Replit's VP of AI, Michele Catasta, on Artificial Developer Intelligence

Full Transcript

Transcript

Michele Catasta: 0:00 Because I would have spent more time learning interesting things about coding rather than setting up my coding environment and debugging various stupid syntactic mistakes. These are things that are honestly not intellectually stimulating about coding. And we're trying to get rid of them as soon as possible so that you can really focus on the creative process of writing software. Building large software still requires a lot of human effort. But allowing a lot of people to be creative will also expand dramatically the pool of good ideas that we explore as mankind.

Nathan Labenz: 0:34 Hello, and welcome to the Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas, and together, we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my cohost, Erik Torenberg. Hello, and welcome back to Replit week on the Cognitive Revolution. For context on this conversation, if you haven't already, I definitely recommend checking out our most recent episode with Replit product designer, Tyler Engert. I spent 10 full minutes at the top of that episode explaining why I believe that Replit could become one of the most important companies in the world, why I already consider it to be one of only 15 to 20 live players in the AI game globally today, and why if I had to pick a single platform where the human AI collaboration economy is most likely to take shape, I'd pick Replit. I also posted a version of that essay on Twitter, and we'll put a link in the show notes. Today, we're going even deeper on the future of AI at Replit with special guest Michele Catasta, Replit's new VP of AI. Michele recently joined the company from Google and made an immediate splash by publishing a new AI manifesto that outlines Replit's plan to support the next billion software developers. They plan to do this by enabling a seamless human computer symbiosis, creating an artificial developer intelligence that will create AI coding agents capable of developing complex software, plus a complimentary economy that rewards human tool creators and hires humans for advanced problem solving when needed. Michele even proposes a new meaning for Repl. In place of the original read, evaluate, and print loop, which describes a traditional interactive computing environment, Michele now unpacks the acronym to mean reflect, evaluate, percolate, and learn. These much higher order concepts are now the core feedback loop at the heart of the company's artificial development intelligence efforts. This is not just one of the biggest AI visions you'll hear from any company in the world today, but one of the most well specified, coherent, and given Replit's track record for execution, one of the most credible as well. While it might sound like AGI, that is artificial general intelligence, Michele believes that Replit's focus on code, which unlike natural human language, has been designed over decades to avoid ambiguity and is also now supported by a wide range of quality assurance tools, will improve reliability enough to allow Replit to unlock huge value without needing to confront the most perplexing challenges related to potentially smarter than human intelligence. Now, to be honest, I'm not entirely convinced that this AGI ADI distinction will prove as meaningful as Michele and team believe. And I will definitely be watching closely for new features designed to protect the next billion developers. After all, I've learned from experience training executive assistants in the art of AI task automation at Athena that the next billion developers will inevitably be very AI dependent. Still, I am convinced that Replit is a phenomenal platform for developers both seasoned and new and definitely a company to watch as the power of their product and the influence of their platform seem almost certain to continue to grow. As always, if you're finding value in the show, we would appreciate it if you'd share it with friends and post a review to Apple Podcasts or Spotify. And I also really appreciate your feedback, including guest and topic suggestions, which you can send either by email at tcr@turpentine.co or via Twitter DM, where I am at Labenz, and my DMs are open. I get a handful of very nice notes every week, and they really do influence the direction of the show. So please reach out. I would love to hear from you. Now, I hope you enjoy this fascinating conversation with Replit's new VP of AI, Michele Catasta. Michele Catasta, welcome to the Cognitive Revolution.

Michele Catasta: 4:43 Thank you, Nathan. Thanks for inviting me. I'm excited to talk about Replit and AI today.

Nathan Labenz: 4:47 Yeah, me too. Regular listeners to the show will know that I'm a big fan of the Replit platform, which honestly even predates the current AI push, but that has also been a pretty exciting thing to see. I want to start with just a couple real big questions, and then we'll kind of come back and ask some smaller, more detailed ones as well. But for starters, you guys have put forward this ambition, I think one of the most ambitious, that you call ADI, artificial developer intelligence. So what is that?

Michele Catasta: 5:21 So that's our, I would say, midterm roadmap. Hopefully, it's not going to take us decades to accomplish that. And the idea is we try to identify an obtainable goal in terms of how we need to evolve our current AI Replit to make developers much more effective. So the reason why we made a play on the name and made it different compared to AGI is I don't want to go on record ever trying to predict the date where AGI is going to happen. I think everyone agrees largely that it will happen, but we're not interested in figuring out when in the next 20 years. What we care about is what can we deliver in the next few years to developers. And I think LLMs have built an amazing foundation to help developers be more productive. GitHub Copilot was a pioneer of that, and we have Ghostwriter and Replit, which is helping you both to complete code, and we also have a chat feature that helps you as a peer programmer while you're developing. I think this is all great, and we're getting a lot of amazing feedback from our users. But what we describe in that manifesto is how can we bring that a notch up to the next level? And the idea is we would love to really empower the next billion software developers in the world. People that don't have any computer science background, they're just amateurs. They land on Replit, and they want to go from an idea to a prototype in the shortest amount possible time. We need an AI to make that happen, and we need a much more powerful AI compared to what we have today. LLMs generate an amazing amount of semi coherent code. Usually, it's aligned with what user wants. At times, it has some mistakes. It has some issues. I think that's another blocker for people that are really experienced developers such as us. So I always know how to find my way out from a bug that GPT-4 or Claude models are giving me. This is not the same for a person who has never written a lot of code. For ADI, our vision is how can we make sure that through several different iterations, and I'm going to go in more detail what that means, eventually we converge towards the right solution that the user was asking for. The way in which we envision that is there will always be a fundamental generative component to that, which could be a much more powerful LLM compared to what we have today. Likely, the field has been evolving at such a fast pace that I can only imagine what we're going to have, say, 2 years from now. So definitely, they're going to be more powerful and more independent. But Replit has a key advantage of offering also an execution environment. And we've been spending years building all that kind of platform. So now you can think of our feedback loop where an AI generates code, then we execute it. We also include back the feedback from user. We include also the errors that came from the execution, and then we learn from this initial step. We feed it back to AI model, generate the code again until we converge to something that actually runs and it is aligned to what the user wants. So this is our ambitious goal as we defined it before. Hopefully, it's not, again, too many years away. I have a feeling that maybe in a couple of years we can crack something pretty powerful. And we've been working heads down towards this goal since I joined Replit.

Nathan Labenz: 8:48 It's amazing how far you already are. I've been a user of the platform myself, and I bring a certain amount of coding knowledge to it. But there's much more that I don't know than that I do, of course. And that's really true for everyone. So even for somebody who knows their way around, the built in autocomplete and chat tools are really useful. But I've been kind of pursuing this myself. Again, listeners will know I'm advising a company called Athena, which is in the executive assistant space. And we're seeing that in line with your vision, the next billion developers, we're kind of seeing that software is becoming much more permeable to non software developers, much more accessible. And we're starting to use Replit and starting to teach what is ultimately coding, or at least it's software development. But, honestly, to people that have never even coded and don't know even the basics of coding. And we skip the traditional stuff these days. We skip the for loops and the syntax and all that kind of stuff, and we just jump right into, here's this platform. You can start to talk to it and try to get it to do what you want to do. We actually do often have them go to GPT-4, which is something that's pretty interesting because for the top level request, obviously, GPT-4 is still the boss model, and they need all the help they can get in many cases. So to get that kind of first skeleton, we do often bounce out of the platform and go to GPT-4. But then once you're in it, you start to iterate, especially more locally, a lot of the built in tools really work well. And it's been amazing to see how quickly people can go from, as you said, kind of wandering in, not even expecting that they were going to be asked to code in this job, to being able to actually manipulate applications and get somewhere. It's crazy. It's only 2023, and this is starting to happen.

Michele Catasta: 10:54 I agree. I mean, it has been an exciting ride. I would have never expected things to move this fast since we saw the first generating models based on transformers. I could see the light at the end of the tunnel. I felt it's about to happen. It's bound to happen, but who knew that the timeline would have been so compressed? And now as a matter of fact, as you're saying, we see a lot of people that have never approached software in their life being very productive. For instance, at Replit, we host monthly hackathons, and the vast majority of them are revolving around AI topics. And a lot of people that come there are not software engineers by trade. We got amazing projects built by PMs who spend most of their time communicating with software engineers. And as a matter of fact, they're extremely good at prompting models. Because what do they do for a living? They put in English requirements and descriptions of how a software should work. So if you put them in front of GPT-4, they do an amazing job at describing exactly what they want. They get the outputs, and Replit helps them to stitch pieces together, helps them to get a quick feedback loop in case there is a mistake in the code. And, internally, you can also have a model helping you to debug the code. And we have a debugger feature in Ghostwriter that does exactly that. When you execute your Repl and you hit a problem, you generate an exception, then immediately, we try to tell you that's how you should fix it. And we are seeing PMs winning hackathons just at their first experience writing code. And I think this is only going to be even more pervasive in the near future. And, yeah, I feel Replit is really enabling this wave more than perhaps any other company out there.

Nathan Labenz: 12:39 Replit has really done a lot of work in terms of teaching people to code in the traditional way. How close do you think you guys are now to a mode where you would say, forget that and let's go this PM route, and then you can kind of fill in those lower level details later?

Michele Catasta: 12:57 My dream is that we're going to go in that direction more and more. Of course, there are obstacles to that. Any AI is expensive not only to train, but especially to serve when you have a lot of users. And I think right now, we just went past the 24,000,000 plus users at this point. So it's not an easy platform to deal with. And, whenever you're giving access to powerful models, you need to think also unit economics. That being said, the trend in AI and especially in LLMs is that of better economies of scale pretty much showing up every quarter. So I do see a future where at least some basic AI features will be accessible to any user, and that will radically change the way in which people write software today. And, that's a welcome change, to be fair. If I teleport myself back in time when I started to write code, I was very young, it was an extremely frustrating experience. I don't regret going through that. It was a lot of fun in hindsight, but if I had Replit when I started, I would have logged in much more. And I think I would be much more advanced today because I would have spent more time learning interesting things about coding rather than setting up my coding environment and debugging various stupid syntactic mistakes that I was doing because I didn't notice that I didn't close a parenthesis or I forgot a quote here and there. These are things that are honestly not intellectually stimulating about coding. And we're trying to get rid of them as soon as possible so that you can really focus on the creative process of writing software. That is here to stay for a while longer. Again, I'm not going to make timeline estimates of when people are going to stop to write software completely. I do think that the creative process still has a place for a while longer, but no one loves to deal with low level tiny mistakes in code. And yeah. And I'm happy to say that the new generation is learning to write code in a completely different way compared to how I did it.

Nathan Labenz: 15:02 Hey. We'll continue our interview in a moment after a word from our sponsors. It really is something mind blowing to see.

Michele Catasta: 15:08 And by the way, it's the same for experts because I've seen notable tweets. I think Andrej Karpathy is saying, I can't imagine myself writing code without GitHub Copilot. I would easily waste 50% of my productivity. And that's coming from a person that's been writing software for as many years as I've been doing. He's a very prolific researcher. They will not get rid of the new AI features in writing code. And I think soon this would be agreement across every person who writes software.

Nathan Labenz: 15:39 Yeah. Totally. So you came, just for a little personal context on you, recently from Google to Replit. And Google and Replit also have a significant partnership, which I think is multifaceted in terms of investment and cloud and now even some product connectivity. Were you involved in that in the creation of that partnership, or were you just kind of separately attracted to Replit?

Nathan Labenz: 15:39 Yeah, totally. So you came, just for a little personal context on you, recently from Google to Replit. And Google and Replit also have a significant partnership, which I think is multifaceted in terms of investment and cloud and now even some product connectivity. Were you involved in the creation of that partnership, or were you just kind of separately attracted to Replit?

Michele Catasta: 16:02 I was involved to an extent. I was there more as a researcher, as a leader of management. So I'm definitely not the person who calls the shots about making such a big partnership happen. I was an internal champion for that. I was a big fan of Replit even before I started to work there. And I was giving demos of PaLM to Amjad and his team back in the day because I was one of the contributors to the code skills of PaLM, which we called PaLM coder back in the day. So I think that was my first interaction with the Replit team more officially. And then I stayed in touch with them. I unofficially advised them for a while, and then I became an adviser and then all the way to joining them full time. So it was a slow cook. I would love to believe that I played a role to make that happen. And after I left, I'm glad that the partnership actually happened. And it's a great partner to have by our side. Replit is a very resource intensive company to run because we give compute to millions of users. So it's something that we couldn't do on our own for sure.

Nathan Labenz: 17:12 Yeah, that's interesting. That might be worth getting a little bit more into the weeds of. Before getting there, in some sense, you've just come from maybe the most formidable AI company in the world, certainly with probably more top notch researchers than anyone else, probably more publications than anyone else, certainly more resources, more compute. And you've gone to, I think, probably the smallest company that I have on my live players list at basically 100 people at Replit, right?

Michele Catasta: 17:47 We're getting close to that, yes. So it is that small.

Nathan Labenz: 17:51 Close to, but still under 100. So how do you think about that trade off and what should we be watching for? Typically, new technology waves are startup favoring. But I've been pretty commonly of the opinion, as I look at different markets, that I think the incumbents are going to do really well in this case, and you're going from one form of strength to another. So how do you think about the trade offs in scale and resources?

Michele Catasta: 18:17 Yeah, so of course, there is a trade off. Some things that I'm currently doing at Replit will be maybe much easier if I attempted to do them at Google just by the sheer amount of resources available in that company. At the same time, there is such a strong benefit in being lean and having less process and being able to move fast and iterate faster and don't fixate perhaps on mistakes because they are considered a sunk cost. At Replit, we can decide this direction didn't pay off. Let's do exactly the opposite. I love it. It's a breath of fresh air. I don't think I belong long term into big tech. I'm very grateful of the experience that I had. I met a ton of amazing people. As you said, the density of AI talent at Google is incredible. But then at the same time, I feel at home at Replit because we have a very narrow focus, and we're very much obsessed with giving the best possible AI to our developers. And it doesn't mean that we necessarily have to build a team exactly like the one in Google Brain. We can be complementary. And the partnership in a sense helps us to make that happen. Because we get access to models and we get access, we can have discussions and we can help each other in the process. And I think there is space for everyone. There is space for incumbents, and I'm glad that they are there because it would be shortsighted to claim that OpenAI and Google haven't been changing the AI landscape, thanks to the APIs that they offer. But then at the same time, there is also space for startups that are capable of building new generations of products that an incumbent wouldn't be making a bet on. So I like to use this example with office productivity. I wasn't surprised to see that Microsoft and Google immediately integrated GenAI in their office productivity. And I wouldn't be going against that kind of product as a startup if the product looked exactly the same. I wouldn't try to clone Google Docs with GenAI because an incumbent will do that much better. Now if you come up with an idea of a completely different way to edit slides or documents, which has GenAI as a first class citizen, that I would tell you, that's a very good company. That's something that will have its own space in the market. And with Replit, we kind of play the same game. We're building something that doesn't exist. We know that a lot of users love that. And we're going to keep building that with the help of the incumbents as well.

Nathan Labenz: 20:52 Yeah, there's even a frame there where I would say, in some ways, while it's small in terms of headcount, in some ways, Replit kind of is an incumbent. And by that, I mean, again, going back to teaching these EAs how to code without really teaching them how to code. When I was introducing Replit to them, I was kind of like, it's going to be hard for you to understand how much better this is than the old flow of having a local development environment and then having to keep your settings in sync and your Python environments all whatever. It's just a total cluster. Obviously, everybody who's experienced it kind of knows that. But if they haven't experienced it, it was like, to you, it's just going to feel like this is how it always naturally would have been, and you'll never know the difference. So I think that in many ways, the strength of the platform itself and just all the provisioning and all of the infrastructure and all of the seamlessness that has obviously been created over time puts Replit in kind of this unique position of being both small and, in a meaningful way, having an existing technology moat that's super complementary to now an AI layer put on top of it. So I don't mean to pitch the business too hard, but as you can tell, I'm definitely a fan of the product.

Michele Catasta: 22:12 I appreciate that. I understand. Otherwise, it wouldn't be where I am today. But I agree with you. That's one reason that drove my choice, because I think that we have such a strong advantage in what has been built before I even joined the company, that I think it's an amazing playground where to apply AI and where to evolve AI. That being said, I think as a company, we always love to keep a low profile in a sense and not become complacent. So as much as I would love to believe, oh, we built something that is unique and is going to help us grow over time. I love the fact that we still all feel the urgency of becoming better and growing and making our users happy. And it still feels like a small startup to an extent. I've been a cofounder in teams with much smaller teams, and it kind of feels this urgent right now in a company that's by some Silicon Valley standards, you could consider it as a company successful, but no one inside feels, oh, we made it. We all talk about, oh, we need another three years to do this and five years to do that. And I love this kind of long term goals that we have in mind to grow and become better.

Nathan Labenz: 23:24 So that maybe speaks to kind of the company's identity. And I wanted to ask a little bit about this because people refer to Replit in ways, in my case, on my list of live players, it's also appeared frequently on lists of most e/acc companies. So do you guys identify as an e/acc company? What does that mean to you? Do you want to wear that label?

Michele Catasta: 23:49 I don't know much of these things, to be fair. I'm more of a spectator on Twitter rather than anything else. Of course, we're a tech company. We are focused on AI. We love progress, and we are providing a novel way of writing code to users. So by definition, we're a company driven by a will of contributing to the technological progress, and we strongly believe that giving development capabilities to the next billion people is going to make the world better. Now I don't think any of us likes labels both as individuals and as a company. As a matter of fact, I think we have a web page for new hires where we say we are apolitical. We don't take a stance. I think we love exceptional talent. We love to work with people that are respectful, and at the same time, we respect any point of view. Sure, I'm sure that I'm just interacting with some people on Twitter. But for those of you who follow him, Amjad himself, I've been knowing him for more than a year right now. I wouldn't be able to give him a label. He's such a unique person in his own beliefs and in his own rational way of seeing the world that I just found him to be an inspirational leader in the company rather than a person that belongs to a certain clique of people on Twitter.

Nathan Labenz: 25:13 I wonder how that kind of apolitical, I mean, there's a lot of different takes on this, right? We've gone through waves over the last few years of highly political, even activist discourse at a lot of companies, including big tech companies. I'm sure you were there for some of that. And then there's been this kind of movement away from that and saying, okay, we're a startup, we want to focus on one thing, and that's our mission, and we don't want to be political. And it sounds like for most things, that's probably kind of a pretty simple, relatively straightforward policy for a company like Replit. But then when it comes to AI, it's kind of tricky again because it is both core to the mission and it's inevitably kind of becoming political. But also, as people that are advancing the technology, you obviously have real responsibility in the ways that you do that. And it would be easy for that to kind of bleed into the increasingly political discourse around AI. So how do you guys navigate that? What is the internal AI discourse at Replit like?

Michele Catasta: 26:20 Yeah. I think that political and responsible are two not necessarily parallel tracks. And in a sense, I'm not a big fan of the fact that politics bleeds into AI. Conversely, I would love us as a community to be very focused on building responsible AI systems. I think that the fact that we partner, for example, with Google and that we work with OpenAI, and we are surrounded by companies that I do believe are posting a certain amount of effort in that. No company out there is perfect. I think there is no playbook on how to do responsible AI today. And it doesn't mean that the researchers working on that aren't delivering. It's more about it's a new field. It's going to take us years to figure out how to do it correctly. If you think about social media, it has been around for 15 plus years at this point. I will argue that we haven't figured out still how to regulate that correctly and how to do it responsibly. So AI is much newer. AI is arguably way more powerful, especially in potential. So it's going to take us as mankind a while to figure out how to do it correctly. Now as Replit, I think we are in a privileged position where, for instance, we don't generate text. We rather generate code. So some of the shortcomings about having a model online that generates fake news, maybe they don't apply directly to us. We are not into image generation and all the issues regarding using stock images that are copyrighted. We don't do face detection. So I would say we dodge some of the hottest short term issues about AI. I don't think we're part of the discourse about AGI because I think it's not in our radar at the moment. We are way more grounded to planet Earth technical challenges to work on before we think of that. And again, I'm glad that there are other institutes and other companies thinking about it. So the internal discourse that we have now about AI is mostly how do we make sure that our users are aware of what we do. And that's the reason why, for example, we even released our LLM as open source, and we use open source datasets. So we try to be as transparent as possible when it comes to our AI features. By all means, I'm not claiming that everything we're doing is perfect. But by means of being as transparent as we can, we get a lot of feedback, and we can correct course easily. So it's better than working in a vacuum at least. We know, and we get to hear what people believe we do wrong, and we try to fix that. So I think, yeah, responsibility is important to us. The political discourse, I don't think we even have the cycles to spend time on that right now. We have so much work on growing that that would be, I think, a waste of time in the short term.

Nathan Labenz: 29:17 You guys recently put out this state of AI development report, and one of the things that jumped out at me was the rise of LangChain, which we've talked about many times in many different episodes. I've used it moderately, but not super intensively, because the biggest projects I've done actually got underway before LangChain. At the same time, it's taken a lot of heat recently online for being, I'm not sure, people suddenly, for a second, seemed like it had flipped and it's not cool anymore. I'm not entirely sure why. I'd love to hear your comments on LangChain in particular and that tool ecosystem in general, and kind of how you would guide developers towards or perhaps away from certain tools.

Nathan Labenz: 29:17 You guys recently put out this state of AI development report, and one of the things that jumped out at me was the rise of LangChain, which we've talked about many times in many different episodes. I've used it moderately, but not super intensively, because the biggest projects I've done actually got underway before LangChain. At the same time, it's taken a lot of heat recently online for being, I'm not sure, people suddenly, for a second, seemed like it had flipped and it's not cool anymore. I'm not entirely sure why. I'd love to hear your comments on LangChain in particular and that tool ecosystem in general, and how you would guide developers towards or perhaps away from certain tools.

Michele Catasta: 30:05 So first of all, I wouldn't guide anyone away from LangChain. I think it's such a fundamental tool in the AI space right now that, at the very least, it's interesting to follow, and I'm grateful of Harrison and the whole team building it. I found it amazing that it's basically collecting most of the research papers that come out. They also have an implementation on LangChain within 24 hours. So it's a treasure trove regardless of how much impact it has in production. Now that being said, I think LangChain is going through the same life cycle of frameworks in general. I remember back in the days when we had the first powerful web frameworks such as Django and Ruby on Rails, they were getting exactly the same heat in the sense that when you force or when you recommend a developer to work in a specific way, some people will love that kind of mindset, and they will be compatible with that. And some people will find it completely counterintuitive, or they will hate the fact that debugging LangChain is, of course, more challenging compared to debugging your own code. So it always comes with pros and cons. And that's the reason why, even as of today, if you want to write a basic web application, you probably have a choice of at least the top 10 frameworks out there. And each one of them is good, and each one of them is powerful. So LangChain is getting beat because it's probably the first one covering that space. I wouldn't be surprised to see a few more popping up in the next months. And I think that the code base, no matter how intricate it could be today, first of all, it can always improve. And, again, we have seen this in other frameworks. Maybe TensorFlow went through something similar, although TensorFlow 2 wasn't exactly a success story. But even in deep learning, we have seen libraries coming out in version 1, collating all the feedback from users, and then the next big release is much better because the developers learn the best practices and the anti-patterns. So let's give some time to Harrison and the team to maybe make a next release on LangChain. But I think this is the nature of AI development today. It's early, it's scrappy, and it's going to be headache inducing. But that's also the reason why it's fun.

Nathan Labenz: 32:29 I think it's a really great answer. Great perspective. And I definitely would encourage everyone to follow the project at least. As you said, they are incredible. You might even call it Replit Speed. They might coin their own LangChain speed. I've seen not once, not a few times, but a lot of times, as you noted too, where they very, very quickly implement some new scheme that somebody published in a research paper. And that's incredibly useful from a developer standpoint.

Michele Catasta: 32:59 Yeah. That's making AI better as a field if you think about it, because whenever a paper comes out, there is always some doubts about how easy it is to reproduce, are those results real or not. And the fact that the implementation either sometimes the authors themselves build it in LangChain so that it's easy for people to reproduce, or someone in the community is going to take care of that at the end of the day. That allows us to find out is that paper actually real, or are those claims completely false. And that helps us to make progress as a field. So even if it doesn't have impact in production, it's so valuable on the research side that, again, I'm glad that we have so many people working in AI today compared to even 6 months ago, literally.

Nathan Labenz: 33:43 It's powerful. That also just makes sense to me as to why I wasn't bothered by any of it because I generally am fine to follow the framework. So, I think I'm the framework target market, I guess. You had said, with all these tools, and especially all the AI tools, obviously, the goal here is higher productivity, greater accessibility, more developers. As I understand your earlier comments on the ADI notion, it's like, we're not looking to replace developers. We're just looking to give them a next generation set of tools that will make them much more productive. Do you have a sense for how much more productive we're talking about here? Like, are we talking 2x more productive? Are we talking 10x more productive?

Michele Catasta: 34:27 I would say the 2x has been already kind of accomplished by code completion. In the process of at least writing code, Copilot has published some metrics where they make developers up to 55% faster. We have something comparable with Replit. We didn't publish the metrics yet, but we see high acceptance rates from our model. So I would say the 2x threshold has been kind of reached. And also, anecdotally, several people tweeting about this fact that they would be wasting half of the time if they were not using modern code completion. I think 10x is attainable, achievable in the next year or so, and that's by means of having better ways of debugging code, more agentic behavior. That's another direction that we can go more in depth later. But this idea that you don't want to use an AI exclusively to generate blocks of code. You want an AI that comes up, first of all, maybe with a plan of what you need in terms of architectural design, and then that can go ahead and build the scaffolding of a project and then gives you the basic implementation of everything that you need. And then as a developer, you just put the finishing touches, you put some glue code, you make sure that the few bugs left are done. That I see easily giving you a 10x improvement. I think Amjad is having a few inspirations on podcasts where he talks about 1000x developer. So that's our north star, which I think is going to take us a while. And the idea there is the moment you have a model that can not only generate what you need, but can also orchestrate tools that are really available. And this is something that I explained in the ADI talk at the beginning. Then all of a sudden, you have a group of developers that work for you. You can think of yourself as a tech lead, and then you have basically the AI below as a set of interns and general developers that can get tasks done immediately. So, if you know that you need a certain pipeline, you need to fetch data from somewhere, you don't have to write the code any longer. Maybe you have Replit in the test that, and the API knows, oh, I should be invoking that service, getting back the results, process them in this way, and then bubble them up to what the developers are adding. So that's the future that I believe will start to give orders of magnitude improvements. But, yeah, I think 10x is something that we should be seeing relatively quickly. The idea that agents are going to become more powerful and more independent, which is far from being an easy task.

Nathan Labenz: 37:13 I'm feeling a little cognitive dissonance around another billion developers, but also 10 to even 1000 times more productivity because I'm thinking it's hard for me to even fathom how much software gets built or what we would be all building in that scenario. If we add another billion developers and make developers just 10 times more productive, then we have the equivalent of at least 10 billion current developers. And that means we have more than 1 current developer productivity per person on Earth. Is there demand for all that software? Taking the other perspective, not thinking so much about the developer experience for a second, but just what does software even look like if there is more than 1 current human developer productivity per person available to be building stuff?

Michele Catasta: 38:15 Right. That's a great question. I think it's hard to make a good prediction about that, but what I can tell is every person who today is using any kind of technology is spending quite a lot of time on tedious tasks. And that's the reason why, for example, you have shortcuts on iOS that allows you to chain a few actions. That's why a lot of people do macros or functions on Excel spreadsheets, and that's why there are variables and rules in Docs today. So you see these hints of programming languages and automation bubbling up in every widespread application. And the reason being everyone who's a bit tech savvy will love to automate even further their life. So to me, having the next billion software developer doesn't mean that 90% of them will be writing the new LangChain or the new Google or whatever it's going to be. It's more about, can we make programming super basic that it becomes a useful skill for everyone? And I think there is a future for that. I've seen people building WhatsApp bots during some of our hackathons for very specific needs, like things that were ordering pizzas for a group of friends organizing a party. So rather than having to download applications and all the startups build that kind of niche product that is never going to have a powerful business model, you're going to build yourself the few niche things that you care about for yourself, for your family, and for your friends. So that's the horizontal growth that I'm talking about, so the billion software developers. Now on the other dimension, that vertical growth, how much a software developer can become more productive. I think the multiplier, the thousand x doesn't apply to the amateur person who's automating their life. It applies more to the John Carmack's of the world or the Jeff Bezos of the world. Like, people that are extremely skilled software developers, they manage very large teams in their lives, And all of a sudden, they are not going to need anymore a full org below them. I think Carmack was in Meta for a while, and Jeff is probably managing thousands of people reporting directly or indirectly to him. I imagine replacing that or part of what they do on a daily basis with an AI and with an agent and with an API. That's the kind of thousand x that we think of. And the goal is not to displace the helpers. It's actually to make even more people capable of building big and great companies. So we would love to see fewer humongous companies and way more mid-sized startups that can tackle very challenging problems. If you think about it, that's going to be a big accelerating factor for mankind. I don't want to go back to the e/acc, but that's the acceleration part that I'm excited about, and that's the reason why I've been doing research in this field and I want to build products in this field because it can really have amazing benefits for mankind if we crack it's probably corrected.

Nathan Labenz: 41:24 Yeah, I certainly see the billion fold opportunity in kind of last mile customization and integration of different things. That's honestly, I think, a big part of what the EAs at Athena will end up doing with a platform like Replit, is they're going to say and I'm developing all these mantras. One of them is copy and customize. And Replit, obviously, is a perfect platform for that, where the goal is to find a good starting point template, fork it, make it your own. And a lot of times those things where you're making it your own are not conceptually complicated. It's just like, I use this messaging app and I want it to trigger a certain thing when a certain thing happens or whatever. Or I use this task tracker, or I use this database or spreadsheet, whatever. Everybody has their own kit, and there's not really a great way to bring those things together in some universal sense, although Zapier certainly and others try. But even then, you've got to make Zapier work. So that part definitely makes a lot of sense to me. I guess I'm reminded of we had Flo Crivello on actually, at one point, I think on the same episode as Amjad earlier, and then also he did his own. And he talks a lot about this single use software paradigm where he holds up an aluminum can and he's like, when Napoleon was alive, he was the only guy that could drink out of aluminum because it was so special, and now we just throw it away. Am I in line with your vision here of just software spinning up in little throwaway moments all around us?

Michele Catasta: 43:10 Absolutely. Yeah. That's exactly what we want. And, in a sense also, we want to empower the creative process of software, not the excruciating aspects of developing complex software because that, for a while longer, it will be still the responsibility of larger companies. It will require process. It will require management. It will require PMs, TPMs, and so forth. So building large software still requires a lot of human effort. But allowing a lot of people to be creative will also expand dramatically the pool of good ideas that we explore as mankind. So I think the analogy that you were bringing before from Flo and drinking from an aluminum can makes a lot of sense.

Nathan Labenz: 43:55 So, as I hear you describe that future, I guess it seems to me like there's some sort of implicit assumption where, for context, I would describe the state of play today as AI has become better than the average person at a random task. And certainly, that includes coding. Right? AI compared to person off the street, AI is going to dominate in code. And AI is closing in on expert performance where there's a pretty well established protocol of what you're supposed to do. So things like answering medical questions, when the answer is pretty well established and you as a doctor are supposed to know, then the AI is also getting really good, closing in on doctor level for just answering a question in the right way. And then there's this third tier of the real human genius or creativity of coming up with the spark of a new idea, whether that's a scientific hypothesis or an application concept or whatever. That basically I don't see AI doing anywhere as of now. And it seems like expectations underlying a lot of what you're saying are like, it levels out before we get there. Am I interpreting that correctly? It seems like we already can do the basic stuff. You want to enable closer to expert so that it can be more and more useful, but you're just not expecting that third tier of breakthrough where AIs actually drive creative process. Nathan Labenz: 43:55 So, as I hear you describe that future, I guess it seems to me like there's some sort of implicit assumption where, for context, I kind of would describe the state of play today as AI has become better than the average person at a random task. And certainly, that includes coding. AI compared to person off the street, AI is gonna dominate in code. And AI is closing in on expert performance where there's a pretty well established protocol of what you're supposed to do. So things like answering medical questions, when the answer is pretty well established and you as a doctor are supposed to know, then the AI is also getting really good, closing in on doctor level for just answering a question in the right way. And then there's this kind of third tier of the real human genius or creativity of coming up with the spark of a new idea, whether that's a scientific hypothesis or an application concept or whatever. That basically I don't see AI doing anywhere as of now. And it seems like expectations underlying a lot of what you're saying are like, it levels out before we get there. Am I interpreting that correctly? It seems like we already can do the basic stuff. You want to enable closer to expert so that it can be more and more useful, but you're just not expecting that third tier of breakthrough where AIs actually drive creative process.

Michele Catasta: 45:35 Correct. That's exactly why we called it ADI and not AGI. The reason is the creative process you talk about, for me, is still exclusively in the realm of human intellect. And being able to have an AI that designs and thinks of the next LangChain, for example, that to me means that we are capable of building a reality that is as smart as a human or even smarter. I don't care about when that happens. If it happens 1 year from today, of course, it's gonna change drastically the way in which Replit looks like, but it will be accessible to everyone most likely. So we will not disrupt only Replit. It will disrupt every single company in the world. It will be a disruptive factor for humankind. But whatever is the timeline for AGI, there is this interim solution where, as you said, we level out, we help as many developers as possible gain, to be more efficient on basic skills. But then the creative process is still something that only humans can do. And 1 example I always use when people ask me, shall I drop out from my computer science degree because maybe there's not gonna be a job for me in the future? And the answer I give you is we still need people coming up with the design of the next library and the next programming language. So if you think about it, what we've been doing in the last 60 plus years in computer science is literally thinking of new abstraction layers. We were working with punch cards and then assembly and then C, and then we started to do operating system kernels and libraries and web frameworks. So every time there is a new paradigm shift, that has been humans working together, learning from the anti-patterns and the pain points of how we're writing software today, and then coming up with the next way we do that. An AI can learn how to do things as we write code today. We can't think of how we're gonna be writing that in 2 years. So that is still something that humans have to work on. And it's great to me that if we reduce the amount of tedious tasks that we have to be working as developers, then everyone can focus on how can we make things better compared to how they are today. So this is the dream that we have. The reason why the, I would say, the AI for code should exist is to make sure that we can put our brain cycles on something more interesting than just writing code as we do today.

Nathan Labenz: 48:10 Just to make sure I understand your outlook, it sounds like you're not, by any means, ruling out that there could be this next advance such that, oh my god, now AI can drive creative process. You're not ruling that out. It doesn't even sound like you're saying it's particularly unlikely, as you said, hey, maybe it could happen next year. But it's basically just like nobody has any idea what to do about that. So we just kind of plan for a more normal scenario because that's all we really can plan for. And normal can still be pretty weird, but it's like near human level developer would be transformative in and of itself. But that's kind of the most transformative future you can plan for. And so that's what you focus on.

Michele Catasta: 48:54 Correct. AI has a history of going through hype cycles and winters. And I think in the early 60s, Marvin Minsky wanted to hire a summer intern for them to solve completely vision. We're talking about 60 plus years ago, and that problem was considered to be trivial. And it took us literally 60 years to have models that are powerful enough to match the vision that Minsky had. It would be, I would say, ill advised for me as a person leading the AI in Replit to say, oh, let's just work on AGI because it has to happen in the next 2 years, and that's the only reasonable path for us to make. If I have to think about the probability distribution of events, all my probability mass is on ADI can happen, and we can build it, and we can bring benefit to our users. And if in the low likelihood that AGI happens before us, we're gonna adapt. We're gonna be there. We're gonna keep our eyes open. We're deeply entrenched in the AI community regardless. So we're gonna be hearing about it because I think people are not gonna be talking about anything else across the world, to be honest, when it happens. But in the meantime, it's much more realistic to say, I'm glad that people are working on more long term problems. And in the meantime, let's focus on something that I feel we can actually do.

Nathan Labenz: 50:14 Where I come out on that, just to put my cards on the table, is I don't think we should necessarily continue hyperscaling orders of magnitude beyond what we have because I can't rule out that a sufficiently powerful system might pop out of there to be truly very, very disruptive. And I would rather give us a little more time to get used to what we have and implement it and enjoy our AI coding assistance before we kind of say, hey, what happens if we scaled up another 1000x more training compute? Where do you come out on that? There's multiple different questions, but before we even get anything like regulation or whatever, what would you advise if you were saying, hey, Sam Altman, do it or don't do it? Do you feel like if they do that, we are at some risk of crazy disruption? Or you feel like it's, nah, that's more just kind of speculative than anything?

Michele Catasta: 51:17 First of all, given that I'm not a person who studied these topics in depth, my answer should not have a lot of weight. So I will just express what is my opinion, but I trust much more people who have been studying this for several years. I'm far from being a well informed person on responsible AI and all the different issues related to that. That being said, I think that no matter what is our opinion, the scaling will happen. I think there have been letters signed by several influential people, and the progress hasn't slowed down at all. We know that even startups at this point have enough capital to build exceedingly large H100 clusters. Inflection is 1 of them. They just announced how big their infra is gonna be. So I think it's pointless to talk about how can we stop that. It's more important to pour our energies on, okay, once that happens, how can we be ready? How can we make sure that nothing goes in the worst possible way? I might be naive, but I tend to be macro optimistic about humankind, in the sense that eventually, no matter how many foul players are there and no matter how weird behaviors we experience in our history timeline, eventually, it seems that as humankind, we'll always find a way to survive and thrive. So I do hope that even when AGI happens, we're gonna be driven by common sense. I think some of the doomsday scenarios that we have been hearing through mass media and Twitter are possibly just detrimental. I don't think it's worth to talk about them. They just give even more fear to people, and we came out from 2 plus years of pandemic fear, and now we just reached the AGI fear. It's not helping anyone, to be fair, to be that pessimistic. I think it's important to think of ways to mitigate that in case it happens, but I wouldn't go around and get on the news and talk about nuclear bombing data centers. That doesn't seem to be a healthy way to start a good discourse.

Nathan Labenz: 53:29 Yeah. And I do agree that that was probably unfortunately phrased at the best, even though I'm fairly sympathetic to some of the argument. But when you get into airstrikes territory, it doesn't do much for discourse, I'm afraid.

Michele Catasta: 53:45 Exactly. Anyway, I think I'm glad that there is discourse going on regardless. I have seen people on both sides starting to have debates and maybe softening their points of view from both hands. And I think that's the kind of discourse that's gonna lead to progress and to us being ready as humankind for that event to happen. I think any extreme point of view, I think, historically, it never helped. So I wouldn't over-index on the extremes, for sure.

Nathan Labenz: 54:16 Okay, cool. Well, let's return then to some near term concerns. I wanted to just give you a chance to talk about, in a sense, what has you excited about Replit's future? But also, there's this constant debate around who has moats in the AI space, what do moats look like? I just spent 3 minutes thinking, what are some moats that I see in the case of Replit? 1 we already talked about is just having the incredible infrastructure that you already have. That even extends to multiplayer mode. When I tell the assistants, by the way, you can jump onto my thing right now and start helping me code in the environment that I'm in. Here's the link. Again, I don't think they appreciate how much. They're not even as mind blown as they should be. Because to them, it just feels like, okay, that's something I can do. But, yeah, historically, people who've done this for longer would be mind blown. User feedback seems like another significant 1. Nobody is getting, I would only identify 2 companies, Microsoft and Replit, that are getting the kind of in-line frequency and volume of user feedback that you guys are getting today.

Michele Catasta: 55:25 Let me correct that. I think we got even more than Microsoft. We have way more telemetry. VS Code is not logging at the level that we do, of course, because we are running in the cloud, and everything is running on our backend while VS Code is mostly a client-based experience. So I would say there are 2 companies that have the data. 1 is Google because they're on their own internal IDE and that looks very similar to Replit as a concept. And there are some papers that explain how the architecture looks like. And the other 1 is Replit, but with a much larger user base. So that is a dataset that I'm extremely proud of, and I can't wait to show models trained on the data that we collect.

Nathan Labenz: 56:08 So a third 1 I was going to say is community. And then within that 1, maybe I'll ask a question around, is Microsoft not logging what copilots are completed? It's not for technical reasons, right? It must be more a matter of licensing or positioning or sort of how they're relating to the customer?

Michele Catasta: 56:28 They do log, of course, Copilot interactions, and that's how they can track so many interesting metrics in the blog post that they publish. But part 2 was more the fine grained interaction that we capture, such as literally keystrokes. We have the complete edit history of what you do in the editor. So not at the level of commits, but at the level of keystrokes. And that is something that VS Code, to my understanding, unless you install additional extension, Microsoft is not logging the code that you're writing on VS Code because that would be a nightmare for them in case it happened. But we do because we offer features such as history and rolling back, and we have multiplayer support. So we need to make sure that we can reconcile edits coming from different users, and that's how we have to maintain all the additional data pipeline. So it's very unique data that 1 day we plan to pour into an AI model and give it back to the users.

Nathan Labenz: 57:32 Yeah. That history feature is pretty cool unto itself. It's even really more than like a Google Docs. It's kind of a playable history of everything that you've done in the environment. And it's another 1 of those things that feels like, man, why wasn't it always that way? It's quite a few of those that Replit has managed to create. So, going back to the community then, you mentioned, it could be a PR nightmare for Microsoft if they're logging all your keystrokes. I mean, obviously, they have enterprise customers. Replit's coming at it from a totally different user base where people sort of maybe more assume that they're using a web app and that data is gonna get logged. But how do you think about the relationship between the platform and the community, obviously, of the marketplace for bounties as well? And that could even be said to be the perfect input output pair for a random user who doesn't know much, asks for this, and this is the code that they got. So I see these kind of compounding dataset advantages. But how do you think about yourselves relating to the developer community? What rights should developers have? This could be specific to Replit or in general. You guys are kind of, I'm sure, training on open source software and there's lawsuits around that. But then there's also, how should I think about if I'm trying to create something original of value on Replit, do I have an opt out? Should I have an opt out? My stuff is going into the next generation of model. Should I feel like that's taking from me in some way? Or what do you think the terms of that engagement and relationship should be?

Nathan Labenz: 57:32 Yeah. That history feature is pretty cool unto itself. It's even really more than like a Google Docs. It's kind of a playable history of everything that you've done in the environment. And it's another one of those things that feels like, man, why wasn't it always that way? It's quite a few of those that Replit has managed to create. So, going back to the community then, you mentioned, you know, it could be a PR nightmare for Microsoft if they're logging all your keystrokes. Obviously, they have enterprise customers. Replit's coming at it from a totally different user base where people sort of maybe more assume that they're using a web app and that data is going to get logged. But how do you think about the relationship between the platform and the community, obviously, the marketplace for bounties as well. And that could even be said to be the perfect input output pair for a random user who doesn't know much, asks for this, and this is the code that they got. So, I see these kind of compounding dataset advantages. But how do you think about yourselves relating to the developer community? What rights should developers have? This could be specific to Replit or in general. You guys are training on open source software and there's lawsuits around that. But then there's also, how should I think about if I'm trying to create something original of value on Replit, do I have an opt out? Should I have an opt out? My stuff is going into the next generation of model. Should I feel like that's taking from me in some way? Or what do you think the terms of that engagement and relationship should be?

Michele Catasta: 59:09 Totally. So we were working on an opt out feature. It's going to be in our documentation very soon. I will send it to you when it's up and running. You're totally right that we are training Volley on permissively licensed code. We didn't want to make the original mistake that Copilot did. I think it's a gray area from a legal standpoint, especially if you are in the US, then there is, you know, you can follow the fair use, and that will, in theory, allow us to keep training models, and perhaps those lawsuits are not going to go much further because of that. In practice, I think it's way more important for us to be correct towards our users. So we're going to be listening to them. I think until now, we haven't received any pushback. And, again, with the opt out feature, we think that we're going to be making it right for some people that want to make sure that their code doesn't end up in the models. By the way, we only train on code that is public on Replit. So if your rep is private, we don't even touch it. And the rationale there is if you are keeping something private, you know, then you might have good reasons why. And then, of course, we don't want our models to be trained on that. Now I don't think that's the end of it all. I think we're moving to ways of attributing contributions to users that are more advanced. And I think you have seen in Bard, for example, that they give you links to the snippets of GitHub code that contributed to what Bard has generated. So we're going to be moving to that model in the future. And I think the best way for us is really to follow the discourse and see what's going to become the industry standard. We want to be like a beacon in terms of how we treat our users, our developers. So we were going to try to adopt those features as soon as possible. And then if the lawsuits go in such a way that will change our mind, then we're ready to adapt. But as of now, I think we are trying to follow what's considered a good practice. That being said, if you want my personal opinion, I do think that the landscape will change as it's changing for image generation. And I think stable diffusion was the enlightening factor to start the discourse. And, you know, maybe there would of course be lawsuits, and there would be even more debates. And perhaps, you know, few years from today, we're going to be finding what is the right balance between giving attribution and still training such models. I would love to see a model where if training data created by user contributes to a lot of code completions or generations, then we find a way to share part of the revenue with that user. It's far from easy to be built. I think it requires a lot of infrastructure both on the AI side and on the payment side. But that's one idea that I was riffing with Amjad back in the days. And, you know, perhaps one day we're going to be able to make that happen.

Nathan Labenz: 1:02:17 I don't know if that would qualify as an interpretability complete problem, but it's close probably, right?

Michele Catasta: 1:02:24 It is close, yes, because a lot of common code patterns will appear pretty much everywhere. What do you do in that case? How can you attribute it to the first person who came up with that? Or is it too trivial that it shouldn't even be attributed to anyone? So it's a very hard problem to solve. And, you know, what we do is we keep an open mind, and we read papers, and we discuss with the community. And I think we all have to adapt in a sense. Companies have to adapt. Developers will adapt as well because, you know, there is progress going on. It was the same, in a sense, what happened with the open source movement back in the Linux days was also another disruptive event. And some people loved it, some people hated it. So, I'm sure this will happen here with AI for developers.

Nathan Labenz: 1:03:12 So, one thing that you guys have obviously also made some news on, but which is not apparently something you think of as a competitive advantage, is training custom models. Or maybe you could correct me, but you're training them, you're open sourcing them. It doesn't seem like you view the models as any sort of secret sauce or kind of core defensible IP. So, with that in mind, I guess I'd love to hear a little bit about how you guys think about training them in the first place. Why do it? Why not just use somebody else's if it's not something you feel the need to ultimately own? Is it a dataset issue? Is it just the desire to show off? Which honestly might be good enough reason. Tell me about why create and open source these original models.

Michele Catasta: 1:04:02 Of course, showing off or proving that we're capable of doing that and we have those skills in house definitely helps because, you know, it helped us to talk with a lot of people in the AI community who then ended up maybe joining us or deciding to help us in the future. So, it's important for a company not to just talk about AI, but rather to prove that they can do it. So that was a way for us to let us stand out. That being said, Replit is a great success story built predominantly on open source software. We love open source at Replit. But vast majority of things we use and the projects we rely on, we either donate back to them or we do a lot of upstream contributions. So Replit wouldn't exist if open source wasn't such a thriving community. So we felt the moment we decided to work on our models, we almost felt compelled to give back. And it came almost as a no brainer discussion. I mean, I wouldn't say that it took us half a day to make the decision. Of course, we talked about it a bit internally. But at the end of the day, there was complete agreement that it was the right move for us to make. Now I don't think that means that we don't see any competitive advantage in the models. As a matter of fact, we released the baseline model, which is trained on an open source dataset. Now we are forever grateful to BigCode project to have open sourced the stack, which is a permissively licensed dataset extracted from GitHub. But the model that we today use in production is a combination of that dataset, and then it's a further pretraining done on Replit data, which is not accessible to others. We're not going to be releasing our users' code bases. And that model comes with a pretty substantial performance improvement compared to the base one. Depending on the language, we have been seeing all the way even to up to 50% improvement compared to the base model. So that's the way we kind of reconciled the two aspects. On one hand, it's great to give back to the open source community, and we got a lot of cool projects built on that. Honestly, more than we ever expected. And then at the same time, you know, we built our own custom model that is in production since early May, if I recall correctly, and it's been doing great. And, yeah, I remember there was another sub question that you asked, which is why did you even do it? Why don't you use third party models? We had in mind a very specific trade off, as in we had a fixed point in terms of latency. We wanted the model, most of the response we want to give them back in 200 to 250 milliseconds. That's the threshold that appears to be basically instantaneous for a human. So we had to work with a certain model size given the GPUs that we had available a couple of months ago. And we decided, okay, given that we want roughly a 3 billion parameters model, can we squeeze the best possible performance out of a small model so that we give a different experience compared to Copilot? You use Copilot, you know that usually the latency there is at least 1 second. And our users love the fact that maybe we generate shorter completions, but they're much faster. So we wanted to build the best possible model for that use case. It wasn't available open source, and we took the challenge and we made it happen.

Nathan Labenz: 1:07:30 How much of a hassle is that? I know that you have worked for at least some of these projects with Mosaic ML, another recent guest. How much of a lift? If you are, I think a lot of people listening here probably work at companies that are trying to figure out what they're supposed to do. And I think for most of them, certainly pre training their own custom model is not going to be the thing. Fine tuning is more in play for most, but even honestly, a lot, they don't even need that. And then, you know, in some cases, just need to set up a vector database and get a chat running against it. It seems like it's getting pretty accessible to do these pre trained things if you have the data and the need. What was your experience like working with Mosaic to actually go through this whole process?

Michele Catasta: 1:08:19 Yeah, Mosaic was a great partner for us. I don't think we would ever dared to work on the project with such a short timeline if we didn't have them covering our backs and giving us access to their infrastructure. So the truth is training a language model today still requires quite a lot of engineering lift when it comes to orchestrating several different nodes and GPUs, writing the right training framework to build your model, doing a lot of ablation studies to find out the right architecture, the right hyperparameters, and so forth. So in the last few months, it's true that we have been seeing quite a lot happening in the open source world. But as you said, the vast majority of progress we see in open source is about fine tuning and instruct tuning. And all of those tasks can usually be accomplished with a couple of consumer grade GPUs. Way fewer companies and people are training from scratch models. So to attempt that task with a very small team, in my case, we are basically 2 people and a half working on this for a sprint of 10 days before we released it and maybe a couple of weeks before doing ablation studies. That happened only because we had access to Mosaic infrastructure and also quite a bit of back and forth with the engineers and the researchers there who helped us a fair share during that process.

Nathan Labenz: 1:09:52 So your job was largely to define what good looks like and assemble the data, and they kind of handle babysitting all the GPUs, orchestrating, obviously, and also that kind of hyperparameter expertise?

Michele Catasta: 1:10:10 Yeah, largely on our hand, we worked on data collection, curation, and a lot of ablation studies with smaller model sizes. And then Mosaic helped us with their software framework that orchestrates multiple nodes. We mostly maybe set the run ourselves in the sense that we were in front of Weights and Biases 24/7, seeing if something was going wrong. But, anyway, when we hit a snag or when there maybe were some harder failures, Mosaic people were on call pretty much as us, 24/7. It was a pretty amazing team effort that we did together. And, yeah, we shared a lot of notes, for example, on what was the best learning schedule to use for this model size. And, you know, we made different choices, for example, compared to the models that are open source. So depending on how much you want to personalize the model, you can go and do it more like a white glove service where you put data in a bucket and you ask Mosaic or other similar companies to do the training for you, but then you get something not customized to your needs. For Replit, we trained our own vocabulary. We had a different architecture. We came up with a different even training schedule in terms of how many times we repeated the data, how many we did on the core dataset that we used. So it was kind of an uncharted territory to an extent, and that's why it was a lot of fun. Felt I was doing research and product at the same time, and that was a fun process.

Nathan Labenz: 1:11:43 Yeah, I love that about AI right now. Blurring of research and development and productization is super fun for me.

Michele Catasta: 1:11:53 People ask me, Do you miss doing research at Google? I said, you know, maybe I'm not doing that full time, but I still do a lot of cool stuff.

Nathan Labenz: 1:12:00 Any highlights that you would give? I mean, it's interesting right off the bat that you trained your own vocabulary. That essentially means you have your own tokenization scheme. Right? So, we just did an episode with a woman named Lily Yu from Facebook, who was the author on the Megabyte paper, where they're kind of trying to get rid of tokenization altogether.

Michele Catasta: 1:12:21 Yeah, I saw some of the early versions of that work. It was amazing, yeah.

Nathan Labenz: 1:12:24 I don't think the transformer is the end of history, you know, as it currently stands. That was definitely a big takeaway. But, so, any kind of notable details that you would want to highlight there? And, you know, any big differences between your tokenization scheme and kind of what people are used to or any architectural decisions that you think people would find particularly interesting?

Michele Catasta: 1:12:42 Two key insights. The first one is we went with a smaller vocabulary, which, by the way, is the same size as the first Llama release. Honestly, I have to read in detail the Llama 2 paper. I would be surprised if they went for a larger vocabulary. But, anyway, a smaller one comes with a price of slightly worse compression in exchange for better inference speed. And, again, given that our model is optimized for being in production rather than benchmarks, we made that design choice of going with a smaller vocabulary.

Nathan Labenz: 1:13:15 Can I just ask how big the vocabulary is?

Michele Catasta: 1:13:18 Oh, sure. It's 32k.

Nathan Labenz: 1:13:20 So solid gold magic carp is right out. Nathan Labenz: 1:13:20 So solid gold magic carp is right out.

Michele Catasta: 1:13:23 Exactly. Yeah. I think GPT-3, for a while, had a 50k vocabulary. GPT-NeoX had a rounded ballpark. The 3.5 Turbo and 4, I think they all have a 100k vocabulary. So that's why we start to say that a token is usually around 4 characters. With the kind of vocabulary size that we use for the Replit model, it's roughly 3 characters per token. So slightly worse compression in exchange of faster inference, but also we specialize our vocabulary on code. We train it on an exclusively code dataset. And that means that specifically on code, it achieves a better compression rate. So what you will miss as a whole if you use that vocabulary for a standard natural language dataset, in reality, you gain that back when you work exclusively with code. And again, we could make all these choices because we knew exactly that the model was meant to do one thing and one thing only. So we're not building a generic LLM, we're building one specialized for code completion.

Nathan Labenz: 1:14:28 How do you think about the world knowledge that underlies a lot of programming tasks? One thing that I've tried, for example, is write me a tic-tac-toe game in JavaScript or whatever, right? If I go to, and obviously that assumes or requires that you know tic-tac-toe, what it is and what the rules are and have that kind of... Maybe there's enough tic-tac-toe demos out there that kind of finds its way into a coding set. But you can imagine a lot of things like that where there's general knowledge assumption in the program specification that a code-only dataset probably doesn't have because that stuff doesn't make it all the way into the code itself, I guess, in a lot of cases. How do you think about that? I mean, that starts to go from the ADI to AGI kind of a little bit as well. But how do you think about the world knowledge that you do or don't want to bake in?

Michele Catasta: 1:15:26 So I think small models showcase a certain degree of world knowledge. Out of curiosity, we ran basic natural language benchmarks from the Replit model, and it turns out that it was, on some of them, even competitive with LLaMA 7B, which is a much bigger model, especially these self-contained benchmarks that didn't require a lot of external knowledge. They were more based on basic reasoning skills. And the reason is we know for a fact that when you train a model on code, the reasoning performance improves quite dramatically. So far from being a model that doesn't know how to do anything except code. But of course, it lacks several skills compared to much larger models, and we're very much aware of that. We never claimed the opposite. It was more out of curiosity. We tried to test our code model as completely clueless about anything else. And the truth is, no. They can do some basic task and there are... I tried some theory of mind prompts that became common with GPT-3, and it was doing a pretty decent job also then. So scaling perhaps is one of the ingredients required to do much better at world knowledge. And I would say also retrieval augmented generation seems to be making the difference there because no matter what is the information cutoff of a model, you will always be lagging behind compared to the present. So it's important to have a way to inject information in the prompt to ensure that the model knows as much as possible. But yeah, I think scaling seems to be the main ingredient to make that happen today and to add new skills. And at that model size, we know for a fact that a lot of skills are going to be lacking. But again, that's the beauty of specializing a model. You know the trade-offs you're making, and you know what you get in exchange. We did it also to show that not necessarily all the effort should be in the mega models. We're glad that people do that because we use them, but there is also space for smaller ones optimized for a specific use case.

Nathan Labenz: 1:17:36 And you will presumably have a whole suite over time of various sizes?

Michele Catasta: 1:17:41 They're coming. Yeah. We're working on them.

Nathan Labenz: 1:17:46 Yeah. Watch this space. So just a couple final topics that I wanted to touch on. One is kind of returning to this notion of safety. And now I'm approaching this in a more kind of mundane way: safety for the user, safety for the platform, safety for other users. I know that you guys are making progress. I'm sometimes amateur, sometimes official red team enthusiast. And so I've tried asking just about every code gen AI to write me a denial of service script. Results vary. Even on Replit, results have varied. In May, when I did that, it just did it. No hesitation, wrote me a denial of service script. I was like, this doesn't seem great. I tried that again this week, and it no longer does it. It now refuses to write a denial of service. I haven't gone all the way into can I jailbreak that or not, but I'd love to hear how you guys are working on that dimension because clearly you are. Do you have a standard process that is specifically dedicated to avoiding harmful code? Is there a sort of standard battery of tests that you measure these sorts of things against? I'm sure you have standard tests for performance, but is there a kind of safety specialist wing of that? Do you have a red team process? What does that aspect of the development look like?

Michele Catasta: 1:19:19 Yeah. We still don't have a red team process by means of how small is our team, unfortunately. So we need to make some choices regarding that. That being said, we rely mostly on third-party providers to make this happen because I believe that they're going to have more cycles and more resources to do a better job than we currently are doing on our own. So that's the reason why you see progress being made, and that's the same kind of progress you see being made today between GPT-4 March versus the June release. We've been seeing a lot of rumors on Twitter, I'm sure, about how the models have been further lobotomized and then OpenAI replying that this is not the case. And I think there is this interesting distinction between the skills of a model and the behavior of a model. So the underlying skills and capabilities are still the same because the base model is exactly the same. What changes is how the model behaves given a prompt of the user. And while back in March, for example, they were a bit more casual replying to you and giving you a DDoS script, now in June, the behaviors are different because the RLHF or, in the case of Anthropic, the constitutional AI approach that they use has been evolving over time. And I think every company is doing that. It's hard to come up with the correct solution to anything. I don't think anyone sees RLHF as the silver bullet that is going to solve the issue as a whole, but at least you can see progress being made. More interesting, now LLaMA 2 has been released and is heavily RLHF'd as you might have seen yourself from the demos online. I believe a lot will also be discovered there. There will be behaviors that are aligned with how we expect the models to behave, and then there will be behavior that are completely misaligned, and people will find ways to jailbreak it, and people will find ways to prompt it in such a way that it will reply horrible things. That's the nature of the beast, and it's still hard to tame in a perfect way, but I think we're all working towards making that better and safer.

Nathan Labenz: 1:21:37 So first of all, when you want to start that red team part of the company, give me a call.

Michele Catasta: 1:21:43 We'll do it. Perfect.

Nathan Labenz: 1:21:45 If I understand correctly, you had a custom model in May and you have another seemingly updated custom model now? Or have you added a sort of intercept layer? Because what I was assuming had happened was you'd updated the model, added some more refusal training, and now it's refusing. But maybe it's different than that.

Michele Catasta: 1:22:04 So our custom model right now is mostly used on code completion. I believe what we were using was Replit chat, which is using third-party models right now and some additional prompting and retrieval that we do on our end. So we try to make prompting on our side more robust, but there are also underlying behaviors of the models that we use that are changing. And we work with Google. We work with OpenAI. So depending on which side of the A/B testing you're on and the kind of changes that we allow that week, you might see a lot of different behaviors. So that is definitely not going to be stable in the near future because we don't like to rest. We always try to find ways to make it better. But yeah, it's a fundamentally hard problem. So I ask our users to be patient about the progress there because it's something that no one has cracked in the community yet. So likely, Replit is not going to be the first one to crack the specific problem.

Nathan Labenz: 1:23:08 For all the complaining that goes on about the models refusing to do stuff and being kind of chiding and annoying, I think that's actually a huge advantage for the OpenAIs and Anthropics of the world that specialize in it, because relative to having to manage that myself, if I'm going to go use LLaMA or whatever and fine-tune it to my ends, I want to ride in their wake of all the work that they're doing. And it's really interesting that you guys are also doing that, right? Probably a lot of people should update their thinking on how excited they should be to go adopt the latest open source model based on the fact that Replit is still using commercial models in their production product.

Michele Catasta: 1:23:59 Yeah. I think there is space for, similar to before we were saying, there should be space for incumbents and startups. I believe there is space also for open source model and models exclusively offered by companies, and each one of them would find their best application area. I don't believe that we should all be using the big boss model GPT-4 everywhere, not only because it's sufficiently expensive today, but also because it doesn't solve every single use case out there in the best possible way. So that's why I'm excited to see progress being made both by open source and by OpenAI and by Google and Anthropic and so forth.

Nathan Labenz: 1:24:36 You mentioned the economics a couple of times. Can you give a rough picture of kind of what users cost you or how you think about what users cost you? I've priced this out a bunch of different ways, even in looking at what is the wattage of an A100 and an H100, and what does that cost given my kilowatt hour price and that kind of stuff. But I think everybody kind of ends up having a little bit of a different cost profile that seems to relate mostly to the workload and a lot of idiosyncrasies of the workload. At the scale you're operating at, I imagine you have a pretty good sense of what the workload is going to look like hour by hour, one day to next. Maybe that's optimistic of me. But how do you guys think about, and you can share actual numbers if you want to, but how do you think about modeling the cost for a user, going from not just containers, but going from containers to now adding on models as well?

Michele Catasta: 1:25:38 Let's put it this way: the State of AI report that you saw published in our blog a few days ago, that kind of growth, we're also experiencing that in our AI features. We have more users that want to use AI. The moment they find it useful, they use it way more extensively than we expected, and then literally Replit becomes their favorite pair programmer, helping them to be as productive as possible. That comes with a price because of course, especially when you use third-party models, you're not being charged by how long you keep a server up and running. You're charged by traffic. So I would say if I give you a detailed answer today, I might have to deprecate it in 2 weeks because that's a constantly moving target. So we try to adapt to that. We try to be financially responsible, but not to obsess too much about cost because the landscape is changing constantly, and you're seeing yourself OpenAI slashing the API cost by like 2 orders of magnitude in the last 18 months. So I expect those economies of scale to improve over time. For now, it's more let's learn to build AI products useful to our Replit users, and then we try not to bankrupt. And we have good margins yet, and then the moment they become even more widespread than they are right now, and it makes sense to have our own models rolled out everywhere for economies of scale, we're going to work on that. But it's interesting to be in a rocket ride that goes so fast, but it comes with some additional headaches of sizing the load and the usage correctly.

Nathan Labenz: 1:27:28 Yeah. So does that kind of cash out? I think this is where most people end up being. If somebody's paying, then it's all good. You don't really have to worry about it. You're going to make money. You might make more on some and less on others, but...

Michele Catasta: 1:27:40 I'm not sure if I agree with that. It depends on what you're exposing. If you give them, say, GPT-4 and they are power users, you can quickly start to lose money on that specific user just after a few days. And that's the same reason why even OpenAI, when you're a ChatGPT Pro subscriber, they limit the amount of GPT-4 messages you can send. I think they just bumped it up, doubled it this morning, if I remember correctly. But we are still in this realm of scarcity, and the scarcity applies to any user, even outside of OpenAI. So it is not that trivial to come up with a proper pricing when there is powerful AI involved.

Nathan Labenz: 1:28:23 Yeah, interesting. So you can lose money on a small percentage of paying customers. Presumably, the average is still healthy. And then the real challenge that you kind of alluded to earlier is, how in the hell might we extend AI features to 20 million plus non-paying customers?

Michele Catasta: 1:28:40 Correct. So we're going to approach the problem gradually. I think we might be starting from code completion, which is easier. We own the models, and we learned how to deploy it very efficiently. So I see a near-term future where we can make that happen. Now putting something at the level of GPT-4 or Gemini in front of users, free users for free, that's going to take a while, I'm quite sure. No matter how fast is the progress on the inference frameworks and even if NVIDIA slashes the cost of H100s by 90%, who knows if this is ever going to happen, it's going to take a while until we're there. But models might become much smaller and cheaper to serve if someone comes with the next amazing architecture, so who knows?

Nathan Labenz: 1:29:31 Yeah. Do you think of pushing inference to the edge in the near term? Like, could you get your code completion to... I don't know if you can get through the browser all the way to the Apple Silicon, for example, presumably in a mobile app you could. But is that a direction you're thinking of pushing?

Nathan Labenz: 1:29:31 Do you think of pushing inference to the edge in the near term? Could you get your code complete to, I don't know if you can get through the browser all the way to the Apple Silicon, for example? Presumably in a mobile app you could. But is that a direction you're thinking of pushing?

Michele Catasta: 1:29:53 We have been brainstorming exactly about that. Maybe not in the way you described it, but we are thinking of ways of either doing that on the edge or doing that in the context of a Replit. With the limited amount of resources assigned to every container, can we also run a smaller model that gives you a basic code completion? These are all the actions we're exploring, especially because you might have seen that there are quite a few local plugins with our Replit Code v1 3B model where people are doing a local Copilot, basically. They take a VS Code extension. They take GGML to serve the model on their MacBooks, and they get pretty amazing speed of completion, on par with what we have on the website, and they're on their own version, their private version of Copilot. We know it's already happening with a powerful laptop. So imagine that we should be able to make that happen in the near future with even less powerful hardware.

Nathan Labenz: 1:30:56 On the near term roadmap for Replit, anything you want to tease that you think people will be excited about? You mentioned more agentic assistance, for example.

Michele Catasta: 1:31:05 Yes. Ghostwriter is going through quite a lot of work under the hood. We're very excited. We started a brand new internal effort to make it not only more powerful, but also way more integrated with the IDE. You will literally start to see it goes writer everywhere across Replit, helping you on several different tasks. That's the first teaser that I would love to give. And the second one, as I already teased before, we aren't done with the first LLM that we released. We plan to be on a semi-regular schedule of releasing new open source models, more powerful ones. We got so much benefit out of releasing that and so much community gathering around us and also giving us back feedback and instruct tuning datasets, for example. We have been seeing quite a bit of them being contributed that we want to keep working in that way for a while longer. So these are the two main teasers I want to share in this podcast.

Nathan Labenz: 1:32:16 Zooming out to this concept of live players that I've mentioned a couple times, Amjad, CEO of Replit, is notorious in my mind. I don't know how many people pay attention to this stuff like I do, probably not many. But he's tweeted a couple of times something along the lines of, Replit is the perfect substrate for AGI. I kind of have to squint at that to try to figure out what it might mean. If I think about what would a self-propagating AI system need, what kind of environment would a self-propagating AI system thrive in? I think about something like Replit as being a pretty natural growth medium, perhaps. Is that something that you've talked to him and said, Hey, maybe let's not tweet about that anymore? Or do you have a sense for, is there a vision there for what it would mean for Replit to actually be the substrate for AGI?

Michele Catasta: 1:33:12 I mean, Amjad and I definitely chat about this pretty much every time we meet and we have a chance. I think it's more about talking about the future of the company and where the field is going rather than about what we should be doing in the short term. That being said, I don't know exactly, first of all, what he had in mind when he tweets because Amjad is a very prolific Twitter user.

Nathan Labenz: 1:33:39 Yeah. You can't track every tweet. That's for sure.

Michele Catasta: 1:33:41 Yeah. I get these notifications, so I try to read everything, but I can't pick his brain. That being said, I agree with that specific tweet, and I will give you what is my interpretation. The reason why a lot of the community today believes that LLMs are not the only way towards AGI is that no matter how good is the world model that they create internally, it is still very limited. Sometimes it lacks common sense. It lacks an understanding. It doesn't do well spatial reasoning and many other shortcomings. So there are a lot of research labs and a lot of institutes working on that on a full time, and I'm sure we're going to be seeing a lot of progress. Now that being said, Replit can become a substrate because think about yourself as a developer. You don't write everything from scratch. You use resources that are available out there. You use cloud compute. You use services. You use APIs. You use tools. The way I see an AGI is something that behaves exactly the way we do as developers. It knows what to use. It knows what to orchestrate. That's what I explained in the ADI, in the Replit AI manifesto when I described the ADI as an AI that not only is capable of building things from the ground up, but is rather also very skilled at picking and choosing what to use. So Replit becomes a substrate even in the future when we have an AGI because regardless, it doesn't make sense to generate anything from scratch. It makes sense to know what to use. So that's the reason why I believe approaches like AutoGPT or BabyAGI have not materialized AGI yet. I don't want to comment on how powerful or limited they are. I think I'm always glad when people do experiments because we learn out of them. But there seems to be a clear bottleneck there because GPT-4 can't materialize intelligence out of nowhere. It can materialize tools and software and cloud compute. So Replit becomes that, I believe. Once you have an AGI, it understands, oh, I can invoke this API from OpenAI. I can call this tool that is being deployed on Replit. I can get CPU power there. I put everything together, and I accomplish a certain task. So that's the way I interpret the tweet, and I do believe that our journey is aligned on this. But next time we have a chat, I will let you know.

Nathan Labenz: 1:36:05 As you survey the scene, I think, obviously, you're somebody with Google and Replit on your resume now who has demonstrated good taste in AI companies. Who else do you see as super influential, possibly, and in a position to really shape the future? I'm kind of struck by the fact that I have this list of 10, 12 companies that I'm like, these companies seem to be the ones that actually might come up with something that really changes the world. And outside of that, I don't see too many. But is there any that you would add to my list? Anybody you think that I should be watching that I'm maybe not currently watching?

Michele Catasta: 1:36:51 I don't know if Perplexity AI is in that list. If it's not, I would recommend to add it. I love what they're doing a lot. They seem to be like an early stage Replit in terms of velocity of delivering new features. Apart from the fact that, of course, the team is amazing and they have a lot of experience, but they're attacking such an ambitious problem in a field that, of course, has a large player that has been dominating for 20 plus years. But I admire their boldness in going in that direction. I think they're executing very well, and I'm very curious to see where they're going to be going. So that's a company that I keep on my radar, and sometimes I chat with Aravind, their CEO, to get some inspiration. That's the first one that came to my mind. I mean, I talk with a lot of founders. I don't want to mention another 10 of them that I also like, but I feel some connection with the Perplexity story because I see them growing at the same speed that we have indeed, and that's inspiring. This has been amazing, Nathan. Great questions. Thanks for making this happen. I think we've been talking for 1 hour and 40 minutes. Didn't even realize. But yeah, it was a pleasure to be a guest in your podcast.

Nathan Labenz: 1:38:03 Well, thank you very much for making it happen. We appreciate all the time. I'm a huge fan. Michele Catasta, thank you for being part of the Cognitive Revolution.