AI, In Conversation.

Mm. A field tag has now domain knowledge is increasing, right? Because now we did the first step, is more people are coming. They are asking couple of things- Mm. -which trust me, like I need to really sit in. Without that, I cannot make... If he says, add this button or something, it's easy. But when he apply couple of things, weights or this, I need to see what comes next, something like that. So if we think in this direction that we get a model and we trained it only for specific purpose, only for specific purpose. S-specific purpose means, let's say I have a field tag. I added all New Zealand defined residential laws, regulations, and I also added their... make it learn him how we are implementing it, and we are fine-tuning it as we are developing the solution. So that model knows it is, it is, it is himself, I say, field tag model. Can- I- No. No, no. It's, uh, I think it's a natural way of thinking, but, uh, in my opinion, it is not correct because- Okay. Mm. -what you are trying to do is you are trying to feed information to a language model and expecting it to be an expert on it- Um, which is okay, but in my opinion, not very correct. See, like large language model understands the language. The field tag, field tag and all related information, whether it's a law or act or business or whatever it is the information that should be put it in a rag. LLM should only be translating your prompts and connecting to this information to form the responses. Okay? See, what you train a model for? You train a model for English, its grammar, its nuances and everything. You can train a model to do a coding. So you feed all that, uh, all that, um, uh, language, uh, say Java, for example, Java. Java language, um, compiler stuff, all that language nuances, how that language was formed, uh, its examples and everything. Correct? Uh, now consider I give this model that code me a device driver for USB, but this model doesn't know what is a USB driver. So what it will do, it will connect to some addit-external information, get the details, and then uses its la- Java learning and code that in Java. Okay? So if field tag and related information is there, you should be connecting to this information to a large language model. You are not training field tag information. Uh, you're not using this information to train the model. They're two different things. Okay. -is a, is a large language model. It understands your English. Okay? But if there is a... Now your domain, field tag is a domain. It has lot of information. You connect these two, keep this information separate, connect these two. Teach LLM how to connect to this guy using MCP server or using a rag way or whatever you want. Okay? And then you use them separately. That's a-- In my opinion, that is a right way instead of... Because here what you're gonna do, even if suppose even if we train, we give all that information and then we train the weights, uh, what it will accomplish? That's I'm trying to understand. Suppose there is a image model. Image model understands the image, JPG, pixels, RGB, CMYK, transparency, and it knows how to form an image, how to decode an image. So when it comes to how-tos, how-tos, you put it into the model. When it come to information, you just use a tool to connect it. In my opinion, thought that, yeah. No. In my opinion, no. The reas- reason being, reason being, um, there are two part of it. Consider there are two part of it. One is, uh, you're trained in neural network and other is information out there. For example, for example, let's say we have a chartered accountant who has studied for six, seven years, understood everything. Right. Now, this guy is trained. You ask him, he might answer something. You ask difficult question, he might go back to his books, read it and give you the exact information This guy doesn't know everything. Now, a new GST comes in, new law comes in. This guy doesn't know, but this guy know. What this guy know? I... This guy is a trained chartered accountant. He will download that new act, read it- And continue it ... and then update itself. Correct. Correct. Correct. So training is one part, information is the other part. Yes, you can do. What you can do, use... Forget about, I don't want to keep it. I will use this entire information and train this. This a heavy, heavily costly process, but somehow I train it, okay? It will answer, but things will keep updating outside. Correct. Update, keep updating, and not necessarily, not necessarily the information this guy is, has got is 100% perfect. Even if you train this huge information with it, there are probabilities that it might be 80% correct, 90% correct or whatever it is. Okay. So train for a skill, keep the information separate. This is a good approach. You can use that vast information, train it, but when updates will come, what will you do? Again, train it, train it. When any law changes or any new layout comes in or any- anything

Right. Right. But let's say, let's say about the car, right? We saw the car Maruti eight hundred, right? Maybe that was also in your time as well. I have seen the initial cars, Maruti eight hundred or Ambassador, right? For me, that was Mercedes that time. But now recently, I looked into the cars for my friend as well, and I looked all the top quality cars- Mm-hmm ... including Mercedes, Audi, or something, right? Now there's a ceiling set. There's a ceiling set. Yes. Yes. Right? So when I look... Previously, when I looked on these kind of high-end cars, right? I saw, wow, I can see. But now when I saw- Mm-hmm ... and I did all test drive, I did not feel that it's Mercedes. I just felt it's a car. No. It, uh, it is definitely it is better, it is good, but after a certain level it was giving me just- I got it. I got- I feel- I got it. So I'm asking- I got your point ... where how you see the model? How-- If we talk about the models, right, and they are working well, how you see how much they go? How much they go? code- Mm-hmm ... it's a very considering artificial general intelligence, Claude model is a very small model. It's specific to coding tasks. It create programs, correct? It analyzes code. Um, it can hack into systems, correct? If we, uh... It understand network protocols, correct? But this is not... This is a very tiny, very tiny piece of, uh, piece of intelligence out there. Consider a general intelligence system which can do anything. It can do anything, correct? Consider like, uh, um, you pick top programmers, top network engineers, top hardware engineers, top scientists, top doctors, top lawyers, top accountants, top every domain on the earth into one- Uh, floor, one common place, and ask questions. Some of them will be able to answer or some will be able to create things, uh, in perfect way. So AGI is where things will evolve eventually. Can you explain AGI? See, consider as of now we are thinking of deploying various you explain- Huh? Can you explain more about AGI? Sorry, I heard this word, but I can't, um- It's a, um... Okay, so, okay, have you seen any this one, fic- fiction movies in which, in which there is a computer say in Star Yeah. Ha, consciousness means the one who feels happiness, sadness. The who, the one who feels pain, um, joy. S- so consciousness is more towards spirituality if we go, correct? The one who lives. Yeah. Uh, the consciousness is the one which makes you live. Mm. This, there is a robot. If I create exactly same repli- replica robot of a Sumit, it is having cameras and eyes. It does things like Sumit. It, it thinks like Sumit. Correct. It is having all those, but is it alive? Correct? Mm. Is it alive? It can imitate that I am sad. It can imitate that I am happy by, uh, changing the motors and then, mm, it makes like this, um, happy face, and then show, shows. It, it emulates or simulates happiness, but it is not happy, correct? So, uh, consciousness we will see. We will see. Uh, this is where the Transformer movie comes in with all the living robots. Correct. They are conscious. They feel the pain of humanity. They, they feel hatred against humanity. And they, so the consciousness is different. Intelligence is different. AGI is where what a human can do, the same thing AI system can do or surpass it in terms Okay. But, um, it still people are trying to figure it out whether it can do research. Can it replace the scientists who are doing pharmaceutical research? Okay? So this is where, uh, the AGI comes in. AGI means that you are basically creating a human brain which learns by its own. It keep learning by its own. It analyzes. It imagines. Um, it create a Leonardo Vin- Vinci in a machine model.