EB-8: Kissan - Transcript
Transcript: Kissan - Part 1
[00:02.984] π©βπ€ Ate-A-Pi: Hey Pratik, what is Kissan?
[00:06.911] π Pratik Desai: So Kissan is a agri -vertical generative AI startup and we are trying to build applications that helps farmers because there is a lot of problems in terms of knowledge decimation from the knowledge that we produce in universities, agriculture businesses to the farmer. And that's one of the main reasons the agriculture industry is very far behind in terms of implementing the base practices and all. So...
what we are trying to do, we are creating multilingual voice base, co -pilots for farmers, which basically they can just have as a conversational tool. They can click a button, start talking to the knowledge base that we have been curating and get the base information that's available for them. And in this, we are, while we are trying to help the farmer, our approach is to actually enable agribusinesses, help them to create this co -pilot. So it's like a,
like enterprise applications basically for agribusinesses where they can actually host their knowledge and through their existing customer they can deploy their knowledge, the products, conventional commerce and all kind of applications.
[01:20.296] π©βπ€ Ate-A-Pi: So it's like a chat GPT for farmers. Would that be a good accurate description?
[01:26.879] π Pratik Desai: It's a, okay, so it's a enterprise, ChatGPT, like the assistant API for farmers.
[01:35.464] π©βπ€ Ate-A-Pi: Right. OK. So here you have, and I guess your primary customer base right now is in India.
[01:44.287] π Pratik Desai: Correct.
[01:45.416] π©βπ€ Ate-A-Pi: OK, so you have a farmer in India can take their Android phone. They can go to the app store. They can download the app. Once they download the app, it is a voice interface or text interface. So they can either type in queries or they can ask queries. They can do that in their local language. Is that correct?
So it can be in Hindi or Tamil or any of like, I guess India has like a lot of languages, like a couple of hundred or whatever. And you guys support what like 20 or 30 right now.
[02:21.759] π Pratik Desai: lonely.
[02:26.943] π Pratik Desai: Yeah, we started with 10, expanding to 15, and we are also adding like a lot of support for the South Asian, Southeast Asia, like Vietnamese and Thai and Malay, Indonesian, Bangla, Urdu, like try to cover the most of the developing nations where we can support language.
[02:46.344] π©βπ€ Ate-A-Pi: So, and then the farmer can start typing a query or start asking a query. And so they go ahead and what is like a typical first query that a farmer would make?
[03:03.519] π Pratik Desai: Yeah, so that's a very interesting question. So, surprisingly, in early days, the farmer will just say hello because they are sometimes. Yeah. Kind of, kind of because it's very surprising, like AI is still like in a stage where, you know, if you get greeted, that is fine. But it's like they don't know what to ask. It's like in early days, it was like a very new tool. But then we have seen where once they realize what it is and how.
[03:11.656] π©βπ€ Ate-A-Pi: Just hello, just is anyone hearing me? Is anyone listening?
[03:32.575] π Pratik Desai: like available it can be or like the vast knowledge it has. We have seen farmer going like literally asking very deep questions and like, you know, like the workflow become very nice for them because they understand now. And some of them there we have seen even the questions where just they started recording but then don't say anything at all because they are just overwhelmed to say anything. So that was used to be the case. But then in terms of questions, mostly they are asking about like a typical
Daily routine questions like hey, I have a problem. I have this space. What should I use or what is the best time to grow this thing? where would I get the best prices for my crop and Like there are many typical queries that we were not supporting like asking real -time weather data and things like that So we are also starting to add all those things But it's like for us it has been learning process to we were working in agri tech. So we knew like what what happens?
Like I come from a greater the background so I know like what's the best biggest problems are but then when actually something goes inside their hand like the intuitively what they ask before they actually get familiar with it that is something was our journey to actually get those thing and that has been helping out a lot like just putting the Application in while let them talk understand what they're trying to say and we are focusing on one demographic so far. It's us like, you know
uh, we are not trying to capture all, all type of demographic, what type of question they will ask and try to model like so topic modeling and all those becoming kind of now intuitive for us. So we understand that this is kind of very much workflow and that's helping us to build the application for businesses.
[05:17.672] π©βπ€ Ate-A-Pi: So roughly, what's the ballpark of the number of users that you have right now? Like 10 ,000, 20 ,000 more?
[05:27.903] π Pratik Desai: We have more than 110k farmers who have been used as a unique user. We have 10 to 15 ,000 monthly active users. And the applications, right. So the application that we built to help farmers, that's basically organically grown. We haven't done anything in marketing or like it just kind of shared somewhere on the social media. It picked up.
But then we also started kind of realizing that we cannot keep expanding the user base. Like it was actually going very rapidly in 30 days. So we have kind of stopped working a lot on the farmer facing website right now. And the reason being like, we want to figure out the revenue model for us, right? It's a startup. We cannot just keep running and go to the level that India has so many farmers. Like there are like 150 million people actually associated with agriculture or not maybe more like directly indirectly.
So my ultimate motive was let's first figure out a revenue model. And then we go back and then start capturing and adding more features because as soon as I add more features, there will be like another like a spike on the users. So I'm like trying to kind of mellow down the growth on the consumer side before we actually start generating revenue.
[06:47.4] π©βπ€ Ate-A-Pi: So you had strong growth in number of users. You got up to 100 ,000, 150 ,000 individual unique users, and then about 10 ,000 to 15 ,000 monthlies. But then you also have the cost base, because the farmers are not paying anything for it. It's a free service for them. So you don't have, as yet, a defined revenue model.
Is it costing a lot to serve the queries? Is the LLM and usage significant?
[07:26.623] π Pratik Desai: Actually, see, I think it's not if you compare to other consumer applications that going on, like we are actually trying to minimize the cost as much as possible. But I think the problem is on the other side, like I haven't bootstrapping this thing whole myself. So I also have to kind of see the growth rate and figure out at what time I will run out of mechanics that.
Microsoft has been very generous to give us basically so the day we log within a week Microsoft found out they kind of give us like a level four sponsorship so I never had the problem of running out of the credit in early days but then I realized if we keep growing then I will never have a chance to actually build a business out of it because we just keep running and from our previous applications that we built for agri tech
in India before Kissan our experience was that farmer cannot be monetized in any developing nations like even in outside in developing nations it's mostly the businesses that you have to work with who already sell like physical thing or physical services with them you have to monetize them not the farmer themselves and that was already our like first rule that you know we are never going to monetize that like it's not possible right now because of the income they have so
So we can keep running it, but I know that with the six or eight months, I'll run out of that credit too, like very quickly. And that's a realization that like, okay, let's start building something for enterprises because we had like a lot of inbound coming out from the businesses. Just like, Hey, we want this thing in our app on website. Can you actually build this thing or can I, can you have a platform? So we now build it. We have revenue coming in. So now I'm confident that now I can actually start going back because the
Ultimate goal is still to educate farmer now either we do it directly from the public knowledge and the things that we curate or we have the businesses who already do and work with them like businesses have the same problem like they produce like So many different type of variety of pesticide the new versions are coming in or the fertilizer or the seeds or like you know equipment but like the farmer never reaches to the farmer like so Helping even them to get to deliver the knowledge. That is something you know our ultimate goal is
[09:36.2] π©βπ€ Ate-A-Pi: So basically what's happened is you have the farmer income levels in most emerging markets are too low to support kind of like subscription software services. So you have to find a third party monetization strategy. And one of the third party monetization strategies is to work with enterprises.
[09:52.127] π Pratik Desai: Correct.
[10:04.584] π©βπ€ Ate-A-Pi: who are already selling goods to farmers to provide, like on their website, they can have this kind of, you know, agricultural advisor that they, you know, so it's kind of a white label. You kind of white label the product and then it's under their branding or something like that.
[10:13.087] π Pratik Desai: I do as I remove, yes.
[10:20.063] π Pratik Desai: Correct. Yeah. So right now for all of them, we are working, we are like collaborating with like powered by Kissan AI. So I don't think they have a problem saying that. But yes, so we are enabling like a few use cases that we learned throughout like, you know, running this for first early months, like the farmers think first question was advisory. Second thing they wanted to actually complete the purchase hook, like, hey, I have a question regarding my crop and something going on inside and you are advising me to do this, this, this, but like, what is the action, right?
[10:26.216] π©βπ€ Ate-A-Pi: I see. I see.
[10:48.127] π Pratik Desai: I want to buy something now, but I don't have anything to buy inside your website. So that's what we now enabling for the businesses. They're like, Hey, you brought advisory and you can also now suggest the product. So it's like conversational commerce enabling inside their app. So now they can actually like reduce two costs. First is running the customer call center for just the advisories, like in every language 24 seven. And second is increasing the direct sale because the biggest problem is
middle man, the middle layers are actually the very big problem for the profit margins for the businesses.
[11:18.824] π©βπ€ Ate-A-Pi: So does that mean at some point in that chat conversation with the farmer, and this is on a, let's say, some farm fertilizer branded kind of experience, and at some point the farmer says, OK, I trust your advice on the fertilizer. I want to buy it. And then you put in a link to the purchase. Is that what happens?
[11:31.679] π Pratik Desai: Okay.
[11:44.287] π Pratik Desai: Correct. Yes. Yes. Yeah. So whatever we're describing, say like, Hey, for your particular problem, this and this can be the solution. This one work base, there's a link here to buy it. And we also identify using like the LLM and, and also like some of the workflow that has been defined by customer that like, if this is going outside of the realm of advisory, we also have like call to expert kind of links. You can provide that.
So if they really don't want to buy from online, they can still call and complete the call. So yes, this is what we are basically enabling for them.
[12:19.336] π©βπ€ Ate-A-Pi: That's amazing. And just in terms of usage, do you see, what is the percentage you see of voice versus text?
[12:30.751] π Pratik Desai: 40 60 at least from the India context. Yeah, India context. We're seeing a lot of voice actually the voice has been so when we started we were native voice But then we started like getting requests from like urban Gardeners or like someone who just trying things out on their terraces like yes We are targeting farmers and trying to solve even though they had they have been doing like practices for so long sometime they have new questions about like because weather and climate is changing very rapidly, so I
[12:32.904] π©βπ€ Ate-A-Pi: 40 % voice.
[12:59.903] π Pratik Desai: But then there are like a whole bunch of people who are doing like in urban areas, doing gardening or like a growing and spending time on terrace. And they don't know mostly anything about any plants, right? They just like a bring something from nursery, put it down. It's like, hey, let's see if it's going to grow some flowers. And it's not going to, because they are not like taking care of it very well. So they are also one of the like a kind of rapid increasing user that we saw.
And for them, they wanted to have a text. So like the graphic was like, we want text. I said, okay, we're going to enable text. But earlier we just wanted to have like a native voice to voice. And we still see like a lot of people using voice, like not majority, but still like 40%, which is like a good amount.
[13:43.432] π©βπ€ Ate-A-Pi: So when you have the farmers and they are using the application, 40 % is by voice. At first, you started off, you wanted to be majority. You wanted to be native voice. But then you had the need for text from urban customers. So now you have both the text and voice options.
Do you think, do you also have like an image option? Are people able to take photos and send it?
[14:16.863] π Pratik Desai: So I think throughout our journey we actually if you I don't know if you're aware like we also started training the Agri Vertical LLMs like because we wanted to reduce cost significantly and we wanted to need to embed the knowledge which is like a challenge but like you know we can go and discuss later but the goal was to take a like a small model we had like a huge amount of data.
So find the great instruction, create instruction, tune model and try to serve users through that. So we can reduce down the cost of LLM and also keep the, uh, our like a knowledge inside in house. So we started training model called Denu, uh, like the symbol I'm wearing. Uh, Dino is like for Camdenu like it's a cow, like a holy cow kind of symbol. It's like there's a little big myth around it, but for, for that.
Next level, we recently also trained the vision model with the 100 ,000 disease images. So the farmers can take a picture and start talking about what's happening. Like there are applications who can detect the disease, but you cannot have conversation. You say like, yes or no, if this, this is exist or not. We wanted to enable the conversation so they can say like, Hey, I have a wide -fledged frustration. What should I do? Something like that. And then we can actually, they can go more. And so we trained it and this is something we are going to enable soon.
for like everyone that they can actually take a picture right now that particular model and because the open source availability of the models for the vision LLM were not great when we trained it now I think with the llama three and other models we are going to have multilingual multi model thing so the multi the vision model that we created for disease was just like English now we are working where we want to have
multilingual, multimodal disease detection. So natively, you can take a picture and start talking in your language. And that's going to be something I see like very useful for.
[16:14.184] π©βπ€ Ate-A-Pi: So you have this new model, Danu, which is a fine tune. I noticed it was a fine tune of QN. Is it?
[16:21.407] π Pratik Desai: The vision model is 5200 QN.
[16:24.008] π©βπ€ Ate-A-Pi: QN, right, which is actually a Chinese model from Alibaba. And why did you choose QN at that time? Because I think the lava models were also available around the same time.
[16:37.951] π Pratik Desai: They are not true. So I think there was a licensing thing, which I saw early days when we selected it. There was like a CC by NC like that, which I don't like the model at all. Those kinds of licenses. Yes, they are. Yes, I want to try to get a release model, but literally it's not useful for open source. Like when you have a CC by NC, it's just like, it's just like a marketing. You say, I see when you're CC by NC, like for enterprise customer, like, Hey, we have released a model. It has good, you well, but you can't use it.
So I really don't like that approach so QN was QN had like a 700 or 600 million user until you go out and reach out to them So like hey, that's good enough right you need to begin with and From the beginning like you know I'm just trying to build something where we don't go and run out running to like this trouble going forward because anyway we're trying to build something like a have a base solid base and build foundation of like we are the agriculture is not running away anyway, so I don't want to get it create like a
quick hype thing and then fall into trouble later on. So selected Qn and I think Qn's like there is a forgot like I'll go and there is something in the Qn and the lava approach in terms of like the clip that the Qn's kind of coming out better. I read like with some analysis somewhere and I'll find.
[18:02.952] π©βπ€ Ate-A-Pi: So clip is the image detection decoder, right? Yeah.
[18:07.903] π Pratik Desai: Correct. Yeah. So there is some encoder level detail. So like, so the team member who has been like working with it, like, you know, there's like, we made a like a choice, not like just on the license, but also some technical aspect. And, but then later on, I saw the lava versions coming out, which were not controlled by CCYNC, I think from based on this trial. So,
We'll see that there's a friend who's working on taking gamma the Google model and they added Siglip on top of it to enable gamma to have a vision for compare capabilities. So and gamma has a like a very huge vocabulary to support Indic languages. So the work is really good. So people are already trying out like many Indian languages natively inside gamma. So my goal would be.
take a gamma multilingual and use the modified version for vision and then add like the add the vision capability on top of it. So it's like so many parallel things we're doing right now that it becomes.
[19:18.312] π©βπ€ Ate-A-Pi: Yeah. So let's talk a little bit about the Danu construction because Danu, you guys ended up, you're fine tuning it. So for three major crops, what are the three major crops that you fine tuned it for?
[19:30.143] π Pratik Desai: So for the vision, we had wheat, maize and rice.
[19:35.144] π©βπ€ Ate-A-Pi: wheat, maize, and rice, and for 10 diseases to start off with, for 10 diseases. And then once you, so it's basically image detection, right? It's basically classification, basically what is the disease type, what is the resource type, and what is the disease type? Would you say that? Is that the result that you get out?
[19:40.191] π Pratik Desai: Yeah, the first one, yeah.
[20:03.231] π Pratik Desai: So, so what we're trying to do is like we are doing a lot of synthetic data over here. So in a sense, we first pass this like, so we already have the like when we created the data set, that's the where the key is like the data set doesn't only have the name of the disease or like some minor condition around it. We are actually creating a lot of synthetic data around the situation and the disease. And when we train the build a whole, uh,
a dataset, we are adding a lot of information beside the disease, right? Like, what are the different symptoms for the disease? Like how to take care of it, how to prevent it? What are the solutions available for this location, this location? So when we create a dataset for every image, the description will be not just the name of the disease, but a lot of other information that actually combined go inside the training. So when someone actually
[20:57.064] π©βπ€ Ate-A-Pi: So it's almost like the input data you prepare is a Wikipedia page for each disease with kind of the image being like the image and then the Wikipedia kind of article.
[21:05.183] π Pratik Desai: Kind of yeah, yeah kind of kind of yes like but we we have like a five criteria that we define and we stick to that and try to kind of add from our knowledge base try to add some information around it So but yeah, this is kind of simple. It's a instruction like, you know when you tune with the instruction you're like a Question and like a answer kind of deliver for it, right? So in the training what we do we take an image and we create like instructions set with the image for question answer question answer around that particular image
So now we have kind of trained the model to answer questions, not just the detected disease.
[21:40.84] π©βπ€ Ate-A-Pi: So when you prepare that data set, is that a English data set, or is that also multilingual? Is that you have all the languages already there in the input data that you throw in?
[21:53.407] π Pratik Desai: For the vision model, as I said, like we just use English one, basically, because for us, it was like a, like first kind of experiment itself that like, can we actually have some additional knowledge embedded inside, like in the model, right? Because QN is like far, far behind from GPTB or goal was say, GPTB is expensive. So if we, like what is our objective? Like GPTB is expensive. We cannot deploy in the world.
And for the farmers who is not going to pay, we have to have like something very cheap. The QN was not there at all. So, and the good thing is our application scope is narrow. Like, Hey, we are just doing for disease. We don't care if it identify bus or the road at the sign or do OCR. We want to kind of have a disease and a condition around this. So, uh, so that's why, like, you know, uh, we curated this data set. Goal was to do it for the English and, uh, then.
As I said, like for the next version, we are going to go for like that was just a 10 disease. We have now 55 disease data set already created. So it's going to cover like a 10 to 12 different disease, sorry, crops. And then the goal would be to say like, hey, first again training, see the improvement, then add the multilingual data. So it's like a granular kind of you go up. But the new version that we released just as a like on llama to with the open happy we collaborated with server that.
We were like bilingual. We had Hindi and English basically as part of the dataset that we curated. And so our like initial progress is going toward creating a multimodal multilingual where we embed our agri -knowledge, we embed the disease knowledge and create like something smaller and efficient for very like vertical specific thing, which may not score well in everything else, but...
[23:23.688] π©βπ€ Ate-A-Pi: I see.
[23:46.686] π Pratik Desai: should be good enough for one specific vertical.
[23:50.408] π©βπ€ Ate-A-Pi: So, um.
What is like, what's your most frequent use case right now that you have with the farmers? You have like 10 ,000, 15 ,000 like regular daily actives. What is the most frequent use case that they use it for?
[24:07.391] π Pratik Desai: So most frequently would be like related to market access or like finding out the queries are related to where to sell, how to sell, where would I get the most price. So it's related to economy, basically the first one. The second one are related to paste kind of thing. Like, you know, they have some kind of problem inside their farm. What should I use it? How to prevent it? What are the initial way of doing it? Are there like some?
Organic way and those kind of things are there like so the mostly first because you know There can be like a ten different type of query regarding market access, but the market access is the first theme second thing around like the health of the crop and Then it will be like just like a basic question about exploring new things about hey What are the new techniques to do this thing? Can I grow avocado in particular? Yeah, like something like that, right? So it's like
Those kind of questions basically come later on, not a lot.
[25:10.376] π©βπ€ Ate-A-Pi: It's interesting because I heard a story on YouTube of a farmer in India who started growing strawberries after learning it from YouTube. And everyone in his area thought he was nuts. And he just was like Googling and YouTubing. And he found out that for some certain months of the year, the climate and the soil was OK. You just need a little bit of fertilizer and.
[25:20.991] π Pratik Desai: Yeah.
[25:40.008] π©βπ€ Ate-A-Pi: He went ahead and he spent his money and he bought these things and he started planting strawberries in India. And of course it's like a hundred times more lucrative than the existing crop. So.
[25:51.391] π Pratik Desai: That's what we are trying to actually enable if you look at it So this knowledge like the person who went he went to Google YouTube He could have understanding of how to use YouTube and Google in perfect way read languages like read description and all like find out the exact way to implement it now that's probably not possible for many of those farmers who are problem literacy languages like all the different accessibility problem, right and
of making this thing, this knowledge available through AI, we can basically have them kind of also learn because now you can ask in Hindi or Tamil how to grow strawberry in this particular area and region and then we will try to get from what we will find out from our knowledge is what varieties are probably possible for that. And we do have like a very similar story that we had a farmer from like Chattisgarh district. This is not very one of the richest one. But and the farmer was like,
Trying to figure out something he was growing this geranium kind of plant which used to create oils it's like a aromatic oil that goes in the processing and then the oil companies just basically buy Like you know get the oil out of it and then then there is a lot of agri -waste out of once the oil is extracted So this farmer has been trying to look out for something he can do with the agri -waste because sometimes he may have to even pay to get rid of the agri -waste so what he can do so
He comes to the app, start asking questions. He had some kind of knowledge that some people in some countries are already doing it or some research has been going on. So he started using it, start asking questions. Like, what should I do? Then he found out that there is a way of making air fresheners out of this agri -waste. Then what are the chemical composition he need to find out? So he do like ask questions, find out, do experiment it, create some air freshener.
And he actually successfully makes it. He tries out with his friends. They love it. He now has a startup registered where he's actually producing. So this has been involved. We've been out for a year and we already have a story like this who like people have become agripreneur like some farmer who can just figure out how to make air freshener from agriwest. And now he's like an entrepreneur.
[28:08.04] π©βπ€ Ate-A-Pi: So amazing. So he starts off growing some geranium derivative, which is used for some essential oil production by some companies. He ends up with a lot of agricultural waste. He normally has to pay someone to dispose of it. He goes to your AI bot, starts talking to it, and he figures out that he can make some...
you know, air freshener from it, and then he tries it out. He makes the air freshener, and then now he has a product rather than just a commodity. That's amazing. So, and this is like your classic, like he was not likely to maybe use a YouTube or whatever because it was just simply too complex.
[28:42.207] π Pratik Desai: Yeah, it's selling on Flipkart and Amazon.
[28:59.871] π Pratik Desai: Yeah, so.
In his particular story, he could have used YouTube. But then the problem is the information overload. There are like so many people talking about it. I don't think there was like very specific information about geranium. But even if there are any, he said, like, according to him, it was so much difficult to find out which one worked best. So he just kept asking AI the question and try to identify, like, what is the base compared between these and these? What is available? What chemical I can buy those kind of thing. Right. So that is something he will have to
read like a five different, 10 different pages and identify the perfect solution to make their question. So it just making this like, it's like using the Google reading 10 pages and using perplexity to find out like, hey, this is a summary, let's go on with it, something like that. And when you add language to that, it's like.
[29:37.256] π©βπ€ Ate-A-Pi: Right.
[29:46.248] π©βπ€ Ate-A-Pi: Right, so it's just too much information for him to effectively identify what is the correct solution. And so kind of it's a data curation, data trustworthiness problem too. There's just a lot of trash data on Google, on YouTube as well.
[29:56.031] π Pratik Desai: Yeah.
[30:05.919] π Pratik Desai: Correct. And also language like this, folks, maybe they can read like English, but like when when you are actually exploring the websites, there will be a lot of technical jargon because that is something I have seen with this whole industry is the people on the top do like very maybe cutting edge research and creating some climate resilient fertilizers or seeds and things like that. But then the farmers are like very, very far away from the reality for them because, you know, they really don't work with the farmers like, you know, for the tech.
It's not that much around like, you know, you can the most of the time folk can read the distance between the research and the people who implement is not that much intact. But over here is so far that if you go to any conferences or like meet any people are working on the research side or because this is kind of research issue in a way like where you can extract some oil and what kind of chemical you add. But when they publish these things like universities and all they are publishing in the jargon that far more never understand.
And they really don't care. So I have seen people just talking and blabbering about climate like a resilient agriculture, researcher doing that work. And when I go and talk to a farmer in the real world, like they literally don't care. Like they don't have any information about what is actually the changing in the climate. But then we have like so much money and the knowledge being produced and money being spent to kind of change the climate. But you are not taking that knowledge to the farmer in the simple language.
like you know in a way they can understand and implement so that is the gap which i see same thing like the this farmer probably uh had a knowledge he can read english but then like you know comprehend like some research thing material coming out and to compare two different chemicals one say hydrogen peroxide and one say something else now how is going to kind of compare the difference between that like what is the base waste to you what is not toxic where should i buy it
Those kind of things can become like easy where he can keep going like, is it toxic? Is it not like, you know, you can actually go back and forth in AI and kind of communicate that this is not possible on the Google site. Right. So this is something I think is like a very big value addition. And now if he can do that by voice in his own language, then suddenly so open up, like just they can just talk for a day. They want basically, you know, kind of kind of scenarios.
[32:22.408] π©βπ€ Ate-A-Pi: Interesting.
[32:27.592] π©βπ€ Ate-A-Pi: Do you think you guys don't have real time information yet like because you said they were asking for weather but
[32:36.287] π Pratik Desai: Yeah. So, so whether on the market. So these are the two things that we have kind of solved in a way we are just waiting for the feature to be released. Like just tasting it out a little bit where we are ingesting a lot of market data and the data using text to SQL to even animal analytics on top of the data. So they can just ask like, what is the base time to
sell this thing, which, which market give me the base price. So that is feature is there. We are waiting for it to enable just like not going through some tasting phase, but in terms of real time, textual and contextual information that we are not adding it like, you know, some using like a sort, which some of the website uses this day, basically, you know, just extract this top Google result and summarize it. So we are not trying to use that one.
because for us, there's the important thing about the data curation part. So we don't want to kind of create summary of a jump. Like if we think the first 10, 15 results on the Google ad jump, then the summary of that will be also a jump. So what we're trying to do is like the industry, the work, like any vertical, very specific thing, it's not, it doesn't change that much, right? The farmer, when they're asking for base practices equation, they're not looking for the news, right? So.
For that purpose, whenever the data go inside, we make sure that we have done like some annotation work, cleaning of the data. So, and then like we are running rag on our model to basically give them the summarized answer.
[34:10.024] π©βπ€ Ate-A-Pi: Right, so.
[34:16.968] π©βπ€ Ate-A-Pi: There is a study in 1997 to 2007 -ish, this guy's in Kerala. There's a guy, Robert Jensen, in Kerala. He did a study of the introduction of mobile phones in Kerala. And then he measured like fishing prices, fish prices. Because in Kerala, what would happen is the fishermen would go out and then come back and they'd sell to whoever was in the area that they'd sell.
And sometimes they'd come back late and they would, you know, the guy, the, the, the dealer who was there would have already picked up whatever you want to pick up and you're left. And so then they would just end up dumping, dumping the fish. So, um, you know, and then that would be a loss for the day. But at other times they could, uh, once the mobile phones were introduced, they could text to see which other dealers had not yet, you know, fulfill their, uh, things for the day, their catch for the day. And then they could go and arrange a meet.
And so the amount of wastage really dropped dramatically in Kerala with the introduction of mobile phones. And it was very, very quick within a few months, like everyone, you know, everyone acquired a phone and everyone was basically using it to, you know, because one last day is a big deal, right? So I wonder if like you've seen this kind of like dramatic kind of pickup in any segment where all of a sudden you see...
this kind of emergent like, whoa, you know, like word of mouth growth and things happen.
[35:51.967] π Pratik Desai: So I have like, so I think it's very good because this has been our observation and the WhatsApp has been like really instrumental in enabling farmer in a similar way. Like yes, we know that they are like, they are not like well educated. They don't know how to market themselves, but WhatsApp have enabled them to create like a small, small groups where they start sharing their produces. So I'll probably put like some story like you said,
When I started working in agritech, the goal for us to bring high tech AI tools in the hands of farmer and just make them access something that we use in US like farm plotting, GIS mapping, market analytics prediction and all. But then we also realized that they're very basic problem. They just wanted to get to the market where they can base prices like how you said in a fisherman wanted to get.
stop agri -waste or like the facing waste and not just throw away their crop. And that was also we were seeing while I am coming from agriculture background. So I'm very well aware of like what happens in this kind of value chain and how the middle layers and the middlemen actually take care or exploit some of the situations like this. So what I wanted to kind of see and work with them, like how they are actually dealing with in the top of the directly selling to the customers.
So I found out that for like regular commodities, they will be just say going to markets. Sorry, sorry for the regular commodities. They will be just going to the market and just like giving away and whatever price they get. But we have seen that like in terms of folks were trying to grow organic. They don't want to go and
give it to the market and get the same price. They were expecting better price. So they started creating this group, what's it looks and adding the folks who are actually looking to buy organic food. So we do not have like a organic market.
[37:58.687] π Pratik Desai: that's why I think you're probably gonna have a great fall allergies
[38:06.728] π©βπ€ Ate-A-Pi: Let's take it. Let's take it. Let's take a second.
[38:08.991] π Pratik Desai: Yeah, yeah, nice. We have like a lot of tree cutting and pruning going on. So, around. So like the pollen numbers are really high.
I'm out.
[39:00.607] π Pratik Desai: I spent so many years of my life on farm but I never had allergies of the pollen but as soon as I land in Ohio and suddenly like I don't know what's wrong with the pollen in US that I have picked up allergy since then so as soon as the spring arrives and if there is something going on today is what? Tuesday right? So maybe there is like a outside there's like a
Cleaning crew or something coming out and like pruning the trees and like that day like our worst for me I think maybe because we do not have a lot of grass in India like everything is like either farmland or Not that many trees probably but it's been surprising So sorry, so I was
[39:46.024] π©βπ€ Ate-A-Pi: No worries. So, okay, I'll do a quick repeat of the question. So basically, I think, there's a 2007 study by a guy called Robert Jensen. And what he found was he was measuring the price of fish in Kerala, in South India. And what he found was the moment you introduce mobile phones,
you saw dramatic changes in the price of fish, which edged towards the price in the market. And the reason was because a lot of times for the fishermen, if they came back a little bit too late, their regular buyer would have already fulfilled their purchasing for the day and would have left. And then the fishermen would often dump or sell the fish at very, very low prices.
And instead, at the moment they had mobile phones, they could check how many other dealers there were in the area. And then they could then meet up with those dealers and then sell on the fish to them. And so they found that quite a dramatic change in the market as soon as you have this introduction of this technology.
So in that sense, as you introduced this AI technology in India to farmers, have you seen this kind of dramatic change in group economics that they do?
[41:22.367] π Pratik Desai: So for our particular have like we haven't been out there for a long to kind of observe like thing but we like throughout our journey building applications for farmers we have seen where Many farmers were trying to differentiate themselves by doing organic or natural farming And trying to get like a better prices in the market because market mostly on the like, you know in India on that level They still do not differentiate between the prices of organic versus the regular
produces so a WhatsApp and what's a group has been some kind of really catalytic the catalytic for them where they are using groups creating finding out people who care for the organic or the nature produces and Selling them directly through that basically so like as you say like you know finding a dealer they are for identifying and finding buyers in what's up groups and suddenly without having like a strength central website or things like that
they are just using mobile phones to distribute, then they are hand delivering them.
[42:28.168] π©βπ€ Ate-A-Pi: So in China, there is a movement through this app called Pinduoduo, PDD. And PDD enabled group buying from farmers. So what you would do is you would have the farmer selling, say, peaches or whatever. And they would do a video on the peaches from the farm. And then a group of people, let's say in Shanghai,
[42:36.703] π Pratik Desai: Yes.
[42:57.704] π©βπ€ Ate-A-Pi: would group together and say 10 people together and they'd each agree to buy, let's say five pairs each. And then the farmer would ship like, you know, 50, I mean, 50 peaches or pears or whatever to one person. And then that person basically distributes out and then, you know, they have the 50 peaches has been bought. And it's really the farmer would get a higher price because he's getting closer to retail. And the buyers would get a lower price because they're getting closer to the farm.
[43:27.711] π Pratik Desai: Yeah.
[43:27.784] π©βπ€ Ate-A-Pi: And so Pinduoduo was quite successful in doing this. I don't know how profitable of a business it is, to be honest, but they were quite successful in some one of the things that China is known for this kind of group buying behavior. Have you seen kind of like that kind of thing happen start to happen in India?
[43:42.879] π Pratik Desai: So we tried like ourself try to do this thing like as a like, you know, the parent company of Kissan, like the one we started to actually help farmers with AI tools, this Ttori platform where we try to create like very similar in a way, like, you know, let farmer create their own store, create their like, you know, marketplace, basically it's a marketplace where they can list all their things and the consumer can find, identify what they want to buy.
Something very similar to that but for India perspective, I haven't seen this thing picking up as China has basically so there are very like Micro patterns basically in buying and purchasing and producing logistics is not still there where a farmer Producing from let's say Western part of India can easily just see anything to the eastern part of the south part
So the transportation logistics and things still not there. So still it's very local. So that's what we realized. We started with this kind of very much similar concept like, Hey, let's enable anyone to buy from anywhere. And then we started running into this problem where who is going to ship? Is it going to be available in two days? There's also trust factor is also very big. Like, like from the China to India, like, you know, the, the culturally, uh, that's a trust factor of someone shipping it.
Then is this are you going to get the right thing or not? And like how do you actually deal with it and with the? Result like also maybe limitations of democracies that like, you know, if you go from one state on the state the law changes significantly So how are you gonna take care of the different state laws? When a farmer actually sent end up sending something wrong, right? Like we take because produce is already kind of rotten while it reaches to from Surat to Bangalore something like that
So there are many, many challenges where, which were actually holding farmers, like consumer to start buying. Basically there were so many inquiries for mangoes and dragon fruit. Like, you know, we can, because we started letting farmers listed, we are like 500 plus farmers who are listing something on the platform, but the consumer then try to just directly reach out to them and talk to them and figure out some way. So there were many trade, which were possible for mostly dry goods, like a turmeric powder and all, but mostly.
[46:02.047] π Pratik Desai: to pick up I think we like at least in India we like this logistic kind of thing kind of missing right now transportation trust factor so it will take some time for sure but that's actually
[46:13.704] π©βπ€ Ate-A-Pi: So it's basically not yet a single market yet. It's still fractured micro markets. And then you have the, I guess the buyers are primarily in the urban areas and the sellers are primarily in the rural areas. And then you have this trust deficit in between the areas. You have different languages. You have different state laws on taxes, I guess taxes and is the pricing regulated as well for products?
[46:17.791] π Pratik Desai: No.
[46:43.167] π Pratik Desai: So yeah, that's actually changes state by state. The agriculture is very touchy topic because of the one of the largest vote bank kind of thing about like the farmers being there. It's a state's subject, not a federal subject. So not a federal government cannot control everything actually. And then every time a state government changes depending upon the vote bank, like a lot of politics actually get played on agriculture and farmers in a way countries like us. So.
That's one of the biggest challenge to kind of just tomorrow start disrupting like as anything on very large level where you kind of kind of claim that, you know, you're going to just remove all the middle layer like overnight because that happened when the current government tried to bring some new farming laws and there were a lot of protests and according to many the mostly the protests were driven by the middle layers who were seeing that like the farmers getting enabled to sell directly.
will kind of cut down their costs. So there's so many things kind of the farmers are also like in the US and like every other countries like we are seeing farming protests in France and Canada and everywhere because that's like a one single group. If you look at it that way, it's like they all think probably similar way. Maybe me too, because I'm also a farmer, but very like to the ground thinking, working with the land. So no very hard dimension.
[48:04.511] π Pratik Desai: Carrefour is all just like simple thing.
[48:05.032] π©βπ€ Ate-A-Pi: Yeah. So you did your PhD in Ohio, right? And then you worked in the Bay Area for quite a while. And one of the interesting things is, let's say for Denu, which is your disease detection AI model, can you compare how difficult it would have been and how expensive it would have been to build the same thing?
let's say five years ago, like, you know, five years ago, would that have been like a $10 million project that, you know, John Deere did? And is that, you know, is that really, is there really a cost compression, you know, from that time to now?
[48:47.711] π Pratik Desai: Oh, a lot, a lot. Like, for example, like just five years back, like this level of conversational AI was anyway not possible at all, right? Because of the...
Like we started using GPT, the GPT in 2020 October or November with access. So we've been using it for a while, but even before that, like only the CNN models were there just like basic training law. The data acquisition costs would be high. The models were not that expensive to build in terms of like just a training because it's just like a regular machine learning, traditional machine learning with data curation cost was there. Like synthetic data generation.
that we are doing right now would be like very difficult to build like you need to have like army to create like 100 ,000 images, something like that. So many things in terms of technology and tooling wise has been possible through the JAN AI era that we see because the capabilities are different, right? Like five years back, 10 years back, if I want to even imagine something which comprehend the question in terms of AI was not possible at all, right?
[49:35.336] π©βπ€ Ate-A-Pi: Mm -hmm.
[49:56.191] π Pratik Desai: because we did not have this level of, you know, the transformer change everything and not just transform. I think GPT is something which actually brought this some level of comprehension of what you were trying to say. Same thing in image. Image detection is probably simple. Like you put an image train through 10 ,000 images of the one crop and then you're okay. We detected in some way, not underweight, like something we do with the face and all. So the premise is change completely then.
just the detecting the disease like, you know, your enabling conversation is something which is more important, like identifying, taking a picture and saying, what should I use for it? Basically, you can detect a disease and ask and provide you the information, go back and forth. And the nature progression is not the reading detecting the disease. Like right now, let's say this is just the first step for us. The goal is like if you if you have seen the vision model where you can actually describe more than a leaf leaf with the disease, right?
We just right now we wanted to see if we can actually do conversation. But next goal is take a picture, a holistic picture of the farmland basically where you have a leaf, you have land, you have weed, you have some other situation going around in the crop or a farm from a drone or from a side or whatever. Right. And now when you actually annotate this images and then kind of train, what will end up happening is that you do not have to run like a very specifically trained.
like how we used to do CNN, but you can actually start talking to the picture of the farmland and say like, hey, why there are yellow spot and there's a weed spot and there's a dry land? What is the like a correlation between this? Right? So this is something will start happening more level on like on a larger level where if you are, if you are running a drone around right now, most of the drone related imaging startups are built on like a, like your CNN models, like, you know, so they have like a lot of pictures annotated.
then they'll send a drone, take a picture, identify this yellow spot, maybe having this disease, right? So those can completely change once we actually start having this one where you do not have to have a one model trained for detecting diseases. You can just have like a situation asking us like, hey, there is a leakage, a water leakage. It's making a big puddle out of it. Like those kind of things can be also enabled. And this is something, you know, can be like a natural progression, which is not possible five years back at all. Like, uh,
[52:19.679] π Pratik Desai: I think this is like we are going into realm of like, doesn't matter how many money you may have spent, they couldn't have kind of that comprehension of the level that we are getting right.
[52:28.552] π©βπ€ Ate-A-Pi: Right on. So regardless of, you know, I mean, five, five, 10 years ago, you could have a drone, you could build out these kinds of specific disease detection models, you could do like an aerial survey, but it wasn't a cost effective solution for smaller farmers, you know, especially in emerging markets. And even when it was, even if you tried to do it, you know, the government tried to do it. The fact of the matter is that
you know, some of the capabilities that you need to implement the product, you know, voice detection, language, etc. were not even possible for any amount of money. And now all of a sudden, within the last couple of years, basically, all of this, you know, wide open space has opened up, right?
[53:16.511] π Pratik Desai: Think about it like a farmer having a personal drone instead of using anything. Just take a picture from the sky every few days. Just then pass this image to let's say our next version of dinner and say like what's going on. And then it will just describe the whole farm fields like this is going on. This is this is what like describing the whole farmland. Right. So that is something and be like, you know, and maybe that can be automated to earlier going forward.
Now earlier that was not possible. Now this will be like very cost effectively possible where just like, you know, take a picture and you can log in what's going on. Like you can take a picture every day and it can give you the nature of progression of your farmland doing the image analysis every day, adding the context window, adding it to context window and see like what is the natural progression of your farm? Like, you know, what, what, what are you doing wrong? What is the right thing that you're doing? Because this is a
like you know most of the thing that we see is like it's not like a static thing that like you created the code and run it and it's over like this is like every day a smaller minor factor of change in weather change in fertilizer amount change in like you know like in Napa Valley changing like a fog can change like you know the yield or yield type or all this thing right so it's like it's like a it's not like a temporal thing that you are talking about and that is why you know the
Some time series based context would be like important. And this can enable for many application, not just like agriculture where every day with the five factors are important for your final yield. Right now, this is all human experience, right? If you are growing grapes for 20 years, you know exactly what is going to happen. But that's like all in your brain with your experience or something that you've been passed down from your father and father and generation, right? And some may have written the book.
But then you don't give that book to a new guy and say like, hey, this is the book, figure out how it's going to work. No, you have to actually work. That can be basically possible. That's kind of see that something probably will take time. But it's possible now, basically.
[55:19.496] π©βπ€ Ate-A-Pi: It's the transfer of tacit knowledge, all of the knowledge not written down in books, which is basically the majority of the knowledge in the world is knowledge which is not written down in books. So I guess this is the part, especially as the models get better in terms of taking in data from the conversation, which actually, I think only a few guys are doing it right now. I think Mid Journey is very specifically fine tuning the next generation model from.
uh, you know, user preferences from the last generation and chat. GPT free is, uh, tuning from, uh, incoming conversation. Uh, but chat, GPT enterprise and the paid version are not tuning on your data. Um, so, so the real question is like, you know, how, how much user generated, you know, you went, we went through this user generated content era of YouTube, uh, and how much of user generated content is actually happening right now. User generated training where we're actually training the models by using them.
It's not very clear. And so I don't think a lot of people have a very clear understanding of how much that actually is happening. Are you guys, for example, capturing the most frequent queries? And then are you doing a quality rating? You have another language model running and saying, hey, did my model successfully fulfill the query or not?
[56:40.031] π Pratik Desai: Yes, yes.
[56:48.072] π©βπ€ Ate-A-Pi: And if not, then what category was it? And then I look at the top categories and then like, oh, okay, what can we do to improve the next version? Like.
[56:54.975] π Pratik Desai: So we're not running a critic model right now live while the queries are happening. Uh, because then like we are anyway, like trying to cut down the cost significantly. So running the critic model with the GPT, but we did use and like the reason for running right now, you know, like earlier I was saying, like, you know, uh, the finding out a model is very important for us. So what ended up happening is the free version that we run is basically for us right now source of the data. Right. So, and that is where, uh,
[57:22.12] π©βπ€ Ate-A-Pi: Okay.
[57:25.279] π Pratik Desai: So if you compare this thing with the, uh, that's a GPT free or mid journey, where if you identify the person who is an artist, who is a software developer or who is a product manager and giving the preferences, then probably you can categorize it better. But what if you're, you cannot actually identify that, right? Then you become like a very generic job, which is fine because from the scale of the computer and the, the talent that, uh, opening I have, like they can figure it out.
What I am focusing on is like my vertical is agriculture. Most of the people that are coming are either farmer or agronomist. So that is actually making it easy for me to do topic modeling from the conversation that we have captured. And that is the some nuances that when the first version of then we trained, which was not the vision model, but just the language model for the Indian agriculture practices, we did use around 100 K of this instruction that has been curated. So.
You know, there are a lot of things which is garbage. You have to just get rid of it. But then the literally, you know, all of us together kind of went through this conversation, identify kind of rated with our nature meant like, no, like literally people who with the agriculture knowledge actually went through the queries and curated this data set. And now that is basically helpful because now we understand what kind of question actually get asked by the farmers and what are the answers that appropriately because that is also important because.
As a as like a tech person who may be looking into like a home gardener may be looking to say like, hey, how should I know like my plant is not giving me any flowers like there's no yield on flowers. So like but then the the word that farmer use would be completely different, right? He would be talking into like the jargons. He'll be talking into like very pinpointed questions. So we have seen like the.
the question that come from like it's very specific. If you're talking to domain expert, the word and the nuances the user are completely different than the same thing that is someone's just learning. So that is what we're trying to capture. And that is what I see that can be like, you know, going forward, like capturing the conversation in your platform. And if you're like very vertical specific thing, you can actually separate out from the mass and say like, Hey, I have vertical specific copilot running. I know my
[59:47.423] π Pratik Desai: Vertical variable all the nuances I have captured in my model so when you are using the new Dino is every model and it understand farmers and agronomist well and not care anything about else about basically So this is what I see like a neat like from my perspective like, you know, we have we are having this debate where the models are getting bigger They're gonna get a lot smarter like, you know going forward But then the thing is like are they going to actually capture this nuances for every verticals or the verticals need this smaller?
models very specifically trained for that purposes right so that is probably going to be the going forward like you know the differentiation like there will be a category of like a very great general purpose model like gpt 456 and there will be like a vertical specific model like harvey's also trying to train gpt4 right with their working with open AI we are working on agriculture model i think there are some other people working on the law model or finance model right
So this is something can separate out where there are smaller models for verticals and the large model, probably the critic models. Like, you know, you can think about it. So they're not very good for deployment in large because they are so huge. So it's expensive. Your smaller model can do the task and this critic model can then say like, Hey, I see what's going on here. This is right. This is wrong. I think that's going to be like a good initial progression that you don't have to deploy on the consumer base, the large models, but they can after the, uh,
curation they can use as a creative model.
[01:01:13.992] π©βπ€ Ate-A-Pi: Indeed. I'm going to just go into something on a tangent here, which is regulation. So internationally, different countries have approached regulation a little bit differently. And I think in March, there was some proposal in India that all deployments of AI models should be registered or something like that.
Like what actually went down? And then, and then I think it was revoked a little bit later. What actually went down in there? What was like the, like, you know, gossip on the ground on like, you know, what actually did the government actually end up implementing something or they implemented and they pulled back or what actually happened?
[01:01:57.727] π Pratik Desai: Okay, so I think so that is a weird story. So it's like there was a tweet from Ministry of Electronics and Telecomics and 90 Rajiv Chandrasekhar and he talks about like
[01:02:10.472] π©βπ€ Ate-A-Pi: Yes, I think Ram Rastogi, and he's the Ministry of Electronics and Information Technology, METI of India.
[01:02:20.255] π Pratik Desai: Yeah, me too. So, so what ended up happening is there is a circular out, which says like, you know, models within like next 15 days or like some 30 days have to kind of register and need to get approval to run. And it's not just the model, even the application that are running on the model. Right. But then everybody like me, we started, so like, Hey, this is not a way I'm not going to kind of keep continuing working because I do not have resources to go and do evaluation.
because this is the first day of this particular tweet out, right? So everybody started reacting and our reaction was like, you know, larger company can go through this regulation because, you know, they have contacts, they have, you know, how the rate taping works and all, it's not very transparent in most of the country and like you need to have a lot of people look around it. So like for a smaller company like us, it doesn't work, then maybe I have to stop working on this thing. Then a lot of people, like in the startup, they reached out, tweeted, and then...
the clarification came that it was just an advisory because some IT law already prohibit people to create an e -fakes and misinformation and things like that. So it is just an advisory. So ultimately kind of then I think they pulled it back and gave more clarification about that doesn't apply to the startup. It doesn't apply for some particular field. It's for mostly and then
It turned out to be something before the election because the elections are going on right now and people may misuse the models to generate misinformation. So it was just some precautionary thing. But I think it was just a premature thing without thinking a lot. It's a knee -jerk reaction of some recent incident that happened with Gemini doing something like talking bad about India. I don't know, I forgot what was the scene. But it did something that was not aligned with how...
the thinking.
[01:04:15.304] π©βπ€ Ate-A-Pi: Was not aligned with the greater power.
[01:04:18.943] π Pratik Desai: kind of yes and like the whole tone of the kind of right now like you know the country goes to like a touching cycle so right there's a different cycle the India is right now so I think the so that was the thing I think then they retracted back and it's like not applying applied for startups or the models do evaluation you are responsible start putting the pop -up say like you know yeah so which is now becoming coming back to like the standard thing that anyway open AI does it anyway and so
That was the thing I think now they have created this think tank about it. My biggest worry for all this think tank is there are not the people who are actually building models and applications are on those think tanks and all these are like a people with the gray hairs been in part of the policy making and things like that who already have contacts and they do not understand the JNA at all. So this was the thing that they say like hey every models need to be evaluated.
before it goes in production. So my I tweeted like do you even have eval set for us to do evaluation like for which works in India like you know like this it need to be kind of in the premise of the rules and regulations of India and the evaluation need to be submitted like who's going to provide evaluation are you going to change eval set every time a government changes or there's a law change in the law who's going to maintain it who's going to be the the authority to actually taste and approve this thing so there is no plan because
Many of the people who are there do not actually understand how the whole system works basically, right? It's just an AI giving some answers and you need to taste it, right? And like going through this, you can kill a lot of startups basically because you know, like you are going to change dynamically. Do I have to keep fine tuning every time you find something, right? Or keep building a new models? Like, you know, how expensive that can be. Like I'm not getting a 100 access this day. I have, I have like a three modeling pipeline to be trained. And I'm like,
knocking doors to get access to GPUs and then if someone comments like hey you passed the UL I'm going to just shut down my app every here and there because they will find something out so it doesn't work like they need to have people who actually are knowledgeable inside understand who has worked, trained, built application to actually you know put the point of the startups or the company or model maker and I think that is not there I don't think it's there in US either like US is also
[01:06:44.831] π Pratik Desai: but US will figure it out US figured it out all the time like you know whichever there are there are like people in their deep state or a swab who are smart enough to figure out things if it's go wrong India need those kind of there are very like a few people who actually have this understanding but I don't think they can do everything so hopefully they find some
[01:07:11.56] π©βπ€ Ate-A-Pi: So in that, like,
[01:07:18.408] π©βπ€ Ate-A-Pi: For the past, I would say, 20, 30 years, computer science has obviously been extremely popular in India. But now with chat GPT, you can start to see that even people of moderate knowledge can start to do some of these computer science tasks and coding tasks on their own. So is there a little bit of like,
um you know is there a move for people to try and do machine learning or um you know kind of not study software or has that has that impact in the in the student community like uh and and also tcs kind of made like its first cuts i think in like 20 years and whatnot like
[01:08:03.359] π Pratik Desai: So right now everything is knee -jerk reaction. Basically, they're just trying to figure out like everybody's like go through the flow. So the flow is like, oh, hey, let's start learning open like this chat GPT calls or this new library, Lama index, land chain.
And build some applications and they are just happy by doing some cause But all of this thing going to get automated so India is going to get affected very well for the There is like, you know, there are very good engineers too, but there is like a whole mid and low layer developers that India produces which works in this Software developing companies, you know like outsourcing companies who can literally easily get replaced like, you know
for many, even today, like even the, the David and David can all of the models are, are the agents are not that capable, but still they can replace a lot of the people that I have in my past work as like a part of someone doing outsourcing and it's like, Hey, no, you can't do anything. So this is going to definitely change. Now, not everybody is going to link the number of software developer anyway, will it reduce, right? But then I see one, one important thing is.
Most of this low level or mid level people who are working in software development. They were not from computer science background Like you know TCS and all they just go to any branch. It's like hey, I'll pick you I'll give you three month training I'll help you with some Java tutorial and now you are a junior software developer and I'll charge $25 to American companies for a one resource
[01:09:23.976] π©βπ€ Ate-A-Pi: Right.
[01:09:36.68] π©βπ€ Ate-A-Pi: You know, it reminds me of those Chris Ferry, like, you know, learn quantum mechanics for babies books. And then you flip like, you know, 20 pages and then at the end, they tell you like, now you're a quantum mechanics guy or a quantum engineer.
[01:09:52.031] π Pratik Desai: Yeah. Yeah. So this is will throw like a smart engineer and nine, not this is every all of them. This is probably is a better one. But they'll throw nine one who are learning things like that. Right. So all those nine probably will not need it for any of the software company. Maybe they can do with the two. Right. And they may still charge for the project price. Now, what is going to happen with those eight? One thing which I see and.
I think going to be important. Like we are already seeing the importance of high quality data. Like there are synthetic data being generated and the domain specific data and all. But there will be always this thing where you are not like, you know, the aura boros stuff. Like I cannot, I can generate a lot of synthetic data, but I'm not going to completely rely on synthetic data. Otherwise, like this AI eats AI data kind of thing can be, I personally believe that.
It's not 100 % going to work out. Now what we do in that case that we will need some domain experts. Now, uh, all these people that we talked about who are working in developing code, but came from the civil engineering or the chemical engineering or textual engineering, they can now go back to their basically field, what they study for four years or whatever, and help out with the curation of this data. Because right now I'm looking for agronomists to help me out. Right. Uh,
But surprisingly the people coming from agriculture degrees are paid well compared to the engineering degrees even in India because there are not that many and most of the chemical engineering companies or the fertilizer base or even the the companies are like trying to run or build businesses. They actually hire them. So they get probably were paid well for more than engineers in India. So all these people coming from different engineering can go back.
bring them domain knowledge and help me out with this data that has been synthetically produced and say like it is right wrong. Do some RLHF stuff, add some nuances, help me out so I can create like very domain specific models. Right now, probably they will not have to write. So we'll not need those many of those, but still there will be like a good amount of domain experts will be needed. And I think that is going to be the natural progression right now.
[01:12:15.519] π Pratik Desai: That is also kind of service and knowledge service. Like, you know, people scale AI has been doing on a large scale, right? Like they are the one of the, the hidden secret behind success of most of the model. But that is something that can be very important in terms of say, like I'll give you the best curated agriculture data because I have figured out how to generate it synthetically have a domain expert and you have millions of thousands of PDFs in your organization.
How are you going to do it? Are you going to hire in California? No. Okay. So there is a company here in India who will do it for you. So that can be a good business model. And I think many people are thinking about it, but then it's like, it's a new, it's also like a different, you know, it's not a status quo of how a services business work. So not many people are serious in terms of even like funding financing people like a SaaS business because SaaS will be a multi -plan and all. But if I say I'm going to create like a some scale AI type of business for one or two or three verticals.
then people will not think about it. It's like, hey, this is like a service business you're doing. So this is something going to be kind of change. Indian will be from a software developer to knowledge workers, but there will be a drastic change in this industry for sure.
[01:13:27.784] π©βπ€ Ate-A-Pi: Indeed. Let's go to, I think, so I want to ask because, so one of the things that I think in the US has been happening is this whole thing about immigration this year. So in terms of like, and you know, I think Sri Ram, Sri Ramakrishna Moorthy at A16Z, he did,
He did kind of a thing with a daddy on his H1B and green card process, et cetera. So are you in that green card process right now? Are you post green card, et cetera?
[01:14:09.087] π Pratik Desai: So we are post green card. So I have I want to do like so everybody in US who migrate their own like immigration stories like everybody is different and all the friends I know have some kind of ups and downs and things like that. But for us, we recently got like I think two years back. But I've been here in US since 2006. Here's my PhD in 2013.
So probably I was eligible for EB1 in the early days. But my the problem I had was that I was always part of a startup, smaller startup, and sponsoring EB1 was like almost impossible at that point. Like I could. And in early days, like right now, EB1 and O1 has become like like another like a student visa kind of thing. Like anybody's kind of people are figuring out they are running seminars and webinars to how to get EB1 and O1.
because it looks like the USC has been very lenient, but early days, even the EB1 and O1 used to get rejected. If you have like, I know some folks who have like a really great like research background and even a patent and they got rejected because the inspector on the other side was, I don't know what was his thinking. But for me, I was in OPT after like, you know, the OPT and I never went to on H1 at all.
because I was in startup. So I worked on this OPT period. Then there was like even EB H1B EAD thing, right for the spouses. So my wife has a H1 and the thing was
[01:15:48.008] π©βπ€ Ate-A-Pi: It's an employment authorization document for the spouse after you successfully apply for a green card or something or H1B. H1B, yeah, something like that, yeah.
[01:15:51.583] π Pratik Desai: Yes.
[01:15:57.087] π Pratik Desai: Yeah, yeah, yeah, it wouldn't be yes. So, so what and so I had a choice either go and get an H1 where I cannot do a startup or I cannot run my own business. But he has this clause where you can actually have your own business. There is no restriction. Each one has this thing where you have to work for one employer. He did not say like, that's a perfect I'll stay on any for my life until we go to a company where like, you know, the big enough to get a sponsor, you won't and all.
So I kept doing my last then, then before I realized like, you know, they're like, I'm ready there to apply for EB1. I got through it when we, through my wife's process that the green car. So it took us how many years? So I was in us 2006 and we got our green car in 2020. So it's like a 16 years after we were in us, right?
probably being illegal I could have got earlier. And the most surprising thing, one of the things, I have no problem, like, you know, I'm okay to wait. I'm staying here legally. Like, you know, in my career, I have worked in terms of like a good project where I was in Ohio, we have like this Wright -Peterson Air Force base, there's an Air Force research lab, and I was like working on outside with the product. So, you know, it's not that like I can go outside and front my resumes, like, yeah, I work on some like a good project for the country.
But I was like, not in a hurry. I was like, always focus on what I'm building. Say, okay, whenever we'll get my wife was in hurry. Like she was like not happy with this eight one process and all. But my biggest thing was going to the DMV in the DMV. When you go in California, anyway, this sucks a lot like that. So like the worst stories, uh, we have Santa Clara DMV just like across my house. Uh, Apple campus is like a two blocks there. NVIDIA is like a four, four, like a two miles on the other side.
[01:17:45.864] π©βπ€ Ate-A-Pi: Right.
[01:17:51.327] π Pratik Desai: Everyone like this is like a Al Camino and Lawrence intersection. So it's like a 10 trillion economy around this one. Everything got invented. And that DM here is just a one single camera for your photo application, like a driving license application, single camera. And it takes four hours because you have to go one line, two line, three line, four line. And there's a camera line with 20 people, 20, 30 people anytime standing there to take a picture. It's like.
We're living in the center of Silicon Valley, the center of the year, we have just one single. So now the thing is when you are immigrant, you your your your license get approved only for some once until your EAD and things expire. So, you know, every every year and every nine months, you have to keep going to renew it, renew it, renew it for hours, four hours. And that's like that's one of the things like, OK, this is something. And then the immigration is something also important, like when you are entering.
[01:18:39.272] π©βπ€ Ate-A-Pi: So your driver's license also expires with your kind of immigration status. Interesting.
[01:18:44.063] π Pratik Desai: Oh yes. Yeah. So, so the story is like my ex, the DMV did not, and then you, they give you like a three month permit. And if it's expired and you still haven't got your driving license, you go and renew the permit. So I renew three times nine months. Then I started like, like almost got so angry in the DMV and he's saying like, Hey, driving is not your right. It's your privilege. And what else? There was something funny. He said,
that I told him like, you know, California has this law. If you are illegal immigrant, you can get a driving license. Actually, no question asked you. So you're living proof. I said like, can I do this thing? Can I just say I'm illegal and give you my living proof? You give me driving. I don't want to come every three months and wait four hours. So it's like, but and then he doesn't have an answer, right? Because it will be easier for me to get a 10 year driving license by just saying like, I'm illegal. I don't have anything. Give me this. My address.
[01:19:14.664] π©βπ€ Ate-A-Pi: driving it.
[01:19:41.567] π Pratik Desai: That'd be easier in California to get a driving license for an immigrant who is like working, living, paying huge amount of taxes. We do not have any privileges. It get every six months or nine months. If you're OPT or EH1 or getting expert, you have to go and renew or keep license. Yeah, that sucks.
[01:19:59.336] π©βπ€ Ate-A-Pi: It's amazing. So, okay, one last question. So you had a tweet, I think maybe about sometime last year, it's been about a year. And you said, start regularly recording your parents, elders and loved ones with enough transcript data, new voice synthesis and video models, there's a 100 % chance that they will live with you forever after leaving physical body.
This should even be possible by the end of the year. And...
[01:20:30.623] π Pratik Desai: Oh yeah, my predictions are very not time bound. I just did not say year. But you see like you are we are we are on a podcast with Avatar, right? Yeah, who's like literally speaking. And now the part that is like also like so so this is what is happening. I don't this tweet actually got very popular, not in public, infamous. I was kind of asked by every big people were reaching out to me in my DMs.
for interview and I got like asked 20 ,000 times to kill myself. But I don't know why we got know but like I'm just posting like I'm like a right like on my Twitter account like if I will I will never publish my Twitter account on my LinkedIn because then they'll I'll never have a customer because I'll just like talk anything that come in my mind.
And at that point was the same thing because I was seeing like all this initial progression of voice synthesis and 11 labs and like, you know, there's a way to leave came out. You're doing some way to leave project at a time where you will literally get a make my mice my like take a picture of me like that. We can do it anyway, like a lip sync thing. This was like April last year. Right. And we run the project just for creating content for India. Our goal was to create a
material where many books are not available in all the languages. So how about can we actually create a transcript? So we're working with some models and the same time I had like this story where my grandmother wanted to meet my daughter because she was like the first girl child in our family after a long long time and she died like two weeks before we reach and I was like this kind of thing which I would like a thing where I
[01:22:10.568] π©βπ€ Ate-A-Pi: This was during the COVID era, like just before the -
[01:22:14.111] π Pratik Desai: No, no, no, before that, before that of 2019. But my regret was that I never recorded her significant amount of voice that I can actually clone it to hear her voice sometime. Now, what's wrong with like so for me, even today, I don't feel that sometime as her birthday and all hearing her voice talking to me is like wrong. But that's like very touchy topic today, even at that point. But at that point, it was probably too early for me to speak now.
It's becoming very regular, right? You are, people are recruiting Steve jobs and all the dead folks and other, and now people are getting okay because you know, this, uh, this like a time kind of change everyone's perception slowly, slowly. They're going to start seeing in the brain and then it will become acceptable for them to talk to the people who already passed. I was just too early to say that. Uh, but it's going to happen. Like maybe.
last year was probably you know one of the biggest thing was many actually model came out we can do very nice cloning like meta has one open I just talked about it but nobody's releasing the weight and if you look at that probably I don't know they read my tweet or not but this was the thing where they are actually afraid that people will start cloning they're like misuse and all things like that but if you can have actually enough
transcript like a recording of anyone's voice you can just literally recreate it and I Don't know this will be possible. Anyway, someone will do it. Like I had many people reaching out with this thing about You know grief like so there are many many many every people deal with the grief differently Some people want to completely forget some people want to slowly slowly go down on that and many people who reached out in DM after that with it's like You know, we like this thing because we also sudden loss
and I could have spent some more time and slowly, slowly go down in the green curve. And it will be possible. I think it's uncanny really. But what else like, you know, right now, the way we are going, it's like our personal choice for anyone to kind of use this technology or not. And there will be like always people on right, left. We have.
Ate-A-Pi (01:24:25.) So yeah, I think the first one that I saw was, I think a concert maybe six, seven years ago, someone had a Tupac avatar and they had a Tupac avatar and kind of, he had an unreleased song that they completed also. And I don't know how they did the voice. But you know, they hand made the voice, right? They just, they hand made the voice and then they put it together. And now of course like that.
[01:24:35.007] π Pratik Desai: Yeah.
[01:24:53.96] π©βπ€ Ate-A-Pi: that is easily doable, that the voice clone is easily doable. The facial clone is also kind of easily doable, but the quality is not there yet. The quality is still a little bit far.
[01:25:05.727] π Pratik Desai: Did you see this new Microsoft model that they announced? So the whole, the wave to lip and the lip syncing thing, they solved it really, really well, basically. So now there is like no bloody lines between the lip when they're talking. So.
[01:25:19.368] π©βπ€ Ate-A-Pi: Well, not only that, but that was real time. So they're generating in real time. So 40 frames per second on a 4090 on a consumer GPU and from single photo. So yeah. Yeah, yeah. So.
[01:25:23.391] π Pratik Desai: See you next time.
[01:25:31.935] π Pratik Desai: Correct. So yeah, so it is it is like, you know, the way we were seeing remember like in Star Wars, like it was normal when people did it in Star Wars, where you have like someone's cassette holding down and you're talking to that person passing some messages, right?
So that is what it is, right? Basically, like you have and the context window is something important too, because as you know, so the last part in the tweet, which was kind of something you think about is like recreating someone's thought like, like, yeah, you cannot, you cannot talk to some person because they're like, it's very complex thing.
But like if you have enough context in terms of how they talk, what kind of word they regularly use, how they greet you and, you know, like your personal conversation, which like, see, a human can be a complex person because he or she talks to 1000, 10 ,000 people. But when they talk to me, there is like a one slice of their brain, which kind of accustomed to talking to me. The same thing in my brain. Right. Now, that part can be smaller.
because like you and my interaction, let's say how many times we have made, we know exactly what we talked about our remasses. So if I have that context and if I continue using the AI model and say like, hey, this is what we were talking last time. What about this? And then it's possible. Like I'm not recreating the whole person for talk like with the memory of 10 ,000 folks they have been in contact with during their life. But for one person, it is basically possible if we have enough context and a captured and it can be very high level generic stuff, right?
and we can do it right now.
[01:27:07.24] π©βπ€ Ate-A-Pi: So it reminds me of the Shoggoth meme of the big creature inside ChatGPT and all you see is this, the face, the nice little smiley face at the front and you don't see the big creature behind it. So in some sense, we are all that way where we only present one facade, one face to someone. And they only see that one face even though there is a complex and multifaceted persona behind that facade.
[01:27:12.223] π Pratik Desai: Yeah, yeah.
[01:27:37.32] π©βπ€ Ate-A-Pi: they only see that single facade. And replicating that single facade might not actually be that difficult. Replicating the complex persona that multifaceted might be extremely complex.
[01:27:46.719] π Pratik Desai: Yeah.
[01:27:50.399] π Pratik Desai: Yeah. And the, the, the whole person, the complex person is not just like a, of your thoughts or knowledge and memory. It's also a lot of time. It's like, oh, so many things are hormone driven, right? Are you going to recreate that part? Because like in a different time, different mood, different, like, uh, your weather changes, your mood, your hungry hunger changes, your mood, and your answer for one question will be different if you're hungry or not. So those things are going to be very difficult to kind of replicate. But me and you, when we have conversation at the.
The blasted view, we can all recreate that one if we want to kind of if you have recorded all of it. So I think that is what kind of man like what like it became very complex. I kind of I don't regret but doing that. But I say I think I could have done it today and nobody have noticed probably.
[01:28:23.688] π©βπ€ Ate-A-Pi: Indeed. Indeed.
[01:28:39.016] π©βπ€ Ate-A-Pi: Indeed. Pratik, thank you so much for your time today, and best of luck to you and to Kissan.
[01:28:48.479] π Pratik Desai: Thank you, thank you very much.