Tech Legend's Advice: CEO & Co-Founder of Modular & Mojo, Chris Lattner, on Tech, Life, & Learning
![Tech Legend's Advice: CEO & Co-Founder of Modular & Mojo, Chris Lattner, on Tech, Life, & Learning Tech Legend's Advice: CEO & Co-Founder of Modular & Mojo, Chris Lattner, on Tech, Life, & Learning](/pic/tech_legend_s_advice_ceo_and_co-founder_of_modular_and_mojo_chris_lattner_on_tech_life_and_learning/VU9DQ2hDM3ZWNGs_.jpeg)
Hello, we are here at the AI Engineers World Fair with Chris Lattner, CEO and co-founder of Modular. Chris has an extremely impressive background. When he was a graduate student at the University of Illinois, he created LLVM. Then he went to Apple in 2005 and helped create various developer tools, including notably he was the main author on Swift, which is Apple's programming language which came out in 2014. Then he worked at various ML and processor companies including Google and Tesla, and at Tesla he was VP of autopilot software. Then in 2022, he started Modular, and Modular aims to make development in AI and ML much quicker and easier, very simple to make, to take the expediency of a complex coding language using a relatively simple to understand coding language similar to Python. So we are so excited to have you here. Thank you so much for meeting with us here. So tell
us a little bit about yourself to start off and how you got into compiler technology and AI. Chris: Yeah, so I mean I represent as developer. I like building things. I like building things in the real world too, including woodworking and stuff like this, but really my passions around computers and programming and systems. And so the thing I love about computers is that unlike in the real world, you don't actually need much to be able to build almost anything on a computer, and you have all these free tools. You have the ability to learn, you have the internet,
and so you have so much at your fingertips. It can be kind of overwhelming sometimes, but it means that you're not limited by having to run to the hardware store and get some gears and bolts and things like this. And so I thought that was really amazing, and so across my career, I've been able to learn and grow in many different ways. And so I've just love software. Harper: That's amazing. And so when was your beginning of your career exposure
to compiler languages? It sounds like long, I mean at least in graduate school. Chris: Yeah, so I fell in love with compilers specifically in school. And so I love compilers because when I was in school, this was a long time ago now, you would learn a course, you would then build a project, you would then throw it away, you then build the next project, you would get a grade, you would throw it away, and you keep doing this. In compilers, it was the first course where you build a thing, you get a grade or you get some assessment, and then you build on top of it, and then you build on top of it, and then you build on top of it. And so the thing that was really fun to me about that is that then you know you're having to live with your mistakes. This is much more realistic, this is much more about what real software engineering is like, and this is what I really loved about that because it became much more about building an organism, building an evolving thing, and then learning from okay, well I could just push this underneath the rug, but if I do that later, it comes back to get me.
Harper: So you have to be really, really thoughtful about every single building block. Chris: Or choose not to be and then decide to fix it later. Which is also a big part of software. And so I love that aspect, and so I got involved in that and then went to graduate school as you said and was very fortunate to work on this LLVM technology, which now has kind of gone on to power a lot of today's computers and systems. Harper: It's amazing. And including Modular, I mean that experience building LLVM, in addition to all of your experience before, has kind of come together to make Modular.
Chris: Yeah, that's right. Well, so I mean a lot of what I love to be doing as a human is always learning. And so what I've done throughout my career, and I advise many of you all to do, is think about first when you get into the market, exactly where you enter doesn't matter so much as how much you can learn. And then when you learn something, you're learning from the people around you. Don't stay in one box for five years, don't stay doing the same thing for 10 years. Yes, look, whether it be within the same team or within the same company or maybe across companies, you look at other opportunities where you can now bring something, yes, but you also get exposed to new things. And so for me, I've been on kind of this quest of learning many different things, and whether it be soft skills like management, whether it be hard skills like how AI convolutional neural networks work, or whether it be compilers or different things across the stack. And so I love that, and that's one of the things that,
as you say, brought me to Modular is a culmination of all of these experiences. Harper: That's amazing. I love that idea of moving what you're learning so that you can have this breadth of knowledge and so that you have so much to bring to your work, to the world. Chris: That's right. Harper: And talking about learning, a lot of the audience here is just new to technology, to AI. And so I guess first we'll start with what
would you recommend to them as a way to break in, just in the most broad sense? Chris: Yeah, so I mean I think that there's really two things. One is the background knowledge that you need to kind of build up. There, I would pick a project, pick something you can build. There's so many good open source tutorial other things, including paid courses and university and things like this, but really building a thing. Get to that "I did it" moment and prove to yourself, but also to other people later. You know, build your resume, build your design book of saying
like, "I did this, I can explain it." People can then ask you questions, and you can actually tell people about the experience. And that experience of building something, I think, is so critical to both getting that like moment of pride where you can show it to your friends, but also it gives you that learning because you have to struggle through the parts that are actually pretty challenging. And so doing that in a way that's directed, so there's an outcome, really helps.
Once you do that, you can kind of go through a couple of different projects and things like this. I'd encourage trying to find an open source community. And so one of the things that is not really taught in schools so much, again depends on where you go, but it's how to work as a team. And so much of software, I mean it's super funny, people think about the lone coder that's just sitting there at the keyboard late at night, you know, drinking Red Bull. The reality is that you're working on a team with other people, and so as you do that, it's all about the social aspect of creating software. And those people are who
you can learn from, the people that can teach you how to work and the technology and the specific skills. And so learning that aspect is also super important. Open source is a really great way to do that. It's a great way to get involved with a team that's already working together. Generally, open source communities are very inclusive, and so it can be a great way to get started. I've been working with open source since the 1990s, so quite a while. LLVM was open source, LLVM is open source. I've been driving an open source project as kind of the, not quite a BDFL,
but is one of the major leaders for 25 years. I've built the Swift programming language, which also open source. And so what I would say is that open source is an amazing idea, but different projects have very different personalities. And so there can be a very big
difference between open design where you can get involved and you can have part of design discussions, whether it's open code but not open community, whether there's community meetups and things like this. And so like the actual nature of the open source project can really vary. Harper: Is Modular working on that kind of community aspect? Chris: Absolutely. Mojo's open source, and so you can work with the Mojo open source library. We have a big community on Discord, and so that's where we kind of pull people together, and GitHub and many of these typical technologies. And so yeah,
open source is a huge part of AI, and so if you're interested in AI, then open source is a really big part of the entire ecosystem community, which is really important. Harper: Wow, it is amazing that Mojo is open source. Let's talk a little bit about Mojo for those who don't know. How would you describe Mojo? Chris: Yeah, so Mojo, there's many ways to look at it. First is, let's talk about Python. Python is amazing. It is the thing that holds together, it's the glue of all of computing, it feels like. But
a lot of Python is really good down to a certain level. And so underneath Python, you have older technologies like C or C++, or newer things like Rust that are really what powers the Python ecosystem. And so the challenge with Python ends up being that it is beautiful and simple until you end up having to go and do something hard, and then you get into this really crazy complexity. And so what Mojo is doing is saying, okay, well, let's take the thing that is beautiful about Python, let's take the Python developers, let's take the tools, let's take the ecosystem, let's take all the packages and all the libraries and all the Python code that exists, and then make it so you don't have to do C or C++ or Rust. And so the way we do this is we use
very fancy compiler technologies and things like this, which you have experience in. So I have some experience in this, I've built some programming languages before, and so I have some experience with this stuff. And say, let's actually take what we've learned, and we as an industry have learned a lot in the last 30 years that Python has been growing, and let's rebuild this whole thing. As we rebuild it, now we can make something that is just like Python at the top level, but it displaces the C and C++ part in the ecosystem. And so for you as a Python programmer, this means you can continue to grow. You don't run to this wall where there's like certain dark magic that you're not allowed to see. You can actually build and scale into these
more advanced topics without having to switch to and retrain to a completely different universe. Harper: Wow, okay. This is phenomenal. So also tell us a little bit about MAX, which is a suite within Modular. Chris: Yeah, absolutely. So MAX is where Mojo is. Let's go make the Python ecosystem stronger and
better and also scale onto GPUs and things like this, which is a big part of what Mojo does. MAX is saying, let's take AI and make it so you can deploy it. And so one of the big challenges that many of us face in the AI deployment industry is that the technologies in the AI space are very messy. And so you've got technologies very focused on training models like PyTorch,
it's amazing. You've got many endpoints out there that many, many of you use like OpenAI and Anthropic and AWS and other things. But if you want to do something custom that's not training, it's inference, it's building application, building AI into your applications, there isn't really a good solution. There's a mess of different things. So what MAX is doing is taking the benefit of learning all the different systems that have come before and building a new thing that can help with that, help get models from research and production faster, make it easier to use. And if you look at what Modular is doing globally, we have a major enemy. Do you know what our
enemy is? You know what we're fighting against? You know who the bad guy is? Harper: The bad guy if you're moving from production to deployment? Or if you're a Python programmer that's struggling with C++? Chris: Python programmer struggle C++. We have a mortal enemy here. And so this is actually really important to understanding who we are, really. Harper: Mortal, okay...?
Chris: I won't put you on the spot. Our enemy is complexity. Harper: Oh, I see. Chris: Well, it's not because of Python. It's because of, if you look at AI, AI is, you know, PyTorch and TensorFlow and these things are almost 10 years old. Pre-Neanderthal, everything has changed. All the stuff is so invested in by so many people. There's this
gigantic mess of systems within Python. Python's amazing, it's beautiful, everybody loves it, but it doesn't get you onto a GPU. It doesn't run fast. And so if you care about something going fast or about cost being low, you get exposed to all this complexity. If you're trying to get, if you're training a model and you're like programming a GPU, one GPU is great. If you want to deploy that thing, suddenly all of this complexity. And so bubbling up, what is Modular about? It's about taking the source of the complexity, the fact that all this stuff got cobbled together, and saying, okay, well we have to do these things like have a programming language that can solve the Python level of the problem. Once you solve the Python level of the problem, now you can solve the GPU problem. Now you can solve the AI problem. Now you can solve each of these things. And this is,
these are very, very, very big problems that the industry is struggling with. But if you do that, then you can actually make a major dent in some of these big challenges that we all struggle with. Harper: I saw that you don't have to use CUDA with Mojo. Chris: That's right. Harper: So it just, it's a replacement as well? It just kind of built into the language? Chris: Yeah, the way I would say is it's an alternate way to program GPUs. So CUDA is one of these technologies which are, again, it's wonderful. It's powered
the industry for 15 years. It's amazing. It's also super complicated. It's changed a lot, very complicated. And also a lot of C++, a lot of libraries, a lot of version incompatibility. There's a lot going on with CUDA these days. And so again, lots of good learnings, bring forward the best thing, put into Mojo, make it Pythonic, rethink all these technologies. Make it, you don't have to deal with CUDA versions of conflict, like just a nightmare, by the way.
And so just by removing complexity, what we can do is we can allow folks to focus on their problem. And then you also bring in other fancy compiler stuff and like all the technology component of this. We think we can do something very significant that then helps people move on with life instead of having to fight the Docker container and these things, which we as AI practitioners have Grown accustomed to, but it should not be that way.
Harper: It takes so much time. It's such a headache. Chris: That's right. And so this is things where, again, it's a very abstract thing, this like destroying complexity concept. But if you map it onto the specific pain that people feel, give you another example. This, if you want to deploy Llama 3, there's probably a dozen different ways to deploy Llama 3 from cloud vendors, from how you run it locally on your Mac, how you do this and that and the other thing. They all use different APIs, how they do function calling completely different, how the model temperature could be completely different. Like all these
little differences can drive you mad when you're just trying to build something, your application. And maybe you want to be able to deploy to cloud for production for your customers, but you want to write unit tests so that you can actually test your app before you deploy to production. And that you want to run locally on your Mac, but you can't run the cloud endpoint locally on your Mac. And so that lack of consistency, again, just drives tons of complexity into our ecosystem. And these are things that we are looking to solve through the world.
Harper: Wow, you're a true developer. You've been doing this for so long. You know the pain points. Chris: I love developing things. And this is the thing about AI, is that people forget just how new all this stuff is. I mean, it's been changing our world just so fast,
but it's also super mature. And this is where we want this whole ecosystem to grow up. Harper: If you don't have to use CUDA, does that mean that Mojo is agnostic to the GPU provider? Chris: We only currently work with NVIDIA. So we only support NVIDIA. So NVIDIA are great friends to us. They're one of our most important partners,
and we think we can do a lot with NVIDIA. We're working very hard on that. Harper: Very cool. I know that there's an open source project working on making Mojo provider agnostic, so like TPUs and GPUs. I forget what it's called. Chris: I mean, this is the cool thing about Mojo being open source in the community and everything else, is that Mojo is a very horizontal technology, just like Python's. And so to me, I think it's really cool that people are building UI library and they're building serving frameworks and they're building all kinds of cool stuff, and including a lot of AI technologies as well. And so this is the thing that's very cool is that
you look at Jeno and FastAPI and like all these things that the world is built on top of, there's a lot of room for improvement. And so this is where again, Modular can't do everything, but we love the people that are investing in doing these things. And so I think that is very exciting. This is again, you talk about developers, I love the passion of developers. I love the
spirit of building things. I love giving people new technologies and new tools so they can be heroes in their own journey and they can build things. And so that's what I find so inspiring about the developer community in general. Harper: I mean, you're so inspiring, frankly. Sorry, I mean, I'm just like, I'm so inspired. Thank you for being here. So as I've shared,
a lot of my audience is new to tech, new to AI, and they have a few questions. I guess first we'll start off with how can they future-proof themselves as technology has changed? There's a natural concern about jobs shifting over time, and so they're wondering how can they best prepare themselves for this world that is coming? Chris: Yeah, that's a really hard question. It really is. I think that a lot of us struggle with that, honestly, including me. Harper: That's good to know. Even you do. Chris: How is the world going to look? We don't know. You don't know. And so you kind of have to make certain bets, but also invest in
kind of general knowledge. And so what I try to do is I try to follow some of the news, and like there's these papers of the day and things like this that people talk about on Twitter or X or whatever you want to call it. Harper: Oh, okay. Chris: But also, and so you can kind of see like what people are talking about, but then also just the fundamentals I think are actually pretty consistent. And so if you care about building models, for example, you need to learn how back propagation and some of those kind of stuff works. And so you can invest in kind of similar, the basic thing. Again, what I would do is I would start from just build something,
build something real, see it through end to end, build a demo. Don't worry about it lasting for, because your life and career is not a demo. Your life and career are the lessons learned, the experiences you build from having built that demo. So the actual details will change and you know,
you'll get to figure out things along the way, but what you're building is a skill set which is really about problem-solving, how all these ecosystem components fit together, who are the people that matter, who are the people that don't. I mean, one of the things I've learned going into a new domain over time is just saying like, "Okay, well just understanding the personality, okay, like the social side of the community of your audience or of who you're building for." Just in the AI space, like okay, who are the big personalities? It's like the People magazine but for AI, almost. And so these things, these people in their back stories a little bit, but you don't want to spend all your time on there. But it can help you understand the map,
understand what's happened, why they're fighting over something on Twitter or something, and things like this. And you can make more sense. But I would, I mean, the short answer is I would not worry about future-proofing yourself. I would focus on building things, learn, gain experience having built things. And what you'll build is you'll build that core knowledge of building, problem-building, products solving problems, working with the technology, getting in there, wrangling with the uncertainty, wrangling with the frustrating parts too. Harper: Absolutely. So pushing through. Chris: I think in the, you know,
having confidence in yourself is a big part. Harper: That's wonderful. So that's reassuring for me to hear because that means I'm on the right track as an educator because I have my 10 days of AI basic series, which goes through the basics of AI, like both in short form and in long form, and then also like a coding series where we code in Transformers, Mojo, you know, and then there will be more in-depth advanced videos like how Transformers work, et cetera. Chris: Yeah, so building fundamentals with Python, that's a really great thing to do. And
because that transcends not just AI but so much of computing these days. And you know, learning Mojo, Mojo is a natural extension of Python, and so now you can learn new kinds of things. And so you can kind of build into the part that makes sense for you. Not you don't have to learn everything. Code is made, software is built by teams, but it allows you to follow your passion and figure out what makes sense to your brain and the way you think. Harper: Exactly. And now there are so many tools
with large language models that make it so easy to debug, to understand code if you're new to the field. Used to be so much more difficult, I'm sure you know. Chris: I mean, I grew up where you had to learn from books, and that was very different than the internet. So unbelievable. Harper: I mean, compilers themselves are already, in my perspective, the hardest type of engineering you could possibly do. So I am just so-- Chris: Sorry, AI is also very hard. So I mean, all these things are very hard. I'd
say that compilers are one of the most obscure and very, you know, there's not enough compiler engineers. People don't fall in love-- Harper: There's a reason for that. Chris: Maybe, maybe both are true. Harper: So another question that I get a lot is, what is a fear that you hear a lot that people have about AI that is just pretty much a misconception, that just should not be a fear? Chris: I mean, one of the biggest misconceptions is that anybody knows what they're doing. Harper: I love this. Let's talk about this. This is a great topic. Chris: So I'll use myself. I won't talk about anybody else. I'll use myself. I have built up a body of certain experiences,
right? I know a few things. I'm very proud of some of the things I've done. You know, a lot of things I don't know, a lot of things either. And so the thing that is very, something you should just know is that you'll often feel imposter syndrome because there's all these smart people in the room. They all sound like they know what they're talking about. They may actually not know what each other is talking about because they've specialized in different parts of this problem. Harper: Exactly. Chris: And so a lot of what happens is you have these very fancy people giving a technical talk, and you're like, "I understand one out of every four words they're saying."
Harper: Right. Chris: They get done with their talk, they sit in the audience, they're listening to the next, and they understand one out of every four words of the next speaker. Harper: Right. Chris: And so this is actually a very interesting thing about AI, is because the ecosystem moves so quick. This is where again, what I would, what I'd recommend is
you figure out kind of what to align with, what technology stack is kind of going in the right direction. And don't worry about the paper of the day or the new model or things like this. Instead, just look at what's useful and be focused on solving problems and learning as you go. Harper: Solving problems and learning as you go. And so we talked about imposter syndrome. I actually want to go a little bit more into that. I'm sure that's something that my audience deals with, that I've dealt with. Perhaps you have dealt with as well, many times. Chris: You know, this is also a consequence of like saying I'm hungry to learn. Harper: Yes.
Chris: Because when you're hungry to learn, you throw yourselves into new situations where you don't know everything. Harper: Exactly. And you know, it's so funny. I have this really strong memory. About four years ago, I was at Meta. I was working with a really, almost a famous machine learning
engineer at Meta. He was remote before anyone was allowed to be remote. And he was the ML guy, and I went and worked with him very closely. We became quite close, and he told me one day, he said, "Harper, I have the worst impostor syndrome." I mean, this man output so much, and he said, "Every day I feel like I'm not enough. I feel like I can't do this. I have no idea what I'm doing." And that really hit me, is, is it-- this man can have imposter syndrome, everyone has imposter syndrome. Hearing that from you is just again, like so--
Chris: I mean, it's extremely intimidating when you're so-- I'll give you a concrete example. So I joined Google, Google Brain team, some of the most elite experts in AI building some of the most advanced, some of the most furthest ahead, the most at scale, like etc., things. All because I want to learn the technology. And so I found myself meeting after meeting after meeting where I'm like, "I don't want to slow everybody down by asking what is an embedding" or something like that. It's like, "I will have to go do some research tonight and understand what that is" and spend a lot of time on Wikipedia or whatever it is to even just understand what people are talking about. But what I've learned is that if you kind of plow through, you invest, you-- a tool I use is particularly when I'm in rapid learning modes, I'll write things down in a little book. And so by writing things down, you're like, "What is an embedding?" Harper: When I'm doing my AI guides, I like to go through the process of debugging and of hitting errors. And I don't try to make it polished. I try to make it real in the sense that like,
once you hit a bug, you go, what do you do? Because I've gone through tutorials before. I was learning React a few years ago. As you say, it makes it too easy because you're following the script and see, actually proven script. And real world is more messy. Chris: It's much more messy.
Harper: And I was following this tutorial and I hit an issue. And then I, I didn't know where to go because the tutorial was so clean, they had no solution for me. This was like pre-large language models, and I was like, "Well, I guess I have to leave. I guess I'm just not going to finish the tutorial. I guess I'm not smart enough" or something. Chris: Exactly. And that's the problem. And this is one of the problems with these overly polished scripts for demos versus "Let me teach you how to think about this. Let me teach you how
this works. Let me show you the error message case. Let me show you these things." Because that's actually your life as an engineer. You end up having a message, struggling through, "What the heck does this mean? How do I solve this? What is this crazy thing mean?" And so again, coming back to destroying complexity, right? So much of the systems of today are connecting together incompatible systems that were not designed to work together. And so you can't really make progress on some of these inherent things unless you clean up the mess. Harper: Do you have any advice for new grads? You know, it seems like you're a very thoughtful person about your life as a whole, to be honest. And so I'm, I haven't asked anyone this question, but I'd like to ask you this question. Do you have any advice for someone starting their journey,
their career journey? It sounds like you said don't be worried about a certain path, don't trust it to be linear, don't hope for an end goal. Is there anything you'd like to add to that? Chris: Yeah, I mean, I would, I would try to find a balance. And so what I would do is I would find a balance between learning the technical skills. Learn how to code Python,
what a for loop is, like technical skills like this. Learn how to work with people, super important. Communities, go to a meetup, go join an open source project and just like go to the forum and see what people are talking about. Again, you're not going to know everything or understand everything they're saying, but you'll figure out who the personalities are. And software is a very social thing. Another is build something. Build a demo, like actually get in, get in there, do something, follow the tutorial, fall off the path, try breaking it, see what happens, and like intentionally do that because that's how you kind of learn how these things work. And then figure out a place where,
particularly you're entering your career, what is a good entry point that isn't just about how much money you get paid or something, but where you can maximize learning. Maximize learning. And if you think about your career as something that you build over a long arc of time, it's not a job. Like think about what is your objective for that first job that you take, and often it should be about like, where is it position you? What can you prove? What can you learn? Who do you learn from? Things like this. That's way more important than is this paycheck bigger than this paycheck. Harper: It's long-term thinking. Chris: That's right.
Harper: Well, thank you so much for being here, Chris. It's really great to talk with you. Chris: Thank you for having me. Harper: I hope you enjoy the rest of the conference. So glad that you're here, and thank you again for all your work at Modular. I'm excited to teach everyone about how to use Mojo. Chris: I'm very happy to be here.
2024-07-26 02:00