If you think using Claude everyday is the same as having AI skills, you are not getting hired in 2026. My guest today, Surya Kari, is a Senior Genitive AI Data Scientist at Amazon. His day-to-day involves working with Fortune 500s and Frontier AI models that are reshaping the job market right now. I asked him what actually gets you hired in 2026, and his answer really flipped how I think about all of this. I mean, a lot of people use Claude or Copilot as autocomplete.

For you to be able to stand out, you should also be able to think through the output. The people getting hired are not doing what every YouTube video is asking you to do right now. So depth in a domain is actually hard for an AI to replace, and that is the actual moat that you build around you, because that is your insurance against a generalist AI. For folks like me that don't have that foundational knowledge that haven't spent so many hours doing these Udacity courses, what do you think people like us miss out on?

How can we get closer to how you interact with these AI tools? By the end of this conversation, I promise that you are going to learn exactly which AI skills get you hired in 2026, and which ones are wasting your time. And if you're a student, a new grad, or anyone trying to break in right now, do not make the mistake of clicking off. Please join me in welcoming Surya to the Ready, Set, Do podcast. Let's get into it. Surya, welcome.

So kind of want to jump off with talking about a generative AI data scientist. I think senior data scientists with generative AI, what's a day in the life look like? You know, maybe pick any of the days last week or something, and talk us through what, you know, just a day in the life at Amazon looks like for somebody at your role and somebody that does what you do. Yeah. Again, thanks for having me here. So I'm a senior data scientist, a generative AI data scientist at Amazon. I've been doing this for just a little over two years now.

My role is very unique within Amazon, primarily because my role is very customer focused. So a lot of my conversations during the day happen with customers around what they want to do with generative AI, what their problems are, what kind of issues they want to solve. I have customers that want to build their own foundation models. I have customers that want to solve a very specific problem with generative AI. I have customers that want to build brand specific chatbots.

I have customers that want to dabble in reinforcement learning. I even have customers that are frontier labs, where they want to build a multi-lingual mixture of expert models. So it varies from one end to the other. And a lot of my day goes in having these conversations with customers. And at least getting them started on this journey of getting into generative AI and building out these workloads. And if it involves us getting in and getting our hands dirty in terms of helping them build it, So that's a very brief 30,000 view of what I do.

We can get a little bit deeper into it if you want, but it's pretty much a day in life. Yeah, that's awesome. Yeah, wonderful jumping off point. And so obviously from here, I realize that you might not be able to take names and such. But when you say customers, right, can you talk a little bit more about who or what persona this customer might be just for us to kind of ground ourselves or provide context around what it is that you're helping them with?

Are these just sellers on Amazon or what are we talking about exactly? No, no, these are not sellers in Amazon marketplace. These are customers that are Fortune 500 companies to really line startups. A lot of them, I mean, there's all kinds of skill levels across companies. Like there are companies that have their own data science teams that are trying to understand if they have the latest or the greatest model that they want to experiment with.

And they have no other cloud provider has that deployed already. And they can't really do like an EPI based testing of it. So they sometimes come to us and say, hey, I want to experiment with that model. What's the best way to deploy it and test it? Right. So that's one way. That's one of the customers. There's some customers that come and say, I want, I have tons of data that I've collected over 20 years, 30 years. I want to build an SLM that is very domain specific or that's very, you know, that learns from the data that I've collected that I can open it up to my own customers.

There's customers that want to create a more cohesive experience for teams within its own There's no real one persona for a customer. And that's where that's where we come in, where we are, we're almost like white glove. We're like a team that gives white glove service to customers that want to get into generative AI. So it's very, it's very customized to every customer's requirements. So that naturally then starts to feel like you're having to wear just so many hats across the board for all of these various types of solutions that you need to provide.

So what is, how do you upskill yourself in terms of, you know, just not only obviously doing a good job at your day job, because that's the most important thing. But yeah, when you say that you're helping build generative models from scratch, also work with the frontier models, that sounds like it spans a huge, very wide bet of things that you're having to do. So is it is most of your upskilling done on the job? Or do you have a separate framework to, you know, continue to teach and evolve your AI skills?

I mean, today, a lot of my upskilling comes from the job because there's just so many things that I need to keep up with. But before I got into generative AI or even before I got into Amazon, a lot of my upskilling happened through, you know, platforms like Udacity, Udemy, and a lot of these platforms are actually great for you to learn, really dive deep into some topics that you're interested in. Udemy is a great path to the, again, I'm not getting paid for this. I'm just trying to tell you what I've experienced, right?

So Udemy has some really great micro courses that you can get started with. Udacity really kind of gets you farther into those topics and they're really great. I don't know how much they're keeping up with it. I haven't checked out the platform in a while. So please take it with a grain of salt. But these days, at least in the last two or three years, things have started to move really fast to a point where keeping up with what's happening has become harder and harder because you have one model today and there's another open source model that comes out tomorrow.

And if customers that come and ask me, "Hey, what about this model?" I'm like, "I've never heard of that model." That's when I learn about a model's existence, right? So, which is why I say a lot of my learning happens on the job today. One of the ways in which I try to keep up is, and I've always had a challenge, for keeping up. Yeah, I think I want to contrast here a little bit what your experience has been as somebody that's obviously in this field, in terms of an enterprise big tech company, where you're day-in, day-out, hands-on working with these tools, and somebody like me, right?

So I'm in a lot of ways the opposite end of the spectrum where all of my, like, first of all, I have no foundational knowledge with AI or data science or anything. I couldn't even tell you the difference between linear and logistic regression. That's just where my level is. However, I think it's interesting, right, that you and I probably, when we're using Claude, are probably doing a lot of the same things. So I guess, really, what I'm getting to is, for folks like me, right, that don't have that foundational knowledge that haven't spent so many hours doing these Udacity courses, building up a really, you know, like a strong fundamental base with this stuff.

What do you think people like us miss out on, I guess, or, you know, and what, how can we get closer to how you interact with these AI tools as somebody that knows in terms of the background that, okay, this is how this LLM tool works? Does that kind of make sense? I feel like I've, you know, went around a bit, but is that question kind of clear? I think there's a fundamental difference in how I use, and I'm not saying that how I use it is the way that everyone should use it.

I don't use Claude, I mean, a lot of people use Claude or Copilot as, like, autocomplete. But I think what is important is, before you start to use Claude or Copilot as autocomplete, you will, for you to be able to stand out, you should want to use it, you know, as a way to, you use the output from Claude, but you should also be able to think through the output from Claude or any other generative tool, right? So you should be able to think through if you're generating architecture, right?

So you should be able to understand, like, is the architecture making sense, right? You should be able to, if you're building large code bases with it, it makes sense that you try to understand, like, what each of the modules you're building flows into, because when there's a production problem, you don't want to be in a position where you don't understand the code that you've generated, right? So the way I try to use Claude or any other generative tool is I try to be as grounded as possible in terms of what it is generating, what I'm putting in production.

I try to have as much documentation as I can, and I try to document exactly how I run that particular set of code that I've created. And I also create meticulous test cases that anybody can access, right? And make sure that the code that you're generating is checkpointed. It has comments, and it has the ability for, even if somebody takes that code and puts it through another generative tool, that generative tool should have some context about what's going on with the code, right?

So that's how I use it, especially for code, is I make sure that it is meticulously documented, Because one of the important problems that I think a lot of enterprises have been flagging is production code is not commented. And when production systems fail, you want people to understand what fails. I want to quickly just pause just a second on what you said around the test cases piece. So there seems to be a lot of directives online that say that these tools are really good at also creating test cases, but it sounded like you do your own, like you still are involved with the test cases that you create. So if that is true, why is that?

Is that by design, and is that something prompt you to do that, or is it just kind of more like a force of habit? I mean, look, obviously, I use generative tools to create code in a faster way. But as far as when it comes to production systems, there is a disconnect between, let's say a module versus that module being part of a production system, right? And because of context lengths being limited, what happens generally in a production system is you have one module being generated in one session and another module being generated in another session.

Now, one of the bigger problems is when you create test cases in one session, those test cases might not actually translate to the whole of the production system. They might translate to the module, but they might not translate all the way down to the production system, right? So unit tests are great, but integration tests will fail, right? So you want to think about the entire production system, and that's where you need to keep in mind that that particular session might not generate the entirety of your test cases, right?

So you would want the model, sorry, you would want the test cases to always think about an integrated testing. So you might also want to think about what kind of integration test cases you want to always create and make sure that you create those integration test cases every time you create a new module, right? That's how I think about it. Gosh, that makes a lot of sense, actually, especially what you said around that.

Yeah, the unit testing seems almost trivial at this point, right? But it's when you start to do the entire integration testing, that's when you would start to see some gaps potentially, and then so naturally from there are so agents, right? Agents are all the rage now. It seems like in many ways agents are now what generative AI was first when it burst into the scene two years ago in 2024. Again, but from my lens, a lot of it just seems to be, you know, people trying to generate a bunch of outrage online just like, hey, I automated my 27 agent system and fired my Now my agents do all of my marketing while I sleep, blah, blah. So obviously this is what I see on X or on LinkedIn, right?

But what I don't get the chance to do normally is again, as I was alluding to earlier, talk to somebody that actually understands this stuff. So where do you land personally in terms of the agent hype train? Is it really as good as people make it out to be? Have you had a chance to play with them at all? What's your general stance on agents? Yeah, I think agents have really taken everybody's life storm. And I think that's understandable. I understand the hype.

I understand the velocity behind why agents have become such a big thing. And I also understand how agents can actually make entire systems more automated, right? A lot of people, I think, have a fundamental misunderstanding about what makes agents special. In my view, and this is my definition again, this is, but agents are different from, say, up a system that's built through a pipeline, like you can have a system that's a pipeline where X, Y, and Z get run based on certain tasks, right? So agents take non-deterministic decisions.

Like an agent, for example, it doesn't see the output based on the task that it is required to do. It can go back and redo some of the tasks. And then it can correct some of its own actions, right? That's the non-deterministic behavior that agents bring to the table, that pipeline systems do not have that capabilities, correct? So that makes sense in terms of like, hey, why agents are such a big hype? Now, agents are still in a very, in my view, again, in a very nascent stage of development.

There are, you cannot always have like an agent or a system that's multi-agent, you know, take complete control of your entire production systems. There's always, there have been instances where someone, and there's news the other day that Claude called completely deleted all of the production databases of an enterprise customer, right? And that's one of the big reasons why I think agents are not fully there yet, in terms of like taking complete control of everyone.

But at some point, I believe that they will get there, because as models get more reliable, as production systems get more, you know, they learn more from human feedback, they will get better in terms of reliability is what I believe. But we're not there yet, right? Agents are at the mercy of the models that power them, models hallucinate, and models also suffer from, so I'll give you an example, right? Long context models sometimes suffer from, if you give a long context model, drag access, and if you're using that as the model that is summarizing the data. Sometimes long context models lose their train of thought. It's a documented, there's papers written about, you know, how long context models forget some information while they're going through rank processes, right?

So, and there's several problems that we haven't identified yet. So are agents ready for prime time? But are agents going to get there? And there's a lot of research that's going into agent memory. There's a lot of open, there's a lot of open sourcing of protocols. You have the MCP server protocol that has been open sourced. You have Google setting industry standards. You have Quad setting industry standards. So there's a lot of industry wide collaboration when it comes to agents.

But I believe it's got a few ways to go before agents can actually take over our entire production systems. But they're still helping. I think if you have the right galleries in place, right, make sure that we do start to spin around in circles and do nothing except eat through all of your tokens, which is, it feels easy to say, but I think it's not as easily implemented in practice. So although there is something you said that jumped out at me about how agents are still at the mercy of the models that exist today. And so this is a bit of a random question.

And you know, I'm fully expecting you to be like, I don't know. But why hasn't somebody tried to make a model explicitly just for agents, given what we know about agents being, you know, just the new frontier, everything will be done by them, blah, blah. So why not just make a dedicated model for agents? Why am I like thinking about this incorrectly? No, we have, no, I mean, all of the models that power agents today are specifically built for agent behavior. I didn't know that. I'll give you an example, right?

So you can, when Claude says, if you ever go through Claude and if you or or chat GPD, you will see that it is sometimes saying, I'm calling X, Y and Z tool. For X, Y and Z purpose. What, what happens in the back end is you've trained a model specifically to call certain tools when it encounters certain costs, right? It can be a tool as a calculator. It can be a tool that will orchestrate your entire back end. It can be a tool that creates music. You build the tools. You give access to the tools. And you specifically build the model.

You fine-tune the model to you teach the model that it should actually look and call these tools. So we are building models that specifically are designed to be used with agents. In fact, all the reasoning that happens, like if you go through chat GPD or Claude, you'll see that, hey, I'm going to chart step one, step two, step three, step four, step five. And in this step, this is the reason why I've come to this conclusion.

So that is essentially models that have been fine-tuned with reinforcement learning to think through their own decisions. And kind of reason, like, why did I take that decision? Is that the right decision? So have we built models specifically for agents? But do those models also hallucinate sometimes? Do those models need guardrails? Because you remember that models for a large extent are very huge black boxes. You can't always understand what they're thinking or what kind of weights generate, what kind of output. So guardrails are really important.

And especially with generative AI and generative models and especially us getting into modalities that are more than just text, you want to have guardrails so people don't generate harmful content. And that's why I'm still saying that agents are still at the mercy of models and models are still not at a stage where we can fully rely on them. That, yeah, I love that last pretty much. A lot of what you just said was just absolute news to me.

I actually had no idea that this is something we were already doing. But it makes sense because we still run into the same barrier of just context limit, whether or not it's being used by an agent or just something else. And that's like the actual roadblock here. It's not the fact that it can't reason or it can't be non deterministic or whatever. It's just that there's a limited context window for it to have fun with. So yeah, that's really cool.

I know you said that you were at this specific, like what you do today, you only joined that a couple years ago. So I am curious about your journey before that point. So what were you doing at Amazon before that? And maybe if you could even touch on your master's journey, how was that like, what were you up to back in the day? Yeah, I don't want to date myself. So I don't want to talk about years of dates. But I have a master's degree in information systems with focus on data science from Oklahoma State. And I started off as an analyst with a bank in the Midwest where we were analyzing treasury accounts.

And we were trying to do all sorts of Monte Carlo analysis on just treasury. I mean, there's treasury reports and all of that stuff. It's been a while. And then I went on to work for a company called SAS, a statistical analysis software. It used to be quite big when I actually joined it. It was actually one of the more desirable companies to work for. I still think it is. It's about a great campus in Keryloch, Carolina. I had a blast working for them. We used to work for the hotel industry. We used to go to software for the hotel industry. And it was really immensely fun to work with them. Then I moved out to Canada for a few years.

Because I was here on NH1B and I couldn't build my own startup, which I really wanted to. So I moved out to Canada where I did a bunch of contracting and I also built up my startup. We built out a startup with a couple of-- I had a great co-founder. I also had my wife who was immensely talented. So all of us kind of create the startup where we were collecting customer behavior within stores using Edge hardware. So we deployed cameras that would connect it to Edge hardware, like in video devices.

because PII laws would not allow us to send this information to the cloud until-- so we couldn't send anything like faces or any personally identifiable information outside of the store. And there was a lot of other push and pull factors that came into it. But it was clear that with what's happening with the shutdowns and everything, there was nobody getting into physical stores. And at that point, nobody really needed this tech at least for a few years. So I shut it down and I moved to the United States.

And ironically, I moved to a team that was doing similar tech. It was a team that was called AWS Panorama eventually where they were building Edge hardware to deploy in retail stores, in malls, and do things like, you know, football counting, trying to understand things like what routes people take before they got to a checkout, what were the most-- so this was then going to be used in planogram planning. Essentially, planogram planning is something that retail stores do to understand where to place the most high-value products. So they could use this to essentially understand because that's what people are going through. So I went into that team. And we built some really interesting pipelines. We used NVIDIA hardware. We used NVIDIA software.

We used Greengrass to orchestrate the low-level Edge connections. We used NVIDIA Deepstream. We used NVIDIA Tau. We used PeopleNet. We used GroundingDeno to automate all of the labeling. So there's a lot that we did. Like, there's a lot of tech that we built for this. But then again, the push-pull factors with chatGPT coming on. And there was a lot of investment in generative AI. And there was a lot of push-pull factors. And then I came into generative AI at that point.

And that's how I came into this new team where I just told you the story about the customers. And that's where it all ended up in. Yeah, that is such a fascinating journey that you've had. I am curious something about, something you said earlier, which is from your entire journey, it's clear to me that, and you said to yourself that a lot of your knowledge around data science, machine learning, those type of things were self-taught.

And I think you referenced it was courses that you took online, mini courses on Udacity or Udemy. I am curious, knowing what you know now and living in the world that we live in today, if today you were trying to upskill yourself from skills of an analyst, which is what you started your career with, to where you are today, which is senior data scientist, how would you go about doing that upskilling? Would that still be the way you chose in your case?

Or would you go down a different path if you had to do it today? I think there's a fundamental difference. I just want to clarify here that my master's degree was, I had a specialization in data science. So my first job as a data scientist was doing exactly what I learned in university. So it wasn't completely new for me. I was actually trained to do it. So what I was going with this argument is, if I were new today, the way I would approach it is, I would build fundamentals first, they're not optional.

Data structures, algorithms, operating systems, networking, databases, distributed system basics, these are the things that AI, you cannot reason away and give that decision to an artificial intelligence system. When your model returns garbage, you need to understand if it's a memory issue, you need to understand if it's a batching bug, if you need to understand if it's a race condition, you need to understand the basics, that's where I would start.

When you go through courses, what I would recommend is learn to use AI tools, but use them deeply, not just on a surface level. What's the example of how one can use AI deep? I think we touched upon this a little bit before, but for example, let's say you're using AI tools to generate architectures or test suites or large code bases, you need to learn, one, what it is generating, you need to be able to understand that my AI tool generated X, Y, and Z modules, what each of these modules does, you need to learn to direct AI effectively, which means that you have the understanding of how to use that tool and how to use what the tool generates, and for that, you need to first understand how to do that yourself.

You need to learn how to trust the output and when to trust the output, you need to learn how to verify the output. I think the third one that I would absolutely recommend is to pick a domain and not be a pure generalist. I think the people that will be most valuable in the future are not people that say, I know Python, I know React, I know something else, but people that say, I'm the person who understands how LLMs are trained, or I'm the person that can build infrastructure around them, or I can understand edge CV systems end to end, I can understand X, Y, and Z more than any other AI system can do, because depth in a domain is actually hard for an AI to replace, and that is the actual mode that you build around you, because that is your insurance against a generalist AI.

The last thing that I would recommend is learn to build things and ship them, because it's important to learn that a GitHub full of real projects is worth way more than a certification somewhere. Your achievements can actually be how much GitHub code you've created, and that's one way to help the world understand that you have felt things, and that's a rare skill, because people can learn what AI does and talk about it, but building something and being able to put it out there for people to use, I think it's invaluable. I love what you said about the AI not being able to replace depth, but it just clicked right in place for me, because you're so right, and we've even talked about this, the way just with context windows and such, maybe this is not

then exactly Apple's to Apple's comparison, right? But it makes sense in my head how it would be easy for AI to just replace somebody that is just kind of jack of all trades, so to speak, because that is kind of what AI is also. So to be irreplaceable, then you would have to just go pretty deep into whatever domain that you choose. So I really love that that definitely is something that's going to stay with me. Moving forward here, what lies next in your career, right? Maybe can you talk through any projects that you're working on that particularly, you know, excite you? I understand, obviously, if you can divulge any details, but just broad strokes, just the kind of challenges that you are facing today, maybe you are, you know, you're aware that you'd be facing in the near future.

I think that would paint a really interesting picture for those of us that are on the outside looking into how big tech enterprises continue to use and evolve with AI models. Yeah, I think off the top of my head, I think there's a few things that I've been noticing as and I'm the team that I work on is worldwide. So there's a lot of conversations I have with customers around the world. And I think one thing that I've understood is the philosophy behind building models in the Western world versus the Eastern, like India or China, it's very different in terms of philosophy. A lot of the customers that I talked to, well, I mean, not talked to, but a lot of things that I noticed is in the Western world, there is, there is this race to bear the most capable, the largest, the most, the latest and the greatest frontier model

that is designed to replace X, Y and Z personas in the real world. There's, you know, there's, there's this race to actually not replace, I think that's the wrong word, but to augment slash, make it easier for certain personas to do their job. While in the Eastern world, there is this race to not bear the largest, like in the, in the, in the Western, in the Western world, there is this race to build the largest, most capable models out there. While in countries like India and China, what I believe is happening is they're trying to build reasonably small models, like maybe medium sized models that are being built for AI as a public infrastructure, which is AI as a way to augment people's lives, their, their interactions with the government, how they can reduce government

bureaucracy, how they can be more accessible to their people. And India has taken some of the first steps to actually create AI as public infrastructure, which I think is something I've been looking at very closely. I think so is China. So the philosophy here is very different. The East looks at AI and frontier models, not frontier models as, as, you know, as models as public infrastructure, while the West looks at trying to build the latest and the greatest and the biggest models. So there's two different philosophies that I'm looking at, and it's fascinating to see both of them kind of, you know, develop side by side. So that's something I've been noticing. I may be wrong. This may, this may change or this may be different tomorrow, but that's something I've seen. So take it for what it is. Yeah. So first of all, that is

really fascinating because this is not something that I was aware of at all. It is just not an idea that not only I'd never thought about, but also never really saw anywhere else. So I'm, you know, currently processing what you just said as, as I say this, but I guess the thing that I'm curious about from what you just said is, so that makes sense to me, right? On a like high level. Yeah. Like there's a reason why there is no whatever Claude Opus 4.6 that comes out of India.

It doesn't sound like anybody is trying to build that. And you said the reason for that is because we're not trying to have that to have any frontier model really that then can integrate into just pretty much everything, right? It doesn't matter if you're an analyst, if you're like a tax broker, right? Claude Opus can, but this is not something that we're trying. We're trying to focus more on the public infrastructure piece. I guess what is still not entirely clear to me is, yeah, maybe you said that this is something that you've been watching. So do you have any examples? So the only thing I can think of is stuff like Sarvam, Sarvam AI, if you've heard of that, is that kind of what we're talking about here? Can you just talk more about the India AI stuff? Like exactly what are

we doing and why? There's a project called Bhashni. There is, there was a great startup, there's several startups in India, which focus on India's multi-lingual capabilities. Yeah. There is the startups that focus on, so India's a lot of languages, I'll give you an example, right? India's a lot of languages. And imagine, and this is how I imagine it would be, imagine that Tamilian going to Punjab and being able to access the same government services, he would have access to in Tamil Nadu, in the same language as he would have in Tamil Nadu, without having the same frictions that somebody today would have, right? Gotcha. You're talking to a system that understands his language, that instantly understands that this person requires service in Tamil

and it understands that person's core problems and provides service back in Tamil. And this is immensely useful for farmers that in rural Tamil Nadu want to deal with a system that's developed by the central government who now have the onerous task to meticulously convert everything, every policy into several Indian languages and rather you have this one foundational model that is now translating all these policies into all of Indian languages in one's broad stroke and it's now a available and there's also, I'll give you another example, right? When one of the things that is a huge friction point for Indian politicians is what if somebody from, you know, from the south goes to the north and tries to talk in a political rally, how would they talk to people without the

language barrier? How would it be if they are able to talk in one language and it translates it automatically into the Assamese possibly. Somebody talking in Telugu is automatically translated into Assamese. So what happens generally in translation models that have been developed in the west is everything translates into English and then from English it translates into other languages, right? But that conversion to English because the languages are so phonetically different loses a lot of that original meaning. So now what Indian, and this is just one of the, this is obviously just one of the projects I'm talking about. There's a lot of phonetical meaning that's lost. So now we're trying to convert from one Indian language to another without having that

intermediary in between, right? You want to teach models how to reason in Marathi and explain in Tamil without having that bridge language in the middle, right? So that preserves some of the cultural aspects of these languages, which is what I was alluding to when I said AI as public infrastructure. Incredible. Yeah, I just had no idea that any of this was being done at all. So that's awesome, you know, and I really appreciate you sharing that kind of as a continuation to that, I guess, because a lot of my audience is actually split across India and the US, you know, it's a bunch of early career professionals slash students. And obviously, as we know, there is a lot of interest in AI, right? So there's I get this question a lot. And I'm assuming you probably do as well

that say you've just graduated your computer science degree will just take a toy example, just doesn't matter somewhere in India, you're fairly savvy, you kind of realize that you like programming, you have all the basics for a person, right? Whose goal is to just become dangerous with AI, right? Dangerous as in just actually have, you know, good fundamental knowledge, be able to potentially do startups or get employed. Just however, you know, use their AI skills, so to say, do you have thoughts around what might be a better route for them? Just and I know it's impossible to answer, right? Because everybody's situation is different. But I'm just trying to, you know, kind of put you on the spot here, right? And and be like, what do you think the way forward

is? Because I talk to folks that go to like that give gate, go to masters colleges in India, do really well for themselves, right? And of course, the US is an option. Europe is also an option. So how do you see this next step in terms of this young person today that's listening to us and is trying to figure out what might be the best path forward for them to learn AI? I think we touched upon this a little bit in the in the previous conversation where we were talking about creating a motor around yourself and you dive deeper into a domain, because you know, it gets I mean, you can always domain train a model, but there's always the more niche your domain is, the less AI can replace it. You know, today, at least AI just replaces boilerplate cord,

you know, replaces some scaffolding and stuff. I think the key observation that I've been seeing is coding is becoming a lot like editing or playing or directing. You know, it's more like, how do you direct an agent or a model to get you where you want to be, right? So, you know, what that has generally created in the honest picture on on somebody who's in software engineering today is, you know, entry level SWE jobs have kind of compressed bit. Yep. Right. I mean, tech in general has kind of pulled back a bit. But that's not true always. But, you know, that's the general understanding for the most part. Yes. But, you know, there's a lot that is, you know, that's, that's, there's also a lot of opportunity that comes out from this particular situation is,

you know, if you are one of the genuinely few engineers that learn how to architect, how to maintain, how to extend, you know, how, you know, you can genuinely create production systems. I think that gives you an edge over a lot of other people. And if I were new today, that's where I would start. Just again, try to get as deep into what domain as possible is how I'm going to look at it. And I think we touched upon this a bit before. But yeah, that's my, my argument still remains.

You know, that makes sense. And I like that it almost strips away the need for location to be a factor at all, right? Because it's almost as if it doesn't really matter where you are, as long as you do the thing you need to do, which obviously we did cover that. So I like that. I think that's a pretty unique perspective on that. But yeah, Surya, this has been so incredible. I feel like, you know, like my brain grew like two inches just by talking to you in the past hour, you have such a way of breaking down like really these complex things in a manner that's digestible by somebody like me, which is not who is not at all, you know, as savvy, right, as the things that you were talking about. And then it's only and in my experience, at least it's only the people that

really know what they're talking about that can do that. Because otherwise, you're just sifting through jargon, you kind of just go from one heavy word to another without being like, what does that mean, which I did not feel like that at all. So yeah, just wanted to share that. And truly, thank you so much for taking the time with it. This has been so awesome. Thank you.