Ultimate AI Masterclass for Founders and Executives 💰| Vaibhav Sisinty | Indian Business podcast

0:00 Sam Altman in the beginning of the year kind of said that AI agents will enter the workforce this year. This year I expect that in 2025 we will have systems that people look at like what agents are the thing everyone is talking about. I think for good reason. Open AAI just announcing an improved model their AI model. They say it has better reasoning. Open AI has got the wind at its back. I think India should be doing everything. I think India should be one of the leaders of the AI revolution. So I can do this without knowing anything about coding. I don't know how to code but I know how to build a product. I managed to kind of raise $5 million. Hi folks, my name is Rebendi the founder Gone School Mr. Webhub. Growth school is where we make you become top 1% professional/founders. And today where

0:45 does growth school stand? I think we've grown 10x the company that started as four five people right now has 150 plus people. Wow. Last month we had learners from 45 different countries. 45 different countries. We are the smallest of all the startups. We also have the least amount of funding of all the startups. A LinkedIn top startup family of 300,000 learners. Amazon has already laid off 14,000 people. Google has already laid off people. And we're seeing this all across the tech industry. As a working professional, what exactly am I supposed to do? Microsoft has laid off people across multiple divisions. 150,000 tech workers have been laid off. of Meta, Amazon and others have frozen hiring Amazon today announcing it's going to cut 18,000 workers in this for every

1:32 problem right there are tools for it one great place for you to find tools for your problem is there's a website called as if a working professional is watching this episode right now what are the first three steps that they're supposed to take in order to make themselves super efficient I believe that everybody has to become something called as an AI journalist And there are just so many people I feel so sorry, so sad for them that they are so stupid that they think AI is all about chubby. If you have read the news, Donald Trump is making it impossible for a lot of countries to buy GPUs. Trump and his team are planning to expand efforts to limit China's tech advancements in pressuring allies to also get in line. We are 6 months to one year away from almost all the code to be

2:19 written by AI. All boss the future programming language is not Python is not Rust is not JavaScript [Music] it's podcast is brought to you by ODO people we've been talking about how AI is revolutionizing the way we work making the average version 10x more productive. But here's the truth. AI is only as powerful as the systems that we plug into. You can't just slap automation onto chaos and expect magic to happen. Because if your stock counts don't match, your invoices are delayed, or your team is jumping between tools

3:04 just to process a simple order, even the smartest tech cannot help you scale. But that's where ODO is here to help you out. ODO is an all-in-one business management software that brings every part of your business together. ODOU offers 45 easy-touse applications like sales, accounting, inventory, HR, project management, all of them into one clean centralized platform. ODU helps you manage your business by keeping all your data and operations in one place. All apps work together smoothly so that things like sales, accounting, and inventory are always in sync, giving you a clear picture about your business in real time. It even comes with built-in AI tools like Chat GPT to help you write website content. It also includes AIP powered themes that adapt to your business goals and smart scanning that

3:51 can automatically extract data from physical invoices into the accounting app. The interface is very very simple and it's very easy to use making it very very quick to learn. And the best part is ODO actually grows with your business so you can add more apps as you need. It also automates your tasks which will save you a ton of time on routine work. And cherry on the cake is that the first application is free for life. So click the link in the description and get started with ODO today. Hi Weber, welcome to the Indian business podcast. It's been a long time since you wanted to have you and finally you're here and uh guys just to give you context I have a confession to make. About a few days back I was having a conversation with web and I told him that yeah bro we use a lot of AI. So he's like yeah that's amazing. And then he started talking and then I realized that um I am an AI

4:38 illiterate. I don't know how to use AI well and I'm an amateur. Weber, you stand at a very very interesting juncture. On one side, you are at the cutting edge of education. On the other side, you are at the cutting edge of technology. You teach AI now just like you used to teach LinkedIn and you are super bullish about it. And like I said, I've been speaking to you and it has been life-changing for me. I don't use these terms very loosely, but I can say that whatever you've told me about AI and about systems and processes, that has been life-changing. And the beauty is I spoke to you after I spoke to Ranir and Reanir said, "Bro, content is not your mode. systems and processes are your mode. And then while I was trying to build those systems and processes, I spoke to you and then it was life-changing because I learned how to build those systems and processes at 10x speed with 100x more

5:26 efficiency. But now I want to understand how does it apply to the working professionals of the world today? Because the US recession is real. We're expecting recession in the next 3 to four months. There is already panic in the market and now everybody's supposed to rethink their position in their own companies. Amazon has already laid off 14,000 people. Google has already laid off people and we're seeing this all across the tech industry. And we're expecting that once US sees a recession, even the Indian tech industry is going to see a recession. So my question is as a working professional, what exactly am I supposed to do? Let's break it down, right? There are multiple aspects to this. One is what happens when a recession arrives? The spending power goes down. When spending power of a consumer goes

6:12 down, what are the biggest companies out there in the US? They're either tech companies or consumer companies, right? Essentially, uh, deep tech companies or consumer tech companies, right? And they all rely on some level of spending from the consumers. When the consumers goes down, when the consumer spending comes down, the appetite for a company to spend on their people directly and indirectly goes down. When that happens, that leads to layoffs. Right? Now, in this case, right, we are in this very interesting time when there's a recession because of multiple geopolitical aspects which I don't want to get into. While that recession is happening, it becomes a very good reason for people for the companies to adopt a process which is super future-paced and

6:58 super efficient. What I'll tell you what I mean by that. There is this massive rise of AI agents that is happening. Uh I don't know if you've seen this. Sam Alman in the beginning of the year kind of said that AI agents will enter the workforce this year. This year what does he even mean by that? If you break that down, right? Essentially every one of us have used chat GPT Gemini and all those tools right and we think we know all about AI just like me right the problem is that there is a gap in understanding what AI is capable of what you're seeing as an assistant where you ask a question and you get an answer and that is where it stops as a result you like oh this is cool I asked for a poem for my dog and it gave a poem cool right AI is cool but

7:44 the real power of AI today as we live is that it has the ability to complete a full task on your behalf. You were just talking about operator. What does an operator do? An operator can use computer on your behalf. For example, a few days back, right? I wanted to get a bunch of ideas on companies that I could probably work on and the best place for me to get these ideas was Y Combinator. So essentially the problem statement was I had to go through 5,000 companies data uh and I had to read each one of them check out each one of their websites to see if there is something that I can build like this in India with my expertise that's a lot of context things that I'm good at the audience that I have the ideas that Y cominator has bagged and can I do it

8:32 in India right a lot of context had to do it manually chat GPT or tools like this cannot do it out of the box because it cannot pull in so much context cannot do things and I want all of this data on a Google sheet to be filled. Operator is a great use case for this. So I went to chat GP operator gave a very nice prompt uh explaining what I exactly want. Left it 9 hours later came back with around I think 25 26 companies that I should look at out of 5,000. Wow. After 9 hours. Yeah. I mean it took I don't remember. Overnight I left it. Okay. Right. Only place where I had to be part of it. It could be done in like 50 minutes or so. I don't know. Overnight I left it and passed out. Only place I had to intervene was I had to log into my Gmail account. It tells you I don't have access to your Gmail. Please log in. So

9:18 I logged into my Gmail account so it can create a Google sheet and that's all I did. I I left my computer passed out. I got 25 startup ideas because it was simple task. AI was able to do it or operator was able to do it. What does operator do? This if I if I had if I wanted a human to do it, right, it's a easy 10day job. H right and it needs a human to do it. But now AI is AI agent is able to this is what an AI agent is. A assistant cannot do this. Assistant can you can ask a question to the assistant it can give you an answer. It cannot accomplish a task but an AI agent can. This is an AI these are the kind of AI agents. For example chat GPD has something called as deep research. What does deep research do? Just like a human being if you had to research on a topic

10:03 what would you do? go out there, read hundreds of articles on the internet, make notes of this and all of that, right? Like it takes a long process. Deep research does exactly that. It takes your command, breaks it down, goes and reads hundreds of articles on your behalf and comes back and comes back and gives you an output. That's an AI agent. Agents like this are going to enter our workforce. And these AI agents, right, are incredibly powerful already. and they're going to only get better, right? And when these AI agents come in, companies have to find reasons to get rid of human beings and get these AI agents to work on their behalf because they can save so much money. Agents can work 24/7. That is what AI operators can do in 24 hours. In a few hours, I would

10:51 not say 24 hours. I don't remember the exact number, but it accomplished it. I don't care how long it took. If you get a human being to do the same thing, it's going to take five, six hours. But the crazy part of this is I was able to replace a human who would be like a you know entry level or like two three years experience basic understanding of startups person to do this job. Instead I could get an AI agent to do it overnight when I'm sleeping right this is the part that people have to realize and this is what Sam Alman meant that AI agents are entering the workforce this year. There are going to be a lot of agents like this like deep research is a great agent which does research like a human being. It when you give a question to it when you give a prompt to it saying I want you to accomplish this it

11:36 breaks down the task thinks like a human just like we do using something called as a reasoning model uh and I don't want to geek out that there's chain of thought reasoning that kicks in here. It breaks down the question, thinks like a human being, researches hundreds of articles on your behalf and finally gives you a report in let's say 50 minutes, 60 minutes, sometimes a couple of hours, right? But that report would take you weeks if not days for you to do as a regular person. Correct? The lot of agents like this which are entering. For example, there are business development agents which can do cold emailing on your behalf. All autonomous. find email ids, send people email id, reply to people's email, everything it can do out of box. Can you give me the name of these tools? So there is one tool that you can use

12:22 called as artician AI which is for SDRs or sales representatives. One more tool where you can build a few agents like this is Lindy AI which is very very no code. Of course if you want to go a little geeky then you can go to N8N N8N or crew AI. What do you mean geeky in a sense if you know how to write a little bit of code or not even know how to write code but if you're comfortable copy pasting code right because a lot of people look at code and like I can't do this boss I don't know how to write a code I don't I don't know how to write a single line of code I know how to control C andrl V do that enough and what will help you chat GP will so the moral of the story is that all these working professionals are supposed to learn how to use operator no the moral of the story with the recession point that we're getting to

13:07 right companies have window to replace humans with technology which is AI agents and that is fundamentally why you will see a lot of these companies like Google and all right are kicking out middle managers senior managers because they are the most expensive human cost that a company can incur right and they have to get rid of it and replace that with as much of AI processes as possible with recession coming in this is going to compound because there's a reason for a company to save money they also have a need to save money and they also O have a potential solution. The companies that are going to do very well even during recessions are probably the air companies like open AIs of the world, Google's of the world while the rest of them are going to adopt it as fast as possible and hence the recession. In

13:53 fact, this second line of thought right uh I'll come to the India aspect of it which you asked me. I think India could be in a very soupy situation. Could be or could not be. Uh there are two variations to it. Again these are perspectives right? I don't know if you have come across this because you're talking about US, right? Yeah. If you have read the news, Donald Trump is making it impossible for a lot of countries to buy GPUs from Nvidia, right? And in fact, for China, he has completely banned it. There are a lot of other countries have banned it as well. India comes in the level two where he allows us to buy some GPUs but not as many as we want. And then level one can buy how many ever they want. Why do you think is the reason? The primary reason here is Donald Trump doesn't want these countries to build AI capabilities as

14:39 good as the US capabilities. Now he'll be like yeah chat GPD to what is the problem? Let other people build. Why is Donald Trump being so like uh you know trying to hold all these GPUs and by the way FYI GPUs are the engine for these AI models to get trained on. So without GPUs there is no AI right? So that is the hardware part of it. Why is he doing this? Because I think he has seen something that most of us have not. You see human capital the reason why US or countries expensive countries have to rely on countries like China, India and everywhere else is because the cost of human labor is so high there that making an iPhone there end to end would cost the would cost basically 2x the cost of iPhone would 2x. So it won't make sense.

15:26 As a result they pass it on to countries where human labor is cheaper. Now with AI coming into the picture, there is intelligence coming in. Machines are becoming more intelligent than they ever were. Now with this technology, whoever controls this technology can control the whole supply chain to a point where you can build autonomous factories. Boss, there is real estate in terms of space to set up factories. You can set up a factory in desert also and there's a lot of desert. They can dig up a hole and put it underground also. That space is not a problem. The problem is people are going to work in those factory. But if there is a technology that is existent or that could be built where humans won't be needed to build a factory or to run a factory then a country like us doesn't

16:12 have to depend on a country like China to build their products. Correct. Because the biggest arbitrage is the cost arbitrage. Yes. And that is controlled with AI. So whoever builds that technology first will control the power. So do you think Trump is trying to consolidate all the manufacturing powerhouse within the United States through AI? I think so. I I that is how it looks like because I'm sure uh he has seen the dark factory videos. Yeah, bro. Those dark factory videos are just insane. Like I saw this video where there was barely any light. Yes. And they're like why are there no lights in these factories? Because there are no humans. There is no need of light. because there are no humans and bro it was mad like it was mindblowing now in

17:00 fact Elon Musk even says that the cost of producing a model S is 65% cheaper in Shanghai than in the US which is why even though Trump has imposed tariffs what he's essentially doing is punishing the consumers of America because he doesn't want them to buy Chinese goods problem that if there's a $1,000 iPhone It's not an uncle in Beijing who's going to pay another $100 for it. It is Karen from New York who's going to pay another $100. And because of this, the US companies are helpless. They cannot do anything because of two reasons. A, even if Chinese phones cost $1,100, they can't do anything about it

17:46 because their products cost on the upwards of $1,500. So, they can't do anything about it. Number two is that they still have to buy raw materials from China and China in return has also imposed tariffs because of which raw materials have become expensive. So the American products have become even more expensive. So tell me something. I understand this whole vision of making America great again. But how does it make sense in the present day world? And more importantly, how does it make sense for India? What are we supposed to do in this drama between America and China? Yeah. uh the tariffs reciprocal tariffs have not gone live yet right across the board and like we were discussing China doesn't have a lot of tariffs yet I think it's a bait card that Trump is

18:31 using at this point of time right it looks it looks like a great leverage point to have conversations with a lot of lot of countries right but I don't know if this will actually go through because it's a net net destroyer of economy then uh uh advantage right that's what economists say based on what I Understand now where does India sit with this equation? What does India exports the most? Software services, pharmaceuticals, auto parts, jewels, jewels, jewelry is number one, diamond. Yeah, d everything around gold, jewelry is number one, pharmacy is number two, pharmaceuticals is number two. The third is IT. The biggest of that is IT. No, no, it's not. It's a third. It's not. It's the third largest actually. It's third largest. Yeah. In terms of exports globally, not

19:17 just to us. I'm talking about global. Okay. Okay. Okay. All right. Global exports not US exports. I don't know of US uh export numbers but it is major. Right. Why were countries relying on India to export it? Cost arbitrage. It always the whole world trades because of arbitrage. Correct. The cost arbitrage is that one engineer in the US will cost you 60 lakhs per year minimum. That same engineer in India will cost you eight lakhs. As a result, companies in India can charge the US companies five times, 40 lakhs and US companies will be more than happy to pay because they're still getting a 40% discount. Right? This was the arbitrage that existed that led to I don't know how many billion dollar industry uh it is today. TCS, Vipro,

20:05 Accenture, all of this, right? That led to the development of China and India. In fact, China became a superpower because of this cost arbitrage. Yeah. It's it's a cost arbitrage that led to everything. Now with just it, right? We are getting into soupy waters. I'll tell you what I mean by that. Again, I'll go back to the same topic of recession and AI agents, right? Have you heard of tools like cursor, wind surf, bolt? No. These are all AI agents which are specialized to write code. OpenAI's new model 03 is so freaking good with code that it can write code like as if it is the top 200 developer in the world. So if it goes into a competition in the world, it will be in it will be ranked top 200 and it's an AI model. That basically means every company can

20:52 have thousands of engineers who are the best engineers in the world. That is 03. That's an AI model. India makes money by writing code and shipping software. If a company in the US instead of relying on India or Indian labor or Indian engineers to build a product, if they can rely on an AI agent which is built on O3 which is one of the best coders in the world, would they need India anymore? Okay, you know what? Maybe they still need India because these AI agents don't operate all by themselves. There is human in the loop. But with an AI agent, you can write code 10 times faster. I'll tell you one interesting thing. In the last weekly business review with my engineering team, I asked a question. What percentage of code

21:39 today inside of growth school is written by you guys? That is you manually write it and it is and what percent is written by AI? The average answer was 65% of the code today is written by AI. And mind you when I say this, the 03 model, which is the holy grail model that I'm talking about, is not available for general public yet. We're using a 2-year old AI model. That's insane, bro. Right. Antropics uh which is which is uh which is a company that runs Claude Antropic. The founder recently said we are maybe 6 months to one year away from almost all the code to be written by AI. All all the code almost all the code to be written by AI as in we as in claude or the world. Okay bro I'm telling you 70%

22:27 60 to 70% of the code in my company today is written by AI. Why combinator did a survey where almost in a few companies 90% of their code is already written by AI and the models that all of us are using today is a two-year-old model. When a new AI model comes in and this agent is only going to get better at doing stuff, you need a human in the loop. Best case, if you think about factories, right, most of our factories had the ability to be autonomous already. Aeroplane can fly without a pilot. Yeah. But we still have a human because we believe in human in the loop. That is we somehow believe that if goes down, pardon my French, a human can save which has happened a lot of times and we've got saved a lot of times as well. The reason why we have humans in a

23:13 factory even though it's fully autonomous is because we wanted a human to turn on the switch, turn off the switch and kind of look at the machine to see that oh everything is going okay. Okay, fine. That's the job. We are getting to a point where we are believing the system to say I don't need a human in the loop in these technologies. But the human in the loop element is going to come in code. Right? Now one person can write code equivalent of 20 people at a quality of a best engineer. Correct? Would companies still rely in Indian companies to write code anymore? Would they need it? That's a big question. Now there's a debate that comes on the other. The answer could still possibly be yes, but then with very less workforce at what cost? H the cost of communication is more expensive than you getting it done. You know today if you want to build

23:58 products in fact I'll send you some links you can link them up right I have built full-fledged solutions I don't know how to write code in two three hours you tell me a problem today that you're facing while running things cool and if there's a solution and you can imagine that web I wish there was this tool that could do this you can build that in 3 hours today 3 hours 3 to 5 hours max a good strong MVP which can do 70% of the things the tools if someone wants to go and explore in your team or anyone are listening to this. There's a tool called as lovable.dev. So I can do this without knowing anything about coding. No. You do you know how to communicate? Yeah. Can you tell what you want? Yeah. That's all it matters. Boss, the future programming language is not Python, is not Rust, is not JavaScript, it's

24:45 English. So you should subscribe to communication masterclass guys. It is English and it is communication quite frankly because there is this word that you must have that has been thrown around all the places right prompting. Prompting prompting what the hell is prompting? Prompting. Prompting is all about articulating what you want in the best form so that AI can understand. What is the language of prompting? English. Got it. Bro, I've got three questions over here and I lay it down in front of you and then you can choose to answer them in whatever sequence. First question. Now I've understood that because of this trade war regardless of what happens one outcome is recession and even if recession doesn't hit the roof at least companies will look at cost cutting. Yes. This cost cutting will trigger an efficiency exercise and

25:31 this exercise will lay off a lot of people and it will push the companies to turn existing employees into super efficient employees. Yes. The instrument to do that is AI. Yes. Got it. Now this instrument of AI for different people are different things. So if a working professional is watching this episode right now, what are the first three steps that they're supposed to take in order to make themselves super efficient? I believe that everybody has to become something called as an AI generalist. Now what are these AI generalists, right? AI generalists are these people who can solve problems using the power of

26:18 AI. What do companies look for in people? They are always looking for people who can solve problems. Now with AI journalists, people will companies will look for people who can solve problems with AI. Why? Because it could be 10 times faster, 100 times more efficient. And if there's anybody sitting in a job today and you're realizing that boss, this is going to hit you like nothing else has ever done. you have to start leveling yourself up as an AI generalist to this right uh I have done this exercise internally with my team with with around 20 25 people over the course of last 9 months so there's a road map that I laid out for the team I'll try to lay that out lay that out for you so that people can kind of pick that up there is a level zero to this level zero is where you kind of define your perfect toolkit I'll explain

27:03 you what I mean by that a tool you have 10 things that you do at your work every day and when you Look at those 10 things, right? You would know that these are seven things that AI I can take use of AI right now to do it much faster, much better to I to solve those seven problems. There are hundreds of tools out there. There could be a chat GBD, there could be a Gemini, there could be a claude which are very general purpose tools to very specific tools. Let's say you do a lot of work on Excel. So there's a tool called as numerous just to understand this better. So the first step is for me to jot down all the tasks that I do. Yes. The second step is to identify merely through human judgment as to which of these tasks could possibly be done by AI. Yes. Then the third step is to go and find tools to do

27:49 each of these tasks. Correct. Correct. So when you go out to look for tools, right, the problem is because there's so many tools out there, you will fall into this pit of which tool do I use where there's always a new tool coming in. So the solution to that is to define your AI toolkit. Okay? Explore a lot of tools. figure out what works for you and hold them very close to you and go very deep into those. But when you say explore these tools, the problem is that the exploration never ends. And I've been into that uh pit where you know every day there is some of the other Instagrammer who'd come and say this AI tool does XY Z and I'm like bro and I'm not saying that they are bad tools. I'm saying that they are great tools but then all of them are great. For example, for a thumbnail designing tool, we looked at so many tools. We just happened to stick with one. But then the

28:35 other day, my editor came to me and said that you know this tool is better. But the only difference between both the tools is that I spent more time with this tool and less time with this tool. Similarly, he spent more time with the other tool than he did with my tool. That is exactly why I'm saying find a tool that you're comfortable with. Go deep. Okay. So rather than losing your focus and say that oh next big thing, next big thing, next big thing because the next big thing is not the next tool. But the problem over here is that GPD3 is not as good as GPD 4. A problem that let's say I choose a tool which is GP3 level and then tomorrow there's another tool that comes out which is GPD4 level. Mhm. But they both are on the same tools. Chad GPT you've got 3, four, 4.5. So you can understand but with these tools you don't know which models are smarter because A you're not a techie. B

29:21 with time the tool which is GPT3 level gets better but when you start with a GPT4 level tool it doesn't look that good. So how do you identify and how do you make sure that you always are at the cutting edge of whichever model is the best. I'll get to that. I will get to that. That's why we're at level zero. Okay. Level zero is for us to find a toolkit that works right now for us. Right now. right now right that solves an immediate problem that is I'm more efficient at work right away in this for every problem right there are tools for it one great place for you to find tools for your problem is there's a website called as there is an AI for that.com okay okay you go there and search let's say excel is your problem you spend a

30:06 lot of time on excel and you want AI to help you with excel you can just search for excel there'll be 100 tools there look at the one which has highest number of votes okay pick that tool up the other Great place for you to identify if the tool is good or not is to go to product hunt. Search for that tool. They must have launched it. Look at the review of those tools. What is product hunt? product.com is a platform where founders go to launch their products. Okay. It's a public forum. Let's say I built let's say I built a thumbnail designer, right? I will go and launch saying that hey I built a thumbnail designer. This is state-of-the-art. With that being said, there will be other thumbnail designers who have launched there as well and their consumers are also here. So people who have used other tools like other thumbnail designers would also rate this and they'll be like I saw this tool good but I found this tool to be better than that. So you'll get a highle picture to

30:53 for a starting point on which tool you want to start with. Let's say for Excel there's a tool called as numerous which is a co-pilot for Excel. Let's say you want to make presentations. A lot of people spend a lot of time designing presentations. You can use a tool like gamma which can generate out of box for research like we discussed there's chgd deep research there's grock research in fact there's a free tool for research which is Google deep research no tool comes closest to that but people don't know about it because really do you think Google deep research is better than chat deep research yeah I'll tell you why I'll tell you the logic of why is the backbone of deep research internet what search engine is chat GPD connected to Bing What search engine is Google uh Gemini connected to? Google. But Gemini is too

31:39 bad, bro. With the 2.0 model that has come in right now, it's phenomenal. Okay, check it out. Okay. Right. And a lot of people in deep research, right? You want every corners to be covered. Well, the quality of output in terms of how it has written the copy, you might like chat GPD better purely because of its text generation models and empathy and all those aspects of it. And 4.5 is very good when it comes to emotional intelligence. Gemini might not be as good. But what do you do research for? You want depth. Nothing can go as deep as Gemini because I have done this activity. same assignment or the same project on Chad GPD deep research on Grock deep search on Gemini uh deep research and on

32:25 perplexity deep research same thing the most extensive documentation I've got is the Google deep research and I'll tell you what chat GPD referred to 40 resources 40 sources Google referred to 350 sources now you might say that web hub we are if doesn't mean that high number of resources would mean high quality of content No, it does. It does. In research, it does because it's a intricacies that you have to capture. Correct. I want to understand an important point that I want you to establish. In fact, anonym is in some cases better than charge or gro which tool is best for what? So, there is no tool that is best for anything. I'll tell you why. Because I

33:10 I'll tell you this is the response that I was dreading. I I I I'll tell you why and I'll tell you how you can find your right answer. Also, the foundational problem that people make is with prompt. Okay? Your output is as good as your input. Right? A lot of people screw it there. Few machines are good with comprehending your question better. As a result, it expands to a better output. But if you give a high quality prompt to everything, that is when there's a level playing field. Okay? Did you understand? Understood. With that being said, as a consumer, what do I value more is what matters. For example, when it comes to writing style, Claude was hands down better than everything else. Was was till GPD 4.5 dropped. GPD 4.5 dropped a

33:57 few weeks back, a month maybe. And it is purely trained to do good writing. It has emotional uh intelligence. It has character. It understands what you mean, right? better than it understands sarcasm all those aspects human elements come into the picture as a result GBD 4.5 right now is the best writing tool but 4.5 is not free you need a paid version of chity for that so what is the best tool right now cla because you can use the free version of it got right same with deep research deep research also goes back into how good a prompt you can write when you write a very detailed prompt on exactly what you want there are two ways of writing a prompt right one is that you say hey do a research on US recession and tell me

34:44 give me a report that is the worst prompt that you can write instead you say I'm I'm making a YouTube video on US recession I want to cover all the sides the impact of US economy or the impact that it will have on Indian economy why is that happening is there a bias of AI is there a bias of war that is happening on the other side is there an impact of Trump was there something that Joe Biden did because of which this is happening. You expand your understanding on what you want to do, right? And then you put a prompt out. You're giving it more surface area to think and give you a better output. Got it? In these cases, I have seen deep research of Google doing better because it goes into those trajectories better. It also uses a

35:30 reasoning model which has a ability to think like a human and reads up and gives you a much more extensive document. Got it. So for research, Google deep research is better. For understanding it's free. Okay. And it's free. So for research Google deep research is better. For writing style understanding claude was better but now Chad GPD 4.5 is much better. And what about Grock? What does Grock specialize in? Grock's unique edge is that it's built uh by Elon Musk who owns Twitter. So no other platform like chat GPT or Antropic or Gemini has access to Twitter data. But how is that a good thing? Because Twitter is full of crap. But it also is the place where the most recent news kicks in first. But who is to say that that news is relevant? Perspective,

36:17 bro. Perspective. But how is that possible? Because bro, I know so many pages which just put out fake news just for living. No, no. There is community notes that have come in that handles it. So I'll tell you what I mean by that. Any platform, who who who are these AI assistants catering to? Humans, right? What is Twitter built on? UGC. What do we like to see? We like and a social network understands what we like to see better than anybody else does. So it understands a human being better than a non-social network platform. I think that is where Grock does well. It kind of somehow right manages to understand your request better and gives you an output that you could relate to. Now is that based on bias? Possible. No, because my argument

37:03 is that uh it makes sense if it is about just telling me what people like. So, Twitter is actually trained on all the propaganda, all the crap, all the hate. And when I ask Twitter a question, I'm expecting it to give me a accurate answer which is by definition supposed to not have crap, not have hate, not have bias and not have fake stuff, which is everything that Grock is trained on. So that is not how it works basically. It's not just the information that it's trained with. First of all, Grock is not only trained with Twitter data. So what happens in these large language models is there is unsupervised learning that happens, there's supervised learning that happens and then there is reinforcement learning that happens.

37:49 I'll explain you what it means by that. Every model could be slightly different but high level there is uh you basically pull out all the content on the internet that you can get hold to books movies everything possible every form of text possible and you just shove that content into the AI model saying read read you don't give context you don't tell what a cat is what a dog is you just say read read and understand what you're able to understand that is basically unsupervised learning Next stage once this is done then comes supervised learning where what you do here is you basically say you give some level of labeling some level of direction this is good this is not good this is right this is not right but highle labeling you say

38:35 basically uh it's like giving it books in some form of fashion right again this is called supervised learning where you're directionally telling what it is rather than just throwing data at it then comes reinforcement learning this is very important step and this is what makes an LLM an LLM Right here what happens is now the switch flips. Let's say you're an LLM. I was just giving you data so far. Unstructured data. Then I give you some structured data, right? In the most layman way possible. Then I'll tell you, you studied everything. Now I'll ask you questions. You give me answers. In reinforcement learning, what happens is I ask the AI a question, it will give me three answers or four answer, whatever that is. And as a human in the loop, there's a human in the loop here with

39:21 most of the other models. Uh the human kind of rates if which of these are the most accurate answers. This is reinforcement learning that happens. And on top of all of this, there is something called as parameters that kicks in. You basically the model okay these are the characteristics that you should have. It's like telling a human being be nice, be kind, this is not right, don't do this, this is right, do this. You those are in a way parameters that you give just because you're trained. Let's say you just because AI is trained with hundred very negative suicidal books. AI doesn't become suicidal because reinforcement learning kicks in where you kind of giving it feedback back saying this is right, this

40:06 is wrong. But in this case with Grock while everybody had finite amount of data, Grock I mean Twitter also has finite amount of information but it has more information than everybody else did which is still valuable because more input more quality input leads to a better model. The second aspect to it is the reason why it chooses to be haywire and chooses to say whatever it wants to say is because Elon Musk decided to have those parameters set like that. If charge wants to be like very rude which Grock is right now right it can be they have to just turn the parameter on saying no filter of being nice being kind being empathetic not not going to talk about

40:52 elections not going to these are the rules that were set to chat GPT these are the rules that are not set to Grock but the quality of output I think you have to give it to the engineers that are working on grock is that at the least amount of time they were able to pull through a model where the quality of output is so good that we're having this debate today. So, long story short, have a toolkit of all the AI tools. I'll just summarize this because it's been some time. Level zero. Level zero, find tasks, find AI for the tasks, research through which AI is the best and then zero down on it. Done about level one. Now, for you to unlock the next layer of win, you truly have to be a problem solver, right? For that, you have to learn a few things. Level one

41:37 and level two is where you learn. Now that you use the tools, it's like an app. You can use it, it's easier, you do all of that. This is where you start going deeper. In level one and level two, your fundamental job is to understand all the AI models out there, not the AI tools. AI models. AI models out there. Okay. Right. Like what how does Chad Gypty work? What is GP 4.5? Just like what we discussed. Yeah. What is 01? You're using Chad GP right now. You don't care which model you're using at level zero. topic, five right here. You have to understand where to use a reasoning model, where to use a non-reasoning model, where to use this agent, where to use a open-source model. If there is an open source model, how do I run that model? Things like

42:24 that. What are the models out there? I mean, essentially here you go deeper, right? Like, uh, it's a good question. I think the solution to this is being extremely curious. Instead of letting chat GP go at uh this saying that default I'm going to use that you start changing those options playing with it asking it different questions reading about it and that will give you direction there is no other way basically iteration iteration and testing like level one this still gives me an understanding of what is the difference between chubity 3 and chity 4 basic understanding answers for example in deep research it takes you um 10 15 minutes to extract the output but once you extract the output that output is just too good okay you can see all those

43:12 sources and quite evidently it is different from the output that charge gibrity 4 would give you right and that input is again different from what charge jubility 3 would give you but you mentioned understanding models what do you mean by that and how can a non- techy person like me do that let's say tomorrow you're meeting a hypothetically an investor who wants to invest in things school okay and you want to be prepared for that. How would you use chat GPD for this? I would first break down why exactly am I having this meeting in the first place? If I'm having this meeting just to understand what is the value of my company or let's say no you've decided that you want to raise capital. I want to raise capital. You want to raise $10 million from this investor. The investor has shown interest. This is your first conversation not a pitch first conversation with the investor and you have never spoken to an investor before

43:58 in your life. The first thing that I would do is research to understand what exactly do investors find value in by asking Chad What where would you ask Chad 4.5. Okay. Why 4.5? Because I know that it's the best. Okay. Got it. Huh? How do they seek value in a particular company? That's step number one. Then I'll get some parameters. Then I'll go and evaluate whether my company stands by those parameters or not. Mhm. Once I understand that then I'll feed all the data about my company to charge and ask what do you think then it would give me some output and then on the basis of that I would say if you are this investor what are the questions that you would ask me beautiful and then I would prepare for those questions then

44:44 I would ask Chad ji to be extremely rude and ask complex questions which I would possibly not know the answers how would you do all of this who would you talk to in this just chat jubety 4.5 and I've done this for a few of my podcast. I'll not take the name of the guest, but you get what I'm saying, right? I would just ask JB to ask me all sorts of debatable questions, complex questions that I would not understand, and then I would try to answer those questions properly. Eventually, I'll ask for judgment. Once that is done, then I'll get into the next step. For example, the investor would ask me, "What would you do with these $10 million?" Sure. The investor will try to bog me down by asking for a lesser valuation. Mhm. the investor will just u ego boost me into believing that uh

45:32 he's the best and I'm the best and we both should do business. So then I will look at all the evil tactics that the investors use. Then I will also look for the red flags in the term sheet. For example doesn't work like that because I'm assuming no assumption assume what if he says no I like it on go on bro 100 crores take it or leave it. Sure you got the term sheet. What would you do? Then I look for all the red flags in terms will look for it. I will ask Chad gi how does the term sheet look like? A B I'll feed sample term sheets and then I'll ask for red flags and then I'll I'll also ask for all the red flags that a term sheet could possibly have. For example, there's a term sheet make

46:17 criteria which which says that the investor can decide to sell the stake of the founder. Yes. After a certain point, if the founder doesn't achieve XY Z revenue, the investor has all right to just take up all the stake of the founder by themselves and they also have the right to fire the founder and deploy new management. So I know all of these things. So I will note down all the red flags. How will you find them in your term sheet? Uh I will feed the term sheet to CHBT. Okay. Now I will tell you a problem that comes. Your term sheet is 450 pages. We can still feed it to CHP. No, you can't. Why? It there is something called as context window. Okay. Uh and there's a cap of context window. If I'm not wrong, chat GPTs most of the models comes with 128k uh tokens. So it will not be able to take more than 40 pages of information. hypothetically

47:04 more than 40 pages it will say the file size is too large or the amount of text is too large and if you try to break it down and send small small chunks it will forget it will forget there is something called as context window of every all of these large language models right context window is something like memory AI has very short memory right it can only remember so much have you ever realized when you're talking to AI on chat GPT sometimes it says the window is getting too long start a new chat why does it say cuz it's going beyond the amount of things it can remember about this window. If you go beyond it, right, it'll stop forgetting the things that you have said in the top. Basically, chibi is like gajjini. Not gajjini. It can remember forever but only up to 120k tokens. Okay. Sorry guys, pardon my

47:49 non-technical language. So, it will not work. Okay. This is why level one. Now, I'll tell you if you knew if you have done level one, level two, what you would have done in level one and level two. Like I said, right, you need to know which model to use and how models work. If you knew how models work, you would have known that you could not have uploaded just the PDF as is because it would not work. If it's a 10-page PDF, it will work or else it will not work. If you would have done level one, level two properly, that is you've gone through that journey. you would have known you should not have gone to GPT 4.5 but you probably should have gone to 03 high or a 01 pro or one of the reasoning models or a deepseeek R1 or a Gemini deep thinking model why 4.5 the

48:35 reason why you need to know the models is GPT 4.5 is a text generation model that is like I said before this conversation the way these large language models work is that it predicts the next word. So it doesn't have the ability to think in a situation of yours which is very very hypothetical. It has to think about multiple use cases on every side. So it has to do something called as reasoning like a human being chat GP doesn't have the ability to think all of this a GBD 4.5 doesn't but a reasoning model does. Oh, when you input the text in a 03 or you must have seen 01 as a model which we never touch saying we don't

49:20 know what that is regular I'm saying purely on regular people right so it does reasoning so the input here should go into reasoning model rather than a regular model right as you talk about it I love the fact what you're doing here is that you said I will ask chat GP to ask me tough questions I do the same but I do it slightly differently you could be doing that as Okay, I'm very lazy to type and listen. So I turn on the voice mode. Correct. I speak for 2 three hours till it breaks and again I resume. Correct. It's a great way to kind of get a balance of what's happening. So there's a reasoning model that goes in thinks through everything, thinks through every equation and goes here and finally you are able to quiz yourself. You know I've done sample podcasts. I I did a sample podcast yesterday of you and me as well.

50:05 Aa is this better? Yeah. Thank god. Uh but Chad Gibb did ask me more tough technical questions which I didn't want you to ask but that's good right uh and then when it comes to if you knew how models work like I said you would know that you cannot just upload a PDF a lot of people actually upload PDF file and they say oh it worked it didn't work it just understood what the v what the file is about it doesn't know the intricacies of it when you upload a PDF because it doesn't remember for you to upload a large file you need to know a concept called as I don't want to go into it but rags rags break down index so that it can read and chat GP cannot do it out of the box. If your work had had a lot to do

50:52 with PDFs you would have known of a platform like Humata which is a rack platform where you can upload a PDF and start chatting with it or you could have just uploaded the PDF large PDF into a different rag system like notebook LM. Have you heard of notebook LM which is by Google? I've heard about it but I haven't used it. H Google notebook LM you can just upload the PDF. It is also a rack system in a way right the feature that notebook LM is very popular for is the podcast feature where people say you can upload any file AI will convert that into podcast and you can listen to it. That is what notebook LM got very viral for. But what it works very well as well for is that you can upload a lot of file aggregate the information inside of it and you can chat about the files that

51:38 you have just uploaded. So I can take up Walter Isaxon's book. Yes. Upload it or five different books of Walter Walter Isaxon and ask what are the common trends that you're seeing about the writer? What are the common traits that you've seen in Steve Jobs's Elon Musk? And it will not the tool the beauty of this tool is that it will not allow it to access information beyond the files that you have uploaded. Oh wow. That is actually very interesting because I tried doing this with the book Prince and I tried to understand if Machi were to give me an advice what would he say? Okay. And this is what I was speaking about floating with power and trying to understand how power consolidation happens. But what happened is that after a certain point it started to go out of the book. Yes. Now if I haven't read the book I wouldn't understand. You would not know. Yes. Yeah. So then that

52:25 becomes deeply problematic. Right. But then if there is a model which has restricted access only to the book and it has it has to do reasoning only within the book then it makes a lot more sense. Only to the files that you've uploaded. Got it. So that's a notebook element. So but then bro there is no end to this. Like you've given me so many nuances. So give me like a thumb rule on how do I understand these models because this is just too much information for me. There are not 100 models out there. No you have to understand what are your top use cases. Okay. And then you have there is finite amount of information here. Once you learn that finite amount of information, you will be hooked to make sure that you're always on top of new things that are coming. Got it. So what's the framework to understand this better? Uh in terms of which model is better for what is I think the only way to do that is by trying is by testing

53:10 things out is by playing with it. There is no better way of learning than actually doing it. But playing with it 4.5, you have to be more curious to understand what I'm doing. Is it the right thing? Long story short 3 4.0. Huh? But that is such an inefficient way to go about it. Why? Because I've got my day-to-day task. What am I supposed to No, you're not going to do this forever. No. For example, today if I want to have a task, see the first differentiation that I told you about is do you go to a non-reasoning model or do you go to a reasoning model once you know where to go when to go there are bunch of reasoning models everywhere you can choose to stick with GP only you don't

53:55 have to go beyond it but reasoning non-reasoning reasoning or non-reasoning so or else you'd be everybody you want to have an edge you don't want to put in the work. No, I'm saying I'm putting in the work, but I have to understand where to put in the work by spending time into understanding where to put in the work. You getting my point? You know the solution to that? That is exactly what GPT5 is going to be. It removes like this is openly spoken about right now when GPT5 comes right this is not your problem. This is everybody's problem. They're like we don't know what to use when you have given us 20 models. what do you use where? So, OpenAI is like you know what the GPD5 is going to be smart

54:42 model which has integration of all the models inside of it. So, you don't have to do the guesswork of which model to use where. I will understand your task and I will use the right model. So, the call to action is wait for GPI if you want to wait. But then this kind of works in everything that you're learning in AI. Everything that you do in life actually you have to be explorative for you to get an edge. You need to have more surface area of knowledge. Understood. Do you think it would be better for example, you ask me this question that uh what if Chad Gibbdy 4 doesn't take up 450 pages. Do you think a better alternative would be to just ask Chad GPD 4.5 as to what am I supposed to do? I want to upload 450 pages of document. Tell me which model to use. Probably but for a question like that, I would go to perplexity. Perplexity. The reason why I would go to

55:29 perplexity is because GPT 4.5 not tell me that perplexity is better. I think there will be biases there unless you call it out. I don't know. It's a good try. These are good experiments. See, this is this is the curious hat on, right? If it does say, then you're like, "Oh, it's a nice model." If it doesn't say, you ask, "Why didn't you answer perplexity?" Then it'll give you a reason why 4.5 Chad Gibb 4.5 explained the yield curve. We were trying to study US recession. It explained the yield curve wrong. Definition of yield curve wrong. the data the example it gave outdated or it gave the complete opposite example was the e curve recent the the data that you're looking for the no no it was just an example it was a conceptual

56:14 explanation that I was looking at and this is like a 50-y old concept and yet GP 4.5 messed up and we understood that while we were reading through the example and we were trying to understand how can GPD 4.5 be wrong so for about a minute or so we were like stunned that uh GPD 4.5 5 said something and because we've trusted the software so many times that we started to question ourselves. There's something called as hallucination. See all of these models all of this AI right basically has been trained from a lot of information on the internet. So it has a character of a human being but it's impossible that it gets wrong information on yield curve because all of the information available on the internet is not 100% accurate. And GBD 4.5 again I go back to the same point.

57:00 It is a text generator. It basically guesses the next word. If you want super factual information which could not go wrong and it's not creative writing, you would go to something which has access to realtime information which is Google deep reset. No, you can go to perplex. This is this doesn't need yield curve is a basic concept. It doesn't need one. In fact, a good hack if you want to be very sure of the information that is coming and it's not you didn't ask for it to write an email because the factory cannot be wrong in an email, right? It's about writing style. But let's say something like this in this case yield curve. I would just tap the button of search on on chat GPT. So before it gives you the answer, it quickly reads up a couple of articles before it gives

57:47 you the answer. Just this just just takes the quality of the output much higher. Level one is understanding the models better and here's where you need to be very good at problem solving. You said yeah I mean essentially level one idea of level one is that you understand what all models are out there which model is good for what and how to work with them the right way by learning how to do prompting rightly know the right models and then choose the right models now let's go to level two level two level two is essentially prompting advanced prompting I would want a framework on how to prompt such that it can give you right outputs and before Before I spoke to An I thought prompting is very basic. Okay. And uh what I found

58:32 deeply interesting is that me and my team were using the same chat jeopardy model and we had the same problem but somehow the answers that I got were so comprehensive that they were like exactly type. Okay. And I remember this very vividly because two months back I asked the same question to an type like how are you getting such comprehensive answers and he had a different framework but I want to understand from you what exactly is your framework because of which you're able to get such comprehensive and complex task done through the operator in just one attempt. Well, it is not just the operator, it's any model out there or any AI out there. The way I treat any of

59:19 the AI agent AI uh assistants out there is how I would how I would treat a intern who has just joined today. Okay. If I give a task to the intern, I would want to be as freaking specific as possible. There are two ways. Let's say I'm hungry. Okay? And I have an intern. You should not get your intern to do all of this. It's just an example. Uh let's say I'm hungry and this intern has joined in today and I'm like, "Hey, I'm hungry. Can you get me food? The intern would be like, "Oh, Weber wants food. Okay, let me go down. Oh, there's a Burger King right here. I like Burger King. I'm sure Weber will like it as well, and burger is closer. I'll buy it and give it to you." And he gets me a burger. I'm like, "Boss, why did you get me a burger?" I go back to the intern

1:00:05 and say, "Why did you get me a burger?" The intern like, "You said you're hungry. I got you a burger. What's the problem?" It's like, "Oh, man. I'm dieting right now. I can't eat a burger." Instead, if I would have told the intern, listen, I'm very hungry right now. I have a lot of work to do. So, I want to eat something which is very quick. With that being said, I'm currently on a diet. Uh, and I've already had a lot of carbs today. So, I want something which is very, very high on protein and high on fiber. So, figure out something that we can get really quickly in the next 10, 15 minutes so I can have this food. Now that I have been a little more and I'm a non-vegetarian, now that I've been a little more specific, the intern will still find a Burger King only, which is closest because I wanted food in the next 15 minutes, but he'll probably order a salad at Burger King with a

1:00:52 chicken breast rather than a burger for me, which suits my needs. So, it's all about input and output. So, the more expressive you are to an AI, the better it gets. So, the way I treat it, what do you call this context? It's not just context. It's a great point. Context is extremely important. I give more things to it. Okay, I call this uh a basic

…