{
  "episodeId": "ep1",
  "episodeTitle": "The AI Supercycle: Models, Moats, and the Future of Building",
  "speakers": [
    "Vikas",
    "Sachin"
  ],
  "speakerMap": {
    "speaker-0": "Vikas",
    "speaker-1": "Sachin"
  },
  "duration": 1617,
  "durationFormatted": "26:57",
  "wordCount": 4461,
  "segmentCount": 20,
  "segments": [
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 52,
      "timestampFormatted": "0:52",
      "text": "Hi, very good evening. Welcome to the very first episode of the AI Valley podcast. My name is Vikas and I have with me"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 61,
      "timestampFormatted": "1:01",
      "text": "Hey everyone, good evening. This is Satya Nair."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 65,
      "timestampFormatted": "1:05",
      "text": "Cool, all right, so this is the brand new episode of our podcast. So we have a lot of topics to cover, Sachin. We've been talking about this quite a lot. The world of AI is buzzing. There lot of new releases every week. As we've seen, companies are rolling out models, applications, and there's a lot happening in this space. So what has kept you busy? What are your thoughts on the new? Gemini 3 release and the excitement that you see around things like Nano Banana and everything else that's happening in the world of AI."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 104,
      "timestampFormatted": "1:44",
      "text": "Wow, that's such a loaded question and maybe we can talk through this over the course of the next few minutes. I think in general, there was a lot of talk, there was a lot of anticipation about Google's next model. And I think it's kind of lived up to it in a way. There's a lot of rave reviews coming out about Google's model. Honestly, just from a personal side, I thought it should have come out with a little bit more fanfare. But that aside, I think it seems like a major reset moment in my mind. I think what Google has done with this model, it has shown the world that we're not anywhere close to plateauing with the so-called intelligence layer or with what these models can do. I think there's a lot more that's there. There's so much white space, I think that Google has shown as a result of this model. And I'm surprised, I mean, while I'm not a big believer in benchmarks, and I've always had my apprehensions of looking at these benchmarks, but there are a few line items on that benchmark which kind of stand out, For example, on the ARC price, I think it's amazing to see the leap that we have taken with these models. And it's almost... in the range of around 50-60 % of a leap from what the earlier models were doing. And it's not just in the leap in terms of what Google is doing, it's also in the leap of what the other models were. And so I think these are good explanations of how these models are doing better. There's humanities last exam, think stellar results, there's maths, you know, some of these questions are even tough for humans. And it's amazing to see what these models are doing. But what I really liked,  and I think Here's where I see Google coming up with its model of how to understand words better. think that is super surprising to me because I think where Google started showing, where I'm really visibly surprised is in the screenshot. Yeah, absolutely. And yeah, mean, why don't we talk about that? that's like a... To me, that's a very big moment because that somehow... tells me that this opens up things like robotics, visual elements, know, these models are always good with text and they'll continue to be, but imagine fixing some of these models to a pair of eyes or a pair of ears, right? And all of a sudden, it's everything's so different."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 248,
      "timestampFormatted": "4:08",
      "text": "I agree absolutely and it is absolutely and I'm glad that you mentioned because I I think it was the great reset in terms of, you know, there were a lot of apprehensions. We just had NVIDIA's earning call and you and I were discussing just about last week and probably the whole of last month that a lot of people have, we know, have been talking about the bubble because of this infrastructure and we thought that it'll show up in the result, but I don't think the market is anywhere near slowing down. In fact, I'll tell you something, the launch of Gemini 3 and anti-gravity, right? To me, I think it was very profound. To me, it replaced the the Wintel era of the 90s, which is about three decades later, As we've seen in 90s, the tech industry was defined by the symbolic relationship known as Andy and Bill's Law. which is basically Andy Grove of Intel would give us a faster chip and Bill Gates of Microsoft would immediately give us software features that consume that new power, right? So because deep down Intel knew Microsoft used those cycles, so they built the fab and Microsoft knew Intel would provide those cycles and they wrote the code. Now, I think for me, why this is profound is we're seeing this exact dynamic play out with AI, right? but at a much larger scale. Take Gemini 3 for instance. When TPU-V7 Ironwood was released in April 2025, the numbers looked like absolute overkill. 4614 teraflops of FP8 performance and 192 GB of HBM3 memory per chip. That was about 5x leap in compute and about 6x increase in memory capacity over the previous Trillium generation. And I think I was following a lot of thread then specifically critics. One of the first thing they were asking is who need the 9216 chip super pods with so much density? Do we have models utilizing this sort of capability capacity? And I believe  Google has shown it and I believe Gemini 3 is the answer. Anti-gravity for instance spins up autonomous developer agents. which holds the entire code base in context and while continuously reasoning on it. So I believe it justifies, when I look at these two products, it justifies every bit of Ironwood's architecture. If Ironwood was Andy's move, then Gemini 3 is definitely the Bill's move. Right? And  we've seen Google, TPU team built out a 42.5 EXAFLOP AI supercomputer because they knew Gemini team was building a frontier reasoning model that would finally crack the 1500 D-Low barrier, right? And it needed every flop of it. So it is definitely special in terms of what's happening. But again, I think that's about Gemini. But I think there were some interesting releases as well. you just last week, think Grok released the 4.5 or rather 4.1."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 441,
      "timestampFormatted": "7:21",
      "text": "Yeah, full party."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 442,
      "timestampFormatted": "7:22",
      "text": "One model and I think that's performing really well as well. The prior week, I think last week OpenAI released 5.1 made their model so much warmer, right? And so I think the pace of innovation is relentless. Now, I think my question to you, Sachin, think given so much is happening, right? How do you see these things play out, right? With all of these capability reaching enterprises, startups, how is the world changing? How are the software and the services that we know of is changing with this new, riding on this new innovation?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 479,
      "timestampFormatted": "7:59",
      "text": "No, think, look, that's a very interesting question. I think what most of these models are trying to achieve with the way they are coming up with all the new innovations, think software and applications as we know it will suddenly have a very different face. And if these models continue to perform the way they're performing and if they continue to scale the way they're scaling, Somewhere I get a feeling that a lot of applications as we know it could become very headless. I don't know, right? Like just say for example, a contact center application. That's the easiest that comes to my mind right now. Think of it autonomously running through voice or running through this whole agentic system. Like if a model is so good at looking at a screenshot of something and saying what the problem is. or understanding the world of that screenshot. The possibilities are endless, Like customer support could become a visual thing more than an auditory thing where it is today. All of a sudden these things can change, right? Like you're talking like, let's say sales calls, for example. You could see people doing sales calls very differently because now instead of just auditory responses, they can have visual responses as well.  And I think this whole anti-gravity that Google has shown, all of a sudden it's packed in the browser, it's packed in coding, it's packed in the agentic system. I mean, it's just insane to think about what are the possibilities and how these enterprises could really use some of these systems. But I think it's also going to be super important to see where are these enterprises actually ready for this. And how much are they ready for this? Because while a lot of these frontier models are doing what they have to do, the data is still sitting in its old messy format. Or in its old format, right? Let's not call it messy. It was doing what it was supposed to be doing. And these data structures were constructed because of the systems that querying these data structures. I think what's important is what we need to figure out now is we need to think back is that this data structure itself is changing. It's no longer the traditional data structures that we were familiar with because these models don't look at the traditional data structures. They don't understand some of these traditional data structures and they have their own structures. So are companies prepared for it? Do they have what it takes to in order to build it for the next generation? And I think that's where we'll see a lot of disruption. because companies that are ready to prepare their data faster in a manner that these LLMs can understand it, I think are going to be the ones that will succeed faster than the ones that will take time to do it. But at least that's the initial reading I'm getting from enterprises."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 643,
      "timestampFormatted": "10:43",
      "text": "Yeah, I think I agree. one of the biggest conundrum that I have when I look at these things is depending on is the model versus application question, right? Given, you know, what we saw with, say, let's say cursor and replate and many other application layer innovation that happened earlier this year and they are accruing a lot of valuation and usage and rightly so. But I think what we have seen in last month is where the value accruing or models getting increasingly popular and people now questioning that if the model companies are accruing more value. So one of the things that is really hard to do when you're solving or disrupting on on some these enterprise use cases is how do you ensure that whatever you're building does not get solved by say next iteration of the model. As model becomes powerful. It's one thing to say that you build with the model, don't build, but it's another thing to sort of. to avoid doing it. For example, I tell you, if you try and put say in your example of contact center, if I and use a voice model like GPT real time or Gemini live native audio or even 11 Labs, chances are that you are solving around or building around a lot of those deficiency around say VAD or tool calls and stuff. And you're doing a telephone integration. What and you are going and putting a lot of engineering cycles and solving some of these things, right? But think of it like this. It is quite possible that next month a model comes around and and solves all of those things natively that you would have put an engineering cycle to fix. Then your leverage, your mode that was uniquely differentiating your solution is gone. So I think this is the biggest sort of question in my mind that how do you when you are if you're not a model builder. How do you interplay with the model? How do you build something that your mode does not get vanished with a new release? And we've seen that play out with cursor. don't know time will tell how the valuations play out. But looking at anti-gravity, you know, it's practically available. What was it? $40 or $100 subscription is not free for everybody."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 788,
      "timestampFormatted": "13:08",
      "text": "Yeah, yeah. Look, I think I have a slightly contrarian view on this because I know some of you know, in essence, a lot of people have downplayed these applications, sometimes calling them rappers and different names to some of them, right? Look, think models, in my opinion, are extremely good at generalizing and they will continue to get better at generalizing. I don't think any of the frontier models are looking at making their models better at specializing. So, I still feel that there is wide space when it comes to certain extremely highly valuable specialized tasks. Now, I think one of that examples are coding. There could be other such examples like marketing could be a very big example. These models could get really good at creating assets, but you still need someone to twist and turn these models in order to make them better at creating assets and creating assets that your brand can use confidently in the open world and generate leads out of that, generate business out of that. And I think that is always going to be a gap. I don't think the frontier models are going to spend so much time and money trying to make their models better tasks because they're chasing a very different game, right? They're chasing AGI. And the whole point of that is to make it generally intelligent. and to make it specially intelligent. So I think I still feel there is room for companies like Kazur or other such companies who will create applications on top of this because look if it was so simple to take a model and fork VS code on top and ask you to choose your model and run let's say code on it. I think a lot of people would have done it. We've seen a lot of commodity IDEs come out but it's not the case. Data has shown us that there have been only few companies that have been able to do this really well. All the front-year companies were eyeing for this. OpenAI tried to place a bet on Winsorff, Google picked it up. Cursor also has investments from Anthropic, has investments from Google. So I think somewhere these companies, the front-year model companies are realizing that this particular use case, is significant dam to the market. There's a big enough market. It's a market that's highly autonomous because developers can take their own decisions. They can buy And as a result of that, there's a lot of money on the table and I think Kerser has done a fabulous job of acquiring a lot of that money, which means But I think what Kerser has done to their credit is they've created more developers in the ecosystem So and that I think is a great place to be which means if let's say Kerser was I don't know what's the best analogy, right? But if Kursor is, let's say, a screwdriver for opening a certain set of tools. You can think of the smaller IDs or other companies that will come in that will become just like a Swiss knife. So you may not need it. You may only need it when you are really looking at multiple use cases. for example, someone like me, Let's say I want to build a small app tomorrow. Don't know the bells and whistles. I don't know much about it. I think... A smaller ID with much, with just the models helping me out would be a much better use case. And I would go to a free tool like this. But if I need, let's say a lot more work on my application, then maybe cursor is the choice. I don't know. Again, I'm just guessing, but I feel somewhere the market will start to segregate and the serious people will go after the software that which helps them solve a lot of problems. And the hobbyists is also a big market and they will go to some of these other tools. Right. it'll start segregation will happen."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 999,
      "timestampFormatted": "16:39",
      "text": "Yeah, I think that's that's not make sense. And I agree. think there is a place for you know, lot of these tools to exist given in an org, there are different personas and all of them come from a different maturity model in terms of their understanding of building and writing code. And so there'll be lovables and bold and replete of the word and cursors and anti-gravity and co-pilots that will coexist. And then there'll be a certain set of users who are comfortable, you know, spinning Gemini CLI or cloud code."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 1021,
      "timestampFormatted": "17:01",
      "text": "Absolutely."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 1033,
      "timestampFormatted": "17:13",
      "text": "in Cloud CLI in the terminal and write code autonomously. Now, let's switching gears, Achan. You've sort of debated on this. You've been a startup co-founder. You've had a successful stint at startups. How do you... And this was obviously pre-AI era. And when a lot of this... The models didn't exist. How do you think when you think about the building startups, building, creating value, how do you think  has building startups become a job of a startup founder or has become easier given there are so much help around with these language models, agents and so on, right? How do you reimagine doing startup or has it become easier? And if yes, then why aren't we seeing an explosion of startups? What's your thought on this?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 1093,
      "timestampFormatted": "18:13",
      "text": "Yeah, think I mean to look first point.  Yes, it has definitely. I think it has opened doors for a lot of companies to get created. But at the same time, I think there is an irony to the situation. The irony is that if both you and I can create the same app with the same idea, which means we'll end up just flushing the market with. the same commodity across and the value of that commodity will drop eventually. And I think that's what's happening. And more and more such companies will get killed. So there will be a massive filtration process that will happen. because you have to think about it this way. if I let's say 10 years back, if I had started a company, there was substantial friction to overcome in order to build a product. I had to hire the right set of people. I had to have the right set of skills. I had to understand how to deploy these applications, all of that, right? And those generally would have taken some amount of time in order to launch the product. Now with AI, I think a lot of that has shrunk. That timeline has significantly shrunk. So what will happen now as a result of that is that you will find a lot of lookalike apps. So... then the challenge is that when you are flooded with so much of lookalike type of applications, how do you decide who is more valuable? And the only answer to that is revenue. Unfortunately, there's no other better answer, right? So if your app has more revenue than mine does, you're obviously more valuable than me. And my app automatically gets killed. So what I think is going to happen is that There are stages of a startup, right? There's idea, there's build and then there is go to market, all of that. I think what's happening is as companies, as startups, we're reaching faster to go to market. So the rubber is hitting the road much faster than it used to, much earlier than it could. the skills of founders are now... So you cannot just build an app anymore and live with it. You have to have a strong GTM structure in order to sell that app. And maybe I can draw a leaf here and say that if I'm sure even Kerser and all these companies, they have a very strong GTM team that's trying to secure customers, right? They're going after these customers. They're not assuming that the market is theirs. And I think that's what is going to stand out eventually, right? Like the one who acquires the market faster, who's able to show revenue on the books. I think that company is going to be valuable more. But let's flip it around a little bit. Let me ask you this question. You've been leading engineering teams for a while now. You've been doing this for a very long time. And I think how do you see that? Because I want our listeners to see both sides of it. While the startup side is one angle and I think that's overblown, it's hyped enough, people know enough about it. But I think what people don't know is and I also want to set some reality baselines here. A lot of people think that because AI is so powerful, they're able to build apps and they will immediately go and start building companies out of it. But I also want them to understand that it's also going to help them in their jobs. That's where, know, so what do you think about that, right? Having held these engineering positions."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 1287,
      "timestampFormatted": "21:27",
      "text": "Sure, yeah, I'll definitely answer that. But to your comment on startups, believe I agree. The way I see this is building products has become easier. Now selling or differentiating them has become a lot more difficult given there's so much competition and probably raising capital, all of those challenges remain, right? So yeah, that's what it seems to me, you know, listening to. Now coming to how I see on how this is panning out on the engineering side of things, right? I think I've gotten a lot more comfortable seeing how this whole thing is panning out and there's a place for an AI, there's a place for a developer. There was a lot of sort of paranoia or people wondering if there's a place for one, how both will coexist. But I think I believe, you know, what we have seen and I think the theory that I like to, you know, quote is from Balaji who says that AI is definitely not, at least today's state of AI is not good enough to do end to end. What it's good at is doing middle to middle, right? Which is essentially, it will probably not do something start to finish. It takes a very sort of talented person to understand the problem and business domain, craft out a prompt, craft a context, provide the right set of data sources to a very capable model, tune it to their output and then let AI do its thing. And then it takes another sort of smart human to understand the output from the model, evaluate whether it's relevant and then sort of put that in product, right? So in the... Context of engineering, I believe there's a Laffer curve of AI, which is essentially the optimal amount of AI in most things cannot be 100 % and cannot be 0%. Because if you are using 0 % AI today, if you're a developer and you're not using AI, if you're a tester, if you're a data engineer, if you're an analyst, you're an ML engineer, if you're using 0 % of AI, you're considered slow. But if you rely on 100 % of AI, then you're producing slop. Right. Which basically means that exact number for your role is situational. Now, know, but knowing that it is definitely not zero percent and a hundred percent, it is it is somewhere in the. And what that heuristic is, is that something that every person needs to figure out that what is the optimal amount of AI in what I do. So I believe that developers"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 1435,
      "timestampFormatted": "23:55",
      "text": "Like a Goldilocks zone"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 1451,
      "timestampFormatted": "24:11",
      "text": "you know, are able to very quickly sort of produce very, very high quality code. And I've seen that in the definitely on the UX side. I have seen a lot of new cool websites coming up overnight and a lot of, you know, cross platform apps being showing up on app stores. But at the same time, I also have my skepticism, right? Because I believe that a lot of the AI generated code. At some point, my fear is a lot of people are becoming so reliant on AI that they are not really doing a PR review when they hit on that accept merge or accept change. They have the trust on AI is becoming so much that at some point if something is going to break, they won't be in a position to break this, right? So which basically means that I'm not afraid of today because the... The people who are in the engineering shops have seen this from scratch. So they are able to guide and maneuver AI in the way they want because they understand the domain because they've understood. But think of people who are just entering. First thing is there is not enough demand for that entry level positions. We've seen numbers, numbers, the demand for entry level positions shrinking because the baseline AI is better than a fresher. But if you let's say cross that barrier and you become a domain specific. So I believe that when this set of people are going to say out of retire or go out of so when they graduate from the job maybe five, 10 years, the number of domain experts required 10 years out are going to be significantly less because people would have who have been very much reliant on AI would not be able to debug this. because they don't know how to code the hardware, they don't know how to code in an editor without the copilot or and as much as won't be able to write a documentation or so on, right? So that's so I'm optimistic but equally worried that how this will all pan out in  a couple of years."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 1578,
      "timestampFormatted": "26:18",
      "text": "so any final ending notes that you want to close on?"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "speaker-0",
      "timestamp": 1581,
      "timestampFormatted": "26:21",
      "text": "Absolutely. again, we will try to because this is our first episode, we'd love to see your support on the AI Valley podcast. This is  Vikas and Sachin. We cover all things AI agents and also what we see in the enterprise space. we'll be covering, we'll try and push an episode every week, at least to begin with. And as always, everything that we speak is our own opinion, does not represent the opinions of our employers. And thank you for listening and have a great"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "speaker-1",
      "timestamp": 1617,
      "timestampFormatted": "26:57",
      "text": "week. All right. Thanks, everyone. Have a good one."
    }
  ],
  "chapters": [
    {
      "title": "Hi, very good evening",
      "timestamp": 52,
      "timestampFormatted": "0:52"
    },
    {
      "title": "I agree absolutely and it is absolutely and I'm gl...",
      "timestamp": 248,
      "timestampFormatted": "4:08"
    },
    {
      "title": "Yeah, full party",
      "timestamp": 441,
      "timestampFormatted": "7:21"
    },
    {
      "title": "Yeah, I think I agree",
      "timestamp": 643,
      "timestampFormatted": "10:43"
    },
    {
      "title": "Yeah, yeah",
      "timestamp": 788,
      "timestampFormatted": "13:08"
    },
    {
      "title": "Yeah, I think that's that's not make sense",
      "timestamp": 999,
      "timestampFormatted": "16:39"
    },
    {
      "title": "Sure, yeah, I'll definitely answer that",
      "timestamp": 1287,
      "timestampFormatted": "21:27"
    },
    {
      "title": "Like a Goldilocks zone you know, are able to very ...",
      "timestamp": 1435,
      "timestampFormatted": "23:55"
    },
    {
      "title": "week",
      "timestamp": 1617,
      "timestampFormatted": "26:57"
    }
  ],
  "quotes": [
    {
      "text": "I think in general, there was a lot of talk, there was a lot of anticipation about Google's next model.",
      "speaker": "Sachin",
      "timestamp": 104,
      "timestampFormatted": "1:44"
    },
    {
      "text": "Honestly, just from a personal side, I thought it should have come out with a little bit more fanfare.",
      "speaker": "Sachin",
      "timestamp": 104,
      "timestampFormatted": "1:44"
    },
    {
      "text": "I think what Google has done with this model, it has shown the world that we're not anywhere close to plateauing with the so-called intelligence layer or with what these models can do.",
      "speaker": "Sachin",
      "timestamp": 104,
      "timestampFormatted": "1:44"
    },
    {
      "text": "But what I really liked,  and I think Here's where I see Google coming up with its model of how to understand words better.",
      "speaker": "Sachin",
      "timestamp": 104,
      "timestampFormatted": "1:44"
    },
    {
      "text": "think that is super surprising to me because I think where Google started showing, where I'm really visibly surprised is in the screenshot.",
      "speaker": "Sachin",
      "timestamp": 104,
      "timestampFormatted": "1:44"
    }
  ],
  "fullText": "[0:52] Vikas: Hi, very good evening. Welcome to the very first episode of the AI Valley podcast. My name is Vikas and I have with me\n\n[1:01] Sachin: Hey everyone, good evening. This is Satya Nair.\n\n[1:05] Vikas: Cool, all right, so this is the brand new episode of our podcast. So we have a lot of topics to cover, Sachin. We've been talking about this quite a lot. The world of AI is buzzing. There lot of new releases every week. As we've seen, companies are rolling out models, applications, and there's a lot happening in this space. So what has kept you busy? What are your thoughts on the new? Gemini 3 release and the excitement that you see around things like Nano Banana and everything else that's happening in the world of AI.\n\n[1:44] Sachin: Wow, that's such a loaded question and maybe we can talk through this over the course of the next few minutes. I think in general, there was a lot of talk, there was a lot of anticipation about Google's next model. And I think it's kind of lived up to it in a way. There's a lot of rave reviews coming out about Google's model. Honestly, just from a personal side, I thought it should have come out with a little bit more fanfare. But that aside, I think it seems like a major reset moment in my mind. I think what Google has done with this model, it has shown the world that we're not anywhere close to plateauing with the so-called intelligence layer or with what these models can do. I think there's a lot more that's there. There's so much white space, I think that Google has shown as a result of this model. And I'm surprised, I mean, while I'm not a big believer in benchmarks, and I've always had my apprehensions of looking at these benchmarks, but there are a few line items on that benchmark which kind of stand out, For example, on the ARC price, I think it's amazing to see the leap that we have taken with these models. And it's almost... in the range of around 50-60 % of a leap from what the earlier models were doing. And it's not just in the leap in terms of what Google is doing, it's also in the leap of what the other models were. And so I think these are good explanations of how these models are doing better. There's humanities last exam, think stellar results, there's maths, you know, some of these questions are even tough for humans. And it's amazing to see what these models are doing. But what I really liked,  and I think Here's where I see Google coming up with its model of how to understand words better. think that is super surprising to me because I think where Google started showing, where I'm really visibly surprised is in the screenshot. Yeah, absolutely. And yeah, mean, why don't we talk about that? that's like a... To me, that's a very big moment because that somehow... tells me that this opens up things like robotics, visual elements, know, these models are always good with text and they'll continue to be, but imagine fixing some of these models to a pair of eyes or a pair of ears, right? And all of a sudden, it's everything's so different.\n\n[4:08] Vikas: I agree absolutely and it is absolutely and I'm glad that you mentioned because I I think it was the great reset in terms of, you know, there were a lot of apprehensions. We just had NVIDIA's earning call and you and I were discussing just about last week and probably the whole of last month that a lot of people have, we know, have been talking about the bubble because of this infrastructure and we thought that it'll show up in the result, but I don't think the market is anywhere near slowing down. In fact, I'll tell you something, the launch of Gemini 3 and anti-gravity, right? To me, I think it was very profound. To me, it replaced the the Wintel era of the 90s, which is about three decades later, As we've seen in 90s, the tech industry was defined by the symbolic relationship known as Andy and Bill's Law. which is basically Andy Grove of Intel would give us a faster chip and Bill Gates of Microsoft would immediately give us software features that consume that new power, right? So because deep down Intel knew Microsoft used those cycles, so they built the fab and Microsoft knew Intel would provide those cycles and they wrote the code. Now, I think for me, why this is profound is we're seeing this exact dynamic play out with AI, right? but at a much larger scale. Take Gemini 3 for instance. When TPU-V7 Ironwood was released in April 2025, the numbers looked like absolute overkill. 4614 teraflops of FP8 performance and 192 GB of HBM3 memory per chip. That was about 5x leap in compute and about 6x increase in memory capacity over the previous Trillium generation. And I think I was following a lot of thread then specifically critics. One of the first thing they were asking is who need the 9216 chip super pods with so much density? Do we have models utilizing this sort of capability capacity? And I believe  Google has shown it and I believe Gemini 3 is the answer. Anti-gravity for instance spins up autonomous developer agents. which holds the entire code base in context and while continuously reasoning on it. So I believe it justifies, when I look at these two products, it justifies every bit of Ironwood's architecture. If Ironwood was Andy's move, then Gemini 3 is definitely the Bill's move. Right? And  we've seen Google, TPU team built out a 42.5 EXAFLOP AI supercomputer because they knew Gemini team was building a frontier reasoning model that would finally crack the 1500 D-Low barrier, right? And it needed every flop of it. So it is definitely special in terms of what's happening. But again, I think that's about Gemini. But I think there were some interesting releases as well. you just last week, think Grok released the 4.5 or rather 4.1.\n\n[7:21] Sachin: Yeah, full party.\n\n[7:22] Vikas: One model and I think that's performing really well as well. The prior week, I think last week OpenAI released 5.1 made their model so much warmer, right? And so I think the pace of innovation is relentless. Now, I think my question to you, Sachin, think given so much is happening, right? How do you see these things play out, right? With all of these capability reaching enterprises, startups, how is the world changing? How are the software and the services that we know of is changing with this new, riding on this new innovation?\n\n[7:59] Sachin: No, think, look, that's a very interesting question. I think what most of these models are trying to achieve with the way they are coming up with all the new innovations, think software and applications as we know it will suddenly have a very different face. And if these models continue to perform the way they're performing and if they continue to scale the way they're scaling, Somewhere I get a feeling that a lot of applications as we know it could become very headless. I don't know, right? Like just say for example, a contact center application. That's the easiest that comes to my mind right now. Think of it autonomously running through voice or running through this whole agentic system. Like if a model is so good at looking at a screenshot of something and saying what the problem is. or understanding the world of that screenshot. The possibilities are endless, Like customer support could become a visual thing more than an auditory thing where it is today. All of a sudden these things can change, right? Like you're talking like, let's say sales calls, for example. You could see people doing sales calls very differently because now instead of just auditory responses, they can have visual responses as well.  And I think this whole anti-gravity that Google has shown, all of a sudden it's packed in the browser, it's packed in coding, it's packed in the agentic system. I mean, it's just insane to think about what are the possibilities and how these enterprises could really use some of these systems. But I think it's also going to be super important to see where are these enterprises actually ready for this. And how much are they ready for this? Because while a lot of these frontier models are doing what they have to do, the data is still sitting in its old messy format. Or in its old format, right? Let's not call it messy. It was doing what it was supposed to be doing. And these data structures were constructed because of the systems that querying these data structures. I think what's important is what we need to figure out now is we need to think back is that this data structure itself is changing. It's no longer the traditional data structures that we were familiar with because these models don't look at the traditional data structures. They don't understand some of these traditional data structures and they have their own structures. So are companies prepared for it? Do they have what it takes to in order to build it for the next generation? And I think that's where we'll see a lot of disruption. because companies that are ready to prepare their data faster in a manner that these LLMs can understand it, I think are going to be the ones that will succeed faster than the ones that will take time to do it. But at least that's the initial reading I'm getting from enterprises.\n\n[10:43] Vikas: Yeah, I think I agree. one of the biggest conundrum that I have when I look at these things is depending on is the model versus application question, right? Given, you know, what we saw with, say, let's say cursor and replate and many other application layer innovation that happened earlier this year and they are accruing a lot of valuation and usage and rightly so. But I think what we have seen in last month is where the value accruing or models getting increasingly popular and people now questioning that if the model companies are accruing more value. So one of the things that is really hard to do when you're solving or disrupting on on some these enterprise use cases is how do you ensure that whatever you're building does not get solved by say next iteration of the model. As model becomes powerful. It's one thing to say that you build with the model, don't build, but it's another thing to sort of. to avoid doing it. For example, I tell you, if you try and put say in your example of contact center, if I and use a voice model like GPT real time or Gemini live native audio or even 11 Labs, chances are that you are solving around or building around a lot of those deficiency around say VAD or tool calls and stuff. And you're doing a telephone integration. What and you are going and putting a lot of engineering cycles and solving some of these things, right? But think of it like this. It is quite possible that next month a model comes around and and solves all of those things natively that you would have put an engineering cycle to fix. Then your leverage, your mode that was uniquely differentiating your solution is gone. So I think this is the biggest sort of question in my mind that how do you when you are if you're not a model builder. How do you interplay with the model? How do you build something that your mode does not get vanished with a new release? And we've seen that play out with cursor. don't know time will tell how the valuations play out. But looking at anti-gravity, you know, it's practically available. What was it? $40 or $100 subscription is not free for everybody.\n\n[13:08] Sachin: Yeah, yeah. Look, I think I have a slightly contrarian view on this because I know some of you know, in essence, a lot of people have downplayed these applications, sometimes calling them rappers and different names to some of them, right? Look, think models, in my opinion, are extremely good at generalizing and they will continue to get better at generalizing. I don't think any of the frontier models are looking at making their models better at specializing. So, I still feel that there is wide space when it comes to certain extremely highly valuable specialized tasks. Now, I think one of that examples are coding. There could be other such examples like marketing could be a very big example. These models could get really good at creating assets, but you still need someone to twist and turn these models in order to make them better at creating assets and creating assets that your brand can use confidently in the open world and generate leads out of that, generate business out of that. And I think that is always going to be a gap. I don't think the frontier models are going to spend so much time and money trying to make their models better tasks because they're chasing a very different game, right? They're chasing AGI. And the whole point of that is to make it generally intelligent. and to make it specially intelligent. So I think I still feel there is room for companies like Kazur or other such companies who will create applications on top of this because look if it was so simple to take a model and fork VS code on top and ask you to choose your model and run let's say code on it. I think a lot of people would have done it. We've seen a lot of commodity IDEs come out but it's not the case. Data has shown us that there have been only few companies that have been able to do this really well. All the front-year companies were eyeing for this. OpenAI tried to place a bet on Winsorff, Google picked it up. Cursor also has investments from Anthropic, has investments from Google. So I think somewhere these companies, the front-year model companies are realizing that this particular use case, is significant dam to the market. There's a big enough market. It's a market that's highly autonomous because developers can take their own decisions. They can buy And as a result of that, there's a lot of money on the table and I think Kerser has done a fabulous job of acquiring a lot of that money, which means But I think what Kerser has done to their credit is they've created more developers in the ecosystem So and that I think is a great place to be which means if let's say Kerser was I don't know what's the best analogy, right? But if Kursor is, let's say, a screwdriver for opening a certain set of tools. You can think of the smaller IDs or other companies that will come in that will become just like a Swiss knife. So you may not need it. You may only need it when you are really looking at multiple use cases. for example, someone like me, Let's say I want to build a small app tomorrow. Don't know the bells and whistles. I don't know much about it. I think... A smaller ID with much, with just the models helping me out would be a much better use case. And I would go to a free tool like this. But if I need, let's say a lot more work on my application, then maybe cursor is the choice. I don't know. Again, I'm just guessing, but I feel somewhere the market will start to segregate and the serious people will go after the software that which helps them solve a lot of problems. And the hobbyists is also a big market and they will go to some of these other tools. Right. it'll start segregation will happen.\n\n[16:39] Vikas: Yeah, I think that's that's not make sense. And I agree. think there is a place for you know, lot of these tools to exist given in an org, there are different personas and all of them come from a different maturity model in terms of their understanding of building and writing code. And so there'll be lovables and bold and replete of the word and cursors and anti-gravity and co-pilots that will coexist. And then there'll be a certain set of users who are comfortable, you know, spinning Gemini CLI or cloud code.\n\n[17:01] Sachin: Absolutely.\n\n[17:13] Vikas: in Cloud CLI in the terminal and write code autonomously. Now, let's switching gears, Achan. You've sort of debated on this. You've been a startup co-founder. You've had a successful stint at startups. How do you... And this was obviously pre-AI era. And when a lot of this... The models didn't exist. How do you think when you think about the building startups, building, creating value, how do you think  has building startups become a job of a startup founder or has become easier given there are so much help around with these language models, agents and so on, right? How do you reimagine doing startup or has it become easier? And if yes, then why aren't we seeing an explosion of startups? What's your thought on this?\n\n[18:13] Sachin: Yeah, think I mean to look first point.  Yes, it has definitely. I think it has opened doors for a lot of companies to get created. But at the same time, I think there is an irony to the situation. The irony is that if both you and I can create the same app with the same idea, which means we'll end up just flushing the market with. the same commodity across and the value of that commodity will drop eventually. And I think that's what's happening. And more and more such companies will get killed. So there will be a massive filtration process that will happen. because you have to think about it this way. if I let's say 10 years back, if I had started a company, there was substantial friction to overcome in order to build a product. I had to hire the right set of people. I had to have the right set of skills. I had to understand how to deploy these applications, all of that, right? And those generally would have taken some amount of time in order to launch the product. Now with AI, I think a lot of that has shrunk. That timeline has significantly shrunk. So what will happen now as a result of that is that you will find a lot of lookalike apps. So... then the challenge is that when you are flooded with so much of lookalike type of applications, how do you decide who is more valuable? And the only answer to that is revenue. Unfortunately, there's no other better answer, right? So if your app has more revenue than mine does, you're obviously more valuable than me. And my app automatically gets killed. So what I think is going to happen is that There are stages of a startup, right? There's idea, there's build and then there is go to market, all of that. I think what's happening is as companies, as startups, we're reaching faster to go to market. So the rubber is hitting the road much faster than it used to, much earlier than it could. the skills of founders are now... So you cannot just build an app anymore and live with it. You have to have a strong GTM structure in order to sell that app. And maybe I can draw a leaf here and say that if I'm sure even Kerser and all these companies, they have a very strong GTM team that's trying to secure customers, right? They're going after these customers. They're not assuming that the market is theirs. And I think that's what is going to stand out eventually, right? Like the one who acquires the market faster, who's able to show revenue on the books. I think that company is going to be valuable more. But let's flip it around a little bit. Let me ask you this question. You've been leading engineering teams for a while now. You've been doing this for a very long time. And I think how do you see that? Because I want our listeners to see both sides of it. While the startup side is one angle and I think that's overblown, it's hyped enough, people know enough about it. But I think what people don't know is and I also want to set some reality baselines here. A lot of people think that because AI is so powerful, they're able to build apps and they will immediately go and start building companies out of it. But I also want them to understand that it's also going to help them in their jobs. That's where, know, so what do you think about that, right? Having held these engineering positions.\n\n[21:27] Vikas: Sure, yeah, I'll definitely answer that. But to your comment on startups, believe I agree. The way I see this is building products has become easier. Now selling or differentiating them has become a lot more difficult given there's so much competition and probably raising capital, all of those challenges remain, right? So yeah, that's what it seems to me, you know, listening to. Now coming to how I see on how this is panning out on the engineering side of things, right? I think I've gotten a lot more comfortable seeing how this whole thing is panning out and there's a place for an AI, there's a place for a developer. There was a lot of sort of paranoia or people wondering if there's a place for one, how both will coexist. But I think I believe, you know, what we have seen and I think the theory that I like to, you know, quote is from Balaji who says that AI is definitely not, at least today's state of AI is not good enough to do end to end. What it's good at is doing middle to middle, right? Which is essentially, it will probably not do something start to finish. It takes a very sort of talented person to understand the problem and business domain, craft out a prompt, craft a context, provide the right set of data sources to a very capable model, tune it to their output and then let AI do its thing. And then it takes another sort of smart human to understand the output from the model, evaluate whether it's relevant and then sort of put that in product, right? So in the... Context of engineering, I believe there's a Laffer curve of AI, which is essentially the optimal amount of AI in most things cannot be 100 % and cannot be 0%. Because if you are using 0 % AI today, if you're a developer and you're not using AI, if you're a tester, if you're a data engineer, if you're an analyst, you're an ML engineer, if you're using 0 % of AI, you're considered slow. But if you rely on 100 % of AI, then you're producing slop. Right. Which basically means that exact number for your role is situational. Now, know, but knowing that it is definitely not zero percent and a hundred percent, it is it is somewhere in the. And what that heuristic is, is that something that every person needs to figure out that what is the optimal amount of AI in what I do. So I believe that developers\n\n[23:55] Sachin: Like a Goldilocks zone\n\n[24:11] Vikas: you know, are able to very quickly sort of produce very, very high quality code. And I've seen that in the definitely on the UX side. I have seen a lot of new cool websites coming up overnight and a lot of, you know, cross platform apps being showing up on app stores. But at the same time, I also have my skepticism, right? Because I believe that a lot of the AI generated code. At some point, my fear is a lot of people are becoming so reliant on AI that they are not really doing a PR review when they hit on that accept merge or accept change. They have the trust on AI is becoming so much that at some point if something is going to break, they won't be in a position to break this, right? So which basically means that I'm not afraid of today because the... The people who are in the engineering shops have seen this from scratch. So they are able to guide and maneuver AI in the way they want because they understand the domain because they've understood. But think of people who are just entering. First thing is there is not enough demand for that entry level positions. We've seen numbers, numbers, the demand for entry level positions shrinking because the baseline AI is better than a fresher. But if you let's say cross that barrier and you become a domain specific. So I believe that when this set of people are going to say out of retire or go out of so when they graduate from the job maybe five, 10 years, the number of domain experts required 10 years out are going to be significantly less because people would have who have been very much reliant on AI would not be able to debug this. because they don't know how to code the hardware, they don't know how to code in an editor without the copilot or and as much as won't be able to write a documentation or so on, right? So that's so I'm optimistic but equally worried that how this will all pan out in  a couple of years.\n\n[26:18] Sachin: so any final ending notes that you want to close on?\n\n[26:21] Vikas: Absolutely. again, we will try to because this is our first episode, we'd love to see your support on the AI Valley podcast. This is  Vikas and Sachin. We cover all things AI agents and also what we see in the enterprise space. we'll be covering, we'll try and push an episode every week, at least to begin with. And as always, everything that we speak is our own opinion, does not represent the opinions of our employers. And thank you for listening and have a great\n\n[26:57] Sachin: week. All right. Thanks, everyone. Have a good one.",
  "generatedAt": "2026-01-18T19:16:00Z"
}