From Writing Code to Orchestrating Intelligence with Raja Subramanian

Speaker A: Welcome to the Software People stories. I'm Shiv.

Speaker B: I'm Chitra.

Speaker C: And I'm Gayati.

Speaker A: We bring you interesting untold stories of people

Speaker B: associated with the creation

Speaker C: or consumption of software based solutions.

Speaker A: You'll hear stories of what worked

Speaker B: and sometimes what didn't.

Speaker A: You will also hear very personal experiences

Speaker B: and insights that would trigger your thoughts

Speaker C: and inspire you to do even greater things.

Speaker A: Before we jump into today's conversation, a Special thanks to PMpower Consulting for encouraging and supporting this show. Right from the start, with over a thousand person years of collective experience, PMpower has helped many organizations, teams and individuals chart out their own journeys of excellence and make progress. Uh, thank you PMpower. Now let's move to today's episode

Speaker C: in the series of AI and Me. We have invited special set of guests. The most impressive part uh, about AI is that we don't know where it's going to take. We all have seen the pacing, the frontier,

Speaker D: lots of conversation between how ah, countries are trying to uh, collaborate as to how to manage, build guard rails as well as create that learning ability in order to harness the AI's ability in the best of the ways. So uh, we ask very specific questions in terms of their aha moments. What is it that you have done in terms of contribution and how AI has made uh, human beings more human and so on and so forth. It's been a very, very interesting conversation with all our guests and I'm uh, sure you're going to enjoy them. In the next uh, version of my conversation with Raja Subramaniam. I had a very dimensional aspect of AI given that me and ah, Raja share a lot of common uh, ground in terms of clean code building uh, easy to accessible code. He talks about how legacy code refactoring has become easier and uh, thinking from coding has been completely separated. He gives several examples of m building my beacon to creating smaller LLMs which is much more easier. And he has this philosophy that we don't necessarily have to build everything from scratch. We can use what exists and make it better. He uses this epithet that countries are not created but just made better. And it is a very lively conversation. I hope you enjoy it. Hi Raja. So happy to have you back on software people stories. Yes, I'm actually going to um, talk in your domain and uh, I have a little bit small amount of butterflies in my stomach when I'm going to ask you questions around AI and you. For the last um, couple of years I think artificial intelligence has become the street cred and every meeting that you have that is what is being spoken. So we wanted to ask you a couple of questions around how has AI changed you? Specifically you as a developer, you as a business owner as well as you as a consultant. And as a fellow human being, he want you to kind of go through some of your interesting use cases. The purpose to understand is that sometimes these complex words, uh, feel very overwhelming for a set of people. And uh, for a set of people it will be like, oh, can we do this as well? So we wanted to kind of uh, throw some light on those areas. So

Speaker E: first of all, thanks a lot for this opportunity. So after a very long time I feel very humble to be part of this conversation. And uh, in fact you uh, wrote a lot about AI than anyone or equally. So I am one of biggest uh, fan I have even. I called you many a times that I like, like, like it a lot. Especially your uh, determination to write something with the natural intelligence, not with artificial intelligence. Many times if you have told this I liked a lot.

Speaker D: Thank you so much Raja.

Speaker E: So in fact it is a butterfly for me in my stomach, not you that I am talking to someone who is very knowledgeable in this area, bringing hands on experience and uh, um, especially in your. It is actually. It's not my domain. It is our domain. That's how I am seeing this. Yeah,

Speaker D: that is true. I think my bad. In terms of saying ours as a collective human intelligence. Right. How much are we actually harnessing this intelligence and where are we going to take it? So write it outset. Right? What is that one AI, you know, tool or technique that you use on a daily basis that you wouldn't have imagined two years, three years back.

Speaker E: So this um, this podcast is now, it is going public. So I don't want to name any organization but most of the engagements, uh, or organizations where I have worked in the past, naturally we have this kind of a thing called as a bad code. Right. You are from ThoughtWorks. You are one of the. You were right.

Speaker D: Yes, yes, yes.

Speaker E: Clean code is one of the things, uh, clean code is one of the drive of Thoughtworks. So whenever I go to various other organizations I would like always I wish to have this code to be like this. If I have this code like this, it gives us lot of flexibility to accommodate changes. That's what if I had to use Robert C. Martin's um, uh, dream state. Here is a code is something that should be very flexible to accommodate changes. The sense that you have to minimize the amount of impact the code will have. But every team definitely will have their own ambition that okay, this is how my code needs to be. But the thing that what uh stopped everyone in doing is that the amount of side effects, right? So if I take a code and if I attempt to refactor that there is a potential that it will create lot of side effects and in the whole process the team will abandoned in the midway. Is it really going to help us uh, in the long run? So thanks to a today I am able to make some of the legacy code into one of the dream state code which every software programmer will dream about. So this is what I'm actually if I had to use a Tamil, if you allow me to use Tamil in uh, uh, this podcast. So that's things. So that's uh. What do you call this? The dream comes true. Right. So those things are what I am able to achieve with AI and I'm able to do that at scale. That's the most important thing. It is not a small thing. We can handle multiple projects and we can uh, refactor those codes and in fact we can make that code into such a great kind uh of a code which I will always like to have. So that is one thing that is what I like a lot that it changes the way that I am um, working with legacy codes.

Speaker D: Wow, that's very cool. In terms of you know you are working with several organization, you are almost time slicing across multiple organizations to create a clean code and that to something that is maintainable. Right. In terms of one of the first things that the clean code talks about is that are you able to make it you know uh, maintainable, scalable and uh, building it for the future.

Speaker E: Because all my forefathers, I mean not my is our forefathers of software industry, what they wrote repeatedly is again and again is how to write a clean code. Whether it is Robert C. Martin or Michael Feathers of working with legacy code. Every book that is coming in the software industry is. Is all a manifestation of some kind of a uh, pain they undergone long back. Whether it is a unit test that has been created over a flight, everything is actually a kind of the pain. So now AI is actually helping us to address that pain. And uh, now another interesting thing is that it helped us to decouple really the design brain of ours from the coding brain. So many times we know right In a pad programming you are bringing all this kind of a tradition. That's what you and SK is always my favorite. So you are all the pad programming came uh from the pad programming world. That is when someone is actually writing A code another person is actually thinking. And uh, when he will get an idea, then he'll grab the keyboard. And then when that guy will actually try to, uh. What is that? He'll try to implement his ideas. And, uh, he wear multiple hats, right? So one is actually he'll wear a domain hat. How to make this particular feature a great feature in terms of the domain and also how to make the code, which is a very clean code. And so both these things that what they are wearing. So most of the time in a coding, it is not a simple mechanical action. Right? So it is just, it is getting all these ideas, kind of an orchestration of multiple ways in which I can make my code a method code. Now, thanks to AI, this is where actually the padding is coming. So I am able to decouple my thought process from out of the mechanical way of creating functions and passing the parameters. And I can do that tirelessly. So once, um, I actually ventured into some path, then I'll get a new idea. Okay, so let us all encapsulate this like this. Let us move this here, let us move it there. All these things I can just keep on doing, doing and doing tirelessly. And uh, whenever I'm making a modification to the code and I am seeing that, okay, this is up and running without any side effects. And also the code is also becoming very clean, which otherwise I have to do with around, uh, 10 people. Or if you have to use a brick, if you are a word, and if you have to do it with some three bricks or four bricks, I can just do that with a simple cloud. I think I answer your question that I am a big fan of plot and uh, I can do that with cloud. I am just telling that the cloud is something that what I am using. I am not underestimating that another LLM is inferior or superior to this. I like this.

Speaker D: Absolutely. There is no right or wrong in that. In fact, all of us have our own preferences and we are not trying to. We are not trying to, you know, keep anything higher or lower. And what is very interesting is in what you're saying is that copilot or any of these things were the terms. But, uh, the way you are saying is I not only work with one agent, I have multiple agents working with for me. And um, they, they sort of do the work and I am the thinker and like the navigator of sorts. Very, very, very cool. Uh, Raja. So one of the questions that I ponder is how much are we contributing to the betterment of AI or you know for the future of AI. I mean I know we are using it and by that indirectly the reinforcement and everything is happening. What about the contribution of what are uh, we contributing back to this ecosystem or the collective itself in your view?

Speaker E: It's a very interesting question. So far I am only a consumer of an egg. I am not a consumer uh contributed to um. I am just using AI to solve some of the problems that what Otherwise it will be very difficult or it will take a lot of time. For example, if you take uh, our own software industry of India,

Speaker D: yeah,

Speaker E: it required Rod Johnson to introduce a spring framework. We have such great organizations in India don't want to name anything in this podcast. We have never contributed to any of a simple Java framework in r uh complete journey, right? So uh, but we cannot say that okay, we have not created anything. We have made um, we have leveraged those frameworks like spring. Whether it is for uh, from now if you take a Java world, uh. Rod Johnson completely changed the way that uh. We were all uh using mindlessly this enterprise Java beans, right? So he created a new pathway that okay, we can use a kind of a framework here. So um. And the second is that I had this idea. How about uh. For example I ventured into one more uh project called as a. My beacon. The My beacon is um. Uh. Is a way to help the kids to understand uh physics in a better way. Right. Today a lot of organizations they are uh, making use of the situation here at least in India to uh, coach kind of thing so that the prices of the simple physics education itself is becoming. The prices are getting escalated. So in order to make uh them easy for them. For example, if someone wants to uh, get hold on the neat kind of examinations or je in mains at least or to j advanced. Okay. So I have created a platform where they can come and they can solve problems that par with JE mainsor need. And uh, we created something called as a path builder. Using the path builder you can uh, break a very complex problem into smaller steps and how the chain of reasoning can be established kind of thing. So we use AII for that. So originally I thought of creating something. I do not have any LLM for physics at least. So I tried to do something with uh, lot of other stuff how we can do this. Later I realized that okay, it requires a lot of money involved here. So finally what happened is that I have just used Alibaba spent 2.5 with the 14 billion parameters that is running in one of my machines here. And uh, so I am more or less I become a consumer of um uh LLMs. I am not into uh creating my own LLM. Uh I had the dream there. I understood that okay it is a very um are a lot of other things that I have to do today with what others are doing. So uh, at least as of now I am going to be a consumer at least for the next two years. Originally had this question why we are unable to even build a simple LLM from India when Alibaba is able to do something from uh which is also E commerce kind of an industry. So let us just to confine this particular podcast. I don't want to jump into something different but uh creating an LLM is something it's not uh a one man thing or even fine tuning an LLM is also not my area using Lora or anything. I'm just going to consume an LLM.

Speaker D: No, I think uh what you are saying Nooraja like how we are using a spring framework and we have made it better. We don't have to. For example I think I have may not have contributed to Python but I've made it better by using it and um providing my best practices in many different forums. Right. The same thing I think with AI also given that you have written so many AI programs as well as made the code better cleaner I think we are contributing in some shape or form and I know you have written so many programs including uh Chai and Guru and Ask Guru and so many other things for PM Power also.

Speaker E: Yeah.

Speaker D: So I think I am very interested to uh learn more about uh the. The thought process that went through into building your uh you know uh huge Alibaba based uh code. Also sometime I think uh, I think our listeners would be very interested if you can post some of your learnings uh in a way that is easy uh and accessible uh to consume.

Speaker E: Correct. This is a very great uh fantastic thing you made me to think it's a fantastic actor. You really died uh when you when you told about I I am using and then I am making this Python into a better environment for people to use. That's a very. It's really impactful statement Immediately what I it is coming to my mind is the. The original author of this MER Soft. Right. The MER Soft we used to have a quick sort of and mergeso uh sort is something that is coming from uh one of the founding fathers of Python when I was doing some kind of a research uh which kind of a sorting is uh better sorting especially comparing across various languages long. I still remember that I, I forget the name of the Mergesa person. Okay. He has been given later for his immense contribution to Python as one of the lifetime achievement award for Python. Great, uh, thing that it is not that we need not to go and create something that on our own. Right? So we need not create a country. We can make the country better.

Speaker C: Absolutely.

Speaker E: In the same way, what we are having today is AI and we can make use of AI to make the life of our people better. Uh, so end of the day it is not, uh, we have to make someone's life better. At least we have to try to turn them towards a positive thing. That is where actually, uh, so I am very. Thanks, thanks for that, the way of, uh, for directing me in that way.

Speaker D: True, true. See, I think all of us have, uh, our own, you know, lenses, right? Uh, in which we are seeing AI when for the first time when Internet came in, the way we were using Internet was more like a, uh, you know, directory of sorts. But now we don't use that lens. I think more and more. It's also a factor of more consumers coming in and how LLMs are also changing based on consumers. Because in the end of the day AI is also learning, as in the speed that is phenomenally hard to understand for all of us. I am actually wanting to understand would there be decisions that you will rely on AI to do for you, or would there be, uh, uh, decisions that you will definitely not go to AI for? For example, I will give you a simple example myself. I will always go with AIs, uh, Google Maps if I have to find a route, uh, to a place which is unknown to me. Right. But, uh, if there is a financial decision, I will, I will probably ask options, but I will take the decision for myself. So those are like two dimensions that I can think of. I will always use AI. I will never use AI for. So do you have such dichotomy in your mind

Speaker E: like that? I consult AIs, uh, for many things, but always I reserve the human judgment for that. I am not going to blindly trust that. Okay. What the recommendations that what AI is giving is. I'm not going. It is not a mistrust. We have to take a tidum nantviradhara vara, we call this in Tamil. So I am not going to blame, uh, a particular AI is direct me in the wrong way. Sorry if I am unable to.

Speaker D: No, I understand. You're saying more like a human in the middle. You will always want to be that person getting options, being a consulting, uh, partner or the coding partner if you will, but you will take the final call.

Speaker E: Even, even today when I am doing a coding, when I'm asking Claudette to give uh me code still I'm going through the code and then I am finding there are a lot of opportunities that uh, this particular code can be written better and then I'll give the directions to that. Then it is actually creating a very good uh code. For example if I had, if I not bore the audience as well as you is that if you ask uh, in those days a GPT or cloud based thing to write a code on Rust, they'll actually keep on create lot of functions and pass lot of things as a parameters to that. So but uh, it does not mean that uh, we have to actually give a direction to that. Yeah. Think of this, put that into a structure for example like a class equivalent of that. Put that into a struct and then try to create some kind of uh, uh implementation for that. Try to think of some kind of a traits to that like an interface equivalent of this. Then the code will automatically become a very meaningful thing. You have to pass lot of things into the parameters kind of a thing immediately. Uh, that uh. LLM understanding this and then it is finishing the task. That means that I have not seen across uh still a code that has been completely created. Uh like the white bottles are thinking that okay, everything will be created perfectly.

Speaker D: Yeah. So I think the next logical question for me is two years from now what do you think AI is going to throw at us? Or what is. What is going to be the vision for you and that this would have. This will be a very cool thing.

Speaker E: I'm um, expecting a lot of local language models uh to come something similar to the Quinn that what I was talking about right. So many times does not require for us to have a very powerful LLM that uh, requires lot of um. Uh. Only certain class of people can access this. Every school require a very simple LLM that is sufficient. For example I have downloaded uh, an LLM from Hugging Face. It is remarkably great in terms of physics, uh, reasoning and chemistry reasoning which I feel that it is sufficient for us to help the education system to some extent. So that will reduce the total cost of ownership. Right. So the recurring cost for example there is some thousand kids are trying to learn something in an, in a school. Then probably they, they can actually think of going for a local LLM. So the GPU costs for this should come down. I think it will come down and to uh, just to accommodate uh a simple LLM, uh with a vram of around uh, uh 30 uh GB that is sufficient for us to solve many of our things. So this kind of local LLMs are uh what I prefer at this moment so that um, at least we can make the money to rotate within India.

Speaker D: Yeah, it's a very particularly in terms of you know social equitable. It is more accessible. I know a lot of ables I'm adding. But it. If you are bringing in some level of localization into uh uh LLMs that makes it lot more easier for people to consume, use and upgrade. Right. I think uh, one of our other colleagues are also saying can AI be used more in the social spaces that will kind of uh, bring in a lot more faster adaptation and problem solving. Very, very cool Raja. I'm going to do a rapid fire now.

Speaker E: Another thing is that when you bring this social, that's what even our Sridharan Webu is um, trying to bring this Arata. Right. I do not know how many of us are really using Arata still. We are resorting to some, the same kind of a WhatsApp or uh, different kind of organization. Right. So for that an AI is not required. So we are such a very great population in India at least if we have uh, some kind of social that can run within India at least. For example why I have to go to a Facebook and uh run a campaign for an organization. Right? So cannot be just run a uh campaign to reach my own brothers and sisters within India. Right. So that kind of uh, thing that what um. That kind of a problem that what India at least has to Indian entrepreneur uh should at least know with the iii I think we can create such kind of uh great uh softwares very quickly or within a much affordable way. So um, we have to start think uh going back to your old question that how we can leverage AI for the betterment of our own people. Right. So now uh, uh in my opinion we are sending a lot of money outside of India and that's creating lot of uh, deficit. Right. And uh, we can um, at least we can. I'm not actually against any of using Facebook or Google kind of uh, stuff. Why don't we do something within India? That's what my mind is telling me.

Speaker D: That's a fair question right? In terms of you know think local, think global but act local and uh, building that local economy and uh, creating more jobs and everything. That's true.

Speaker E: Correct, exactly.

Speaker D: So uh, now I'm going to do ask you a rapid fire one or two words. So first uh, Question is, what is that one AI tool you love? I think I can guess that it will be. Claude, uh, is There any other AI tool that you love?

Speaker E: The Quinn.

Speaker D: Okay,

Speaker E: 2.5.

Speaker D: All

Speaker E: right. That's the LLM because it is what is running in my, uh, RTX 6000 GPU, and it is helping me to do a lot of stuff.

Speaker D: Excellent, Raja. And, uh, what is that one thing that you will never delegate to AI I think you mentioned about the code itself. With final code, you'll never delegate anything else. You'll never delegate to AI

Speaker E: Human judgment or discussing about private things. Uh, for example, asking AI to do certain things which we are supposed to do. I won't delegate that to this.

Speaker D: True. And, uh, one word on your relationship with AI I know it's a very personal question, but I will tell mine if it's easier. All encompassing. It's there everywhere, so around me. So that is just my perspective. Thank you so much for this amazing conversation. It

Speaker E: was very great talking to you. Thanks

Speaker D: a lot. Thank you. Bye. Bye.

Speaker E: Bye. Bye.

Speaker B: We thank Siddharth for the music and Anita for promoting the Software People stories.

Speaker C: If you like this episode,

Speaker A: please subscribe on your favorite podcast client

Speaker B: and spread the word in your network.

Speaker C: If you'd like to share your story,

Speaker A: contact us at. Uh, podcast.