How To Go From No Job Offer To Amazon's AGI Development Team In 18 Months - w/ Kunal transcript

I still remember I have completed my internship at home tap and I have not got the full-time offer from the company where I interned at because they were not sponsoring H1B and the market was pretty bad back then. That's Kunal Mishra. 18 months later he was building Amazon's frontier AI models. By the end of this episode, I promise you will know exactly how he did it and how to get hired at Frontier AI roles in big tech. Let's build an LLM. So I built LLM from scratch that is a 30 million parameter model. I spent like $50 to $100.

Honestly, I'll say that has landed me job in my first company as well as in Amazon. Amazon. The exact project playbook plus the entire Amazon loop from the inside. Lead code don't forget it. It's still relevant guys. I have solved around 450 problems for Amazon. I believe 40% is behavioral. So please focus on that. That's the mistake most people do. and exactly what's happening behind the scenes in Amazon's AGI or at this point. point. So I actually work in developing the frontier models for Amazon. I'm in the post training team.

Please join me in welcoming Kunal Mishra to the Ready Setu podcast. Without any further ado, let's get into it. Kunal, welcome. Yeah, thanks S. Thanks for having me. So excited to pick your brain here today. And where I want to start off really is this remarkable 7-year trajectory you've had where it seems like at least off of your LinkedIn about 7 years ago, you were doing a summer internship at DRDO. Fast forward to now, you're involved with AGI at a global massive big tech organization like Amazon. when you're forced to kind of retrospect just how much of a whirlwind these past seven years have been, what are some thoughts, you know, or overarching emotions that come top of mind? mind? Yeah. Um, actually that's a funny question because I have contemplated like back uh in 2019 that I wanted to be an ML engineer. So starting uh from my

career uh after college placements I took some time like uh couple of months uh preparing for ML interviews. So I was selected in a couple of companies uh in the college itself but I didn't want to join them because they were software developer roles. I wanted to be an ML engineer from the very start. So in 2019 itself I took a course I think it's supply day course prepared ML engineers end to end and then I got my first job as an ML engineer and then from there uh luckily for me AI has also evolved as soon as my I mean I got the job and so the career transition starting to happen from a uh machine learning engineer to senior machine learning engineer and then masters in the same like information system with concentration in machine learning and then from there uh US exposure in the home tab where I was

also building machine learning models. Uh then finally Amazon happened and now I'm building the frontier models for Amazon. So yeah quite quite a journey I believe. believe. No. Yeah truly. Yeah I mean quite a journey is really an understatement just it's just incredible how much you've covered in terms of ground right just in these past seven years. And then when you let's say you're just out and about right you're at a restaurant and you meet somebody they're like okay hey Kunal what do you do [clears throat] so how would you describe working on AGI at a company like Amazon because as I'm sure you're aware everybody and their grandma seems to have a different definition of AGI what it means to them but I am curious as far as Amazon goes or maybe even just big tech goes exactly what is AGI when when it pertains to your job.

So if I have to explain a grandma what is AGI, I will say that when uh the computers are able to do your job as the way it is, it's an AGI but that's a wins very layman terms AGI um for me uh and for our like where we work around in general because I live in the Bay Area and here most of the talks are around AI because it's a trending topic. Uh for me AGI is basically when uh uh the models that we train uh can get to a certain level of intelligence they can that they can cooperate uh with humans uh in their work. So they can be your personal assistant your add-on. So when you say that hey this job is given to me you uh dictate those terms to your agent uh by words by speaking it out and it can do it for you without you being like um kept keeping an eye on top of it and it gives you the ultimate result. Hey I

have done it that's the goal of Asia in my mind. There can be like multiple goals but that this I'm telling a short-term goal. Yeah. No, I think that makes total sense and I think it's a great way to describe exactly what it does where you throw a problem at it and you just like go on a walk or something and then when you come back it's just done for you exactly the way you would have liked which I think that last part is where maybe outside of at least the way I interact with AI. I think that's not quite there yet. Um, but I'm curious, would you agree like I guess would you are you um a proponent of the idea that AGI is already here or do you think it's still coming? I believe there are still some improvements that needs to be done. Um, human brain is much much superior than AI as of now and it can do things. It

can create new things on its own. uh AI basically all the LM models they are the just internet you can say zip file of an internet combined so they are just making whatever is present in the internet right now but human brain it doesn't work like that human brain can think beyond the internet human brain can uh create new things discover new things so that is something we are still not there yes but for the tasks that are deterministic and that are mundane yes you can do that with AI yeah I like that um I like that analogy where It's internet in a zip. I haven't heard that one before. So I I do like that a lot. Um we will talk more about Amazon obviously, but I kind of want to go back to something you had said, you know, earlier when I asked you to reflect on your journey. Why was it so important for you to specifically be an

ML engineer? So I'm just putting myself back in my own shoes when I had first graduated college. For me, any [clears throat] job that paid a decent salary, I would have jumped at, you know, like I couldn't care less. I was like, I don't know anything. Just give me a job. Give me money and I'll do it. But it sounded like you knew very early on that you wanted to go a certain way. I am curious if you were to pull that version of you in the past, why was it so important for you to only go down the route of a machine learning engineer?

Uh, yeah. I even if I go back and if you call me what was that aha moment for me when I decided to go for that decision I cannot recall it exact same time but uh one thing I learned over the time was u in my undergraduate days I think it was in third year of my college uh I was doing some projects and this was pretty hot topic back then when people were exploring these things back propagation was a thing and like a couple of very new terms that are like very old if you see in this context they are like uh what are you talking are you grandpa or something but they were very old at uh that time and uh those things fascinated me because they were new uh they were like predicting something like earlier we used to predict the house models Kaggle was a thing kegle is still a thing um and but that was a very big

thing back then and you create those big models from Excel sheet and they are predicting values and thing that fascinated me so uh while I was placed um uh in in some roles in in my college but I wanted to be an ML engineer. So that's why I uh took that decision that hey let's try this thing and when you're young you don't uh feel too much about the repercussions you just say that hey let me take that step and we'll see what happens and maybe I didn't think too much and just took that step.

Got it. And then at what point when you were first working your first job out of India did you first start consider or considering getting a masters in the US? Were there other options? Were you also looking at Europe maybe? Um, talk me through the decision process that brought you to Northeastern. Yeah, sure. So, uh, when I was looking into uh, my like higher educations, there were a couple of options. So, in India, you can do a gate for masters and um, and if you want to come here then you can definitely uh, research a bit of universities and apply to a couple of them and then you can get admissions. uh my idea was that uh the real development or the tech heart is here. So uh it's not like India is very backward. India is very superior in tech but the core algorithms uh or you say the tech mind tech heart is here. So I wanted to

explore that part. I wanted to work on supercomputers that have like thousands of GPUs and I've read about Northeastern that it's a great university to be in. I researched about the program. There were uh the good part about my program was um there were 51 electives. So one mandatory 51 electives. So you can uh choose your own trajectory. You can be a front- end engineer, back end engineer, cloud engineer, AI engineer, you can be anything. So that gave me a couple of like so I took a couple of subjects in software engineering, a couple of in ML, a couple of in cloud so that I can get best off on some worlds. I can be good at multiple things and that's why like I took that decision to come to uh come for the masters in the northeastern university and yeah so when you were going through this process and this is one of the questions

that I get a lot I'm sure you get this as well uh can you help guide somebody that's trying to figure this out right now in terms of is a is a masters in computer science better or worse than a masters in like machine learning or AI specifically And I know there's some other courses that are kind of parallel which I'm sure you will know much better than me. But if somebody's trying to make that decision in in terms of what program should they go for, what are some helpful maybe questions to ask or like frameworks that you can share for our students that are currently considering going down that same path?

Correct. Uh I have been like asked this question a couple of times. Um I will say uh that uh the world still works in fundamentals. So a CS degree and the core CS concepts are very much needed even right now as well as they were before. OS concepts um if I can go into details as well like how every LLM technique or how every large angle model whatever it's using it be inference it's some kind of primitive things that you learn in computer science at the same time with right now I'll say first build your fundamentals if you have done undergraduation in computer science that's that's fine then you can go with ML here I was an EC major electronics and telecommunication so uh for me there are only a couple of subjects in computer science So I built my foundations as well. But I'll still say having a degree in computer science is

still relevant. Having a a degree in computer science with some electives in machine learning, you can do wonders. A complete degree in machine learning is also uh useful if you know the foundations. So it depends on the use case. That's as I always say, but you should evaluate yourself like a student should evaluate himself before coming here and he should understand where he stands at right now. just going by the trend or be a bandwagon fan won't help you here. Uh foundations the world still work on foundations. That's what I say.

Yeah, I think that's a great takeaway and I also like that because as you said a lot of grads that come to the US are pivoting from you know computer science adjacent streams into computer science and for them it just makes in my head at least a lot of logical sense to start there right and then if you want to further niche down within AI or what have you that is an option on the table but it can be a bit of a drastic jump to go directly from like an PCE straight to an AI and where again the student might be familiar with the fundamentals but it's still good to make sure that you have those covered at least you know from a direction point of view. So generally is would you say that's like the right approach for this?

Yes. Yes. totally agreeing on your point uh that a student in EC must get some foundation knowledge because ultimately when you start working uh and so let's say an EC engineer jumps into an AI course and he starts writing uh code or something or he creating started GitHub repositories GitHub repo or something like that uh AI can still do it right now uh with the level of expertise AI had cloud code opus 4.8 8 models can write you the same repo with the same expertise. But if you have foundational knowledge and you can connect the dots, then AI is still not there. So you should build your foundations, take your time. Uh yes, the world is running at a very fast pace right now. Uh last like one or one and a half year have changed most of the things but still there is scope. You still have a lot of time and explore

things uh in the foundations and then jump to ML. Love it. So once you were done with your graduation, can you talk us through your first job and the process of finding that first full-time job after your graduation? What was that like? What were some really helpful, you know, tips that you can share with, you know, job seekers in the current market? Obviously, as you and I know, it's probably the worst market ever. So I understand that it might not convert apples to apples just in terms of you know direct conversion because it was just a different time but still I'm sure as you said the fundamentals are always important. So yeah with that in mind can you share kind of your journey with getting your first job and maybe as you do that also share any frameworks or helpful tips for current job seekers to employ for their first job out after

their master's degrees. Got it. Yeah. Uh I still remember u I have completed my internship at home tap and I had I think 5 months to graduate and I have not got the full-time offer from the company where I interned at because they were not sponsoring H1B that's separate thing but I had to uh do things from scratch to find a job uh and the market was uh market had started going bad uh it was like 2024 I wasing graduating in May 24 and the market was pretty bad back then. Agreed. I do remember that. Yeah exactly. So how I started was um I uh first thing is you have to believe in yourself uh that that was true that is still true and that will be true like 100 years later uh that you'll do something you are here uh you are capable you will do something and the next is uh think about what is the current thing that you can build that

the agents cannot so I started working on this project so how I how something came in my mind is like uh what is trending right now an LLM. Let's build an LLM. So I built a model LLM from scratch that is a 30 million parameter model from scratch. I spent like $50 to $100. I had an idea of how to connect the components and uh honestly I'll say that has landed me job in my first company as well as in Amazon [snorts] because when you implement something from scratch you know everything inside in and out. My interview was more like a discussion around my project rather than an interview. My manager and my interviewers are very happy with the work that I have done and they were asking question definitely they'll ask cross question you but if you have done something from scratch uh you are very well prepared and you can answer any

questions. Uh so this yeah this project that I built like it's called it was called tiny tales GPT and it was as simple as that you put the beginning of any story for your kids like Jim was a curious kid and just hit on generate and number of stories three it will generate three different stories for you and that thing I have done from scratch. So built a from model to data set everything like data set was there definitely I pulled in trained the model trained it for like 10 12 hours couple of iterations and then yeah I published it on LinkedIn uh then I started applying application is still a process even you are the best of the world you have to apply uh you should meet people connections so I believe 60 40 or 70 focus on yourself 30% focus on connections u send people your resume send people the project you

have done because ultimately if you have done it uh people will recognize you and they will help you out. So I believe in my journey that project was uh useful. So let's say if I want to do it something right now uh there are many things you can do like uh create create an agent from scratch maybe yes how can you you create um let's say you are asking for the technologies right so there are there is vlm and sz lang right they are [snorts] used for inference so I'll give you a problem like how how can you increase the throughput of uh any model just say right now inference is used anywhere anywhere uh there are a couple of things like if you want to go into details. There is Nvidia, there is AMD, there is Ryzen, multiple things. U how can you con take a model and make it work on any system.

How will you do it? There are multiple things that needs to be explored. We are just touching the surface of it. And still, yeah, I believe if you want to have an if you have expertise in something, most of the jobs are still welcoming you. Dude, that's crazy. Like I actually can't believe you just made an LLM. That's so wild. Can you Yeah, I honestly don't even know what to ask you about that or where to begin, but here I'll ask. You can ask me anything. I build I build that. that. Well, no, no, no. So, I meant more so than the technical aspects because honestly, a lot of your responses will literally fly over my head. But so, but I'm still trying to, you know, double click on that. I guess from your experience of building that something you said jumped out at me where you said you just learned the entire inner

workings of any you know frontier model or any model really like you understand how it works. So for those in my audience which is basically like 98 plus% people that are pretty savvy that know their way around like Claude can get their things done maybe can you share one or two learnings or findings that is not super obvious until somebody goes out and does something like you did which is literally build a model from scratch. Yeah, that I can share a couple of uh blockers that I like it was it was like two months or I spent two and a half months building that. Uh yes, there I hit a like a bit of blockers then I like go over it and then again find something. something. Uh so a couple of things that are on top of my hand are so that's a 30 million.

How will you decide the parameter size? It's a 5 million model versus a 10 million model versus 30 million parameter model. If you give it to cloud code right now, maybe it will take up the GPUs, run the experiments on top of it. But cloud code doesn't care about your money. U you have to be so you are renting the GPU. GPUs are very expensive right now. Per hour you have to pay like $6 or $7. You have to experiment multiple times and when to stop what are the epochs that are deciding and when you are building from scratch there's nothing u like there are multiple assumptions that you have to make. Those assumptions, those reasoning are something that you develop after seeing the results. So I rented the GPU 10 first let's start with 10 million parameter model. Oh, I see in the second epoch the uh inference is not so good.

So I go back 20 million parameter model. Let's change the uh attention layer. Why? Because I have an intuition that attention layer is something that kind of gives more complexity to the model. Uh this is something you get to know when you read about things when you become subject matter expert for some say like on a lower level but still you cannot ask clot to do it because there are there are knobs that you need to tune clot can tune all the knobs but it will take money from you like you have to pay for it right so the same thing can be built on $2,000 $3,000 $5,000 but I built it on $50. So that's something um if you want to go frugal, if you want to be like uh get the best out of it and since if you have money you want to trust you and um definitely claude uh if that clot was at that day I could have used it for

debugging purposes for sure. Uh Uh switching frameworks working on Jupyter notebook then training it and I actually hosted it. So uh I created a website and anyone can run their queries on top of it logging it keep that active 24/7 making sure that there's no fishing attacks multiple things it's a end to end product so uh one thing I should parameter second thing is like how will you train on two H00 GPUs minimum GPUs so that's there are a couple of things that you need to learn and unlearn and on the way you like get a final model yeah that's very interesting so for anyone any of our listeners that are looking to you know really learn more about this project cuz I know I do. I would love to like you know read something or if you have it you know anywhere hosted in terms of like the knowledge base do you is there somewhere

that you can point us to that we can all go and look at it by chance? Uh so u it's I have also made it public it's in my LinkedIn for sure. Uh awesome. I'll link that then. Yeah. Yes. And then uh the code if you want uh the code then I think I have not made it public. Uh but I can do it like uh I can share the code. It's not also you can look into Andrew Karpath's uh llama. C file. It's the Python interpretation of that with some changes uh with some changes based on that. So um definitely anyone um getting into I will say that uh read hacker news get to know uh read Twitter uh follow your favorite uh ML um ML guys whatever you follow like there are couple dozens of them you can follow anyone and get to know the like the basics as well as get to know the trend where where the ML is going right that sounds great yeah I will be linking

those resources in the show notes for anybody to check out so guys feel free to do that um continuing from there can you share kind of how the Amazon opportunity came about, how you felt when you first you know found out about it and like kind of walk through the the interview process itself if if you could please for us. Yeah. Uh so I applied it uh I think on LinkedIn I saw that it's called software dev engineer and machine learning that was a complete title. I read the responsibilities they were very same that what I was doing before I applied it without any referral. Uh but it's good if you have a referral from someone. Um and then I applied it. U I got a call the next day from a recruiter that hey they like my profile and want to uh schedule some uh first the coding round. So first was the coding round.

Then uh I think I gave the coding round in two days because yes lead code uh don't forget it. It's still relevant guys. Uh lead code um I have solved around 450 problems and uh it it will always help you. uh I won't say that's the best way to uh like it's the only way to get in a company but most of the companies are still doing it so uh that's a thing and understand the problems if you can uh need code 150 is a very good resource if you can solve the problems you are in a good shape to start your journey and then uh so I'll tell you my after the two rounds I think they were somewhere related to uh need code 150 problems and then I got a call from riotra again that hey u you have been selected and then there are five interview rounds. Uh there were two data structur lead uh one two system design round and

in Amazon there is something called bar raiser or behavioral kind of round. So in the five rounds were very intense uh like um they can you maybe share a couple examples of questions from each five if if at all that's possible. Sure. Uh I cannot give you the exact question but I can tell you the format. Yeah. Exactly. Yeah. Exactly. U I can tell you the format it is for in Amazon. In Amazon first of all uh they uh kind of put some focus on the leadership principles. There are 16 leadership principles. In every interview you will be asked to uh pure a scenario uh based on one or two leadership principles and there then there will be cross questioning like um in the inner details because uh they really wanted to know that what was yours your contribution into that particular situation and everything. So those are really important if you're

preparing for Amazon. I believe 40% is uh behavioral. So please focus on that. That's the mistake most people do. My I have suggested this to my friends as well. Then uh the comes two part one is system design and the second is uh data structure and algorithms. So there were I believe two rounds for me for data structure algorithms and the two rounds for u system design as well. For data structural algorithms you will be asked to solve a problem similar very similar to lead code. Then you I think you need to uh run the dry code and there is a coder pad also. So you have to like write it in make make sure that it compiled and then if you can run by the examples that's even great. Uh they can there can be follow-up questions as well. I can give an example for example if a given problem is only positive numbers. So the

interviewer might ask what if I give negative numbers as well. So you'll be able to think on your toes right away that how does if you understand the logic you'll be able to do it for sure. Yeah. Yeah. I'm just giving an example and then similar some other and um for system design it really depends since I was an ML engineer the system design was mostly out to an ML system on some problem. Then they go into intricacies of this like uh mostly system design rounds are for understanding your thought process and understanding why you took decision A instead of decision B. What were the pros cons of both the decisions? uh what's the trade-off you made and that's what important that's what I do in my day-to-day job as well make decisions on which uh framework which uh workflow anything like every at every day you're

making some decisions and like that build over time and you like uh work in as you work in the company and that's how you like grow in the company. So that's an initial version of it. So system design is basically about taking decisions u what databases what um workflow uh how do you write the code how do you define the packages and everything so that's some of the tips I would like to give to the students who are watching this podcast and lastly behavioral again so uh they go went through my resume asked like my projects whatever I have written in my resume in detail uh like in depth and uh my especially my professional experiences and then finally Again it was two leader. So in every round there was two leadership principles and then rest of them were DSA and system design.

Got it. For somebody that's preparing specifically for the system design part of the interview. Of course as you said it helps to just go through some like toy scenarios. I say toy but not you know literally toy but yeah put yourself in the shoes and drive through decisions. But outside of that, can you recommend any resources maybe like books, any YouTube videos, whatever that helped you that you can recommend to somebody else that's preparing as well?

Sure. U there are two books I recommend everyone. Uh Chip Huen's uh first book is design machine learning systems iteratively. That the first one second book is AI engineering same by the same author Chip Huen. Then if someone wants to go into LLMs uh then there is LLM handbook uh I believe Axim leone I'm not able to pronounce it correctly. Yeah yeah yeah will this will all be linked yeah in the show notes. So yeah yeah yeah and uh this is especially if you want to like uh understand the training part of the models and LM engineers handbook is a great resource uh then you can follow blogs of Sebastian Rashka. He's a wonderful author. He has two books also.

How to build a LLM from scratch. How to build a reasoning model from scratch. Uh it if you want to just go through it the GitHub is free of cost. You don't need to buy the books. You can just go through it. Um I can give more into agentic part. Some in agentics are like um there is one certification from claude called claude code developer or claude code architect. U I'm not able to recite. Yeah. Yeah. Yeah. I know. I know of it. Yeah. It's the architect one. Yep. Yep. Correct. architect one. So you'll get to know how the agentic workflow and the harness works. So agent are basically model plus harness. So a model you'll get to know the resources that I've told you. Harness is something you need to develop like how does the model works? How does it does the tool call and most of the things uh yes these are the resources on top of my

head and again uh one thing that everyone should every every techie every student if you are in tech you should learn or you should read daily is hacker news. So there's a section called showen. I kind of read it. Showen is basically when uh someone does something they put it on show their project. So just read it. Uh see how people are who are interested in any tech. It's not only AI are building things. It's a fun place to be in. I just spend 5 to 10 minutes there daily to see what people are building. Maybe if there's interesting blog I might read through it but skim through it daily. Hacker news is a wonderful place to like read daily.

Awesome. Yeah, that is another recommendation that I haven't heard before and I love it. Yeah, and I'm not even a hacker or you know like an ML engineer, but even the books that you mentioned, so like for full disclosure, I do work with AI a lot. So, but all of these are sounding like great resources. Even if you're not nowhere close to even trying to appear for an ML engineer interview, I still think that if you're loaded up with the information in these books, I feel like at least it's sounding like you you will be able to just do much much more be way more productive with the you know agents or like the models at your disposal. So, really appreciate you sharing that. The other question that I get a lot especially around Amazon is specifically around the bar razor piece. So I I've heard that it's supposed to be it's like

designed to be really stressful you know it's like they they put you on edge. So to the extent possible um what can you share for somebody preparing for that bar raiser specifically for ML engineer roles what what is the best way for them to prepare? Yeah I think I can I have a good news and a bad news for all of those watching. Good news is uh sorry I'll say the bad news first. Bad news is all the interviews are stressful not only bar is uh a little stressful. The good news is if you keep your confidence uh things will work out. So uh they just want to judge your reasoning skills. They just want to take uh hit on your decision skills. So uh prepare fully be confident. uh if you are giving any answer you should be able to reason um behind that answer it's okay uh if you don't have a reason for something just

say that uh you think it that way but give it otherwise uh the interviewer is also like trying to help you in some way he will be push you to the limits but if you keep your calm and you just uh ultimately the manager or the senior manager or the bar raiser who is taking the interview they wanted to check whether you fit in the team or not and if you work in the team it's not only your expertise. It's not only how knowledgeable you are. It just can you work with the team. While you're working in teams, there are multiple places where uh you are saying something else.

The teammate is some saying something something else like disagreeing thing. So you should be able to handle that situation and you are just getting glimpse of it. So keep your calm uh reason behind your decision whatever you're saying and uh let's say if you say that this is what I believe in and why you will be good. Uh there is nothing else. [clears throat] Just be prepared. prepared. I love that. I just love how general that is in terms of this is not something special that happens at the end. It's all stressful. So why don't you just I'm telling you to that.

No. No. And that's literally the whole point of this podcast. And you know, again, so appreciative of you laying it straight for us, not mincing words at all, and just being so transparent. Um final question here before I let you go. I saved the best for the last cuz with this I realize that again it's totally okay if you can't speak too much to this cuz obviously I understand the nature of this but when it pertains to AGI within Amazon are you just straight up building like a frontier model that Amazon is about to drop maybe you can answer that maybe you cannot but any information that you can share in this general atmosphere of things would be super helpful cuz I'm sure for anybody listening this question has to have been at least close to top of mind like what is going on here behind the scenes.

Yes. So I actually work uh in developing the frontier models for Amazon. I'm in the post training uh team. We have actually a couple of models frontier models that were actually launched beforehand. It's called Nova. I can talk about it because they are available in public. So if you go to nova.amazon.com nova.amazon.com you can see models. We are still uh improving them. We are working on different techniques to make it state-of-the-art models. They're already state-of-the-art but to even make them improve to have general public use it to make sure that they are on top of the leaderboards we are working towards that and yeah I actually build like I and my team definitely build those frontier models for Amazon. So sorry if this is an ignorant question but is like why is there no like why isn't there more hype around Nova?

Does that make sense? Yeah. Yeah. True. uh see how um LLM model makes hype is how useful two two reasons I will say uh how useful it is for general public and how useful or how good it's scoring at the benchmarks there are a couple of benchmarks MMLU GPQA GPQA now there are two lawn finance agent multiple of them codebench also does it does it I think codebench I think is another bench codebench yes all of them uh So the uh ILLM model um should be good for public use as well and these benchmarks as well. There are multiple frontier models right now like there are 50 or 60 different companies I can remember on top of my head that are building models.

Uh it's uh and you'll remember only and you chat with only a couple of them uh four or five top the top ones. We are getting to the top. Uh we are good at and for a good public use case it should be good at everything from coding to English reasoning most of the things. So we are getting there. Once we get there yes we'll definitely create the hype and you will hear about Noah for sure. Got it. That's so interesting. Yeah. No I mean that firstly I feel like that's such a great challenge to work on right where you have just worldass resources.

you're all aligned on, you know, a problem and you're trying to make something truly great that literally has the potential to cuz I can't think of something that Amazon is not involved in. It's such a huge huge company in terms of impact. So, yeah, man. This is so cool. I I can't believe how much ground we've covered just in in these past few minutes here. Really want to appreciate you taking the time. I'll ask you the question that I love ending with. I I know I said last question, but I lied. So, sorry. This is the actual last question. But yeah, if you had the platform to share something with the entire world, it does not have to do anything with what we talked about, but if you could share a message that everybody could magically receive in the world, what would that message be?

Uh, be healthy. Uh, rest of the things will follow. Huh, I love that. Uh, a a sick Wait, I'm trying to think of the quote. I'm going to butcher it, but I think it's something like uh a healthy man wants a 100 things. A a sick man only wants one. Uh yeah, that's one way of putting it. Yeah, I don't have that. Yeah, the reason I said this code because uh like after this I have to go to gym. So yeah, but uh I can give one more quote. Uh let me think real quick. Uh maybe. Yes. um take one step at a time and uh be grateful wherever you are right now and yes these two things yes love it thank you so so much for taking the time this has been such an enjoyable experience experience yeah thank you so much Nam I really enjoyed this conversation as