56:38.833 --> 00:56:43.800
It's like, you know, there's an episode, know, my favorite series is Star Trek.
00:56:43.800 --> 00:56:48.032
You know, that one episode, the the original series called wink of an eye.
00:56:48.032 --> 00:56:48.766
Correct.
00:56:48.766 --> 00:56:49.965
You can't see them.
00:56:49.965 --> 00:56:52.666
That accelerated humans.
00:56:52.666 --> 00:56:56.266
And they will do things without you being able to sort of like see them.
00:56:56.266 --> 00:56:57.000
Correct.
00:56:57.000 --> 00:57:08.766
So, I mean, it ends in a bad way in that particular episode, in the sense that, you know, people figure out the origin of our that, you know, people are accelerated and they want to take on whatever those issues.
00:57:09.400 --> 00:57:19.833
But in the wink of an eye, you know, this is startup, so that doesn't have, you know, for solving, say, the case log in India, correct?
00:57:19.833 --> 00:57:28.300
No other country has this dubious distinction of having a billion population and millions of unsolved cases, backlog.
00:57:28.300 --> 00:57:31.500
You know, and you should really up to it, you know.
00:57:31.500 --> 00:57:36.932
So anyway, so I didn't want adding on to that sir I have gone by several, I think.
00:57:36.932 --> 00:57:38.800
What do you think the world was wondering?
00:57:38.800 --> 00:57:39.932
What's your take on that?
00:57:39.932 --> 00:57:47.733
Like, the perplexity and Openai thing Okay, for 200 dollars What's your take on this?
00:57:47.733 --> 00:57:48.632
Oh, yeah.
00:57:48.632 --> 00:57:49.632
What do you think, Guru?
00:57:49.632 --> 00:57:52.500
And, yeah, it's first of all, it's too expensive.
00:57:53.666 --> 00:57:55.599
And the deepseek come in.
00:57:55.599 --> 00:58:00.699
So I see perplexity was created as an alternative to Google. Correct.
00:58:00.699 --> 00:58:09.632
But it still relies on Google search results to be able to give you or maybe Bing and how deepseek to be able to search and give you what it does is we run a query.
00:58:09.632 --> 00:58:11.199
This is my understanding.
00:58:11.199 --> 00:58:19.465
And the guy behind it is from IIT madras And so you should clarify, it's on stage, but it ran 20 different queries and then beat it out.
00:58:19.465 --> 00:58:21.500
You know probably a has and so on.
00:58:21.500 --> 00:58:34.266
System weights and biases, you know do all of these things and then gives you somebody I've used it I found, you know, I have my own biases as well as to what my favorites are correct for all of these things.
00:58:34.266 --> 00:58:40.500
But you know, as a replacement for Google, you need to have a Google kind of search capability.
00:58:40.500 --> 00:58:40.833
Correct?
00:58:40.833 --> 00:58:43.300
Even today, the default is still Google.
00:58:43.300 --> 00:58:55.000
And it may take like, when you it's like, let's see, you have an email which looks like an essay versus receiving an email there that is like one sentence long, my biases is to look at one sentence long, correct.
00:58:55.000 --> 00:58:56.432
Not like an essay.
00:58:56.432 --> 00:58:56.632
Right.
00:58:56.632 --> 00:59:08.266
So this is how it looks like, when it comes to, the O-3 research, it's extremely priced $200 to do Phd level kind of stuff.
00:59:08.266 --> 00:59:10.965
So is Phd going to stop tomorrow.
00:59:10.965 --> 00:59:12.666
I don't know that.
00:59:12.666 --> 00:59:27.599
So, you know, I've tried out some of the reasoning models to, repeatedly question itself to give, you know, I've got some, yeah, I've learned some very esoteric queries like, and, and some very pointed queries on it.
00:59:27.599 --> 00:59:29.532
Queries actually come out right.
00:59:29.532 --> 00:59:39.565
The philosophical and like, what is the nature of, the universe kind of question, you know, and it's not 40 to, if you remember the Hitchhiker's Guide. Correct.
00:59:39.565 --> 00:59:40.932
definitely not even there.
00:59:40.932 --> 00:59:42.233
It had deep thought.
00:59:42.233 --> 00:59:42.532
Correct.
00:59:42.532 --> 00:59:50.300
Really very deepseek like deep thought, you know, it would give I don't think that these are competing, but it's too expensive.
00:59:50.300 --> 00:59:53.400
You know, actually, looking at a place like India for some of these things.
00:59:53.400 --> 00:59:54.699
I don't think so.
00:59:54.699 --> 00:59:58.733
And I don't think the Prime Minister needs to go to sort of like get some on these technologies.
00:59:58.733 --> 00:59:59.465
Correct.
00:59:59.465 --> 01:00:01.733
You look at what's homegrown first. Correct.
01:00:01.733 --> 01:00:03.500
To do all of these things as well.
01:00:03.500 --> 01:00:11.032
and necessarily homegrown technologies have grown with the limitations of the place as well.
01:00:11.032 --> 01:00:11.800
Correct.
01:00:11.800 --> 01:00:21.432
So if you hand India is at a fantastic stage where you can really grow some of these companies and technologies, and it has a wide enough, implication.
01:00:21.432 --> 01:00:24.266
So sort, all of these things as well.
01:00:24.266 --> 01:00:27.565
What does perplexity or O-3 make a difference in India?
01:00:27.565 --> 01:00:29.300
I'm not particularly sure.
01:00:29.300 --> 01:00:30.333
What do you think, Guru?
01:00:30.333 --> 01:00:36.032
Or if not many people have that much money to spend on a monthly basics Okay.
01:00:36.032 --> 01:00:48.333
If you want to do some research kind of thing, you generally spent six months or one year on doing research on a topic fairly and, you know, something like this comes up maybe to, make it faster.
01:00:48.333 --> 01:00:58.565
But more important for researchers is the effort that is involved in performing it so that it, it drives the thought process, it refines the idea.
01:00:58.565 --> 01:01:03.000
So it's about looking at a, two different conflicting comments.
01:01:03.000 --> 01:01:03.565
Yeah.
01:01:03.565 --> 01:01:06.233
It's like also the training given in schools, right.
01:01:06.233 --> 01:01:11.599
It's like, you know, even schools today, in the higher levels.
01:01:11.599 --> 01:01:19.599
Can I still teach like, you know, they don't teach Python, you know, or they may teach Python, but it's more Java based than anything else. Okay.
01:01:19.599 --> 01:01:23.632
In reality, it's just opposite in the real world, isn't it?
01:01:23.632 --> 01:01:28.699
So, I think there is also a fundamental disconnect somewhere that needs to be.
01:01:28.699 --> 01:01:31.965
I mean, there are too many things that we can keep going on and on.
01:01:31.965 --> 01:01:37.032
But does do those things really matter in India at this point in time?
01:01:37.032 --> 01:01:38.099
Probably not.
01:01:38.099 --> 01:01:39.833
You know, if you really ask me.
01:01:39.833 --> 01:01:41.000
Right.
01:01:41.000 --> 01:01:44.865
India some very specific, you know, use cases as well.
01:01:44.865 --> 01:01:59.565
And, you know, the, services mindset certainly does not also help, with product innovation, you know, and product innovation, the copying is treated as product innovation to a large extent.
01:01:59.565 --> 01:01:59.865
Right?
01:01:59.865 --> 01:02:02.032
Actually innovation, I'm sure it does.
01:02:02.032 --> 01:02:03.132
But there is a lot of.
01:02:03.132 --> 01:02:04.065
Yeah, not to me.
01:02:04.065 --> 01:02:05.032
Not too many.
01:02:05.032 --> 01:02:05.965
Correct. Yeah.
01:02:05.965 --> 01:02:14.400
So and things like that, like as an example, you remember, Guru sometime back They built a new rocket motor.
01:02:14.400 --> 01:02:14.766
Correct.
01:02:14.766 --> 01:02:21.065
Which is which actually flew as well, you know, based on, on newer technologies than anything else.
01:02:21.065 --> 01:02:22.666
I forget the name of the company. Correct.
01:02:22.666 --> 01:02:24.199
That was pure innovation.
01:02:24.199 --> 01:02:25.032
Right.
01:02:25.032 --> 01:02:27.266
And that was like, come on, 3d printed model.
01:02:27.266 --> 01:02:27.699
Yeah.
01:02:27.699 --> 01:02:49.733
3D printed model module, which I think isro had not gotten to that stage as it got into, and things like that, some of those fundamental I'm not saying no research based on, suddenly on but an AI you know, that you can hand count you know, a who are the guys or or actually doing deep research in India.
01:02:49.733 --> 01:02:53.300
So, I think that's where things but go on krithi.
01:02:53.300 --> 01:02:59.666
So, one last question before we wrap up What is this kumbhakarna moment?
01:02:59.666 --> 01:03:02.833
like that people are talking about?
01:03:02.833 --> 01:03:05.599
They what are they what exactly that means, right.
01:03:05.599 --> 01:03:08.000
Let me, let me rephrase that question. Okay.
01:03:08.000 --> 01:03:13.365
Now, now, you heard about what Chetan thinks and, what guru thinks and things like that.
01:03:13.365 --> 01:03:21.699
Right now you are exposed to AI right from sort of like youth, right from college onwards.
01:03:21.699 --> 01:03:22.599
Correct? Yeah.
01:03:22.599 --> 01:03:24.833
What do you think really?
01:03:24.833 --> 01:03:26.300
I mean, how do you use AI ?
01:03:26.300 --> 01:03:27.099
Let me ask.
01:03:27.099 --> 01:03:31.266
So basically we has just passed out.
01:03:31.266 --> 01:03:32.465
Well not really passed out.
01:03:32.465 --> 01:03:38.699
But has been assigment I mean where they give you a assigment you just put the questions and we just get the answers.
01:03:38.699 --> 01:03:43.965
We are not like putting a efforts to literally go through the books and get the answers.
01:03:43.965 --> 01:03:44.300
Okay.
01:03:44.300 --> 01:03:48.333
It's better to read that way and also videos.
01:03:48.333 --> 01:03:49.599
And there's a thing.
01:03:49.599 --> 01:03:54.365
AI also wher u put up a pdf to get all the answers that related to the book.
01:03:54.365 --> 01:03:57.632
If you would be able to read the whole 30 pages.
01:03:57.632 --> 01:04:01.199
So it's like time consuming when we listen to it.
01:04:01.199 --> 01:04:01.465
Okay.
01:04:01.465 --> 01:04:06.932
So this is where we are using like so as an example, you know for this right.
01:04:06.932 --> 01:04:13.833
The future generation looks at times it's not necessary for you to learn everything you know.
01:04:13.833 --> 01:04:16.965
And then because each generation builds on top of the other, right.
01:04:16.965 --> 01:04:22.465
If you go back and try and build the satan rocket today, they don't have another tools and dices Correct.
01:04:22.465 --> 01:04:25.000
And you don't know when can you use that thing?
01:04:25.000 --> 01:04:29.365
What do you use Guru that, before calculators came into the thing.
01:04:29.365 --> 01:04:30.500
Don't know.
01:04:30.500 --> 01:04:33.900
And and no, no, they use some scale I forget, you see.
01:04:33.900 --> 01:04:34.132
Yeah.
01:04:34.132 --> 01:04:41.099
I, you know, I have no, no, no, they have this, you know, to calculate log table. Right.
01:04:41.099 --> 01:04:47.865
They won't do the anyway. Sorry So you know, now now look even we don't remember some of these things, okay.
01:04:47.865 --> 01:04:55.965
Even though it may have been like a generation or two away and you are at least five generations of a four generations of it. Correct.
01:04:55.965 --> 01:05:05.699
So things like this, why, why we say this is it's not necessary for you to actually you have to like this that remembers how things were done.
01:05:05.699 --> 01:05:06.632
Correct.
01:05:06.632 --> 01:05:14.300
And then, you know, you have a tool to sort of, look back, correct, and actually learn how it could be done if necessary and things like that.
01:05:14.300 --> 01:05:17.965
It's more of a memory to enhance memory to as well. Correct.
01:05:17.965 --> 01:05:23.266
We lost the art in India to manufacture mud buildings as an example.
01:05:23.266 --> 01:05:23.565
Okay.
01:05:23.565 --> 01:05:25.599
Which existed like 100 years before.
01:05:25.599 --> 01:05:26.333
Correct.
01:05:26.333 --> 01:05:30.766
For that matter, making, or making iron, which does not rust.
01:05:30.766 --> 01:05:38.932
Because if an ally of whatever the case is right, you iron pillar as an example, the ability to create the stone temples in Karnataka.
01:05:38.932 --> 01:05:40.000
you can't.
01:05:40.000 --> 01:05:41.032
You lost on that.
01:05:41.032 --> 01:05:45.365
But you have the ability now to store all of that generation stuff going forward.
01:05:45.365 --> 01:05:45.699
Correct.
01:05:45.699 --> 01:05:48.500
So that means that you don't need to know something known.
01:05:48.500 --> 01:05:49.599
Pythagoras theorem.
01:05:49.599 --> 01:05:50.333
Correct.
01:05:50.333 --> 01:05:55.166
Should only know that, you know, at the end of the day, can tap into something that can utilize that.
01:05:55.166 --> 01:05:57.333
You have the broad aspects of it.
01:05:57.333 --> 01:06:00.000
That's how exactly you should be using this.
01:06:00.000 --> 01:06:00.266
Correct?
01:06:00.266 --> 01:06:04.400
It's not possible to download an entire LLM.
01:06:04.400 --> 01:06:09.300
Basically they will be made once and that would be for example Yeah.
01:06:09.300 --> 01:06:09.900
You video.
01:06:09.900 --> 01:06:13.400
For 15 minutes of and we just put it on it goes on.
01:06:13.400 --> 01:06:22.032
Well it is about borders on it and it's not necessary to go through like 20, 22, 23 years of education, correct?
01:06:22.032 --> 01:06:22.532
Yes.
01:06:22.532 --> 01:06:28.233
You don't I don't think I think that's look, that seriously needs to get crunched at some stage as well.
01:06:28.233 --> 01:06:28.565
Correct.
01:06:28.565 --> 01:06:44.532
To be able to say, you know, you have a lot more, you know, I mean, you may lack the experience, but that doesn't mean that you cannot tap into a higher experience, thing that also the way that, generations are going to utilize this is to build stuff, right?
01:06:44.532 --> 01:06:46.833
Which would never have been attempted before.
01:06:46.833 --> 01:06:47.565
Correct.
01:06:47.565 --> 01:06:49.032
And they know that is correct.
01:06:49.032 --> 01:07:02.900
If you are able to review a document in seconds versus you send it to email somebody, have somebody look into it and go through it and do their commands with their biases and whatever it is.
01:07:02.900 --> 01:07:03.565
Right.
01:07:03.565 --> 01:07:05.465
So it takes a lot of time. Right?
01:07:05.465 --> 01:07:09.432
So you don't need some of those things anymore.
01:07:09.432 --> 01:07:12.699
If you really if you want to skill yourself, correct.
01:07:12.699 --> 01:07:19.699
You shouldn't be scared and machine learning, you know, as an example, because you can train models for exactly what it is, right.
01:07:19.699 --> 01:07:28.233
Using an auto rickshaws, as is, which have existed for 75 years, you know, it will continue to be driven by humans in another ten years.
01:07:28.233 --> 01:07:33.233
You've got the technology today to do autonomous know, like the way Tesla does. Right?
01:07:33.233 --> 01:07:36.166
It's done as a combination of real time.
01:07:36.166 --> 01:07:36.632
Yeah.
01:07:36.632 --> 01:07:44.865
And because the technologies also include because you can do all processing in real time, you don't need massive cloud connection to do all of these things.
01:07:44.865 --> 01:07:45.833
You don't.
01:07:45.833 --> 01:07:48.865
So that these are all the things right, that you can do.
01:07:48.865 --> 01:07:51.833
One click they have this parking thing.
01:07:52.800 --> 01:07:53.166
Yeah.
01:07:53.166 --> 01:07:56.465
And wouldn't you like a driverless auto rickshaw? Yeah.
01:07:56.465 --> 01:07:57.000
What do you think?
01:07:57.000 --> 01:07:58.532
Yeah, I for one.
01:07:58.532 --> 01:07:59.333
Yeah.
01:07:59.333 --> 01:08:06.199
Now, can I account for all the auto rickshaw drivers who are gang up and say this is the wrong thing?
01:08:06.199 --> 01:08:09.565
I need a job for this because have not prep for this.
01:08:09.565 --> 01:08:10.532
Probably not.
01:08:10.532 --> 01:08:10.932
Yes.
01:08:10.932 --> 01:08:14.065
You know, it's like the bank employees who resisted computerization.
01:08:14.065 --> 01:08:16.265
And today it turns out that the bank is a computer.
01:08:16.265 --> 01:08:16.632
Correct.
01:08:16.632 --> 01:08:23.166
So, you know, it is that, you know, so I'm saying bank is in phone now the bank is a phone.
01:08:23.166 --> 01:08:39.233
You know, at the end of the day, you know, for all of these things, you know, so, and I mean, it's more of a philosophical discussion of human creativity and all of these things, but also human creativity is a subject of inherent bias.
01:08:39.233 --> 01:08:43.666
Is is it also okay, guru Can you go on the environment?
01:08:43.666 --> 01:08:49.399
But I want to say I have hardly know the answer to it.
01:08:49.399 --> 01:08:50.765
Exactly. Okay.
01:08:50.765 --> 01:09:06.600
So subject matter of your learning, learning and so on, you know, because your are also biological by nature that means that, you know, you will also have a set survival senses kicking in it.
01:09:06.600 --> 01:09:09.033
It's not necessarily always going to be.
01:09:09.033 --> 01:09:15.966
So there's a system architecture, you know, at the latest stage you know, how can evole all of these.
01:09:15.966 --> 01:09:20.632
And then there's some other things you that's subject matter for another discussion.
01:09:20.632 --> 01:09:26.166
And I think we've and some thoughts on design about how things evolve.
01:09:27.300 --> 01:09:46.533
back to kumbhakarna moment back to kumbhakarna moment its like indian goverment wake up and thinking, oh, I need all this GPUs I have a lot of intellectual matter, you know, of all of them, to be able to sort of, it's almost all the some like outdated and so on.
01:09:46.533 --> 01:09:50.733
And so anyway, you don't have an intellectual.
01:09:50.733 --> 01:09:54.765
man or women power because of lake of skills Yeah.
01:09:54.765 --> 01:10:10.166
But that's the kumbhakarna moment where they wakeup and figerout And the rest of the world is sort of like overtaken, and and I'm still in the services economy of 50 70 years back, and I have not done anything to encourage innovation. Correct.
01:10:10.166 --> 01:10:10.466
Yeah.
01:10:10.466 --> 01:10:16.199
And that's a subject matter of so many other things, which has been so obvious.
01:10:16.199 --> 01:10:18.966
But the kumbhakarna moment is on and on.
01:10:18.966 --> 01:10:27.199
And now, you know, I think companies say we say clear LLM for us you know, that is not followed by some deep thought.
01:10:27.199 --> 01:10:35.966
That's what you want to do with LLM correct, you know, and the use, you know, when you building something you need to have like an end goal in mind.
01:10:35.966 --> 01:10:47.065
We've seen that and we call it the kumbhakarna movement because, you know in the Ramayana you have kumbhakarna who could sleep for how many years at the time 12 years at the time or something, right.
01:10:47.065 --> 01:10:51.500
You know, and you would, wake up for 12 years and sleep for 12 years.
01:10:51.500 --> 01:10:51.865
Okay.
01:10:51.865 --> 01:10:55.466
And during that time, so many events would have changed, right?
01:10:55.466 --> 01:10:59.365
And that, he was always lagging behind by 12 years.
01:10:59.365 --> 01:10:59.966
Right.
01:10:59.966 --> 01:11:03.033
And looks like the current state of affairs.
01:11:03.033 --> 01:11:06.932
And it's up to the it's always been up to the private sector to take things forward.
01:11:06.932 --> 01:11:12.033
But you know, this, it's also a good thing that kumbhakarna has been woken up suddenly.
01:11:12.033 --> 01:11:27.699
The out of all of these things that are handing out freebies and other things, suddenly people have figured out that now going forward is going to be can become an existential issue of whether people are going to have the necessary skills or the jobs to be able to do all of these things.
01:11:27.699 --> 01:11:33.365
But from a technology perspective, at least woken up and said, okay, we want to do all of these things as well. Correct?
01:11:33.365 --> 01:11:45.100
Saying so, I think it's a good step that they've taken and hopefully, know, they will sort of like look into, growing some of these things should have been done a long time back.
01:11:45.100 --> 01:11:45.966
Correct.
01:11:45.966 --> 01:11:53.332
And our own experience when we tried to explain our company and AI to a lot of people as we were growing, I'm still not correct.
01:11:53.332 --> 01:12:00.932
People don't understand or are at a skill level where they can really see how things can move.
01:12:00.932 --> 01:12:01.765
Correct.
01:12:01.765 --> 01:12:19.533
To really spend time, you know, with them and then suddenly becomes like, oh, I'm going to lose my job and I want to block this, correct But I want my son or daughter to go abroad and work in Microsoft and make a lot of money so that that that conundrum is not not being solved, as that does not change.
01:12:19.533 --> 01:12:21.666
All of that is going to stop.
01:12:21.666 --> 01:12:30.399
I think it's about time south, like focus on building apps, on using AI and locally grown AI to make sense.
01:12:30.399 --> 01:12:37.399
Okay, you know, to do this now, do you need a bank to do you think that AI today. Correct.
01:12:37.399 --> 01:12:42.899
You know, can you know, how long do you think it's going to take before my phone can talk to his phone?
01:12:42.899 --> 01:12:52.065
And, you know, basically sort of exchange money, data or whatever it is that is necessary, you don't need a central bank for any of these things.
01:12:52.065 --> 01:12:58.432
You don't need, you know, at some point in time, there's a need for growing decentralization.
01:12:58.432 --> 01:13:06.565
You know, luckily for us, you know, central Bank introduced the, crypto, currency equivalent of central bank.
01:13:06.565 --> 01:13:08.432
But where are other use cases.
01:13:08.432 --> 01:13:10.033
Mean this is what we're saying. Correct.
01:13:10.033 --> 01:13:21.100
We may have you may introduce some forward thinking this, but it's not being taken by people because there's a lack of exchange and yeah, and skill levels.
01:13:21.100 --> 01:13:25.899
Now this can seems controversial, but this is where it is and it's not going to change overnight.
01:13:25.899 --> 01:13:33.399
It may change with some generational thinking and saying, okay, we are doing this to solve real issues, okay?
01:13:33.399 --> 01:13:42.966
But is it sufficient for me to sit today and say, I want to create a New LLM for India and then do what we don't have a end goal in mind, correct?
01:13:42.966 --> 01:13:46.199
I think this is where the gap is correct.
01:13:46.199 --> 01:14:04.233
And if you have civil engineers who want to sort of like become testers, and if you have, structurally engineers, you know, wanting to do software, then you know that there is a real services element, which is really taken the take on the economy in a very, very wrong direction.
01:14:04.233 --> 01:14:04.600
Correct.
01:14:04.600 --> 01:14:08.699
So, these are all things that we can talk about.
01:14:08.699 --> 01:14:11.199
People know but can’t solve?
01:14:11.199 --> 01:14:14.500
But with AI you stand to have a real chance, isn't it?
01:14:14.500 --> 01:14:17.632
You know, that is the way That's the way to adopt things.
01:14:17.632 --> 01:14:20.865
So, so when.
01:14:20.865 --> 01:14:21.800
Thanks.
01:14:21.800 --> 01:14:34.699
As I said, once again, this has been like a monologue because there are a lot of things that we wanted to cover and a lot of the products that we mention, you know, where I need stuff, I get some more clarity and things like that as well.
01:14:34.699 --> 01:14:38.466
thanks, krithi for moderating this long session.
01:14:38.466 --> 01:14:39.899
Also, Guru, okay.
01:14:39.899 --> 01:14:44.233
For, for this and for the team for getting this together, thanks for breaking it down.
01:14:44.233 --> 01:14:48.932
So both of you Chetan and Guru It was really insightful looking through it.
01:14:48.932 --> 01:14:49.533
Yeah.
01:14:49.533 --> 01:14:52.600
So the LLM is never going to call you sir okay.
01:14:52.600 --> 01:14:53.300
Right. Okay.
01:14:54.733 --> 01:15:00.466
So you can train it to call you that, but that's an example of, how it can be done.
01:15:00.466 --> 01:15:01.033
Right. Thank you.
01:15:01.033 --> 01:15:01.666
Thanks so much.
00:00:00.133 --> 00:00:06.266
Welcome to the 2025 podcast is our first, podcast in 2025.
00:00:06.266 --> 00:00:10.699
So myself Krithi and this is Chetan and this is Guru.
00:00:10.699 --> 00:00:28.365
Okay, so so we'll be starting off with the podcast, So, sir there's been a lot happening at Quoqo and especially, outside the legal space So basically the Qfin and Syn-Apt seem to be making waves, oh, so what's the perspective of yours on this?
00:00:28.365 --> 00:00:42.399
I think before we get into this, Guru I think in 2025, it started with a big bang correct ... more In terms of AI and events across AI and also, events at Quoqo Correct.
00:00:42.399 --> 00:00:48.865
And I think we should sort of like give viewers a perspective about what's happening in the world.
00:00:48.865 --> 00:01:01.765
I'm sure people have, heard of Deepseek and, what's going on in China and, how it is open source and is cheaper and can be so many things and things like that.
00:01:01.765 --> 00:01:09.965
So, I mean, where we left off with Code and Counsel in 2024, was we were we were not even at an open level.
00:01:09.965 --> 00:01:10.433
Correct.
00:01:10.433 --> 00:01:13.965
When it came to sort of OpenAI and things like that.
00:01:13.965 --> 00:01:28.165
But I mean, just to give, viewers a perspective in terms of large language models, in the size of these, if you go to Hugging face, Guru or if you, generally investigate, you will find a few large players, right?
00:01:28.165 --> 00:01:31.933
got obviously OpenAI, which is open or close sourced ...
00:01:31.933 --> 00:01:33.566
depends on whom you ask.
00:01:33.566 --> 00:01:36.965
If you ask Elon Musk ... he will say, I'm sure it's open source.
00:01:36.965 --> 00:01:40.299
But, anyway, there's, Claude by Anthropic.
00:01:40.299 --> 00:01:44.132
The 3.5 sonnet is is a super LLM ... right?
00:01:44.132 --> 00:01:55.799
And, then you've got Mistral, then you've got Microsoft's phi some math specific modules, there are many, many correct for all of these things.
00:01:55.799 --> 00:02:06.433
And then you have, the ability to run some of these locally, you know, using Ollama, whatever you need ... Qwen And I mean, you know, by the way, China also has its own models.
00:02:06.433 --> 00:02:07.566
You know, you've got Qwen ...
00:02:07.566 --> 00:02:13.233
you've got, Deepseek V2, you know, which came out, and the hints around some of these things as well.
00:02:13.233 --> 00:02:22.332
But, what's happening, I believe, is, you know, depending on how you encode, LLM can either be, text generation or reasoning LLM.
00:02:22.332 --> 00:02:27.432
And most of the OpenAI ones are reasoning LLM and things like these.
00:02:27.432 --> 00:02:41.432
So, and what Deepseek has come up is potentially, an O1 competitor and, you know, we, you know, I told you like, the day it got launched thing that I tested this and I thought that, you know, it may or may not be production ready yet.
00:02:41.432 --> 00:02:58.665
At least, you know, from where we, looking at things and we might need something much more, you know, and I thought that you know, and my first, inclination was, OpenAI still had a lot of edge over some of these things from a production perspective in terms of, our products and things like that.
00:02:58.665 --> 00:03:02.199
Why I'm giving this context is how things have changed.
00:03:02.199 --> 00:03:18.199
You know, you've suddenly gone from a paid model and deployed only on Azure and to, fully open source model where you can deploy on, on your servers provided you have the ability to sort of like deploy these.
00:03:18.199 --> 00:03:26.699
But, you know, most people don't also realize saying that if I want to I mean, you have different quantized versions of Deepseek as well.
00:03:26.699 --> 00:03:27.466
Okay.
00:03:27.466 --> 00:03:34.400
You may be able to run like a 1 - 1.2 Billion on a local comp, let's say, an eight core GPU.
00:03:34.400 --> 00:03:45.832
But you know, if you want the full length you know, if you want like the 64 billion or whatever that run it with or 72, you need some real, really powerful machines.
00:03:45.832 --> 00:03:46.165
Correct.
00:03:46.165 --> 00:03:55.233
So it's not something which, saying, is it production ready, is it something that you can sort of, like immediately adopt it?
00:03:55.233 --> 00:04:01.665
And, I would not necessarily sort of like immediately jump at some of these things. Correct.
00:04:01.665 --> 00:04:03.765
It really depends on the context.
00:04:03.765 --> 00:04:17.333
You know, then, you know, definitely take some time to understand how the models are performing, how it can be used for certain applications, etc., but at least Deepseek has been a ground breaking and it shatters all beliefs even the mindsets of people.
00:04:17.333 --> 00:04:20.100
All you need a lot of resources to train a model.
00:04:20.100 --> 00:04:29.132
Oh, it's not possible for anyone outside the Europe or the USA to develop models So if something that got shattered, going in the direction I think, you know.
00:04:29.132 --> 00:04:29.733
Yeah, yeah.
00:04:29.733 --> 00:04:45.165
So, but you know, this I was talking to, know, a VC recently, and I mentioned this thing, and I had read somewhere saying and it was a proper sort of like encapsulation of the moment thing, an LLM has become a commodity. Correct.
00:04:45.165 --> 00:04:56.432
And it's basically on top of what you build is what really matters and what other what you build is essentially like a wrapper, but you know, you don't need a VC to what I've had happen, right?
00:04:56.432 --> 00:04:59.033
You know, you don't need it requires an adoption.
00:04:59.033 --> 00:05:02.966
That's what I, I think that's what we've done for Quoqo for a long time.
00:05:02.966 --> 00:05:06.632
But, so does it change the mechanics overnight?
00:05:06.632 --> 00:05:10.699
I'm saying does it make it accessible or is it something that, that I can do it?
00:05:10.699 --> 00:05:19.966
before Deepseek maybe not at a level of a reasoning model, but, there are others as well which did the tasks pretty competently Correct.
00:05:19.966 --> 00:05:23.399
We have benchmarked, we've got our own QLLM which we did.
00:05:23.399 --> 00:05:30.165
We did not need, huge, quantized model to be, utilized we trained on some things.
00:05:30.165 --> 00:05:40.300
And we thought that whatever the current state of, requirements are you know, from a user in a production scenario perspective was more than sufficient with what we have now.
00:05:40.300 --> 00:05:58.466
Right now, if you if you're looking at a PhD, there were like comments just yesterday, OpenAI came up with something, very, very for a GPT pro user spending like $200 a month, you know, you've got like a PhD level, output window that, research reports and.
00:05:58.466 --> 00:06:04.065
Yeah, but then, you know, human ingenuity is something that you may not be able to replace.
00:06:04.065 --> 00:06:22.966
And I think Deepseek if you read that paper and you know the way that they've gone about sort of like, you know, adopting a non cuda, architecture to be able to sort of human sort of like develop that model, you know, gone above and sort of like use a very different way of looking at things.
00:06:22.966 --> 00:06:23.399
Correct?
00:06:23.399 --> 00:06:28.432
as you know, like if you're using GPUs, these are usually Nvidia ones.
00:06:28.432 --> 00:06:41.632
the reason why AMD is also been left behind is, you know, the world adopts Cuda architecture for, programing and, you don't find the same adoption in AMD or an Intel GPU kind of like a scenario, right.
00:06:41.632 --> 00:06:44.865
So now you have gone about and created something else.
00:06:44.865 --> 00:06:45.199
Correct.
00:06:45.199 --> 00:07:03.333
Which can work in parallel, you know, looks at Cuda as a separate, if it looks at a GPU as a separate, sort of like an entity from which it can tap in and tap out I mean, this is at a very simplistic level, and not that I also that I'm an expert at how this how they've come about doing this.
00:07:03.333 --> 00:07:16.466
But this involves sort of like, some, some element of deep research But how does this all sort of like add up and Kriti’s question correct, saying, got a lot of things that are coming up.
00:07:16.466 --> 00:07:19.766
And what is it, you know, you've got some new products also coming up.
00:07:19.766 --> 00:07:20.165
Correct.
00:07:20.165 --> 00:07:21.899
And what is it that, mentioned?
00:07:21.899 --> 00:07:37.466
Qfin and I Syn-apt So just to give users, an overview, if you go to our website www.quoqo.com then at that point in time you will sort of, you know, there are links to non legal tech applications.
00:07:37.466 --> 00:07:44.399
So Qfin is is a set of financial agents, almost 30 or 32 of them that we built using our expertise on.
00:07:44.399 --> 00:07:49.865
Q bot, that solves things like, financial consolidation of subsidiaries as an example.
00:07:49.865 --> 00:07:50.632
Okay.
00:07:50.632 --> 00:07:52.000
It's a huge problem.
00:07:52.000 --> 00:08:01.300
And also, you know, especially in large enterprises, it also works with audits sort of, with with auditing requirements in mind.
00:08:01.300 --> 00:08:13.899
So there are quite a few, you can also visit the sub-site qfin.guru So and then and then look at it Syn-apt is also on the agentic AI process.
00:08:13.899 --> 00:08:23.665
we can assist with some custom development of, agentic AI Very specific to a company as requirements as well, during development of Qbot what if you remember?
00:08:23.665 --> 00:08:33.765
Qbot is, specific to legal tech in terms of having a conversational interface for legal specific data sets, right.
00:08:33.765 --> 00:08:36.732
So you've got Qfin, which is also agentic AI Okay.
00:08:36.732 --> 00:08:41.799
Not exactly in the same lines of Qbot but Syn-apt, goes a little bit further.
00:08:41.799 --> 00:08:42.232
Correct.
00:08:42.232 --> 00:08:44.133
And uses different models.
00:08:44.133 --> 00:08:48.966
and different agents to be able to accomplish a number of different things as well.
00:08:48.966 --> 00:08:57.932
Qfin is advanced in the sense that, it contains a lot of technology and a lot of different LLMs to be able to accomplish what it does.
00:08:58.932 --> 00:09:08.033
When you're doing financial apps, as you know, Guru our agents, the, accuracy is a lot more important.
00:09:08.033 --> 00:09:08.899
Yeah. Important.
00:09:08.899 --> 00:09:09.200
Correct.
00:09:09.200 --> 00:09:17.299
You try and run a LLama I think .2 latest on and mathematical sort of like reasoning, you know, like, you know, it doesn't do well.
00:09:17.299 --> 00:09:17.799
Okay.
00:09:17.799 --> 00:09:26.066
So, and it's not I mean, you mentioned Llama, but and depends on, and we have our own thoughts around these.
00:09:26.066 --> 00:09:31.566
I don't want to make these because on these, proprietary and how we test some of these things as well.
00:09:31.566 --> 00:09:41.232
if you remember, in the early days of, GPT2 you could convince, the GPT that two equals one, you know, two equals zero.
00:09:41.232 --> 00:09:47.000
And you could just load it up with so many things that it's not like.
00:09:47.000 --> 00:10:01.033
I mean, I would be there, but, I wanted to give users the ability to try out Syn-apt and Qfin these are cutting edge, agent AI and, and fintech specific products.
00:10:01.033 --> 00:10:08.299
These are as cutting edge as probably Deepseek you know, when it comes to functional sort of like use cases.
00:10:08.299 --> 00:10:13.799
And we believe like we are the first ones in the world sort of like create all of these things as well.
00:10:13.799 --> 00:10:28.166
And by the way, you know, with, with Q fin, we can integrate into a number of, different architectures in terms of, also integrations, things like these, so that you can just plug and play.
00:10:28.166 --> 00:10:30.265
So this is our new thing.
00:10:30.265 --> 00:10:37.033
Same, Qfin and Syn-apt new things for Quoqo and developed.
00:10:37.033 --> 00:10:53.100
by utilizing the knowhow we do things on the legal tech side, I wish some of you And by the way Guru you want to speak about how enterprises are adoption these two things Yeah, definitely.
00:10:53.100 --> 00:11:11.100
You had demos with a few large companies you know, you know, like the day after, when they were looking at these use cases and they can, you know, decided evolution, all little, you know, all of them went to their own, yeah.
00:11:11.100 --> 00:11:16.200
And how they can, use something because most of these activities are really valuable to them.
00:11:16.200 --> 00:11:24.033
And, you know, the, the level of some of these, you know, artificial them will be very interesting, you know?
00:11:24.033 --> 00:11:24.466
Yeah.
00:11:24.466 --> 00:11:31.966
And, on that doesn't mean that it's probably going to be deployment soon as been. Okay.
00:11:31.966 --> 00:11:53.799
And, so, we had the opportunity to demonstrate a very large, in financial services company that did the work being done by about 24 analysts, you know, is essentially sort of like can be automated to an extent, with, lot more efficiency.
00:11:53.799 --> 00:12:07.865
And, you can and, and also it is quite a, I could execute tasks within, say, 20, 25 minute window, which is like, just like an eternity in computer processing times.
00:12:07.865 --> 00:12:08.700
Correct.
00:12:08.700 --> 00:12:16.033
Which, people would take like that twenty four people would take like a month to execute, essentially. Right.
00:12:16.033 --> 00:12:23.033
So it's reaching a phase saying, you know, it is possible to sort of like bring in a lot of efficiency.
00:12:23.033 --> 00:12:25.000
Cost savings. Right.
00:12:25.000 --> 00:12:27.865
Into the process.
00:12:27.865 --> 00:12:32.166
Not that, you know, that the human element is taken out of the mix, right?
00:12:32.166 --> 00:12:39.133
But the human element can be focused on better task of analyzing what comes out of it, rather than trying to dig for the information.
00:12:39.133 --> 00:12:56.700
You know, so I you yeah, I think the, the better way to do this is if you remember, Alta Vista which was the forerunner of, Google you know, if you try to run search in Alta Vista that used to be a big thing, like saying that you are able to sort of like look at stuff and search for it.
00:12:56.700 --> 00:13:01.332
I think that is the that is the use case in many enterprises right now.
00:13:01.332 --> 00:13:07.399
So the way they do things is at least about 35 to 40 years old. It has not been changed.
00:13:07.399 --> 00:13:10.732
And for a long time.
00:13:11.332 --> 00:13:29.633
And when you expose on these things, it's basically like it's like, it's a paradigm shift in thinking saying, oh, there are better ways to do things because as technology advances, reallocate resources, time and money, correct, for all of these things as well?
00:13:30.633 --> 00:13:46.966
So that is, that is how some of the agenitc AI systems are also sort of like scaling up, but sorry, it's a longwinded and sort of like, long monologue from my think it just also outlined how, how much we are using these applications in our day to day.
00:13:46.966 --> 00:13:48.432
activities.
00:13:48.432 --> 00:13:50.000
yeah, that's an important thing to highlight.
00:13:50.000 --> 00:13:59.500
Also, I think within Quoqo and Krithi you need to speak how you are using some of the, some of the agents that you have correct in your day to day work as well. Right?
00:13:59.500 --> 00:14:12.432
So for that, so like you mentioned, like the deekseek and o1 So I, I'm hearing many things from many people that, you know, they're comparing both of them like what is the main difference.
00:14:12.432 --> 00:14:25.299
between deepseek and openAi I think basically Deepseek is also trained on openAI LLMs.
00:14:26.566 --> 00:14:46.533
you know, responses, you know, they claim that only about 830 of them were used, from an API Correct, to the level that they use a different way in which, they use, reinforcement learning. Correct.
00:14:46.533 --> 00:14:55.633
They use a method called called Mixture of Experts it's like a developing smaller LLMs and integrating them with those bigger umbrella.
00:14:55.633 --> 00:15:04.633
And based on the question that you ask, it basically, you know, activates that particular agent, which, gives you the more pointed response.
00:15:04.633 --> 00:15:10.533
That's always the so would it be right to say Guru is like, you know, I remember root learning nowadays in schools.
00:15:10.533 --> 00:15:12.100
That is not encouraged.
00:15:12.100 --> 00:15:17.666
But let's say you are to root learn tables from 1 to 20, correct.
00:15:17.666 --> 00:15:22.399
That is, that is used to basically develop your first, language model.
00:15:22.399 --> 00:15:23.533
Correct.
00:15:23.533 --> 00:15:25.500
And then you know how to multiply, right?
00:15:25.500 --> 00:15:34.265
Then you teach a child or, computer to go to the 21st table or the 35th table.
00:15:34.265 --> 00:15:35.000
Correct. Right.
00:15:35.000 --> 00:15:40.700
And then you extended, you know, and then it figures out that you can do it, you know, for infinity.
00:15:40.700 --> 00:15:41.500
Correct as well.
00:15:41.500 --> 00:15:42.133
Correct.
00:15:42.133 --> 00:15:47.033
So and the way that you can learn the table is multiple ways you can do an addition.
00:15:47.033 --> 00:15:48.566
You can do multiplication.
00:15:48.566 --> 00:15:49.365
Correct.
00:15:49.365 --> 00:15:50.732
Can use squared methods.
00:15:50.732 --> 00:16:02.533
You know depending on so that these are so Deepseek basically explored one of the methods similar to this in which it's like a thinking Method too. Is that right Guru Is that right.
00:16:02.533 --> 00:16:06.166
Another way of looking at it is let us say, there is a closed room in which there are multiple experts.
00:16:06.166 --> 00:16:09.466
The finance guy and a marketing guy and there is a sales guy.
00:16:09.466 --> 00:16:13.700
And, you know, all this, you know, all these different types of people.
00:16:13.700 --> 00:16:15.033
Now you are asking question.
00:16:15.033 --> 00:16:21.765
standing outside the room and you will get the correct answers, because the person who is actually having expertise in that field is going to answer you.
00:16:21.765 --> 00:16:43.432
So that's basically how you are able to sort of like, yeah, but you know, earlier you do is instruction based, sort of like, you know, your you would have to change the context for the LLM or through prompting to sort of like put it in a different frame and make sure that and but now even the LLMs also sort of like you have the ability to have memory.
00:16:43.432 --> 00:16:55.000
So it's all and also one of the things is, you know, if you use, for instance, like 3.5, the early version of OpenAI ... not the turbo one correct, as an example from OpenAI. Correct.
00:16:55.000 --> 00:16:59.000
You know, it would quickly lose context, right, right.
00:16:59.000 --> 00:17:00.500
And then it forget.
00:17:00.500 --> 00:17:07.099
So, so all of those things are being tackled on a, much but on a much better way. Yeah.
00:17:07.099 --> 00:17:13.665
And have larger context windows that have better output, output token sizes.
00:17:13.665 --> 00:17:13.965
Okay.
00:17:13.965 --> 00:17:14.732
I will come to that.
00:17:14.732 --> 00:17:16.066
I have my own things on that.
00:17:16.066 --> 00:17:17.299
Yeah, yeah.
00:17:17.299 --> 00:17:23.032
It's because I think it's is it's very hard to implement in the local production.
00:17:23.032 --> 00:17:23.500
Yeah.
00:17:23.500 --> 00:17:26.400
I, I have my own views on this. Correct.
00:17:26.400 --> 00:17:27.865
Do you want to go first Guru?
00:17:27.865 --> 00:17:28.665
or should I.
00:17:28.665 --> 00:17:28.932
No.
00:17:28.932 --> 00:17:29.500
Go ahead, go ahead.
00:17:29.500 --> 00:17:30.232
Okay okay.
00:17:30.232 --> 00:17:33.665
So see when you deploy locally. Correct.
00:17:33.665 --> 00:17:37.532
You are assuming that you have servers.
00:17:37.532 --> 00:17:37.900
Correct.
00:17:37.900 --> 00:17:41.066
Which are capable of running an LLM.
00:17:41.066 --> 00:17:47.432
So let us say you remember the, you know, the logerithmic skale in school.
00:17:47.432 --> 00:17:47.799
Correct.
00:17:47.799 --> 00:17:50.566
So you have let's take the number two.
00:17:50.566 --> 00:17:55.965
You have to two squared, 2 cube, Two to the power of four to the power of eight and so on. Right.
00:17:55.965 --> 00:18:00.965
So if you take two squared, you know, you get a smaller number.
00:18:00.965 --> 00:18:04.865
If you take to sort of eight it's going to be large.
00:18:04.865 --> 00:18:15.965
If you take to the power of zero it becomes one The point being that these are all quantizations Correct.
00:18:15.965 --> 00:18:21.400
So the larger so you can take a very big model and take a sample of it.
00:18:21.400 --> 00:18:24.932
and you can run it on a smaller level computer.
00:18:24.932 --> 00:18:29.532
It only is like I only know two to the power of two. Correct.
00:18:29.532 --> 00:18:34.900
But if you take two to the power of 64 or whatever, I'm just simplifying this.
00:18:34.900 --> 00:18:43.432
You have a very large number to be able to compute If you do two to the power of 64 on your calculator app, on your computer will give you an error.
00:18:43.432 --> 00:18:44.799
It will give you an infinity error.
00:18:44.799 --> 00:18:45.165
Correct.
00:18:45.165 --> 00:18:55.099
So because beyond a certain point, the floating points it goes beyond you know, it goes beyond the possibility to create such a large number.
00:18:55.099 --> 00:18:56.000
Correct.
00:18:56.000 --> 00:19:03.465
But what happens is at that scale, you have a lot more information easily, which is not compressed.
00:19:03.465 --> 00:19:08.365
So for from the LLM’s perspective, you can access that information.
00:19:09.165 --> 00:19:20.432
there was a visualization from China about how an LLM then sort of like upgrades, different layers and how it's able to train and figure out what it is. Right.
00:19:20.432 --> 00:19:24.465
You know, in an uncompressed model, it has access to all of these.
00:19:24.465 --> 00:19:34.165
And then basically depends on the GPU's power to sort of like, chunk out and, be able to arrive at the computation that makes sense.
00:19:34.165 --> 00:19:34.532
Correct.
00:19:34.532 --> 00:19:40.865
So, it is like dealing with uncompressed files and dealing with compressed files.
00:19:40.865 --> 00:19:41.532
Yeah, yeah.
00:19:41.532 --> 00:19:46.900
And if you need to deal with a compressed file, it's going to take time for you to extract kind of open it.
00:19:46.900 --> 00:19:50.000
It may not have the same color accuracy.
00:19:50.799 --> 00:20:00.099
it may have removed a bunch of repeating colors as an example in a graphic and things like that.
00:20:00.099 --> 00:20:00.833
Right.
00:20:00.833 --> 00:20:05.133
But, you know, in an uncompressed format, correct.
00:20:06.799 --> 00:20:17.400
every pixel is a representation of an actual graphic Our video editors, as an example, color accuracy is very, very critical for them.
00:20:17.400 --> 00:20:20.165
So they have to deal with uncompressed formats.
00:20:20.165 --> 00:20:25.066
And when you actually view this podcast, it's a compressed version of whatever it is.
00:20:25.066 --> 00:20:26.665
It's not the original file.
00:20:26.665 --> 00:20:32.432
Similar to the LLMs, you will have differences in quality when you run all of these things as well.
00:20:32.432 --> 00:20:36.599
Now to back to your question, saying, is it easy to run all of these things?
00:20:36.599 --> 00:20:43.200
What I said is If you want to run the uncompressed version, it's going to take some very serious hardware.
00:20:45.865 --> 00:20:50.700
you need to have the proper servers, the GPUs, the data storage.
00:20:50.700 --> 00:20:55.465
And it's also a question of how fast do all of these can, do a token output.
00:20:55.465 --> 00:20:56.200
Correct.
00:20:56.200 --> 00:21:02.900
If you are in a production grade environment where people sort of, let's say you get a token output of 20 tokens.
00:21:02.900 --> 00:21:03.833
Correct.
00:21:03.833 --> 00:21:08.066
But your output window is, 6000 tokens.
00:21:08.066 --> 00:21:09.732
How long is it going to take?
00:21:09.732 --> 00:21:15.032
You figure 20 tokens per second to figure out something, which is $6,000?
00:21:15.032 --> 00:21:18.932
Each token and say is about 13, 14 letters, correct?
00:21:18.932 --> 00:21:21.000
You know, on an average.
00:21:21.000 --> 00:21:22.833
So it's going to take some time.
00:21:22.833 --> 00:21:23.066
Okay.
00:21:23.066 --> 00:21:28.532
If you use our products If you take Quoqo products, it's literally instantaneous.
00:21:28.532 --> 00:21:28.833
Correct.
00:21:28.833 --> 00:21:32.000
Because we have the infrastructure to run some of these things as well.
00:21:32.000 --> 00:21:41.333
So but the short answer is unless you have the hardware necessary to run all of these things, and the hardware keeps progressing.
00:21:41.333 --> 00:21:41.833
Correct.
00:21:41.833 --> 00:21:46.900
It's not that older hardware can't run or newer hardware is the only way forward.
00:21:46.900 --> 00:21:51.932
Really depends on what kind of application that you want and to whom are catering to.
00:21:51.932 --> 00:22:00.766
You know, for someone who is sort of, know, is able to wait correct, then you know, it is not mission critical, then you can wait.
00:22:00.766 --> 00:22:03.333
You know, maybe it'll land in your inbox one hour later.
00:22:03.333 --> 00:22:04.232
As an example.
00:22:04.232 --> 00:22:05.099
Right. All right.
00:22:05.099 --> 00:22:07.633
And, and then the results come over the next day or so.
00:22:07.633 --> 00:22:14.833
That may be a case where you may want it on this You know, on, older less powerful slower servers, things like this.
00:22:14.833 --> 00:22:30.299
But then if you want to run this, production grade environment and conforming to Gen Z standards of wanting everything instantaneously, right, then you need something very, very spontaneous and very, very powerful infrastructure.
00:22:30.299 --> 00:22:33.299
Yeah, yeah, yeah, I think that's very much so.
00:22:33.299 --> 00:22:34.865
And also to add to that.
00:22:34.865 --> 00:22:35.599
Okay.
00:22:38.133 --> 00:22:46.200
depending on what you run you know, if you run things locally, you're also limited by something called an output window size.
00:22:46.200 --> 00:22:46.500
Correct.
00:22:46.500 --> 00:22:48.833
It really depends once again on the infrastructure.
00:22:48.833 --> 00:22:50.400
Yeah, you can change that.
00:22:50.400 --> 00:22:54.232
But you know, more stuff running locally.
00:22:54.232 --> 00:23:01.665
You know, if you use open source stuff to be able to run and then locally limited to over 2000 tokens, whether 2000 tokens.
00:23:01.665 --> 00:23:04.200
takes for ever to generate also depending on the hardware.
00:23:04.200 --> 00:23:05.032
Correct.
00:23:05.032 --> 00:23:17.032
And if you want, if you want like, OpenAI research tool to connect to generate like 15 pages, it's gone to take time even for a OpenAI it will take time Easily.
00:23:17.032 --> 00:23:17.833
Some 15 minutes.
00:23:17.833 --> 00:23:21.165
Ten minutes even for them So, yeah.
00:23:21.165 --> 00:23:33.165
And if some of those things I don't think locally can run, you know, a need like an entire, Sever farm to be able to sort of like, be able to handle complex tasks.
00:23:33.165 --> 00:23:47.532
Also, you know, for shorter things, like you want a conversational bot as a fancy to run locally, I think it's great to have, a locally running LLM, but anything more than, you know, as an example.
00:23:47.532 --> 00:23:47.799
Correct.
00:23:47.799 --> 00:24:04.200
We have like Diligence project We have a product called Diligence, where it can review so many things together and basically, you know, come with a conclusion or a help you arrive I want to conclusion or another or an analysis on a transaction as an example.
00:24:04.200 --> 00:24:04.900
Correct.
00:24:04.900 --> 00:24:08.165
It may not be possible to run it with current output windows.
00:24:08.165 --> 00:24:12.700
You know, it's like it's like getting cut off mid-sentence.
00:24:12.700 --> 00:24:14.465
Okay.
00:24:14.465 --> 00:24:16.200
So you'll get a report.
00:24:16.200 --> 00:24:22.133
But you know, what's going to happen is the computer is going to wait for some time for this, output to come through.
00:24:22.133 --> 00:24:27.633
And, you know, the computer will wait for the time allotted and then say, okay, this is it.
00:24:27.633 --> 00:24:29.500
And then it has to complete the task.
00:24:29.500 --> 00:24:32.032
You are dealing with it at the end of the day of machine. Correct.
00:24:32.032 --> 00:24:34.732
The machine has like a start loop and an end loop.
00:24:34.732 --> 00:24:37.833
And this is essentially how it works in real life.
00:24:37.833 --> 00:24:42.566
But, that's what I think Guru. Guru is a lot more, completely agree with.
00:24:42.566 --> 00:24:43.465
Yeah, yeah.
00:24:43.465 --> 00:24:44.532
So a different use case.
00:24:44.532 --> 00:24:47.099
It really depends on the use cases that you are working with.
00:24:47.099 --> 00:25:01.799
Do you want the repository kind of a tool that and AI repository like Quoqo legacy where you can upload your documents and you expect, you know, documents to be available, maybe with, next day or know a couple of days later.
00:25:01.799 --> 00:25:08.365
So now we can now, use local models with, very high level quantization. Correct.
00:25:08.365 --> 00:25:17.932
But then if you want instantaneous results, Chatbot or also some of those other use cases, some of these things, may not be suited for the use case, basically.
00:25:17.932 --> 00:25:18.400
Okay.
00:25:18.400 --> 00:25:25.066
So maybe that is the short response on the higher the quantization level, the lower the quality it becomes.
00:25:30.965 --> 00:25:41.500
Speaking about the, infrastructures and, Quoqo has been doing some cool things, like micro data, you know, endorser centers like throught Quoqo what you're supposed to get around.
00:25:41.500 --> 00:25:44.400
It's like it's the way, like this.
00:25:44.400 --> 00:25:51.099
So I think, you know, when Guru I doing when we distribute as well.
00:25:51.099 --> 00:25:51.833
Okay.
00:25:51.833 --> 00:25:56.133
And I mean, how difficult you know, run some of these things.
00:25:56.133 --> 00:26:22.066
And you also, you know, you either have ambani scale of, you know, hundreds of data center that is yet to be built, versus, you know, you have public cloud like Microsoft or AWS and, Guru is an expert, at, and I always these guru saying he is always over buget by 5X Using the cloud.
00:26:22.066 --> 00:26:22.965
Right.
00:26:22.965 --> 00:26:31.232
And and you know, what is lets say your are teching a model you want to, Guru my trainng said the other day was 6MB.
00:26:31.232 --> 00:26:38.500
Correct But, you know, if I put it through a public cloud in India, it cost us almost ₹55,000.
00:26:38.500 --> 00:26:39.566
Correct. That's done.
00:26:39.566 --> 00:26:42.333
I mean, there were so it was cleaned.
00:26:42.333 --> 00:26:47.432
it was done something, you know, I wanted some special stuff on it or whatever the case is, right?
00:26:47.432 --> 00:26:48.333
It never got.
00:26:48.333 --> 00:26:55.833
And then after having burned through all that, it came back saying that there's an error in the file which it could not compute. Correct.
00:26:55.833 --> 00:27:08.465
So it's not very easy to solve because public cloud infrastructure, though they everyone thinks is very easy to train your model, clean your data set and collect the data set and all of these things.
00:27:08.465 --> 00:27:09.432
Right.
00:27:09.432 --> 00:27:12.400
And it's very, very expensive. Right.
00:27:12.400 --> 00:27:26.365
And, you know, for a infra company like Quoqo, and Quoqo is focused on legal tech, and, you know, you have different use cases and sometimes, you know, you may have data privacy restrictions on utilizing, data.
00:27:26.365 --> 00:27:32.000
And if you look at all the LLM right now, correct, that very focused on the US, as a law.
00:27:32.000 --> 00:27:38.865
Correct Because, know, if you want India specific stuff correctly are to create those data sets yourself and India.
00:27:38.865 --> 00:27:40.700
so unique that the legal correct.
00:27:40.700 --> 00:27:45.432
As an example, it uses the US templates it uses the UK templates.
00:27:45.432 --> 00:27:47.000
It uses India templates.
00:27:47.000 --> 00:27:48.965
All the common law stuff. Correct.
00:27:48.965 --> 00:28:05.133
But if you want to create India specific templates, there's, you know, it's only, and if you ask a human to do it, lawyer to do it today, let's say to draft and agreement, they have an inherent biases to it because they're only trained on documents outside of India.
00:28:05.133 --> 00:28:06.000
Okay.
00:28:06.000 --> 00:28:10.932
So the endeavor is to remove the biases in some of the LLms that we have, correct.
00:28:10.932 --> 00:28:15.633
An example of biases is using English to do this podcast.
00:28:15.633 --> 00:28:19.400
So, you know, we could have done it in any other Indian language, which we know.
00:28:19.400 --> 00:28:20.266
Correct.
00:28:20.266 --> 00:28:23.766
But so those kind of biases coming okay.
00:28:23.766 --> 00:28:31.266
So to remove that and clean all of these things take a lot of a lot of time, effort and energy.
00:28:31.266 --> 00:28:34.766
And it's not that, you know, you can simply have data put it together.
00:28:34.766 --> 00:28:40.000
and do a chat bot you know, we have seen those days where it gives you like, rubbish results.
00:28:40.000 --> 00:28:40.665
Isn't it Guru?
00:28:40.665 --> 00:28:41.665
Oh, yeah.
00:28:41.665 --> 00:28:44.299
And, when you do it, you know, there's a requirement.
00:28:44.299 --> 00:28:53.432
It's a we also have some patent technology, where it's not necessary to always train, but I won’t get, into very big details on that.
00:28:53.432 --> 00:28:58.865
It is, of course, you know, our internal IP for all of these things as well.
00:28:58.865 --> 00:29:11.432
But suffice it to say that, you know, if you're pulling through all of these different bits and pieces, you are going to consume a lot of power and want to consume a lot of computing data and, private cloud.
00:29:11.432 --> 00:29:11.965
Correct.
00:29:11.965 --> 00:29:16.932
I sorry, on public cloud, you will burn a lot of these things.
00:29:16.932 --> 00:29:21.700
And I don't know how many people are successful in getting you.
00:29:21.700 --> 00:29:26.032
You know, you will you will still get a Tesla T4 and never get h100.
00:29:26.032 --> 00:29:31.133
And why is it that you will not get an allocation and simply not to get an allocation?
00:29:31.133 --> 00:29:37.266
And also, if you're looking at India, correct, the concept is all about Services.
00:29:37.266 --> 00:29:40.465
It's never been about products and innovations.
00:29:40.465 --> 00:29:41.400
Correct.
00:29:41.400 --> 00:29:46.066
People underestimate, you know, it's like trying to get a chartered accountant. Correct.
00:29:46.066 --> 00:29:47.066
You will get one.
00:29:47.066 --> 00:29:51.799
Sorry for, you know, all due disclaimers on this.
00:29:51.799 --> 00:29:52.465
Correct.
00:29:52.465 --> 00:29:59.532
You will get somebody depending on how much you pay and you will get you know, it's across the spectrum.
00:29:59.532 --> 00:30:02.532
And this is true of lawyers as well and true.
00:30:02.532 --> 00:30:06.532
But India such a service is denominated economy. Correct.
00:30:06.532 --> 00:30:17.865
That the innovation piece and ability to actually do deep research, and do deep things is underestimated today. Right?
00:30:17.865 --> 00:30:23.566
So, you know, we've got a new product, on the hardware side called Qcore.
00:30:23.566 --> 00:30:28.665
I was, I think about two years back correct Guru about a year and a half back.
00:30:28.665 --> 00:30:33.032
We started experimenting with edge devices, perfected it. Correct.
00:30:33.032 --> 00:30:34.232
It was very difficult to get.
00:30:34.232 --> 00:30:45.932
We were probably the first to get, and we know that Nvidia what I'm saying that, you know, we were we are running, you know, actually LLMs on edge devices, which we had a very specific use case.
00:30:45.932 --> 00:30:48.032
I don't want to get into that.
00:30:48.032 --> 00:30:50.700
But it was possible to run tasks on.
00:30:50.700 --> 00:30:56.133
And at that point in time, you know, Nvidia's only company that made a GPU on an edge device, right?
00:30:56.133 --> 00:30:57.365
So available.
00:30:57.365 --> 00:30:59.432
So, it still does.
00:30:59.432 --> 00:31:04.900
I mean, you know, you you know, it's like many people that are running on Raspberry Pi, I can go ahead.
00:31:04.900 --> 00:31:09.900
Try, you know, see what you get like two tokens a second, you'll be happy.
00:31:09.900 --> 00:31:10.232
Correct.
00:31:10.232 --> 00:31:13.400
So but that it's an edge device.
00:31:13.400 --> 00:31:15.165
It is not a Raspberry Pi.
00:31:15.165 --> 00:31:15.500
Correct.
00:31:15.500 --> 00:31:23.799
So I an edge device needs to have a certain amount of output capacity to be able to run some of these things as well.
00:31:23.799 --> 00:31:26.432
And it needs to take into account concurrency.
00:31:26.432 --> 00:31:29.032
Network availability and so many other things as well.
00:31:29.032 --> 00:31:29.333
Okay.
00:31:29.333 --> 00:31:32.732
Sorry, I'm getting a little technical here, but that's how we started.
00:31:32.732 --> 00:31:33.432
Correct?
00:31:33.432 --> 00:31:35.566
Other damn thing would consume so much power.
00:31:35.566 --> 00:31:40.465
Also, it's not that an edge device, you know, is so energy efficient and things.
00:31:40.465 --> 00:31:47.833
So it's like running, in an old style, filament light bulb, but, 200W. Correct.
00:31:47.833 --> 00:31:48.833
So something like that.
00:31:48.833 --> 00:31:52.766
So and that can consume a lot of power very, very quickly.
00:31:52.766 --> 00:32:04.900
And the devices get hard because it's not necessary that you have cooling systems in, you know, run it in a, place in somewhere in India without adequate cooling and that thing with the smoke.
00:32:04.900 --> 00:32:05.266
Correct.
00:32:05.266 --> 00:32:06.299
So after some time.
00:32:06.299 --> 00:32:10.400
So we've gone through all of these, you know, I've learned a few things also, isn't it.
00:32:10.400 --> 00:32:12.700
So it's happened as well.
00:32:12.700 --> 00:32:16.566
And why I mentioned this is the Qcore is our data center.
00:32:16.566 --> 00:32:24.000
It's a micro data center, where you can run, your applications and inference tasks very, very quickly.
00:32:24.000 --> 00:32:24.665
Correct.
00:32:24.665 --> 00:32:26.900
Because it's made for application.
00:32:26.900 --> 00:32:33.865
And you know what the output levels look like and it can scale to a certain number of users and things like that.
00:32:33.865 --> 00:32:39.165
And it has a certain power capacity and tier 1 networks associated with it.
00:32:39.165 --> 00:32:47.232
It does not meant to become like a huge data center, because in India, people in the moment you go to a very huge data center, correct..
00:32:47.232 --> 00:32:53.000
And you know, the problem becomes one of cooling, of management, of power.
00:32:53.000 --> 00:33:05.465
And tier 4 for data centers, you need to, to have at least, two power sources in India unless you are in, in what you call as a Gadi bhaga (boundary area), which is a border area. Correct.
00:33:05.465 --> 00:33:10.965
You know, you will not have power supply from two discoms Correct.
00:33:10.965 --> 00:33:18.599
electricity board, most states in India have only one discoms So they will get the second discoms Right.
00:33:18.599 --> 00:33:20.432
And the second discom is the sun.
00:33:20.432 --> 00:33:20.865
Okay. Right.
00:33:20.865 --> 00:33:28.566
So that's also limited to approximately 12 hours a day and it's never gonna run on like a gigawatt capacity.
00:33:28.566 --> 00:33:30.333
And on that it's not gonna happen.
00:33:30.333 --> 00:33:39.833
So, you know, you have to do I think in most, places, you know, you can do like a tier 2 different tiers of data centers, no that It's better.
00:33:39.833 --> 00:33:41.333
What do you have?
00:33:41.333 --> 00:33:43.799
Tier 2 or 3 capability as well.
00:33:43.799 --> 00:33:44.599
Correct.
00:33:44.599 --> 00:33:50.032
And tier 3 is quite adequate, you know, to be able to run what you require.
00:33:50.032 --> 00:34:03.133
You know, you don't necessarily need to pay for, hundreds of acres of, you know, you know, to run like cutting, you know you need micro data centers, correct.
00:34:03.133 --> 00:34:09.766
In addition to it, I just also have to add that we want to charge customers in Rupees instead of Dollars .
00:34:09.766 --> 00:34:10.532
Exactly.
00:34:10.532 --> 00:34:11.300
Right. Correct.
00:34:11.300 --> 00:34:16.500
If you look at AWS, Asur or, any other global cloud.
00:34:16.500 --> 00:34:16.766
Yeah.
00:34:17.132 --> 00:34:22.333
So you pay them in dollars, but then that those charges have to be transferred to the customers.
00:34:22.333 --> 00:34:22.599
Yeah.
00:34:22.599 --> 00:34:25.266
That's something I think I think, you know, you're right.
00:34:25.266 --> 00:34:25.965
Well isn't it.
00:34:25.965 --> 00:34:27.333
It's a joke, right.
00:34:27.333 --> 00:34:29.599
People sitting in an next road in Bangalore.
00:34:29.599 --> 00:34:33.599
which charge in dollars to which we think is a joke.
00:34:33.599 --> 00:34:36.365
I mean, people buy it Also, okay, okay.
00:34:36.365 --> 00:34:39.932
So, yesterday I was, talking to someone in Chennai.
00:34:39.932 --> 00:34:40.565
Correct?
00:34:40.565 --> 00:34:44.199
They bloody sitting in Madras and right is charge in dollars.
00:34:44.199 --> 00:34:44.900
Correct.
00:34:44.900 --> 00:34:45.766
You know what? That.
00:34:45.766 --> 00:34:46.565
What's the fun?
00:34:46.565 --> 00:34:46.932
Correct.
00:34:46.932 --> 00:34:48.666
You have just adopted some.
00:34:48.666 --> 00:34:53.099
And if you look at the USD to INR it's indicating a 92.
00:34:53.099 --> 00:34:53.833
Yeah.
00:34:53.833 --> 00:34:56.432
Here it soon becomes unaffordable in India.
00:34:56.432 --> 00:34:57.833
and it's a service economy.
00:34:57.833 --> 00:35:02.333
How many people actually know guru that you need such huge data centers right.
00:35:02.333 --> 00:35:04.333
I know you think about it right.
00:35:04.333 --> 00:35:05.300
Doesn't make sense.
00:35:05.300 --> 00:35:05.632
Right?
00:35:05.632 --> 00:35:08.032
It's making way for the US based companies.
00:35:08.032 --> 00:35:15.465
to take over the indian company And also, you know, to large extent data center is real estate and power play.
00:35:15.465 --> 00:35:15.865
Correct.
00:35:15.865 --> 00:35:22.400
You need to have a real estate to be able to have scale, and you need to have power to be able to rent.
00:35:22.400 --> 00:35:22.565
Okay.
00:35:22.565 --> 00:35:28.699
Now how many one gigawatt stations, you know, per data centers are there in India, Guru at least to my knowledge.
00:35:28.699 --> 00:35:30.400
Zero. Right.
00:35:30.400 --> 00:35:34.500
You know, you top out, say 250 or 300 K.
00:35:34.500 --> 00:35:35.666
We are and things like that.
00:35:35.666 --> 00:35:41.199
And that's the max that you can, you know, I'm not maybe I'm you know, it's kind of stand corrected on these things.
00:35:41.199 --> 00:35:54.565
But to get a gigawatt of power correct requires like, you know, can you imagine the power, supply for some when you need to have, like, coal based nuclear power, nuclear power source?
00:35:54.565 --> 00:35:55.233
Okay.
00:35:55.233 --> 00:36:00.632
Fusion is something which the Chinese and the French have not yet perfected since I did come to India. Correct.
00:36:00.632 --> 00:36:02.699
So they have a lot of things to do?
00:36:04.032 --> 00:36:04.333
Yeah.
00:36:04.333 --> 00:36:21.632
Hopefully it's, you know, we have our own internal jokes where, you know, I feel that, you know, for, for product companies, especially in the AI space and India and want be research and I think you will need these micro data centers.
00:36:21.632 --> 00:36:27.365
And I think the government has an ambitious plan of procuring, thousands of GPUs.
00:36:27.365 --> 00:36:31.500
And, you know, anyway, Nvidia has annual event is coming up.
00:36:31.500 --> 00:36:35.666
So we we know that there's going to be another major technological leap.
00:36:35.666 --> 00:36:36.500
Correct.
00:36:36.500 --> 00:36:42.266
You know, I mean, India wants to subsidize GPUs, what,₹64 an hour? Correct.
00:36:42.266 --> 00:36:43.766
That is the thing.
00:36:43.766 --> 00:36:48.632
But I am asking some people saying, well, that may be an approach towards availability.
00:36:48.632 --> 00:36:49.032
Correct.
00:36:49.032 --> 00:36:50.166
And affordability.
00:36:50.166 --> 00:36:50.932
Right.
00:36:50.932 --> 00:36:55.565
And Who the people who are building these applications right.
00:36:55.565 --> 00:36:58.733
You know, people are living longer hours, right?
00:36:58.733 --> 00:37:00.500
Who is actually doing the deep tech guru?
00:37:00.500 --> 00:37:01.500
I mean, isn't it.
00:37:01.500 --> 00:37:07.900
You can read it as number of research, how research is happening in India for very, very less amount of research.
00:37:07.900 --> 00:37:10.865
Now, what will you do with a grasshopper scale?
00:37:10.865 --> 00:37:16.599
You know, GPU and one one rack is probably enough to run so many things.
00:37:16.599 --> 00:37:23.032
I read a report today that India's fastest supercomputer has, 800 GPUs. in It.
00:37:23.032 --> 00:37:26.833
Okay, you know that maybe older scale or whatever it is.
00:37:26.833 --> 00:37:45.132
We know we've been to enough stuff on, you If you go to a Taiwanese trade show or an Nvidia show, correct, for some of these things, you know, you will have some rack servers and fully contained ones that can probably run like a quarter of India, isn't it?
00:37:45.132 --> 00:37:55.733
You know, in terms of just a small rack, you know, it's no longer about like having a acres and a acres because the, the computing has reached such an extent.
00:37:55.733 --> 00:37:58.365
It's like one, it's like one, 2 or 3 nanometers.
00:37:58.365 --> 00:37:58.599
Right.
00:37:58.599 --> 00:38:07.532
Even the CPU sizes and GPU sizes and the ability to sort of like colitis and weld the technology has reached a stage where you can or GPU, right?
00:38:07.532 --> 00:38:17.532
I mean, the if you look at the, the grasshopper versions that I go to, they basically take like two dice and then Weld in the GPU.
00:38:17.532 --> 00:38:22.199
And, you know, they have the ability to set off like really it's not as simple as just taking these two.
00:38:22.199 --> 00:38:35.900
Obviously I'm oversimplifying, but if you really look at the technology and the ability to interconnect all of these things, the ability, you know, to be able to sort of, you know, have the bandwidth in the GPU to process so much.
00:38:35.900 --> 00:38:36.632
Correct.
00:38:36.632 --> 00:38:38.166
It is decreasing day by day.
00:38:38.166 --> 00:38:48.800
you know, there have been examples, people say I can run, LLM's correct, let's say so many tokens per second on interconnected, Apple Mac mini.
00:38:48.800 --> 00:38:49.965
Correct? Correct.
00:38:49.965 --> 00:38:57.833
And we are on the fourth lesson about M4 pros, but the M2 Pro has more bandwidth and more cores for the GPU, correct?
00:38:57.833 --> 00:38:58.932
Than the M4.
00:38:58.932 --> 00:38:59.766
Okay.
00:38:59.766 --> 00:39:02.565
You know, so you're going to be looking at that kind of throughput.
00:39:02.565 --> 00:39:06.699
It's not necessarily like the older generation is bad.
00:39:06.699 --> 00:39:07.865
And therefore cannot.
00:39:07.865 --> 00:39:14.465
It's more in terms of what is the use case, what does it require to run all of these things?
00:39:14.465 --> 00:39:15.199
Yes.
00:39:15.199 --> 00:39:19.800
And, do you require some huge things to be able to do this right?
00:39:19.800 --> 00:39:22.532
But what is the use case in India for this scope.
00:39:22.532 --> 00:39:22.965
Yeah.
00:39:22.965 --> 00:39:32.065
When we look at OLA, which probably was the first to sort of like AI turning, how many users are there stuff to do all these things.
00:39:32.065 --> 00:39:38.166
And, you know, the question is, you build the wrapper you dont need the kind of, yeah.
00:39:38.166 --> 00:39:39.400
You don't need need it.
00:39:39.400 --> 00:39:45.766
And India is in level of Creating wrapper is indian is not deep tech correct.
00:39:45.766 --> 00:39:47.666
And that's the point, right.
00:39:47.666 --> 00:40:02.800
And the other question was before you started recording this podcast, he was saying, you know, this happened that, Can’t innovate necessarily started people moving and etc..
00:40:02.800 --> 00:40:19.266
But, you know, if you look at someone like you running a company, correct, there's a skill level of people that you get sometimes and sorry, it's it's, you know, it is not, you know, people who say, oh yeah, jobs that are not enough jobs at the end.
00:40:19.266 --> 00:40:30.800
You know, you've been, you know, with all of these things, skill levels have not reached the level which all the knowledge that it was. Okay.
00:40:30.800 --> 00:40:35.865
And there's a gap for this, and this is a huge gap.
00:40:35.865 --> 00:40:36.532
Okay.
00:40:36.532 --> 00:40:42.199
Now, in the government, you know, the don't give it to people like you had them.
00:40:42.199 --> 00:40:50.632
It was a good idea for for a particular property people look at me like this that technology has been for 25 years Yes.
00:40:51.632 --> 00:40:52.065
Okay.
00:40:52.065 --> 00:40:56.132
Correct 20 years atleast maybe not 25.
00:40:56.132 --> 00:40:56.733
Right.
00:40:56.733 --> 00:41:02.266
So and they're saying this which I mean, this is a huge gap.
00:41:02.266 --> 00:41:08.666
And in one industry players and this is and only.
00:41:08.800 --> 00:41:36.432
No, no, no really guru, I mean, you know, it's like, you know, in this particular episode So they seem to have gathered the most for all of these things, but these are all things which I thought that, you know, if I'm doing this, I might as well say it or probably also wasn't saying, today, the, government calls out for all these people to bring to build like and huge GPU cluster, which is maybe going to get outdated by the time it’s going to surface they come through.
00:41:36.432 --> 00:41:47.500
And then you want to build in in India specifically, that's where the huge skill deficit, where people tend to figure, you know, when we build QLLM I put out on LinkedIn post for saying we all laughed at guru.
00:41:47.500 --> 00:41:50.766
We did not even have the resources, right, you know, to do it.
00:41:50.766 --> 00:41:57.965
But we would sort of like, use ingenious methods, putting so many things together to get things running correct.
00:41:57.965 --> 00:42:07.166
And that seems to be the real question mark, saying, you know, one, are you at a skill level where you can do all of these wonderful things?
00:42:07.166 --> 00:42:09.532
The answer is absolutely not correct.
00:42:09.532 --> 00:42:12.900
You are the skill level as a country, correct?
00:42:12.900 --> 00:42:13.733
To do services?
00:42:13.733 --> 00:42:15.266
Absolutely correct.
00:42:15.266 --> 00:42:17.199
Are you at a skill level to build products?
00:42:17.199 --> 00:42:18.532
Absolutely NOt correct.
00:42:18.532 --> 00:42:21.699
So I think that sums it up, isn't it Guru?
00:42:21.699 --> 00:42:25.965
You know, in fact, when we go on sort of like do stuff internationally, correct.
00:42:25.965 --> 00:42:28.632
know, with our innovation, sometimes people are surprising.
00:42:28.632 --> 00:42:29.432
Are you in India?
00:42:29.432 --> 00:42:31.900
Why not come here Why not move here?
00:42:31.900 --> 00:42:34.565
You know, all of those things are just sort of given out.
00:42:34.565 --> 00:42:41.065
But if you look at most of the companies that exist today, they're just copies of things that have come up.
00:42:41.065 --> 00:42:42.099
Yes.
00:42:42.099 --> 00:42:45.932
I think it's just copies of stuff that has come on. Now.
00:42:45.932 --> 00:42:46.833
Where is innovation?
00:42:46.833 --> 00:42:50.233
is the innovation in the UI because it looks nicer.
00:42:50.233 --> 00:42:58.833
Mine is, mine looks my app looks nicer than yours kind of stuff, which is not innovation And I know that right is the innovation in getting things.
00:42:58.833 --> 00:43:01.266
it's like, deliver on time.
00:43:01.266 --> 00:43:02.132
Correct.
00:43:02.132 --> 00:43:04.666
You've got a huge population, which is jobless, correct.
00:43:04.666 --> 00:43:25.132
You know, but you know, therefore utilization But then there is a real, you know, you know, where is a really a vishwarya moment where you can, you know, build, a Dam, or you can really think of, you know, maybe, you know, early 1900s by the British where the Maharaj of Mysore decided to sort of like give a grand opening.
00:43:25.132 --> 00:43:28.733
IASC maybe the scientists at ISRO. Correct.
00:43:28.733 --> 00:43:32.699
That really sort of like, innovation correct to get things done.
00:43:32.699 --> 00:43:35.000
Maybe not saying that it doesn't exist.
00:43:35.000 --> 00:43:36.766
Maybe in certain portions. Correct.
00:43:38.000 --> 00:43:47.132
where some of these things come up, but can we still do, like, booster, the rocket that comes back and hitch itself to, you know, the launch pad?
00:43:47.132 --> 00:43:49.000
No, no, we're not there yet.
00:43:49.000 --> 00:43:58.532
So that requires some amount of innovation, some amount of skill giving some amount of, some sort of dedication to actually take the country forward.
00:43:58.532 --> 00:43:59.666
Well, correct.
00:43:59.666 --> 00:44:08.500
Easiest thing is to leave the country, go outside, and try do something because, again, you don't need a lot of resources to be able to build.
00:44:08.500 --> 00:44:11.599
But you certainly at some point, I don't think there is ever.
00:44:11.599 --> 00:44:19.166
Have you ever faced a situation where we could not buy or do something Guru You know, there's always I think, the difference also is in the mindset.
00:44:19.166 --> 00:44:20.432
Yeah. okay.
00:44:20.432 --> 00:44:27.065
To build, a LLM we need only big companies like OpenAI or Microsoft or Google, etc.
00:44:27.065 --> 00:44:38.132
but then came the deepseek and shattered all the AI and also early attention made here, also saying, we don't you only need like a Mini LLM why should we need a proper LLM?
00:44:38.132 --> 00:44:39.599
Why should we do this?
00:44:39.599 --> 00:44:48.565
We are some interesting, ex company, like three lakh people doing services, you know, you know, the power of AI.
00:44:48.565 --> 00:44:49.800
I don't need to say more.
00:44:49.800 --> 00:44:50.666
Correct.
00:44:50.666 --> 00:44:58.565
But, you know, the early attention made to sort of like, you know, also hijacked the, on the headlines and things like these as well.
00:44:58.565 --> 00:44:58.932
Correct.
00:44:58.932 --> 00:45:17.733
It requires a certain amount of, also brute sort of, brute headedness, you know, to be able to sort of like saying, I'm doing this for the, for whatever it may be, may or may not turn out to be monetarily, you know, successful it may or may not, but it certainly gives you a, I think, guru.
00:45:17.733 --> 00:45:32.166
Can you also say that in the last year we worked on some so such critical technology that has given us so much, happiness that we actually can build this, you know, this this can only come from when you actually build something.
00:45:32.166 --> 00:45:32.865
Correct?
00:45:32.865 --> 00:45:40.699
When you invent something new, when your not just important stuff and trying sell it here, there not like an import export business.
00:45:40.699 --> 00:45:46.199
where import some and try and sell it here Yeah, it's not, not that it is a bad thing. Okay.
00:45:46.199 --> 00:45:52.800
You know, a joke, like very recently, you know, we asked for some, Exim bank support.
00:45:52.800 --> 00:45:53.532
Correct.
00:45:53.532 --> 00:46:01.965
name a particular banking institution and the state of affairs saying I need some support in this part of the globe because we are doing here.
00:46:01.965 --> 00:46:09.166
But, you know, it usually gets delayed because of, you know, the country's tax policies and things like that.
00:46:09.166 --> 00:46:14.900
Can you say, you know, the the answer we got back to Italy was, if you are a chappal manufacturer, we can help you.
00:46:14.900 --> 00:46:20.400
I'm like, you know, this is the place that this is where made in India currently stands Correct.
00:46:20.400 --> 00:46:24.900
And this is where, you know, the state of affairs is when it comes to innovation.
00:46:24.900 --> 00:46:26.965
And we don't say this enough. Correct.
00:46:26.965 --> 00:46:33.166
And even at Quoqo you know, we strive for you know, we like you know, please, innovate.
00:46:33.166 --> 00:46:38.065
And it's not that we don't have a lot of these things cut out or, you know, built correct.
00:46:38.065 --> 00:46:40.800
It's not that you need to use something which is over there.
00:46:40.800 --> 00:46:41.099
Correct.
00:46:41.099 --> 00:46:49.965
You create yourself, you create your own data centers, and you are able to sort of like be able to do that, but you need to have some critical mass for all these things as well.
00:46:49.965 --> 00:46:51.000
just say this.
00:46:51.000 --> 00:46:55.032
Sorry, that's my, that's my gross of the season also, isn't it?
00:46:55.032 --> 00:46:57.532
I'm sure a lot of founders also feel the same way.
00:46:57.532 --> 00:46:58.599
They want to change.
00:46:58.599 --> 00:47:09.733
What do you think? Guru What do you think Krithi LLM , when you were talking and LLM you see LLM the minimal resource, if that, you know, could realistic.
00:47:09.733 --> 00:47:14.766
More like how is it like more realistic using LLM the minimal resources.
00:47:14.766 --> 00:47:18.099
people are talking that LLM are the new when you was bringing with it.
00:47:18.099 --> 00:47:24.599
So you mean building LLM and everything and LLM minimal resources like how is it more realistic?
00:47:24.599 --> 00:47:40.632
So basically, you don't really need, very large number of resources to build a model that can cater to certain, know, specific now, use cases, let's say you are only dealing with legal documents, for example, to take or, talk about what?
00:47:40.632 --> 00:47:42.266
That's been our approach as well.
00:47:42.266 --> 00:47:43.599
Is that okay? Yeah.
00:47:43.599 --> 00:47:59.333
So you don't really need a model that conquer’s the world then that would just need to be able to understand your documents, try to go to, solve the problems if you are working on, so with respect to resources, you don't really need a great number of resources, but the process itself is quite grueling.
00:47:59.333 --> 00:48:04.800
You need to go through the data collected, probably because garbage in, garbage out, you have to clean it properly.
00:48:04.800 --> 00:48:09.333
You have to ensure that the document formatting is proper, etc.
00:48:09.333 --> 00:48:19.333
and then you need to, train it so we need some amount of GPU power to be able to train a model that does a decent, job giving you the response Yeah.
00:48:19.333 --> 00:48:33.000
So just to add there, you know, we've taken an approach where you have a specialized data sets, like actually, you know, and some of these, can be accessed as demo ones, like we have, all india Supreme Court cases.
00:48:33.000 --> 00:48:34.432
We have legislation.
00:48:34.432 --> 00:48:35.065
Correct.
00:48:35.065 --> 00:48:43.065
we have tax bots for member states like these that you can use to query very, very specific data sets. Correct.
00:48:43.065 --> 00:48:55.400
And some of these come together, you know, let us say, but, you know, to create a mini language model for a very specific task that you can potentially serve and apply on each devices and things like this.
00:48:55.400 --> 00:48:59.900
You know, your iPhone 15 is quite capable of running some of these things as well, right?
00:48:59.900 --> 00:49:01.500
It's got enough computational power.
00:49:01.500 --> 00:49:03.632
Not that you lack any of these things today.
00:49:03.632 --> 00:49:06.199
It's just the use case. Correct.
00:49:06.199 --> 00:49:18.032
And also when a lot of these things come up and you have a large enough data set and you also need to sort of train them, you will need a combination of real data, synthetic data.
00:49:18.032 --> 00:49:20.532
And there are synthetic elements on it suddenly.
00:49:20.532 --> 00:49:26.132
So it's also called as bull shit data and it's because it's like completely synthetic data. Right?
00:49:26.132 --> 00:49:28.965
You're just generating random numbers.
00:49:28.965 --> 00:49:30.565
Correct, for some of these things.
00:49:30.565 --> 00:49:30.833
Right.
00:49:30.833 --> 00:49:38.266
So and so when you reach that scale, when you have a language model where you want it to sort of train and give an example.
00:49:38.266 --> 00:49:38.733
Correct.
00:49:38.733 --> 00:49:48.599
You know, let us say you have in India, you have, backlog of millions of cases people avoid court correct for what, reason.
00:49:48.599 --> 00:49:56.032
Guru you go there until you file when you are 20 and you get out when you are 60 I'm saying this if you are lucky, correct.
00:49:56.032 --> 00:50:10.965
There are cases running for, tens of years with no, you know, it's because for lack of trained cadres lack of in sense that, the number of judges may not be sufficient for the population.
00:50:10.965 --> 00:50:11.865
Correct.
00:50:11.865 --> 00:50:16.365
And, you know, we follow a procedure which is not necessarily Indian.
00:50:16.365 --> 00:50:26.032
It's anglicized because we adopted a process when we were, you know, when we are colonized, trained, we were, you know, for a number for hundreds of years.
00:50:26.032 --> 00:50:26.333
Correct.
00:50:26.333 --> 00:50:42.766
So there's a process that is coming to this, which may not necessarily actually match the Indian ethos as well today, but let us say that you've got judgments or the thinking of 200 different judges over the last 200 years. Okay.
00:50:42.766 --> 00:50:47.800
And you want to resolve this issue of cases, correct?
00:50:47.800 --> 00:50:54.733
we can train the, LLM exactly sort of like analyze cases the way depending on the genres or other things.
00:50:54.733 --> 00:50:55.365
Right?
00:50:55.365 --> 00:50:59.000
I mean, I have to do this so that, you know, the resolution.
00:50:59.000 --> 00:51:20.365
Correct for a case can come in a format, where, you know, you may have a sitting judge or someone else who looks at it and is able to pass, you know, the process, the documentation, you know, if it is streamlined and you know, you can have a resolution in a matter of days versus 50 or 60 years, right?
00:51:20.365 --> 00:51:22.365
So that is my thinking on this.
00:51:22.365 --> 00:51:29.666
But, you know, that is the that is the ingenuity and that is a deep detail that is required for all of these things as well.
00:51:29.666 --> 00:51:33.333
I think, as Quoqo we are in a position to execute some of these things as well.
00:51:33.333 --> 00:51:37.065
But it also requires saying, is there a political will to do this?
00:51:37.065 --> 00:51:51.365
Is there technological skill and or and evaluated by LLM, you know, can I you know, at the end of the day, you need a resolution you know, it is not necessarily a mediation that you require for some of these things as well.
00:51:51.365 --> 00:51:54.000
that's my thought process. Longwinded.
00:51:54.000 --> 00:51:54.733
Yeah.
00:51:54.733 --> 00:52:00.333
I in addition, that there might be a few other, limitations to local models. Is it.
00:52:00.333 --> 00:52:07.900
I spoke to again the output kind of, output window how would you can actually get as an output it already?
00:52:07.900 --> 00:52:13.965
So if you are building smaller models, will be adding some outputs or it may not be able to do that.
00:52:13.965 --> 00:52:18.465
Some of the limitations that might be there for, you know, models that are. Yeah.
00:52:18.465 --> 00:52:25.132
So I'll give an example to take your judgments in a this and it's like 20, 25 pages long.
00:52:25.132 --> 00:52:33.199
You know, the reason being you have to, you know, a judge has to be justice has to be seen fair and not just sort of like be fair. Right.
00:52:33.199 --> 00:52:40.199
So there's and you go through law school, you know, so all of these things come up where you have to be seen to be fair and square, right?
00:52:40.199 --> 00:52:45.199
Like if you walk into if the police intervene in a conflict, they're not going to immediately take sides.
00:52:45.199 --> 00:52:53.932
I mean, notice this, because the police has to sort of, like, appear at least be fair they will never take side immediately.
00:52:53.932 --> 00:52:56.166
And it has a process of its own. Correct.
00:52:56.166 --> 00:52:57.333
So similarly on this.
00:52:57.333 --> 00:53:02.333
So one is when you are doing this data, it has to be clear of biases.
00:53:02.333 --> 00:53:06.065
India has a history where the judges have been biased.
00:53:06.065 --> 00:53:06.900
Okay.
00:53:06.900 --> 00:53:09.699
You know you have judgments either which way and across.
00:53:09.699 --> 00:53:22.132
You know for whatever it is, if you go down LLM through a graphic program and see and today we can map or if we go for which judge you know, what is it that you can expect as a result? Right.
00:53:22.132 --> 00:53:24.400
You know that technology is available.
00:53:24.400 --> 00:53:26.166
it does not require you.
00:53:26.166 --> 00:53:39.099
And that's how some of the funds there and litigation funds that invest, correct, will invest, use this technology to figure out whether, you know, there's stand a chance if they pick up, a piece of, litigation. Correct.
00:53:39.099 --> 00:53:41.865
And on it on its own, others on its course.
00:53:41.865 --> 00:53:54.000
So this technology is available, but the other thing that access to justice has to be cheap, you know, to solve major issues like justice and the backlog of cases.
00:53:54.000 --> 00:53:58.099
And at the end, you want something with an enforcement.
00:53:58.099 --> 00:53:59.000
Right.
00:53:59.000 --> 00:54:18.432
And if you don't want something like drag on for 50 years correct, and you you have to go to the court but even this has been the problem for the last 100 years now, you have really thing Deep technology can solve these things?
00:54:18.432 --> 00:54:22.599
So I think you have a commercial to dispute, you dont have to go with the court.
00:54:22.599 --> 00:54:32.500
And then I want to basically say, okay, you know, I wanted to be paided, you know, for all of these things, you know, that can be done very, very quickly.
00:54:32.500 --> 00:54:33.900
traffic fine.
00:54:33.900 --> 00:54:34.565
Right?
00:54:34.565 --> 00:54:58.266
your traffic fine board is very, very, very, great and, you know, people prefer to pay for to the government rather than more litigate, you know, traffic fine, including, unfair locking up my bike or car or whatever the case, when there no road and forced to park you cannot come and lock it up basic common sense so these are things that are required.
00:54:58.266 --> 00:55:02.099
Fundamentally, outer look I would think that.
00:55:02.099 --> 00:55:08.599
Ai in india is fantastic, Legal is fantastic, is’t it Guru? right?
00:55:08.599 --> 00:55:14.233
You know, you are able to solve so many things, so many inefficiencies.
00:55:14.233 --> 00:55:40.000
And one of the things that also Krithi you know, when you use this here, it applies, you know, AI it doesn't take a site we where on a demo recently not very long ago, there it basically said stuff on the financial side where I could have hesitated it's like, the most straightforward of things, you know, it's emotionless as well.
00:55:40.000 --> 00:55:44.233
But at the end of the day, it's also going to tell you exactly what it thinks.
00:55:44.233 --> 00:55:51.065
And it never sort of, I mean, that's, original thinking when we get there is another thing altogether correct.
00:55:51.065 --> 00:55:59.000
This is this can be used as an exploitative technology, or it can be as a transformational technology, like any piece of technology that you have. Right?
00:55:59.000 --> 00:56:11.065
At this point, the countries are going to exploit through technology, like dark stores dark pricing dark this stuff that you've got competition commission coming in and doing all of these things.
00:56:11.065 --> 00:56:14.432
It's about time to really think about it as transformative technology.
00:56:14.432 --> 00:56:16.132
And we are very specific to India.
00:56:16.132 --> 00:56:23.233
What this, and yes, I'm, I feel that I am entitled to speak on some of these things as a practitioner, as well. Correct.
00:56:23.233 --> 00:56:33.365
and these are things that require you to some sort of I would think that I would use this at every village level, taluk level and district level.
00:56:33.365 --> 00:56:34.166
Correct.
00:56:34.166 --> 00:56:38.465
You know, a lot of the cases can just be solved in a few days basically.
00:56:38.465 --> 00:56:38.833
Yeah.
00:56:38.833 --> 00:56:43.800
It's like, you know, there's an episode, know, my favorite series is Star Trek.
00:56:43.800 --> 00:56:48.032
You know, that one episode, the the original series called wink of an eye.
00:56:48.032 --> 00:56:48.766
Correct.
00:56:48.766 --> 00:56:49.965
You can't see them.
00:56:49.965 --> 00:56:52.666
That accelerated humans.
00:56:52.666 --> 00:56:56.266
And they will do things without you being able to sort of like see them.
00:56:56.266 --> 00:56:57.000
Correct.
00:56:57.000 --> 00:57:08.766
So, I mean, it ends in a bad way in that particular episode, in the sense that, you know, people figure out the origin of our that, you know, people are accelerated and they want to take on whatever those issues.
00:57:09.400 --> 00:57:19.833
But in the wink of an eye, you know, this is startup, so that doesn't have, you know, for solving, say, the case log in India, correct?
00:57:19.833 --> 00:57:28.300
No other country has this dubious distinction of having a billion population and millions of unsolved cases, backlog.
00:57:28.300 --> 00:57:31.500
You know, and you should really up to it, you know.
00:57:31.500 --> 00:57:36.932
So anyway, so I didn't want adding on to that sir I have gone by several, I think.
00:57:36.932 --> 00:57:38.800
What do you think the world was wondering?
00:57:38.800 --> 00:57:39.932
What's your take on that?
00:57:39.932 --> 00:57:47.733
Like, the perplexity and Openai thing Okay, for 200 dollars What's your take on this?
00:57:47.733 --> 00:57:48.632
Oh, yeah.
00:57:48.632 --> 00:57:49.632
What do you think, Guru?
00:57:49.632 --> 00:57:52.500
And, yeah, it's first of all, it's too expensive.
00:57:53.666 --> 00:57:55.599
And the deepseek come in.
00:57:55.599 --> 00:58:00.699
So I see perplexity was created as an alternative to Google. Correct.
00:58:00.699 --> 00:58:09.632
But it still relies on Google search results to be able to give you or maybe Bing and how deepseek to be able to search and give you what it does is we run a query.
00:58:09.632 --> 00:58:11.199
This is my understanding.
00:58:11.199 --> 00:58:19.465
And the guy behind it is from IIT madras And so you should clarify, it's on stage, but it ran 20 different queries and then beat it out.
00:58:19.465 --> 00:58:21.500
You know probably a has and so on.
00:58:21.500 --> 00:58:34.266
System weights and biases, you know do all of these things and then gives you somebody I've used it I found, you know, I have my own biases as well as to what my favorites are correct for all of these things.
00:58:34.266 --> 00:58:40.500
But you know, as a replacement for Google, you need to have a Google kind of search capability.
00:58:40.500 --> 00:58:40.833
Correct?
00:58:40.833 --> 00:58:43.300
Even today, the default is still Google.
00:58:43.300 --> 00:58:55.000
And it may take like, when you it's like, let's see, you have an email which looks like an essay versus receiving an email there that is like one sentence long, my biases is to look at one sentence long, correct.
00:58:55.000 --> 00:58:56.432
Not like an essay.
00:58:56.432 --> 00:58:56.632
Right.
00:58:56.632 --> 00:59:08.266
So this is how it looks like, when it comes to, the O-3 research, it's extremely priced $200 to do Phd level kind of stuff.
00:59:08.266 --> 00:59:10.965
So is Phd going to stop tomorrow.
00:59:10.965 --> 00:59:12.666
I don't know that.
00:59:12.666 --> 00:59:27.599
So, you know, I've tried out some of the reasoning models to, repeatedly question itself to give, you know, I've got some, yeah, I've learned some very esoteric queries like, and, and some very pointed queries on it.
00:59:27.599 --> 00:59:29.532
Queries actually come out right.
00:59:29.532 --> 00:59:39.565
The philosophical and like, what is the nature of, the universe kind of question, you know, and it's not 40 to, if you remember the Hitchhiker's Guide. Correct.
00:59:39.565 --> 00:59:40.932
definitely not even there.
00:59:40.932 --> 00:59:42.233
It had deep thought.
00:59:42.233 --> 00:59:42.532
Correct.
00:59:42.532 --> 00:59:50.300
Really very deepseek like deep thought, you know, it would give I don't think that these are competing, but it's too expensive.
00:59:50.300 --> 00:59:53.400
You know, actually, looking at a place like India for some of these things.
00:59:53.400 --> 00:59:54.699
I don't think so.
00:59:54.699 --> 00:59:58.733
And I don't think the Prime Minister needs to go to sort of like get some on these technologies.
00:59:58.733 --> 00:59:59.465
Correct.
00:59:59.465 --> 01:00:01.733
You look at what's homegrown first. Correct.
01:00:01.733 --> 01:00:03.500
To do all of these things as well.
01:00:03.500 --> 01:00:11.032
and necessarily homegrown technologies have grown with the limitations of the place as well.
01:00:11.032 --> 01:00:11.800
Correct.
01:00:11.800 --> 01:00:21.432
So if you hand India is at a fantastic stage where you can really grow some of these companies and technologies, and it has a wide enough, implication.
01:00:21.432 --> 01:00:24.266
So sort, all of these things as well.
01:00:24.266 --> 01:00:27.565
What does perplexity or O-3 make a difference in India?
01:00:27.565 --> 01:00:29.300
I'm not particularly sure.
01:00:29.300 --> 01:00:30.333
What do you think, Guru?
01:00:30.333 --> 01:00:36.032
Or if not many people have that much money to spend on a monthly basics Okay.
01:00:36.032 --> 01:00:48.333
If you want to do some research kind of thing, you generally spent six months or one year on doing research on a topic fairly and, you know, something like this comes up maybe to, make it faster.
01:00:48.333 --> 01:00:58.565
But more important for researchers is the effort that is involved in performing it so that it, it drives the thought process, it refines the idea.
01:00:58.565 --> 01:01:03.000
So it's about looking at a, two different conflicting comments.
01:01:03.000 --> 01:01:03.565
Yeah.
01:01:03.565 --> 01:01:06.233
It's like also the training given in schools, right.
01:01:06.233 --> 01:01:11.599
It's like, you know, even schools today, in the higher levels.
01:01:11.599 --> 01:01:19.599
Can I still teach like, you know, they don't teach Python, you know, or they may teach Python, but it's more Java based than anything else. Okay.
01:01:19.599 --> 01:01:23.632
In reality, it's just opposite in the real world, isn't it?
01:01:23.632 --> 01:01:28.699
So, I think there is also a fundamental disconnect somewhere that needs to be.
01:01:28.699 --> 01:01:31.965
I mean, there are too many things that we can keep going on and on.
01:01:31.965 --> 01:01:37.032
But does do those things really matter in India at this point in time?
01:01:37.032 --> 01:01:38.099
Probably not.
01:01:38.099 --> 01:01:39.833
You know, if you really ask me.
01:01:39.833 --> 01:01:41.000
Right.
01:01:41.000 --> 01:01:44.865
India some very specific, you know, use cases as well.
01:01:44.865 --> 01:01:59.565
And, you know, the, services mindset certainly does not also help, with product innovation, you know, and product innovation, the copying is treated as product innovation to a large extent.
01:01:59.565 --> 01:01:59.865
Right?
01:01:59.865 --> 01:02:02.032
Actually innovation, I'm sure it does.
01:02:02.032 --> 01:02:03.132
But there is a lot of.
01:02:03.132 --> 01:02:04.065
Yeah, not to me.
01:02:04.065 --> 01:02:05.032
Not too many.
01:02:05.032 --> 01:02:05.965
Correct. Yeah.
01:02:05.965 --> 01:02:14.400
So and things like that, like as an example, you remember, Guru sometime back They built a new rocket motor.
01:02:14.400 --> 01:02:14.766
Correct.
01:02:14.766 --> 01:02:21.065
Which is which actually flew as well, you know, based on, on newer technologies than anything else.
01:02:21.065 --> 01:02:22.666
I forget the name of the company. Correct.
01:02:22.666 --> 01:02:24.199
That was pure innovation.
01:02:24.199 --> 01:02:25.032
Right.
01:02:25.032 --> 01:02:27.266
And that was like, come on, 3d printed model.
01:02:27.266 --> 01:02:27.699
Yeah.
01:02:27.699 --> 01:02:49.733
3D printed model module, which I think isro had not gotten to that stage as it got into, and things like that, some of those fundamental I'm not saying no research based on, suddenly on but an AI you know, that you can hand count you know, a who are the guys or or actually doing deep research in India.
01:02:49.733 --> 01:02:53.300
So, I think that's where things but go on krithi.
01:02:53.300 --> 01:02:59.666
So, one last question before we wrap up What is this kumbhakarna moment?
01:02:59.666 --> 01:03:02.833
like that people are talking about?
01:03:02.833 --> 01:03:05.599
They what are they what exactly that means, right.
01:03:05.599 --> 01:03:08.000
Let me, let me rephrase that question. Okay.
01:03:08.000 --> 01:03:13.365
Now, now, you heard about what Chetan thinks and, what guru thinks and things like that.
01:03:13.365 --> 01:03:21.699
Right now you are exposed to AI right from sort of like youth, right from college onwards.
01:03:21.699 --> 01:03:22.599
Correct? Yeah.
01:03:22.599 --> 01:03:24.833
What do you think really?
01:03:24.833 --> 01:03:26.300
I mean, how do you use AI ?
01:03:26.300 --> 01:03:27.099
Let me ask.
01:03:27.099 --> 01:03:31.266
So basically we has just passed out.
01:03:31.266 --> 01:03:32.465
Well not really passed out.
01:03:32.465 --> 01:03:38.699
But has been assigment I mean where they give you a assigment you just put the questions and we just get the answers.
01:03:38.699 --> 01:03:43.965
We are not like putting a efforts to literally go through the books and get the answers.
01:03:43.965 --> 01:03:44.300
Okay.
01:03:44.300 --> 01:03:48.333
It's better to read that way and also videos.
01:03:48.333 --> 01:03:49.599
And there's a thing.
01:03:49.599 --> 01:03:54.365
AI also wher u put up a pdf to get all the answers that related to the book.
01:03:54.365 --> 01:03:57.632
If you would be able to read the whole 30 pages.
01:03:57.632 --> 01:04:01.199
So it's like time consuming when we listen to it.
01:04:01.199 --> 01:04:01.465
Okay.
01:04:01.465 --> 01:04:06.932
So this is where we are using like so as an example, you know for this right.
01:04:06.932 --> 01:04:13.833
The future generation looks at times it's not necessary for you to learn everything you know.
01:04:13.833 --> 01:04:16.965
And then because each generation builds on top of the other, right.
01:04:16.965 --> 01:04:22.465
If you go back and try and build the satan rocket today, they don't have another tools and dices Correct.
01:04:22.465 --> 01:04:25.000
And you don't know when can you use that thing?
01:04:25.000 --> 01:04:29.365
What do you use Guru that, before calculators came into the thing.
01:04:29.365 --> 01:04:30.500
Don't know.
01:04:30.500 --> 01:04:33.900
And and no, no, they use some scale I forget, you see.
01:04:33.900 --> 01:04:34.132
Yeah.
01:04:34.132 --> 01:04:41.099
I, you know, I have no, no, no, they have this, you know, to calculate log table. Right.
01:04:41.099 --> 01:04:47.865
They won't do the anyway. Sorry So you know, now now look even we don't remember some of these things, okay.
01:04:47.865 --> 01:04:55.965
Even though it may have been like a generation or two away and you are at least five generations of a four generations of it. Correct.
01:04:55.965 --> 01:05:05.699
So things like this, why, why we say this is it's not necessary for you to actually you have to like this that remembers how things were done.
01:05:05.699 --> 01:05:06.632
Correct.
01:05:06.632 --> 01:05:14.300
And then, you know, you have a tool to sort of, look back, correct, and actually learn how it could be done if necessary and things like that.
01:05:14.300 --> 01:05:17.965
It's more of a memory to enhance memory to as well. Correct.
01:05:17.965 --> 01:05:23.266
We lost the art in India to manufacture mud buildings as an example.
01:05:23.266 --> 01:05:23.565
Okay.
01:05:23.565 --> 01:05:25.599
Which existed like 100 years before.
01:05:25.599 --> 01:05:26.333
Correct.
01:05:26.333 --> 01:05:30.766
For that matter, making, or making iron, which does not rust.
01:05:30.766 --> 01:05:38.932
Because if an ally of whatever the case is right, you iron pillar as an example, the ability to create the stone temples in Karnataka.
01:05:38.932 --> 01:05:40.000
you can't.
01:05:40.000 --> 01:05:41.032
You lost on that.
01:05:41.032 --> 01:05:45.365
But you have the ability now to store all of that generation stuff going forward.
01:05:45.365 --> 01:05:45.699
Correct.
01:05:45.699 --> 01:05:48.500
So that means that you don't need to know something known.
01:05:48.500 --> 01:05:49.599
Pythagoras theorem.
01:05:49.599 --> 01:05:50.333
Correct.
01:05:50.333 --> 01:05:55.166
Should only know that, you know, at the end of the day, can tap into something that can utilize that.
01:05:55.166 --> 01:05:57.333
You have the broad aspects of it.
01:05:57.333 --> 01:06:00.000
That's how exactly you should be using this.
01:06:00.000 --> 01:06:00.266
Correct?
01:06:00.266 --> 01:06:04.400
It's not possible to download an entire LLM.
01:06:04.400 --> 01:06:09.300
Basically they will be made once and that would be for example Yeah.
01:06:09.300 --> 01:06:09.900
You video.
01:06:09.900 --> 01:06:13.400
For 15 minutes of and we just put it on it goes on.
01:06:13.400 --> 01:06:22.032
Well it is about borders on it and it's not necessary to go through like 20, 22, 23 years of education, correct?
01:06:22.032 --> 01:06:22.532
Yes.
01:06:22.532 --> 01:06:28.233
You don't I don't think I think that's look, that seriously needs to get crunched at some stage as well.
01:06:28.233 --> 01:06:28.565
Correct.
01:06:28.565 --> 01:06:44.532
To be able to say, you know, you have a lot more, you know, I mean, you may lack the experience, but that doesn't mean that you cannot tap into a higher experience, thing that also the way that, generations are going to utilize this is to build stuff, right?
01:06:44.532 --> 01:06:46.833
Which would never have been attempted before.
01:06:46.833 --> 01:06:47.565
Correct.
01:06:47.565 --> 01:06:49.032
And they know that is correct.
01:06:49.032 --> 01:07:02.900
If you are able to review a document in seconds versus you send it to email somebody, have somebody look into it and go through it and do their commands with their biases and whatever it is.
01:07:02.900 --> 01:07:03.565
Right.
01:07:03.565 --> 01:07:05.465
So it takes a lot of time. Right?
01:07:05.465 --> 01:07:09.432
So you don't need some of those things anymore.
01:07:09.432 --> 01:07:12.699
If you really if you want to skill yourself, correct.
01:07:12.699 --> 01:07:19.699
You shouldn't be scared and machine learning, you know, as an example, because you can train models for exactly what it is, right.
01:07:19.699 --> 01:07:28.233
Using an auto rickshaws, as is, which have existed for 75 years, you know, it will continue to be driven by humans in another ten years.
01:07:28.233 --> 01:07:33.233
You've got the technology today to do autonomous know, like the way Tesla does. Right?
01:07:33.233 --> 01:07:36.166
It's done as a combination of real time.
01:07:36.166 --> 01:07:36.632
Yeah.
01:07:36.632 --> 01:07:44.865
And because the technologies also include because you can do all processing in real time, you don't need massive cloud connection to do all of these things.
01:07:44.865 --> 01:07:45.833
You don't.
01:07:45.833 --> 01:07:48.865
So that these are all the things right, that you can do.
01:07:48.865 --> 01:07:51.833
One click they have this parking thing.
01:07:52.800 --> 01:07:53.166
Yeah.
01:07:53.166 --> 01:07:56.465
And wouldn't you like a driverless auto rickshaw? Yeah.
01:07:56.465 --> 01:07:57.000
What do you think?
01:07:57.000 --> 01:07:58.532
Yeah, I for one.
01:07:58.532 --> 01:07:59.333
Yeah.
01:07:59.333 --> 01:08:06.199
Now, can I account for all the auto rickshaw drivers who are gang up and say this is the wrong thing?
01:08:06.199 --> 01:08:09.565
I need a job for this because have not prep for this.
01:08:09.565 --> 01:08:10.532
Probably not.
01:08:10.532 --> 01:08:10.932
Yes.
01:08:10.932 --> 01:08:14.065
You know, it's like the bank employees who resisted computerization.
01:08:14.065 --> 01:08:16.265
And today it turns out that the bank is a computer.
01:08:16.265 --> 01:08:16.632
Correct.
01:08:16.632 --> 01:08:23.166
So, you know, it is that, you know, so I'm saying bank is in phone now the bank is a phone.
01:08:23.166 --> 01:08:39.233
You know, at the end of the day, you know, for all of these things, you know, so, and I mean, it's more of a philosophical discussion of human creativity and all of these things, but also human creativity is a subject of inherent bias.
01:08:39.233 --> 01:08:43.666
Is is it also okay, guru Can you go on the environment?
01:08:43.666 --> 01:08:49.399
But I want to say I have hardly know the answer to it.
01:08:49.399 --> 01:08:50.765
Exactly. Okay.
01:08:50.765 --> 01:09:06.600
So subject matter of your learning, learning and so on, you know, because your are also biological by nature that means that, you know, you will also have a set survival senses kicking in it.
01:09:06.600 --> 01:09:09.033
It's not necessarily always going to be.
01:09:09.033 --> 01:09:15.966
So there's a system architecture, you know, at the latest stage you know, how can evole all of these.
01:09:15.966 --> 01:09:20.632
And then there's some other things you that's subject matter for another discussion.
01:09:20.632 --> 01:09:26.166
And I think we've and some thoughts on design about how things evolve.
01:09:27.300 --> 01:09:46.533
back to kumbhakarna moment back to kumbhakarna moment its like indian goverment wake up and thinking, oh, I need all this GPUs I have a lot of intellectual matter, you know, of all of them, to be able to sort of, it's almost all the some like outdated and so on.
01:09:46.533 --> 01:09:50.733
And so anyway, you don't have an intellectual.
01:09:50.733 --> 01:09:54.765
man or women power because of lake of skills Yeah.
01:09:54.765 --> 01:10:10.166
But that's the kumbhakarna moment where they wakeup and figerout And the rest of the world is sort of like overtaken, and and I'm still in the services economy of 50 70 years back, and I have not done anything to encourage innovation. Correct.
01:10:10.166 --> 01:10:10.466
Yeah.
01:10:10.466 --> 01:10:16.199
And that's a subject matter of so many other things, which has been so obvious.
01:10:16.199 --> 01:10:18.966
But the kumbhakarna moment is on and on.
01:10:18.966 --> 01:10:27.199
And now, you know, I think companies say we say clear LLM for us you know, that is not followed by some deep thought.
01:10:27.199 --> 01:10:35.966
That's what you want to do with LLM correct, you know, and the use, you know, when you building something you need to have like an end goal in mind.
01:10:35.966 --> 01:10:47.065
We've seen that and we call it the kumbhakarna movement because, you know in the Ramayana you have kumbhakarna who could sleep for how many years at the time 12 years at the time or something, right.
01:10:47.065 --> 01:10:51.500
You know, and you would, wake up for 12 years and sleep for 12 years.
01:10:51.500 --> 01:10:51.865
Okay.
01:10:51.865 --> 01:10:55.466
And during that time, so many events would have changed, right?
01:10:55.466 --> 01:10:59.365
And that, he was always lagging behind by 12 years.
01:10:59.365 --> 01:10:59.966
Right.
01:10:59.966 --> 01:11:03.033
And looks like the current state of affairs.
01:11:03.033 --> 01:11:06.932
And it's up to the it's always been up to the private sector to take things forward.
01:11:06.932 --> 01:11:12.033
But you know, this, it's also a good thing that kumbhakarna has been woken up suddenly.
01:11:12.033 --> 01:11:27.699
The out of all of these things that are handing out freebies and other things, suddenly people have figured out that now going forward is going to be can become an existential issue of whether people are going to have the necessary skills or the jobs to be able to do all of these things.
01:11:27.699 --> 01:11:33.365
But from a technology perspective, at least woken up and said, okay, we want to do all of these things as well. Correct?
01:11:33.365 --> 01:11:45.100
Saying so, I think it's a good step that they've taken and hopefully, know, they will sort of like look into, growing some of these things should have been done a long time back.
01:11:45.100 --> 01:11:45.966
Correct.
01:11:45.966 --> 01:11:53.332
And our own experience when we tried to explain our company and AI to a lot of people as we were growing, I'm still not correct.
01:11:53.332 --> 01:12:00.932
People don't understand or are at a skill level where they can really see how things can move.
01:12:00.932 --> 01:12:01.765
Correct.
01:12:01.765 --> 01:12:19.533
To really spend time, you know, with them and then suddenly becomes like, oh, I'm going to lose my job and I want to block this, correct But I want my son or daughter to go abroad and work in Microsoft and make a lot of money so that that that conundrum is not not being solved, as that does not change.
01:12:19.533 --> 01:12:21.666
All of that is going to stop.
01:12:21.666 --> 01:12:30.399
I think it's about time south, like focus on building apps, on using AI and locally grown AI to make sense.
01:12:30.399 --> 01:12:37.399
Okay, you know, to do this now, do you need a bank to do you think that AI today. Correct.
01:12:37.399 --> 01:12:42.899
You know, can you know, how long do you think it's going to take before my phone can talk to his phone?
01:12:42.899 --> 01:12:52.065
And, you know, basically sort of exchange money, data or whatever it is that is necessary, you don't need a central bank for any of these things.
01:12:52.065 --> 01:12:58.432
You don't need, you know, at some point in time, there's a need for growing decentralization.
01:12:58.432 --> 01:13:06.565
You know, luckily for us, you know, central Bank introduced the, crypto, currency equivalent of central bank.
01:13:06.565 --> 01:13:08.432
But where are other use cases.
01:13:08.432 --> 01:13:10.033
Mean this is what we're saying. Correct.
01:13:10.033 --> 01:13:21.100
We may have you may introduce some forward thinking this, but it's not being taken by people because there's a lack of exchange and yeah, and skill levels.
01:13:21.100 --> 01:13:25.899
Now this can seems controversial, but this is where it is and it's not going to change overnight.
01:13:25.899 --> 01:13:33.399
It may change with some generational thinking and saying, okay, we are doing this to solve real issues, okay?
01:13:33.399 --> 01:13:42.966
But is it sufficient for me to sit today and say, I want to create a New LLM for India and then do what we don't have a end goal in mind, correct?
01:13:42.966 --> 01:13:46.199
I think this is where the gap is correct.
01:13:46.199 --> 01:14:04.233
And if you have civil engineers who want to sort of like become testers, and if you have, structurally engineers, you know, wanting to do software, then you know that there is a real services element, which is really taken the take on the economy in a very, very wrong direction.
01:14:04.233 --> 01:14:04.600
Correct.
01:14:04.600 --> 01:14:08.699
So, these are all things that we can talk about.
01:14:08.699 --> 01:14:11.199
People know but can’t solve?
01:14:11.199 --> 01:14:14.500
But with AI you stand to have a real chance, isn't it?
01:14:14.500 --> 01:14:17.632
You know, that is the way That's the way to adopt things.
01:14:17.632 --> 01:14:20.865
So, so when.
01:14:20.865 --> 01:14:21.800
Thanks.
01:14:21.800 --> 01:14:34.699
As I said, once again, this has been like a monologue because there are a lot of things that we wanted to cover and a lot of the products that we mention, you know, where I need stuff, I get some more clarity and things like that as well.
01:14:34.699 --> 01:14:38.466
thanks, krithi for moderating this long session.
01:14:38.466 --> 01:14:39.899
Also, Guru, okay.
01:14:39.899 --> 01:14:44.233
For, for this and for the team for getting this together, thanks for breaking it down.
01:14:44.233 --> 01:14:48.932
So both of you Chetan and Guru It was really insightful looking through it.
01:14:48.932 --> 01:14:49.533
Yeah.
01:14:49.533 --> 01:14:52.600
So the LLM is never going to call you sir okay.
01:14:52.600 --> 01:14:53.300
Right. Okay.
01:14:54.733 --> 01:15:00.466
So you can train it to call you that, but that's an example of, how it can be done.
01:15:00.466 --> 01:15:01.033
Right. Thank you.
01:15:01.033 --> 01:15:01.666
Thanks so much.
