Myntra Founder Is This The End Of Software Engineers Sparx

TITLE: Myntra Founder: Is This the End of Software Engineers? | SparX CHANNEL: SparX by Mukesh Bansal DATE: 2026-04-18 ---TRANSCRIPT--- I’ve always felt like from my early days of career that meetings are the curse of corporate life. But in the past, engineer used to be the most expensive thing. You need just few of them. So, what do you need the money for? What happened to companies with thousands of thousand employees? I think the world has completely changed. We are no longer in the world we were 3 months ago. At least 90% code is right now AI generated. If you have a compelling idea, just build the whole damn product in a week. It wouldn’t be surprising when it is in a day. Today’s guest is Piyush Ranjan, former CTO at Flipkart and VP of engineering at Google, now building the future of AI-powered education with Furmi AI. If I’m an employee, I want to be highly desirable hire. What should I be? You know, most people who are in the knowledge work business, especially people in the management team, product manager, engineers, they are going to get completely wiped out. Even the non-engineers in the team, that everybody, you have to go and start using AI in a daily basis. Everybody’s going to be AI managers. All our jobs will be how do you manage this resource given to you? If innovation becomes a commodity, then distribution becomes a precious thing. So, as [snorts] a startup, if you don’t have distribution, that’s one of the biggest challenge.

[music] Everybody is worried that using AI causes cognitive death. Building software, if you are into AI, is almost like an addiction. Memory is another big problem. If you build any of these agents, you’ll quickly realize that the memory is not You know, sometimes we like that. The AI will get to a point where it can automatically improve itself. And when it can do that, then it’s a runaway effect. There is no stopping. So, from normal AI to super intelligence, you know, it’s just going to happen in the blink of an eye, literally. Hi Piyush, welcome back to Sparks. Thank you. Thank you. Since we last talked, last 6 months and Sparks we’re mostly focused on deep tech topics. Deep tech is starting to become so much important. Most of deep tech is AI. And I’m really looking forward to hearing your conversation. I know you’ve been building all kind of amazing stuff. So, really looking forward to this conversation. Yeah, yeah, no. I have been watching all the episodes, every one of them, including the one which came out yesterday. It’s like fascinating what is happening in AI with all the different perspectives which Sparks has brought. It’s a lot of fun. See, you and I have worked together for now what, you know, 12 years? But let’s start with you have this really cool chief of staff, Enrico, who’s making more waves than you these days. And I don’t know, maybe I should have invited Enrico. So, what’s the deal with Enrico? Is it uh real? Oh, yeah, yeah. Enrico for sure is real. I think Enrico is the most productive AI employee in our company, Furmi, uh right now. Productive AI employee or productive employee period? Uh productive Well, I I didn’t want to put the other employees down, but truly it is actually one of the most productive employees and and the most knowledgeable, actually. So, yeah, it’s it’s fascinating. I think What is Enrico? How did you build it? So, what happened was uh you know, as as you know, when I was at Google working on Google Assistant uh and then we were thinking like, okay, well, what are the various things one could do. And after I left, I was thinking that there is assistant for work would be a big thing. Uh even drew up the architecture around it and uh talked to a bunch of friends uh and it was clear the need was there. And then when we started working on different startups, the time was never right. And [clears throat] then I had to take a 2-week break uh when I was on a cruise and I was like, thanks to Starlink, it was there. I started building that architecture which was there. But you know, these days, like building software, if if you are into AI, is almost like an addiction. Yeah. Right? Every builder I know is addicted. They use actually the term addicted. is a wrong word. It is an addiction. I agree with you. is And you know, like I I think 3 months ago or so, I had I had said on LinkedIn that if you’re a builder and you’re not feeling addicted, you’re doing something wrong. And if you feel addicted, you know what I’m talking about. And I have seen like you have also seen this, right? Like everybody’s I ignored it for 2 months, but I finally succumbed to it 1 month ago. And it’s mad right now. Yeah, it is it is fascinating. Just yesterday I was talking to a person who has been a long builder in Microsoft forever. And I was telling him that you should get into hands-on building things. I just showed him for like half an hour. And then we were at dinner table and he’s on his phone. And I’m like, what’s going on? He’s running cloud code on his phone. [laughter] I’m like, what’s going on? Anyway, so the thing was I was there for 2 weeks. Then I started building out uh this uh because, you know, in our startup we don’t have admins. Like everybody manages their own thing. But the workload is increasing regularly. So, I started building out this chief of staff. Open claw had come out. But it was doing certain things. But there’s a lot more which I wanted to do and there was a whole architecture in mind. So, ended up building that. And the productivity within the company, my own productivity, of other people who need things from me, the way we work together as a team, has suddenly gone to another level as a result. So, that is what it is. And you’re continuing to build, upgrade Enrico. Every day. Every day. It’s actually every day there’s a new thing which we can think of and we can kind of put in there. It’s all about uh you know, the gap between and and this is what I will say, that there used to be a world where if you thought about something, there used to be so much more you have to do before you could build it into reality. Today, the gap between imagination and instantiation has shrunk so much that every day when I think of something, that, oh, wouldn’t it be great if Enrico did this? It’s like Yeah. 10 minutes later. I want to share something with you. You know, initially, I was obviously hearing about all these um coding platform getting so good and you were sharing what you were doing. But I was just kind of watching. I was reading, but not doing anything. But a month ago, just inspired by you, I started building my own AI assistant. And the first day was rudimentary version. Next day was better. By third day, it was doing real stuff. I’m like, wow, this is you know, incredible. Then I started building another project. And then So, first 2 weeks were about learning. But in 2 weeks, I got a pretty good hang of what we can do with this. And then about 10 days ago, I started building a product for capacity, which is a new incubation we are doing. And I just thought, let me see how far it can I can take. And it has kind of gone exponential curve. I I’ll show you after this once we are done. But basically, I think I have built the V1. I have done uh re um re-architected twice for scalability, etc. Built multiple skins. And it is fully functioning and I’m talking about team about launching with that. Now, contrast is the fact that, you know, last year, I wanted to build capacity, but we didn’t have team. So, I hired an outsource agency. I paid them $100,000, if you can believe. And it failed. So, that $100,000 write-off. Then we started building our team. Now, team is slowly coming together. In between, I also talked to Neurix if they can build. Again, I was ready to offer them between 100 to 200k. Now, the truth is I think I’ll build a better product than that in less than 10 days. About 4, 5,000 dollars of credit burned. And so much learning in the process. I think the world has completely changed. It was unthinkable and like I’m obviously preaching it out every single day, but I don’t think people still realizing. It’s probably a matter of months or quarters, but it just hits people in the face. We are no longer in the world we were 3 months ago. That’s actually fascinating. I mean, you know, I would love to uh really see more of what you were saying. There are a couple of trends which we are seeing. If you remember in in November when we did the Furmi offsite and we were telling like even the non-engineers in the team, that everybody, you have to go and start using AI in a daily basis. We are actually seeing that thing, that the people who have curiosity and agency, like the example which you gave, right? You’re curious about, okay, what can I do with this? And you had the agency, you just went and did it. Once you put those two together, the people who have the in the classic world, the expertise, they are now no longer, unless they attach curiosity and agency to that, they are no longer anywhere close to, you know, the kind of work which can be done in in work now. deep realization in the past, see, organizations are great, but they’re also very complex. There was so much human dynamics at play. And I’ve definitely been a builder all my life. But to build anything, I needed an organization. Recruit people, take care of their motivation, people dynamics, which is I still not been able to solve, you know, that always first goes wrong before it will go right. So, you need this whole org infrastructure to do things. But now you can take it out of the way. As you’re saying, if you have curiosity, agency, you know what you want to do, you can just go very far, you know, it is I’m still not fully grasped it, but I think the you know, the at least software engineering world or the building world is going to completely get upended. I think there’s enough evidence and I was reading this recent uh this thing, a story which is getting a lot of publicity. There’s New York-based startup that in GLP business. Mhm. There’s two guys. They built a $1.8 billion sales business. Yeah. And all with mostly AI work, some contractors. Yeah, actually the the revenue per employee, which is what how even high-quality classic companies like Google and Facebook used to measure themselves with, which are very profitable companies. Those numbers look little Mhm. compared to what the current, you know, the vertical AI native companies are are trying to do. I think that they’re like at least three to five x more. Instead of 1 million per revenue per employee, it’s like uh early-stage companies like 3.5 5 million uh you know, per employee kind of setup. And And the thing which going back to the people thing, I think the way old companies were set up, like I you know, Fermi is a small company, right? There are like only 10 12 of us. But the amount of work we are able to do is like five x more. But the beauty also is that the kind of people who are doing that and the way the team is set up is not how a five x larger company would have been set up. There is no org structure. Actually, Jack Dorsey was talking about this uh earlier that he talked about why companies are set up this way in hierarchical manner as they become bigger is for information flow. Now, with AI what is happening is information is like in Enrico example, right? The Enrico is there, has all the information in the company. It’s accessible to everybody. So, the way we have set it up is there is no hierarchy, but instead there are problem-solving loops. And everybody who is, you know, motivated about that problem and is self-driven is sitting in there and there are AI tools available and they’re solving that problem in that loop. And that’s a very different way of doing things than You mentioned revenue per employee. Mhm. And if you look at entire IT services industry in India Mhm. uh I think uh let’s say company like TCS, they probably 700 800,000 employees and um close to 30 billion dollars revenue in that range. So, that translate into about $40,000 $50,000 per Mhm. employee revenue. But from what you’re saying, that can be easily with 10 x or 100 x. 100 x. I think so. That’s just I mean, absolute game-changing not only for these companies but also for Indian IT service industry. Well, that’s a big question mark looming over. We have 6 million IT service employees generating the revenue of 20 30,000 per employee. Yeah, I mean, I think first of all if if it is not done by the incumbents, it’ll be done by somebody else with disruption. Being Yeah, being done by somebody else already. So, the thing is it is going to happen because that is lot more efficient. But then it also but and I have a very different point of view than many people who have said this. There’s a different question. What happens to the 70,000 employees or what happens to people whose jobs are, you know, being taken away? And I actually feel that this unlocks the whole uh opportunity for those 70,000 people to go do things. Now, the thing is it’ll definitely sift the I think in our lot last podcast we had talked about that there’ll be people who will like, “Oh, this is not working for me.” And there’ll be like, “How do I utilize this?” And the people who utilize this, like just like we were talking about the examples where we felt like, “Oh, how do you use this?” And suddenly got so much more out. Those 70,000 Maybe 7,000 of them will create 7,000 startups. Right? Things like these will happen. Maybe the uh you know, the these kind of software companies like Infosyses and Wipros of tomorrow will be built by some of those 70,000 80,000 people. Yeah, one of the things I was thinking in uh you know, you know, in Miraculous we are incubating these companies and um also invest uh in the past year I had a certain mental model that three to six months for thesis building and the three months for prototyping and so on. I think all the out of the window. Now, if you have a compelling idea, just build a whole damn product in a week. Mhm. And if the product is looking good, then have a next discussion whether we should build like it’s almost like build it first and then see whether we want to do anything or discard or open source, right? And that’s a Yeah. And that’s all the pitch will become if or now I don’t think I’ll take any pitch like I said, “Do not show me PowerPoint. Build a product. Let me play with the product and then see what you want to do with this.” Actually, it’s so funny you said that in a week, I feel like it won’t be surprising when it is in a day very soon because if you can think through like in the fog this far software can come and meet you there. I’ll I’ll tell you my journey from November 2024 to what I saw just two days ago, what has happened. And that last thing I haven’t used yet, but I told the Fermi team that, “Look, so November 2024 I suddenly started to buy it by the time and actually all of this I had posted on LinkedIn also just to share like the point of view. But it was like I am no longer writing code. Everybody was using cursor and all that, but it was like I’m using cursor just to specify and to test. So, I was like, “Oh, code review and testing is the only job, right?” Then obviously Andrej Karpathy then talked about why coding a month later. Then by uh a little bit later, I think by October when Sonnet 4.6 came out I was like, “I am not opening the IDE also.” Yeah. I’m just only inside the Claude code terminal or something. IDE is also out of picture. In that I was looking at the diffs going through. So, I was like watching, “Okay, these are the diffs and okay, this is what has happened.” But I’m talking to this person as a pair programmer observing their code. And that was the situation where we we did the offsite, right? It was like, “Okay, everybody go use it.” Then the the most That is the time when I said like, “Oh, everybody’s going to be AI managers.” Which I think actually will happen. Like all our jobs will be how do you manage this resource given to you, right? As opposed to human resource. Now, you are a AI manager. The better you are, the more you can actually make them productive and all that. So, I felt like that is going on. Then it shifted to uh we were doing this thing we called the agent tango, but now it’s like, you know, not even look at the edits. Claude, you edit, another model critique. So, builder critique kind of a dance. And we did that a lot. Just yesterday and I want to try this out. I found that there is actually now open source which has come out which actually is a harness where you can have the AI come in and you will have emails and all that and it will try like 10,000 different prompts test all of them and figure out which one performs better according to the email and then kind of put that there and now suddenly your AI whatever the uh you know, the application you had built is suddenly better all because it actually built itself to the loop. Yeah. What you need is eventually just more data from the real world coming so that you can keep refining yourself. I think that will be the input which will allow this cycle to keep getting better. Yeah. Yeah, no, I think I’m just thinking that um you know, most people who are in the knowledge work business, especially people in the management team, product manager, engineers if they are today not kind of calling time out on whatever they are doing Mhm. and absolutely committing to reinventing themselves, they are going to get completely wiped out. It’s just a it’s you know, you mentioned about managing the powerful resources. That’s also so counterintuitive. Even as a manager we would have seen we’re all used to managing people of certain capability plus 50% minus 50%. Mhm. Suddenly you have a star in your team who is five x better. Mhm. Most people struggle because the star guy needs also more autonomy, has more independent, he can override your thought process. Some people are very good at it. That’s why this management adage that I want to work with people smarter than me. Mhm. Easier said than done. Most people don’t want to work with people smarter than them. Now, this thing is going to be I don’t know, million times smarter, billion times smarter. Mhm. So, how do you use this resource which requires very different kind of thinking. So, what you’re calling AI manager Mhm. I think it’s a brand new skill. Like very little from the past extrapolates into that. Cuz you know It’s a very different way you work with an infinite resource. It’s like I’m continuously finding people are not asking enough questions. They’re not still expecting that they’re a limited expectations is what they’re asking this thing to do. Whereas it’s totally unshackling that I want the whole world right now and I’m going to specify in a structured manner. Yeah, actually if you look at that, the the next second-order effect of that is that the companies which hire the right kind of people who can do that will succeed so much more. Because what is a classic classic companies’ interviews where do you know know this expertise? Like come do it on the whiteboard or walk me through cases where you have handled this, that or like we will give you a some sort of an intelligence smartness kind of test or proxy. But now if you look at that, if you are given a resource which is much more intelligent than you are going to be, how much your brain can hold compared to what this can do what are those people who you want to bring into the team now? Which is very different from how we used to interview before. And I think when when intelligence is utility what is really at premium is the question, right? And that’s where I bring the curiosity and agency. Like we if you do a new brand new company you want people who actually are willing to dig deep and be curious about like how to utilize this the best and also have a sense of agency. They’re not waiting for somebody to give direction because you have this super horsepower, you know, machine available to you. So, it’s a very different hiring also. So, I think we have made the point about that this this you know, kind of nearly infinite resource that’s available to all of us, how we use based on our mindset, how we are retraining ourselves and so on. Yeah. But I’ll I’ll come back to that later. I want to a little bit go inside the engine as well. See, the things I’m building right now, biggest challenge is happening is you know, I’m not sitting in front of my desktop for 24 hours. So, at some point I will step away. And whatever work I give it, it finishes sometime 30 minutes, 60 minutes. So so much idle time. If I can think through and anticipate and have the feedback ready for agents. Then I mean I don’t think I have come anywhere close to what they can build in 24 hours. It’s my inability to plan and give it that much work. Actually I’ll I’ll share a funny anecdote because we talked about Enrico earlier. So after I was using this Chief of Staff for a while, I wanted to build more into it. But I didn’t know what. And the same problem that walking away from the desktop. asked Enrico to build itself. No. So I started Claude Code and gave it the credentials to talk to Enrico. I said, “This is your user. Go conduct an interview to understand what his needs are.” Then write a PRD. Yeah. And then go implement that. Then ask Codex to validate it. And then ship it over there. And then after like I think 2 hours, Enrico had suddenly new capabilities which it had wanted. Look, I think this this feedback loop in AI I think again, you know, highly underappreciated. If you go back the whole original thesis around singularity which Ray Kurzweil popularized from 2005 onwards, his whole point was at some point and he actually, you know, predicted 2029 that AI will get to a point where it can automatically improve itself. And when it can do that, then it’s a runaway effect. There is no stopping. So from normal AI to superintelligence to God knows what, you know, it’s just going to happen in a blink of an eye literally. And now I’m also hearing that all these models with the Claude Code or OpenAI, at least 90% code is right now AI generated. And it’s probably a matter of time which is going to be 90 and 100% with just humans just basically setting up those feedback loops and just watching the show. What has changed? Why are these models able to write such complicated code, almost think at system level? Actually it’s so funny in our Slack in in the company, we have a channel. Mostly these AI agents will only respond when you talk to the name. So we have a channel. So when Enrico was built, right? Six other people in the team said, “Oh, we also want something like this.” So we replicated. We have a channel where we let them talk. Okay. But Enrico is supposed to detect are these talking too much? If they start talking too much, is this timeout? And then everybody goes quiet and cool back. Yeah. So the because these things are very eager to you know generate tokens. So I actually So I can tell you my view of what actually has happened. Of course I’m not working inside of Frontier Labs. So I do I do look at it from outside. And and So obviously the the core of it was the the generative part was the prediction of what is the next token based on the context you have given. Right? So that was the thing. Now of course one big thing which happened is Well, two big things happened which made things super easy. But one of the biggest ones was the context window was huge. Right? And the fact that there’s so much coding code data which is available out there. And the context window is huge. And the code is fairly structured. And it’s a translation problem. You know, you give in English and how do you translate that into code? So that coding became the first high quality uh killer app, if you will, that I can code with this. Yeah. But the when you started coding and then people obviously did things like chain of thought and all that and reasoning models which could actually go through the loop and kind of understand think through how they’re doing and then use that to you know, go further. That started appearing in uh coding models with more complex code, planning and then you know, execution and all that. That was the whole angle which is all model based. Then on the side, the second thing which started was the tool use. That going from generating content to becoming agentic. And agentic really means that you are an agent of the caller and you can use tools on their behalf. So either going and using the tool or giving you a function call which you can call depending on the security model or all. But when that happened, when you came to the coding world, then all of a sudden a lot of tools made things much easier. What they found is like simple things like doing grep was so much better on code than to inject the code into the model. Now the context became lot more efficient because you had these tools. And taking all of this and then memory started showing up. And then multiple agents could be called from an agent. So it became like a team of developers. And all the harness around making these things work better has given exponential capability. Given that you sit in Silicon Valley and I don’t know what you hear from the grapevine, but what what to expect in the coming months and quarters in terms of further improvements in capabilities? I think the overall one of the biggest things which which I have heard from people in the very near term is like how long can an agent keep doing things autonomously while staying on task? Like the example we were talking about. Uh Can you give it a really complex problem which takes an agent a day to do? Might take a human a month to do. And it will actually stay on task and get it done. Just as a model as opposed to like you know, harness on top which is like doing 20 other things which is also not a bad thing. But that is the more model can do, the more you know, things can happen. So model can have kind of built-in feedback loop. It can watch its output and then based on keep generating. Yeah, it does happen, right? A chain of thought is an example of that. Like it like before you actually kind of gets stale very quickly. Yeah, exactly. So how do you stay on task? Because the further you go, how do you not you know, forget about it? Memory is another big problem. Like if you build any of these agents, you will quickly realize that the memory is not No, my Chief of Staff right now struggling with that, you know, like whatever memory architecture I keep building it keeps breaking. So I wish it can just you know, sometimes we like “I already told you that, you know.” Like why are you you know, not remembering it? You know, the the thing which I feel is and I had read this in a book somewhere that consciousness is actually really having global memory among all neurons. So if if if your brain and my brain could join through neurons, then the consciousness will go up a level. It will be the global memory which we would have. So that is what it is. And I’m starting to see for example with something like Enrico is like I’m trying to give it global memory and a company-wide memory. And it is very interesting how its behavior changes as a result. So I think there is a lot of work which is happening in memory. A lot of people are doing. It’s so funny that a lot of people have tried different things. But it seems like simple files is the it turns out the simple solution is actually working very well. So that is there. Embodied AI is another one, right? Like the world How does the world operate? Because all of this is happening still behind a screen. But robotics is coming. I mean it’s all over the place. And how does what these models do and how does it operate in the world? Like you know, can it make a cup of coffee? Right? It can definitely print out a recipe to how to make a cup of coffee. But the the edge which connects to the world and there are a lot of startups which are in that space. That is huge. Then of course in vertical areas, you know, for me is in education. There is so much which is happening over there because everybody’s worried that using AI causes cognitive debt for the students, right? Like the the the atrophying of you know, cognitive capabilities. MIT has published a paper on it. There is clearly divergence happening between grades and and test scores. Every student Students are the power users of ChatGPT kind of things. But that is hurting them. So what do we do? So they’re like you know, and and you know, we are building a solution we think is world class which is how do you preserve the struggle but not just give the answer. But yet should be enjoyable enough that the person doesn’t switch tab to ChatGPT. What does it mean for for me in particular but overall company building in general? I don’t imagine now startups needing to raise millions of dollars. I don’t know for what. Perhaps for customer acquisition. But in the past engine used to be the most expensive thing. And maybe great engineers will still be even more expensive. But you need just few of them. So what do you need the money for? Yeah, actually there is one other thing which also is great here is that the way past customer acquisition used to happen is very different now. Um and I I can talk about why it would be different. Or maybe I think maybe I’ll make a point on that also is the way it always happens in Silicon Valley is very different than how it happened in India. Silicon Valley is always on build a product and it’s better. And word of mouth virality, product hooks, you know. Silicon Valley companies I never heard spending gazillion of dollars in customer acquisition. Maybe recently. But Indian B2C were built by customer acquisition means you know, you burn $1,000 1,000 rupees per customer and eventually see which Flipkart will eventually see after 20 years, you know. Well, there are times when I have seen that in Silicon Valley especially if it is that winner takes all market happens, yeah. Then they throw so much money at the problem. Like Uber Lyft for example this is a good example of that. That they know that if the leader will take away with so much that they’ll have the pricing power and everything eventually. But also the other big change in how software was done and how it is done now is the marginal cost is non-zero. Right? Earlier what used to happen was I put all these billions millions of dollars, hire this team and I build the software. But marginal cost is almost zero. But marginal cost is not zero. So now you have to think about it from day one that how do you architect a more efficient product which is higher in quality, higher in performance, lower in cost and yet is world class going out there? If you manage that, then you don’t need that much money. If you don’t manage that, then you need money till you figure out that thing. So, I think it depends on the team. It’s not given that everybody will need money. When you see the same outcome, some people will need lot less, some people will need more depending on the way they have been operating. But the best ones will not. I I think you’re right. You You shared this open-source project. I think it’s open-source called Paperclip, which where basically you can hire AI employees or any You can hire AI CEO, AI CFO, AI engineer, product manager. I don’t know how well that is working at the moment, but I think that’s also basic this able to hold long-term context is what a good employee needs. So, as the models get there and maybe till then we build some hacks around that. But that also has to be around the corner. I mean, we can all hook together a some kind of employee like you have been kind of custom built and recall. Uh so, it’s kind of somewhat already possible, but I assume it’s just around the corner where any of us watching this can hire one or 10 or army of expertises and let them loose on whatever you want to get done. Yeah, I think that is true. Definitely, even today a lot of that can be done. Like when you are trying to work with your chief of staff or these agents or I am trying to do that. You can quickly tell in in your own domain that no no no, you got to do it this way, right? And then it becomes very clear that oh, it is because I have had this lived experience. So, going back to the CFO and those examples, how good those are, it’ll be hard for me to tell as I’m not a finance expert. But somebody else can actually look at, okay, these are the kind of decisions you want to make the CFO take and these are the kind of decisions you want to go to the reals. Yeah. So, I think that that line is not clear. So, I’m thinking someday things like Paperclip will be there and obviously a lot of companies are finding useful. Uh I I get very hesitant when I rely on AI for expertise which I don’t have. Mhm. Because there is always this It’s also like, you know, it’s a It’s one of those Indian uncles who knows all the answers, but you don’t know which ones are right, which ones are wrong. So, um that is the thing which I worry about. But I think see, in the software world, even few months ago a lot of developers were not trusting AI-generated code. But I am seeing the transition. Like now people are starting to trust more and more and because it generates so much. If you’re going to verify everything, anyway you’re not going to go anywhere. So, you get into the whole heuristic that you mostly trust. When something goes wrong, then you deal with it. And you build checks and balances. It looks like, you know, that’s where the things are going. I think so, too. I mean, if a lawyer trusts a legal AI, then I trust the legal AI more. Right? I I I agree with you that yes, that is of which People are also seeing it in the medical advice, right? I think right now I honestly get much better medical advice from GPT if I give it enough context, all the prior reports. I have not found anything where I feel it is hallucinating or misleading and it’s all the generally fairly reasonable of saying, okay, this one you should really consider doctor and not take my advice. Yeah, actually I read a article, I think it was by OpenAI exec who was talking about that she had some medical issue and she went to like one of the top doctors in Stanford or something and they prescribed a medicine. But she had ChatGPT and she was telling her this stuff and ChatGPT flagged it saying, no, you have this other condition. Do not take this medicine. And she showed it to the doctor and the doctor was like, oh, thank you for telling me because in all the things he had not realized that one additional thing was there. And and I I forget. I mean, it’s it’s published as a blog, so we can find it. But he he missed that. But this thing did not and it was going to be a complication, serious complication if she had taken that medicine. Yeah. So, it’s like definitely there is more to be done. It’s uh I think we are going to see that kind of expertise available for uh you know, pennies on the dollar very soon. So, does this all Look, let’s zoom out. What is a big picture here? Does it transfer into just unleashing of entrepreneur energy or on the other hand, concentration of all innovation, etc. in very few places, you know, people who are model owners. How does this pan out? It’s a very strange animal, you know, very difficult to think. At least I have realized I’m not going to draw too many parallels with, you know, whatever electricity or internet or mobile phones. It’s different. It’s not that kind of step change. I I have some parallels which I have been thinking and maybe we can Yeah. debate and like kind of like try to uh think through like um So, one thing I have a very uh I’ve not been able to shake this, but maybe it’s because I’m a optimist by nature. That the the entrepreneurial spirit is going to be a blaze. Why? A lot of people who had great ideas did not have the ability to instantiate it. This is why engineers were commanding such premium, right? Because you worked with a There used to be a concept non-technical founder, technical founder. No longer. Which means if you have a vision, if you have a mission, if you have a idea, the cost of trying it out is very low and you can now really go forward. Which means entrepreneurial spirit. So, anybody can go do a startup. They don’t have to have the skills because there is something with skills available. So, skills and expertise is relatively easy to get now and I think that that should increase entrepreneurial spirit. The distribution thing, I don’t Do you agree with this because before we go to the distribution? No, I 100% agree. I think, you know, yeah. I mean, I just give you the example, you know, I was giving you about earlier I mean, I have lots of resources, yet I was depending on other people to build you know, this V1 product of capacity. And now I’m able to do on my own in a matter of days, which is incredibly empowering. Yeah, it it There is also what I was, you know, contemplating also. There is a bit of a learning curve, especially, you know, you have been in technology all your life. So, for you it’s somewhat it’s somewhat smooth transition. Yeah. Last time I coded was 20 years ago. But I still have a reasonably I think I would say a technical mindset and able to think, you know, I was product manager and so on like But yet I had a initial hesitation. I think you were telling me. Remember in December you sent me a message saying, give me 2 hours, I’ll show you what can done. I was like, yeah yeah, we’ll do we’ll do. Right? And until I started and you know, So, I was almost thinking, you know, maybe this could be a miraculous initially also, almost like a 6-week course where we handhold people like online and get them from I have never written a line of code to what you know, I published this fully functional thing in 6 weeks. Otherwise, you know, that lot of people will be lost in that. That it’s It’s not trivial that I think you are right that there is when enough people are doing that, others will that hesitation will go away, but there is a little bit of that activation catalyst needed. So, that definitely could be It could be like an entrepreneurial factory kind of thing, right? Like come in and on the other side you’ll come out with very low cost. You would have tried Yeah. I absolutely think And also the factory which takes non-technical person one end and the person come out and that thing I can build anything and I know what I’m doing. Yeah, actually uh I’m starting to become the mindset that that non-technical is no longer even an adjective. In people’s mind. Yeah, I know. I I know it’s in people’s When enough of them are doing it, then it would not be then it would actually be It’s not that it would not be just the bar will keep rising. Right? So, there will be a place above which you better understand that technology. Like if you were thinking of like how do I apply quantum computing to cryptography, then yes, it might be that you still are. Uh but the bar keeps rising. I think the other challenge which people have been saying, which is the second question that will people with, you know, models and and product surfaces will be the winner take all. Now, because it is so easy to try anything out, generally startups are all about innovating faster than large companies can distribute or distributing faster than large companies can innovate, right? As you said, big companies already have distribution. If innovation becomes a commodity, then distribution becomes a precious thing. That whoever has in some ways almost the winners get locked in. Like, you know, yeah, you can innovate, but this guy can also innovate, you know, tomorrow. So, as a startup, if you don’t have distribution, that’s probably the biggest challenge. And we have to also I mean, we’ll see in the Fermi and other things we’re building cuz there, you know, AI is not going to help that much. We still need to sign 1,000 schools and acquire students and so on. Yeah, some collateral, some assets. So, it’s almost like it may flip, you know, the distribution becomes a rare commodity. Yeah, I I think distribution becomes a rare commodity. My thinking is that it still does not mean that an incumbent who has distribution is the winner. I’ll tell you why. And this is where maybe there’s a kind of like And this is just theory, but maybe where we will end up will look like a very not sad, but very different than today’s big tech world. But the here AI is interactive. AI will get personalized. I mean, the whole There was that democratization of personalization thing which we had talked about once, right? That the whole benefit of ChatGPT is that it actually personalizes to you. Otherwise, you could do a Google search, right? Or like Gemini personalizes to you the further you go into it. So, as it is personalization, eventually you a product market fit of one. For me, when eventually using, it gets to a point that it actually is a product which you will use throughout your life to learn everything because it is the best teacher coming along with you. If that is the case with everything, then what is the benefit of having wide distribution if I have captured a corner already? Because you cannot switch out. And if that happens, then if imagine a lot of startups start and they all go and they all can build super fast because innovation is cheap or like execution on vision is cheap. And then they distribute and now people are stuck. Then you what you really are don’t have big tech, you have law firm style. Like you have these small 10 engineers with a billion dollar cornered in a multi-billion dollar market and then 10 engineers with their a billion dollar cornered and then they’re like a lot of them. Yeah. Maybe there’s a roll up, but it’s not a that’s a very interesting vision and I would like I would definitely like things to play out that way. I know you were at Google, perhaps Google fan still, but I am not a fan of big tech at all. I think it’s just too much concentration in too few hands. But I think this is where a little bit of human nature also plays out. We all like big brands and brands are about trust. And we also want, you know, uh everybody operates like I want what you have. You know, people don’t want what they have, right? That’s the reason I I think Hollywood will always be super big. If let’s say million people can create their own, you know, movie in the corner, people will still want to watch the blockbuster movie cuz everyone start talking about it. People still want to read the top 100 books. Books have long tail, but you know, forever only top 100 books get all the numbers and so on. So, that I think there is something about that and even you have 10 equivalent startup, the one that has a brand cachet trust and that’s where existing guy I think can have fundamental advantage that I if I can but uh they’ll have to fight the cultural shift cuz the people who have distribution they’re also slow moving. They are not very good in, you know, taking risk, launching innovation fast. So, somewhere I think they may not recognize their strength and use it well because that require them to unlearn. Unlearning as a human being is so hard. Unlearning in organization is next to impossible. So, that’s probably, you know, counterbalancing force there. Yeah, I actually I do think that the the I I feel like the pre-2022 companies operate very differently than pre-20 uh post-2022 companies. It’s hard for them to unlearn. I’m mostly thinking that maybe the brand is a good point and the cultural side of like having that common language. Like, okay, we all talk about the common things. Uh I was feeling like if I’m buying a product of any type and it is really serving my needs because it can fit to me like a hand in glove. Yeah. Will I switch out of it? Even if the other one fits, will I take the risk of switching out because I’m so comfortable with it? And if I will not, then there are 10 glove vendors who have showed up in the market in different areas and captured the corner. And now they are stuck, which is very different from how earlier because it used to take a lot of capital to go build things out and acquire. And now it is like it’s not that way. So, now you have 10 people offering you products and you happen to be with the one, but you are so happy with it that you’re not going to switch over. Something drastic will happen for you to switch over. Um Even if the other person has a brand which is more well known, you the risk of disrupting your life might not allow you to change. Yeah. Like People who use Windows, they don’t switch to MacBook. That’s right. They’re so used to their Windows. And even with the, you know, earlier the BlackBerry to iPhone transition, a lot of people held on to their BlackBerry for 4-5 years. You know, Yes. the it was stopping to work, the keys were coming off. They were very reluctantly, you know, so there is a bit about human And here the assumption is that these are iPhone 1 versus iPhone 2, right? It’s like like Yeah, yeah. That is the So, I think Yeah, let’s uh How does this does this worry you the fact that you are talking about all this innovation, all the acceleration, but this is on the back of all the state-of-the-art models. You may have used your ingenuity in building uh Enrico, but ultimately all the power is coming from the engine under the hood. 100%. And engine is owned and controlled by very few entities. And some of them have massive distribution. What the and they are watching whatever you are building? They’re watching whatever I’m building, they’re watching. So, what does it mean in terms of prolonged innovation versus ultimately, you know, every thing ending up in their control? So, I think that two three things which are going to stop. I think by itself the models are not enough because going back to the memory problem, right? These models can be super powerful, but if you have to teach them every time something new. They’ll build memory. So, they will build memory, but then the actually memory is the one which I worry the most about. But let me give you an example in real world like Furby, for example. We have built this product which is out there and people are using it and it’s powered by multiple models underneath. But we have tried to make sure the assumption is these models will keep getting better. So, what we have said is our product has to be high quality, has to be fast and has to be accurate and as low cost as possible. Now, with those equations, we test all of them and we switch between them as needed. Then as the frontier of these models keeps moving forward for the task which we are performing, eventually there are cheap models which have become good enough. Similarly, there will be open source models. Like Google came out with Gemma 4 just yester- yesterday or 2 days ago. That is a very small model which can actually do more thinking and is agentic. It can be used in multiple purposes and it’s open source. You can put it on your local machine with GPU also or for much cheaper go somewhere else. And now there’s a class of tasks for which you can go there. You don’t have to go to the higher end models. So, things will keep uh evolving and you will be able to bring the best solutions to your best horsepower to your solution. Now, the problem there of course is that if these models make you captive in any way, like for example, the memory. Now, there is of course, you know, like end consumer products are doing that. If you use ChatGPT, if you use Claude, if you use Gemini, they all have different ways. They personalize the experience for you that it’s hard for you to switch around. So, that definitely is a end consumer a problem. When you are building the model, it you have to architecturally think about like, you know, am I getting beholden because I’m getting a service? What is my migration path? All of those things. But those are architectural decisions we made in the past also, right? If you went to GCP versus AWS versus Azure, you know. What is the role of open source models? Have you been playing with it in the the products you are building? I just started using some open source model and I was quite surprised the performance at least for my limited task are nearly at par at 1/10 to 1/50 the cost. Most of the open source models, at least the best ones are right now Chinese model, which is again very strange phenomena that it used to be the other way around, but that’s where we are today. What is the role of these open source models as a startup thinks about? Are they better off saying, “Okay, let me build with the state-of-the-art model, eventually cost will come out” or say from day one, “Let me look at everything and be very cost efficient”? It depends on the My theory is it depends on the complexity. I have tried uh three different uh obviously when Deep Seek came out, went and used Deep Seek and and tried that, but uh most recently uh MiniMax 2.7 and Gemma 4 were the ones I was playing around with. But the basic approach which I generally have taken is for the most high leverage tasks where I do not like a mistake will be very expensive, I’m going for the most expensive model. Like all coding is like Opus 4.6. Um And it is expensive, but it is the highest leverage. But if I’m running a cron job on on something and it is doing like that would be possible to run on something simple. If it’s a highly multimodal uh you know, task which I have, then I definitely rely on Gemini. I’ll tell you an example of this. Uh so, one of the things which is one of my favorite things which, you know, as an entrepreneur, you know that there are times when you are building a business, you have problems and you’re like, “Well, how do I think about this problem?” Then you go to a bunch of advisors and you talk to them, right? What I ended up doing is creating a virtual board of advisors. And I told Enrico, “I want a virtual board of advisors and these are the five people I want.” And the five people were Steve Jobs, Peter Thiel, Jeff Bezos, Charlie Munger, and Sam Altman. Why? I wanted five different views of looking at the same problem. But how do you get that information? So, I ended up choosing Gemini 3.1 Pro. Why? Because Google has the most information in multimodal form with the interviews they have given, books which have been written, everything. I’m like, “That is the model which actually will know.” Then crafted prompts to extract that persona out. And then Enrico essentially sets it up, says, “All right, let me instantiate these five agents.” And they all know that they are in a board of advisor meeting. And who else is there? And then because Enrico knows so much about the company it is able to constitute a lot of read me pre-read packets. So it actually sets these pre-read packets and then ask me, okay, what questions would you like and are there other things you would like to include? So then I’m like, well, this is what I’m thinking, this is what I’m thinking. Oh, by the way, those two documents, remember we had talked in a meeting? Like I have the meeting notes, all right, pull that. Right? And then it constitutes and I’m like, all right, send the pre-read. But because I use Gemini 3.1 Pro and it is then able to utilize all the published information of how say Steve Jobs thought about a problem or Charlie Charlie Munger thought about it. And then it looks at it from multiple ways. And then they debate each other. And then like and it is interesting to see how the convergence happens or doesn’t happen. And at the end I have looked at a problem from five different sides, from five different ways and made a very Who knows what is the right or wrong decision? But that is the hard thing. And that model, it’s very hard for anybody else to do with open source and all that. If Google charges a lot, I will have to pay it because the value is very high. This board of advisor sounds fascinating to be able to get real-time feedback from personas of people who have been very unique thinkers in many different area. But just want to push on that a bit more is See, it sounds very fancy and you know but how do we differentiate between gimmicky stuff versus real stuff? Does it actually change the decision you’re taking for me? And more importantly, because of that, are your odds of success dramatically higher? So odds of success we will find out because this is very fresh, but I will give you examples of the kind of decisions which have gone there. And what I have seen. So for example, we are trying to figure out the pricing for our pilots. We are going from free pilots to paid pilots. So what kind of pricing should be? Here is why I don’t think it’s gimmicky. First of all, it’s not like I have put a prompt and said, okay, now act like Steve Jobs. That’s not what it is. The amount of context going in is including like every bit of every important meeting we have, its meeting notes actually get distilled down to key strategic things. And there’s a treadmill of stale information, new information and that information packet then goes to the person. And that goes without me having to do the effort. This is the other thing. I don’t have to do I’m living my day like I am. I’m talking to people, I’m working on things and all that. And that information, the body of knowledge is being built, the group memory for the company. And goes in. So that is one. Second thing is it’s not a fire and forget. It’s a board of advisor meeting. So they will look at it and then they will ask question. Hey Piyush, so what do you think? Do you think that you’ll be able to turn customers like this into customers like that? Or what would happen if they don’t turn? Would you be willing to abandon them? And that is the answer I have to give. And that is not something I said that asked, but that the model looks through and says this information I don’t have and comes back. But when I say I will do this, all the other advisors are also listening. So then they say, but you said this, but this is what it means. So it is a very robust conversation. I’ll I’ll share the transcript with you and just like just reading the transcript is fascinating to me. This is how it happened. So but we have made decisions to to Yeah, this explains few other things also. I think until two months ago you you used to reach out to me a lot for inputs. Now I guess if you can get advice from Sam Altman and Bezos, why would you reach out to I actually think that the these things have been the crutch given that you were out for a month or so. Maybe necessity is the mother of invention. [laughter] Excellent. I hope it translate into much better odds for for me then it’s all worth it. Yeah, we we we will see. Actually, more seriously, we we had a leadership meeting just before this. And I brought the answer like this is what we should do. And these are all the leads in the room. And I said this is what we should do. And it actually was a robust conversation and nobody was able to poke simple holes at it. Right? It is as if I collected information from advisors and brought back to the team. So I think it is So we have framed this changing world fairly well. We have this I don’t know, superhuman type of resource at our disposal. It requires us to work very differently, upgrade ourselves so on. In this evolving environment for a company whose ideal employee and we can look at from both point of view for an employer, who should I hire? And if I’m an employee, I want to be amazingly highly desirable hire what should I be? I think as an employer, you have to architect your company right, which would be identify prob- problems and problem-solving loops. Within that the people you would hire, I’ll go back to the earlier thing, two kind of people. Either the people who have lived experience, which is not showing up anywhere else. If you’re just going for expertise, which is easily, you know, aggregated, then that does not work. But if you’re going for lived experience, this person has done these things before and there are nuances here. That person or people who are generalist in in attitude, not generalist by qualification, but generalist in attitude, which means they are willing to go do what it takes. They might be an engineer, might be a UX designer but does not feel like this is my job and this is not my job. So those are the two kind of people. If I was and I’m, you know, in Fermi, that’s what we are doing. But third aspect also flexibility, right? Because a lot of people who have lived experience in the past, they’ll become fairly rigid. I never liked working with experienced people because they’re always, I know how it is done, been there, done that and so on. But today’s environment that is going to be such a big blind spot. And I think let me also mention this new thing you’ve been doing at both Mirabilis and Fermi about this whole idea work trial and that itself has turned out to be a filter because a lot of senior people say, I don’t want to go through work trial. You know, this is not how people interview. I used to do it when I was entry-level engineer or whatever. But today if you don’t do it, you don’t qualify for the new world. So that’s the flexibility. Yeah, actually you are absolutely right and I had made that assumption. But it’s not just the lived experience. The second part is required across the board. You are right that the person has to To me, there are no more managers. Everybody is an IC. I mean like I’m writing tons of code and I’m Everybody’s doing what is needed. They’re working in that problem like what is the problem? How do I solve it? And we’ll do what is needed. So there is no managers of humans as the primary job, at least in the companies at the scale which we are and I I don’t see it being a very big role anyway in the long run. There is decision-making in face of ambiguity, but that is like a like what Amazon used to say like single-threaded leader kind of thing. There’s one person. It doesn’t have to be managerial job. So I think that there is that and that can be a fairly, you know, high agency junior person, too. So which means therefore a lot of people it really is a time for really consider reinvention, which is very very hard. How do you reinvent? I am quite determined. I’m We’ve kind of always looked at each other Piyush is a technical guy. You probably looked at it Mukesh is a business guy. But I’m no longer going to be that. I want to, you know, keep pace with you. If you can build your chief of stuff, I want to build your equivalent or better or whatever, right? So but kind of changing my primary mental model that I’m a technical person, I can build anything technically and then all the other layers, right? And primarily fundamental IC. I don’t want to manage anyone. I’m not a manager. That’s a bad skill at this point. I don’t want it for myself. Yeah, I I actually don’t know how to get others uh to see that, but there is I mean this connects back to our earlier point, right? If you’re a builder, you should feel you are addicted. If not, you’re doing something wrong. And I feel like there are Once there is a critical mass, it will happen. I think there are people There are many people around me who have been managers, general managers, CEOs and they are all now hands-on building and cannot let it go. So one is So I think your critical mass point is valid, but I think we are also in this accelerating curve. I’m realizing every day I’m better than the previous day because the more you work, you see patterns, you realize what you can do and so on and it kind of builds on itself. And within few weeks or months, you can be so far ahead. The other people think, you know, how can you know you are able to do so much? So I think it’s also don’t wait too much also for the whole mass phenomena play out. Maybe. Actually, this is this thing which you said, when I saw that happening, this is why I say that we’ll all be AI managers because this is a management skill. That how do you get these things to go do what you want? So I I realized like, oh, I’m actually building a management skill for these virtual employees of mine. And and that is Yeah, the longer the faster the the sooner somebody starts, the more they will learn about What happened to companies with thousands or tens of tens of thousand employees? Will we need those companies or most of them will become few employees, few maybe single digit, double digit plus thousands or millions of AI agents? Knowledge work. I think blue collar work you will need a lot of human beings on the floor or the relationship going to work. Humans want to talk to humans, they hate talking to AI. But all the knowledge work, how do the future organization look like? By the way, humans are loving talking to AI, too. You know, character.ai, which was all about talking to other AI It was the fastest growing AI company at that at that time. I think it’s still an example. I think now that they’re nowhere. Yeah. Yeah, but the humans are talking to like people are using these things as therapists. They’re using them as companionship. Yeah, mostly on chat, not on voice, I think. Yeah. Today, one metric of an industry, maybe connected to the revenue per employee thing also, but we think big established companies have to have thousands or tens of thousands employees. It’s starting to make less and less sense to me. Yeah. I don’t know what will be because company culture is the hardest one of the culture is one of the hardest things to change, right? So, we started here. Yeah, so [laughter] Here there was that book Switch. Mhm. Which was about like how hard it is to change culture. The thing which uh Therefore, today if there’s a 10,000 person company I don’t know how they will have to reinvent themselves. Uh of course, uh you know, Sledgehammer. Huh? Sledgehammer. That could be I mean, IBM Lou Gerstner wrote that book, right? Making elephants dance, right? So, I mean, people can do that, I guess, but but a company today which is starting I don’t know what it would do with 10,000 people when they join. It can probably become a This is I going back to the law firm example. Yeah. You might have a law firm with 10,000 lawyers in it, but they’ll be all over the world and they all have their own thing. And there is no like the the There is no common asset which necessarily is they’re deriving all from other than probably the brand. I don’t know. I’m just making an assumption. We don’t know what it’ll be like, but I keep coming back to some sorts of parallels like that. I think people watching this may feel that we are probably beating the drum of AI a bit too much in terms of, you know, how everything is spectacular about this. I do think, you know, by and large it is. But, what do people also get wrong about AI? Or are there areas where we are over projecting or over promising? I think that uh we are optimistic. I mean, going back to the part of it is coming from being optimistic, but there is clearly uh failure modes here, right? We talked about it like yes, AI is a great productivity tool, but the biggest user base is using that and causing cognitive debt. The more you use AI, the less your brain works. The less your brain works, the less you are actually, you know, sharper at it become dependent on it. That is a very simple one. The harder one which I feel This is the example of the paperclip one which I worry most about is am I over-relying on something which I don’t understand is accurate? Mhm. Now, I know it happens in real world also. You can go to a lawyer for advice and the lawyer can say something and you will take that word for gospel and you’ll come back and the lawyer can be wrong, too. But, I have this thing that with AI we believe that oh, no, no, it is always going to be right because it knows so much about so many things. Yeah. And that over-reliance can result in mistakes. And that’s most scary one at work which I feel is you become intellectually lazy and instead of you commanding AI, AI now commands you. I think my translation of that is a couple of skills which are super important. One is just this critical thinking. Mhm. Which is not very easy. You know, it’s a critical thinking is you are trying to be both constructive some level, but also you are being skeptical simultaneously. You’re questioning everything and yet trying to take a bold leap. So, that’s a I think a lot of people don’t practice it. Second is even harder perhaps in the whole system level thinking. I always struggle in how do you like in the work at least you were doing until recently, the architects are able to think at a system level. What are all the building blocks? How different components relate to each other? What are the second order effect, third order effect? And those skills are going to be super important. How do schools, colleges teach? If you’re a working professional, say most working professional get siloed into function. I am a good designer. I’m a you know, good programmer. I’m a good lawyer. But, that is not a framework. You know how to execute a even a perhaps a complex task. But, how do zoom out, think at abstract level and transform to system level thinking? And those people can then do magic with AI. I think that to really make that happen, we have to even go into the education system and the training system because I don’t think people have been told the more the higher you go in education, more specialized you become. And it’s kind of very built into the system that you become more specialized and that is your trade, right? Whereas the system level or like looking at very disparate things, I think it is very important that even in education, maybe that is the other thing which we should like figure out with the AI at the doorstep or not at the doorstep, inside the living room for everybody. Uh how does our education in high school and colleges change? So, that it is not teaching you things which you don’t really I think it’s very clear this space is moving too fast. I think uh people have to get their hands dirty, pay very, very close attention to what is happening, what is the art of possible because this is probably a very, very unique moment in human history. We are definitely It feels like to me that we are going through a phase transition. And things will look very different. If you have to just zoom out now, Piyush, and maybe say in 12 months from now, if let’s say we are having this conversation in 12 months from now, what could look dramatically different compared to today? I think uh the the example the bigger problems which we are talking about uh uh where we feel like we are still applying our mind more would also get tackled. Uh like you know, the architecture of level things or like the systems level thing. Uh the other thing which I think will definitely happen, I don’t know if it’ll like in in US or India, but it’s already in places like China. It’s like uh you talked about like people physically will do things. I think even that will start to go away. Uh like I am looking at what uh things like the robots today which are not out, but what they are able to do I think like there’s this blue blueberry sorting problem, right? Like okay, can a robot sort blueberry because it’s a real world problem. Otherwise, seems pretty easy. Uh but, they were not able to. And now they’re getting to the hardware and the dexterity and everything. And when those things are solved, now all of a sudden this proliferation one can say we have not hit the GPT moment yet. Actually, I think the So, sensing, planning, actuating. Earlier, planning was very hard. But, planning I think will become easier. The missing part of the planning is the world aspect of it. Actuating and sensing now the delicate part of it. Like sensing something like texture of a blueberry and actuating such that you only hold it with that much force and not any more as opposed to a marble, right? That is a hardware problem getting solved and the planning problem. So, these things can And you know, like sometimes you know, for decades nothing happens and then decades happen in weeks, right? So, so the thing is that I I feel like a year from now, I would feel like it is there. I mean, I think that it would not be everywhere, but it would be within arm’s reach. Yeah, it’s like future is already here, but not evenly distributed kind of situation. My last question to you is Piyush is um Right now, if you were to take inventory of your time, I see builder time versus manager time. For me, it has gone from almost you know, 90% manager time to now 50/50. And I’m pretty sure I might My goal is to bring my manager time to less than 20% and 80% be IC builder time within few months. I think it has drastic One of the benefits which I get of being in a completely different time zone from a lot of my team is that my manager time was the morning 3 hours. And I’m shrinking that down to 1 hour. So, I generally, you know, I work almost every waking hour that way other than if I go to the And mostly as a IC builder most of the time. as a IC. So, I think out of 12 hours, I was So, yeah, less than 10%. How much code you are contributing to Formistack? I’m actually contributing Well, Plot code is contributing, but I most recently I was contributing to like some of the biggest our patent-pending feature was built by me. Also, some of these are because our senior-most engineer left. And this is something which was there and date was coming. So, like So, you’ve been to CTO and senior engineering roles at Flipkart, at Google Pay and so on, Google Assistant. I imagine you were not writing code. Uh no, I actually was not, but I am a engineer at heart. So, like even in Google Pay like something was happening, I was looking at the code, I was debugging. I was like, “Oh, this is what is going on.” But, actual the amount of work which you had to do required the kind of time which you could only do it at night then because the whole day was meetings, right? I’m This is why it is addictive. Sitting in meetings is not addictive. It’s like realizing your vision is even the meetings which are now and don’t appear like that because everybody is discussing the problem and they’re all going and making progress and next time they’re discussing the progress. Yeah, I’ve always felt like from my early days of career that meetings are the curse of corporate life. I think we are on the verge of being delivered from that curse and that itself is a something very profound to look forward to. Yeah, it’s fascinating. Great. Let’s watch this space closely and I don’t know, maybe we will have to come back in 6 months at the way things are going but thank you it’s a great to great to catch up. Yeah, I know it’s great to catch up in office. Absolutely. Thank you. Thanks for having me. Okay, thanks Kush. Let’s show you how to create an AI agent using Claude code. Step one, install Claude code. Download the Claude desktop app which has it built in. Step two, create a project folder. This is your agent’s workspace. Drop in any context it needs, brand docs, business info, goals, whatever it should know. Step three, create your Claude.md file. This is the brain of the system. It becomes the persistent system prompt every agent in this workspace reads on every session. Step four, build your first agent. Type /agents in the terminal, select create new, then generate with Claude. Describe the role in plain English. Claude writes the agent file. Step five, give your agent skills. Skills are reusable instruction manuals. How to write a blog post, how to build slide deck, how to do keyword research. Install Anthropic’s official skill library via /plugin. Step six, connect MCP tools. This is where agents stop being chatbots and start doing real work. Connect external tools, Gmail, Notion, Google Analytics, Slack, whatever you use. Final step, and run. That’s the full system. Build one agent first, get it working, then stack the rest. At Sparks, we aim to bring to you stories of exponential impact. We share in-depth analysis of what goes behind success stories. If you find our conversations interesting, you can join us by subscribing to our YouTube channel. You can also listen to Sparks on Spotify, Apple podcast, or any other audio platform of your choice. If you have any suggestions on who we should invite or what topics we need to cover, just let us know in the comments. We are always listening, looking for ways to improve, and keep getting better as we go along.