theCUBE + NYSE Wired: Physical AI & Robotics Leaders | Howie Xu, Gen & Michele Catasta, Replit

Clips
More from theCUBE + NYSE Wired: Physical AI & Robotics Leaders

Howie Xu

Chief AI & Innovation Officer

Gen

Michele Catasta

President

Replit

play_circle_outline Empowering Non-Developers: Michele Catasta Leads Replit’s AI Revolution with Innovative Software Creation Tools

play_circle_outline Replit's Evolution: From Developer Platform to AI Agent Pioneer and the Role of Vision in AI Development

play_circle_outline Rapid advancements in AI model capabilities since 2022 influence software creation timelines.

play_circle_outline Emergence of unbundling SaaS trend using Replit for custom internal tools.

play_circle_outline AI’s role in cost savings and accelerated prototyping for businesses.

play_circle_outline Evolution of Replit tools from traditional coding environments to AI-driven interfaces.

play_circle_outline Vision of Replit as a potential next-generation tool for knowledge workers, akin to Excel.

Info
Transcript

Howie Xu, Gen & Michele Catasta, Replit

Howie Xu

Chief AI & Innovation Officer Gen

Michele Catasta

President Replit

Michele Catasta, president at Replit Inc., joins Howie Xu, chief AI and innovation officer at Gen, during theCUBE + NYSE Wired: Robotics & AI Infrastructure Leaders 2025 event to explore how AI is transforming software development. The conversation highlights Replit’s vision for enabling technical and non-technical users to build and iterate with unprecedented speed.

Catasta shares the story behind the Replit agent, a tool designed to democratize app creation by leveraging intelligent code generation and real-time testing. With a focus on model fores... Read more

explore Keep Exploring

What is the background and recent developments related to the AI team at Replit? add

What was the founding vision of the company Replit regarding AI and software development? add

What factors influenced the perception of the timeline for advancements in AI and the development of agents? add

What is the trend regarding the replacement of SaaS software with custom-built solutions using tools like Replit? add

What are the benefits of empowering non-technical personnel in a company to build prototypes? add

What was the evolution of Replit and the features it has developed over time? add

What is your vision for the future of spreadsheet applications like Excel? add

bolt Powered by CUBE AI

Howie Xu, Gen & Michele Catasta, Replit

search

>> Hello, everyone. This is Howie Xu, chief AI and the innovation officer at Gen. Among other things I do at Gen, we developed the AI native browser, Norton Neo. Check it out. But today I'm welcoming my friend, Michele Catasta. He actually is the president of a company called the Replit, 45 million people using Replit. So tell us a little bit about yourself first.

Michele Catasta

>> Thanks for inviting me, first of all. Michele, I'm president of Replit. I am in charge of everything AI. I started at Replit around 2022 initially as an advisor and as a consultant. Then I built the AI team zero-to-one, helped to create a lot of features on the product. And in September 2024, we launched Replit agent, which is a way for everyone, especially non-technical people, to create software from scratch with the help of AI. And the last nine months have been an incredible ride for us. We've been growing very much both in terms of revenue usage and we're learning a lot from our users.

>> Yeah. In the last three years, AI is the talk of the town in Silicon Valley and the entire world, but in the last nine months the AI agent is the talk of the town. So tell us a little bit about the agent you are working on, meaning that you joined two years ago as a consultant. At that time, Replit is still a developer platform. Before you joined, it's not like AI agent company, right?

Michele Catasta

>> Correct.

>> So the company is transforming itself in AI agent company. Can you give us a little bit more details? What happened?

Michele Catasta

>> Yeah. I think the vision of being able to create software completely aided by AI with an agent was already there since the inception of the company. Replit was founded in 2017. If you go back and check the decks that Amjad, the CEO of Replit used to raise money, you read a couple of slides where he made it clear that one day software would've not been typed manually by users, but rather there would've been an AI to create it on their behalf. It took a while, so it had a lot of vision already back in the days. And in parallel myself, I was starting to do research on those topics in 2015. So I think both of us had asked ...

>> 2015 you were with Stanford?

Michele Catasta

>> Yeah, I was about to join Stanford and then I spent a few years there, spent a few years at Google training LLMs, always with the vision that eventually that would've been the key building block to then create agents that would've allowed us to give software creation capabilities to everyone in the world.

>> So what did you see as a tipping point of that? Okay, the company or even the world projected that at some point a software will be done by AI, but now this is the right time. Two years ago, what did you see, the model or what?

Michele Catasta

>> Definitely the model. The fact that I was working really closely with them, like training models and using them on a daily basis and trying to push them to the limits allow me to see not only what we're capable to do at that point in time, but also the trajectory and how fast they were improving. So I reached a stage where I felt if I extrapolate from today, is not going to take long until we can create agents. In all honesty, I've never been great at doing time predictions and I would say almost no one in AI is great at that. We keep compressing them because more incredible people are working on these topics. More capital has been invested on it as we speak, so I think our timeline is moving faster for everyone. But it felt around 2022, I left Google basically a couple of weeks before ChatGPT was released. So we knew that there would've been some level of product impact in the world. I don't think anyone expected it to explode that fast. And I gave myself a timeline, say, three to five years from when I joined Replit to make Replit agent happen. It took 15 months for the very first release. And to an extent of course thanks to the fact that frontier models kept improving at a much faster pace than we expected. And at the same time as a company, we made a big bet on the fact that it was about to happen. So we started to build agent and to build infrastructure necessary to do that way before it was working as well as it does today. I think that's, I would say, the key recommendation I will give to any startup today. Always built for the future capabilities of models rather than what you can use today. Otherwise, it will be way too late.

>> Yeah. I mean, apparently when you and I knew each other, you were still the vice president and now you're the president. Clearly the company has a lot more conviction in AI. AI is the center of everything for Replit. So can you share with me between September last year when you released the first version of the agent to now, what is the surprise you found during the nine months?

Michele Catasta

>> Where do I start from? I think it's a long list of surprises. The first one, I would say the variety of what our users build with agent, ranging all the way from small toy projects, personalized software, all the way to internal tools that today are running inside companies. They're deployed on Replit, they're driving tens of millions of dollars of revenue, so the range is ...

>> Can you share, even without naming the exact name, what's the nature of the software that's actually producing or impacting tens of million dollars in the company?

Michele Catasta

>> On one end, a lot of SaaS software is being replaced by tools be with Replit. So there is this phenomenon called unbundling SaaS. I think even Paul Graham talks about the idea that we tend to go out on the market by maybe seven figures SaaS that we don't leverage very much in companies. You can rebuild as much subset of it with Replit for a fraction on the price and tailor it exactly to the needs of your company. This we see happening more and more often the more the agent becomes

>> So the motion is instead of me going to buy some SaaS company service, now I just build it myself and then run it for myself. I don't have to sell other people, but I actually save the SaaS cost.

Michele Catasta

>> Exactly. So this leads to cost savings, like substantial ones. And at the same time, the fact that we are empowering everyone was also non-technical in the company. Those tend to be the people that know exactly what should be built. There is a reason why we have product teams in our company. Those are the people that drive the product roadmap. The fact that they don't have any more to rely on engineering resources and they have to figure out when they're going to be capable of getting those resources. They can rather go sit down on their computer, open Replit, build a prototype zero-to-one just with the ideas that have in mind. That means that as we speak today, companies are prototyping and testing way more ideas in a much shorter window of time. And some of those ideas are winning ideas. And so they turn out to be something that goes, maybe is deployed initially as a prototype internally, and then they decide to put it in production.

>> So one is cost reduction, but now you are talking about speed to revenue?

Michele Catasta

>> Yeah, generating revenue. Speed to revenue also ...

>> What kind of companies are doing that?

Michele Catasta

>> Our best target today are with SMB and mean markets. We are in talk with a lot of enterprises, like in Fortune 500. As you know, enterprises also have a long list of enterprise requirements that we're building towards that, making sure that all the data governance aspect and security aspects are going to be fulfilled. I think some of them already happy where we stand today. Some others are looking at our roadmap and they're going to be adopting Replit very soon.

>> So maybe this is a good time for you to give a high level description about Replit, what Replit does from a clouded development environment to the AI agent, but what exactly can people do? Can you give our audience just a description of that?

Michele Catasta

>> Yeah, I think I'm going to describe the evolution of Replit. We started as a development environment in the cloud, making sure that we were saving time to the users that beforehand had to set up the development environment in local. Now they funded on Replit. We were taking care of installing dependencies, we were taking care of supporting pretty much any programming language out there. And over time we had all the most used developer tools that you would've in your tool chain, debuggers, LSPs, search functions, et cetera, et cetera. We kept doing this until basically 2023. At that point, we also started to add some AI features, say auto completion, integrated chat bots, and so forth. What we built in the end were a set of tools that today agent is harnessing. So we had all the tools that you will have to give to our coding agent today built in the past years. So the moment we decided to double down on this bet and build the agent, which just had to build the agentic workflows, the product surface, but the tools were already available for us. And so you have seen this shift where the development tools were built for humans until not long ago and now we're building them predominantly for agent. To the extent that the user interface of Replit has changed radically. The first thing you see on our website is a prompt box. We're asking you to prompt what the agent should be building. We don't put in front of a code editor.

>> So English is the programming language?

Michele Catasta

>> English is the programming language for Replit. Correct. And when we launch agent, if I recall correctly, three users out of four were generating code only with AI. They were not editing the code at all. At this point is almost nine out of 10. So the vast majority of our users, they've never opened the editor. We even hit the file tree so that if they really want to go and look at the code, of course it's there. Whatever the agent generates is owned by the users. So if they wanted, they could do whatever they wanted with it, but the truth is it's not easy anymore for a lot of apps.

>> When did that start happening from the September release you had the last year?

Michele Catasta

>> It's been gradual.

>> Gradual.

Michele Catasta

>> We literally ship a new version of the agent every single day. So every day there are some improvements, every day it gets a bit better until we earn the confidence that we reach a stage where you don't have to change the code that often any longer and then we started by adding the file tree. And now for example, rather than showing you the diff that the agent generates, we just collapse it by default.

>> How do you differentiate yourself against some of the vertical players like Lover Bowl or Hey Boss.xyz, They're all doing some kind of the more vertical websites kind of thing. You can do things very horizontal play, but what's your differentiation? How do you think about it?

Michele Catasta

>> So on one hand I would say is the technical depth. We allow you to build not only prototypes with a beautiful design, but rather Replit is a tool that you can use for the entire software development lifecycle. All the way to taking an application, bringing it in production, and even deploying it on Replit directly. So there's this aspect of the technical depth and at the same time we're also focusing a lot on B2B and enterprise. We are adding more features as we speak to make sure that we can cater to the needs of companies and that's where we think that our product is going to keep growing even further. And to be honest, it's also the place where I get more excited to receive user feedback. The users tend to be very advanced. They have very specific needs, so learning today what works and what doesn't work on Replit agent is extremely valuable because in a sense it's determining our roadmap for the next few months.

>> Right. So last time we talked, you said that the Replit is tailored for a knowledge worker and the knowledge workers cannot program. So can I assume that knowledge worker in your definition is consumers or presumers, not like the programmers, professional programmers? It's two different worlds in your opinion?

Michele Catasta

>> Yeah. What I love to say is there is a category of products out there that they're all fighting for the same, say, 30 to 50 million people in the world, depending on how you define a software developer,

>> Yeah, the Cursor, Augment, Factory. We all interviewed in the recent days.

Michele Catasta

>> They're all building amazing products. Like we use Cursor Rapid, for example.

>> We'll get to that in a bit.

Michele Catasta

>> Yeah, they're very important and very necessary, but they're built with a different ICP in mind.

>> People who actually have taken computer science one-on-one at least.

Michele Catasta

>> Correct. Yeah, yeah. They need to understand the very least the basics. For us, the north star is to serve knowledge workers, which is roughly 1.5 billion people in the world. So a much larger market with very different needs. And it's also forcing us to think about our product in a very different way because their level of technical skills tend to be lower. I'm not talking about their raw skills. We have met a lot of people that are extremely proficient in using agent. They just don't want to learn how to code or they don't have the time to do that.

>> They don't want to have a systematic way of thinking, right?

Michele Catasta

>> Correct. If anything, it makes building Replit agent even more challenging, to be honest, but I think it is a challenge that we are excited to work on because it's a much wider market and there are no tools for those people. There is a richness of tools for the developers in the world and there is practically nothing for every other knowledge worker out there except maybe Excel. Excel is the standard de facto for every knowledge worker in the world. I do believe there is a near-term future where Excel can be complemented with something much more powerful and this is what we're building.

>> It sounds like your vision is to be the next generation Excel.

Michele Catasta

>> Yeah, that's my personal vision. I pitch it internally at the company all the time. I think we can definitely become that. But for Excel to go from a few users to maybe 1.5 billion today, from what I know, it took four decades. So we're going to take ours with time, too.

>> But everything is compressed in the AI age.

Michele Catasta

>> Everything is compressed. Yeah, if we get it in five years or 10 years, I will be more than happy.

>> Right. And then what's the future of Excel at that point?

Michele Catasta

>> That's a great question. Not something for me to answer. I see a lot of our users uploading spreadsheets on Replit and then doing advanced data analysis with it. So I think part of the complexity that today lives in Excel will be moved towards agents, including Replit agent. At the same time. I think the Excel interface ...

>> Cursor Augment generators.

Michele Catasta

>> Yeah, coding agents in general. Yes, yes. But I think the some UIs are here to stay, like document processing, spreadsheets. They're so familiar with the knowledge workers that are not going to go away anytime soon. We have to teach knowledge workers on what is the UI of the future, which in my opinion is also not going to be a chatbot.

>> Okay, so vision is great. Let's talk about realities. What's the reality? Meaning that if I want to do some complex, whether small-scale SaaS or not, how good is the Replit AI agent is for that? Because one of the criticisms of the AI coding agent is if it's buggy, how do you know? Is that really toyware or even if you intend this to be a series software, how do you know this is serious software?

Michele Catasta

>> First of all, I don't think it's toyware. Not because I'm personally invested in it, but rather because you see what our users are capable of building.

>> But this is a wide spectrum, right? Toyware is here, mission critical is here. Where are we?

Michele Catasta

>> Definitely not on the mission critical side if you want to build everything exclusively with an agent without having a system understanding. I do know that if you use an agent in a large code base to help you to create PRs, you can definitely build on a mission critical system. I consider Rapid to be a very large scale software that has to be extremely reliable. We use agents, we use Cursor. So there are definitely ways to build mission critical software helped by AI. With Replit, the challenge is even harder because we do the zero-to-one and we allow you to keep building on top of it, again, without having any coding background. So in that case, I would say it is not as powerful yet to create something that is going to scale to millions of users without any help, but it's powerful enough for you to create tools that are going to be used by thousands of people in your company with high level of reliability.

>> But this is even without millions of user or even with 200 users, you want the software to be reasonably stable, reliable, compliant. How do you make sure that the Replit generated software meets that sort of the goal?

Michele Catasta

>> So you have to follow a different mindset compared to how we've been building software in the last decades. What we did since the very first release of agent, we had the support for rollbacks. So every single time you make progress on your project, we keep the entire timeline, the entire history of all the work that you have done. And here the approach is more you explore how much more you can do and, in case you go down a path that breaks your application, you can always go back. We recently made a feature even more powerful. You can do instant previews of every point in time of this review application. So you can run it, you can see if it's more stable compared to the one that we were using. So there is some level of user testing rather than software testing that helps you to make progress and make sure that they're running a stable version of the

>> So this makes sense, but how do I know this snapshot of the code is good enough?

Michele Catasta

>> You either go back and you basically keep in mind which version was working correctly. What is coming really soon, this is going to be our next release of Replit agent, we're bringing in testing in the loop. Both at the software level, the whole school way in which we test software, but also the level of user interactions of an application. So we have a lot of internal progress that I'm not going to disclose everything today because I don't want ...

>> But conceptually, right? Yeah, we still wanted to understand. As a software engineer, I get it. I use Augment, the Cursor and then I think about testing strategy, but since I took CS 101 or 401, I know how to do it. But with knowledge worker, I mean, what is corner case? What is the common path? They don't even speak that kind of the jargon. So you have a Replit generated software. How do I know, hey, I can give to John in a different department to start use it or no, no, no, I need to do more testing. How am I going to, like Replit is going to help me, or how to think about it at a conceptual level.

Michele Catasta

>> Yeah, the next version of the agent is going to help you exactly with this problem. So it will automate testing. Again, not only at the level of just running unit tests, but rather at the level of ...

>> System level.

Michele Catasta

>> System level, like exercising specific code paths, like going through an endpoint to add the user and maybe added some data.

>> But me as a user need to tell the system?

Michele Catasta

>> No, no, the agent already internally where a version it comes up with a rubric of all the testing that should be done.

>> Got it.

Michele Catasta

>> And then executes it and it also can execute that in isolation so that you have your dev environment where you can run all these testing and make progress. And then at the same time, we're going to provide a production environment that is going to be stable with a separate database. So all these features are being worked as we speak and this is part of something that we launched already back in May and that is going to be our focus for several months. We call it safe by coding. It's very important the progress we made in the last months as an industry to make coding more powerful, but it really is our user asking at the same time for web coding to be safe. But they want exactly the guarantees you're talking about. The fact that they make progress, but they don't break, everything else that is there. Now, I'm not going to guarantee you that it's going to be 100% perfect. There is always some level of ...

>> It's a journey.

Michele Catasta

>> It's a journey, but it's a journey that we already started that with the same extrapolation of the pace of progress that we were talking about before. I don't think this journey is going to take us a decade. I think we're going to see already something very powerful by the end of the year and we're going to keep working on this for the short term future.

>> Okay, so a lot has been discussed about Replit system. It sounds like a lot more to look forward to because from my point of view quality control is really the key. Without quality control, of course it's a spectrum, but it's going to be towards the toyware side of the spectrum than the other side. Right? Sounds like with quality control you move to the other side of the spectrum. Very much looking forward to it, but let's actually talk about how you cook this Replit, right? You mentioned that you actually also use Cursor, Cloud. Can you give a little bit more details? Maybe one of the questions I have is what percentage of the code is written by bots versus human these days for the new code?

Michele Catasta

>> I don't know the exact number. I don't think we keep track of that. I know that the large companies are actually declaring around 30%. I think it's much more, in our case. Maybe it's not code that is being generated directly, say, by Courser or Devin or other tools that we have internally, but definitely a lot of the code might be even coming from Cloud AI or ChatGPT. So I see pretty much every single developer in Replit is very much in tune with AI. There is a

>> So you would've seen majority of the code generated today, this week is actually more from the bots than from humans at this point?

Michele Catasta

>> I'm fairly confident of that. If you take a look around our office and you see the screens of our developers, there are at least like a couple of AI tools open all the time. And they all start from a template, a scaffolding coming from one of the tools. And then of course they do some manual interventions on that, but the starting point happens to be always AI.

>> So when was the tipping point at Replit? I'm curious because I'm pretty sure when you joined it two and a half years ago, that's far away from trees.

Michele Catasta

>> Yeah, yeah.

>> So even though there was Microsoft Copilot of the world at the time, but when did that tipping point happen?

Michele Catasta

>> I would say when chatbots online started to become good enough, so I'm talking about maybe early, mid-2024, like ChatGPT and also the Anthropic models, especially when Sonnet 3.5 was released, they were proficient enough with code that I would say every single engineer at Replit realized how powerful they were. And so it has been maybe one year since the adoption has been skyrocketing internally. Today, I would say every single engineer is spending much of their time, again, shuffling between GitHub and any of the AI tools that they prefer. To the extent that I'm happy to give access to pretty much every single tool that developers ask for. I don't have any strong preference on what should be used. If anything, I'm also supportive of them testing everything that is out there because I believe every tool has good ideas.

>> So you are saying that they're using variety of AI aging tools, not just the one vendor?

Michele Catasta

>> Yes. When we onboard a new engineer, we give them Cursor and we give them Cloud. But then if they request for anything else available out there, I'm 100% supportive for them to test. Even if we found out that it doesn't work well in our environment, usually we still have something to learn by using them.

>> So would you say your software engineer at this point is supervising the bots more than writing code at this point already?

Michele Catasta

>> I would say so. Maybe some of them wouldn't love to hear this statement and especially those that have more seniority because they ...

>> Okay. My next question is how did you convert people who actually were resisting a few months ago, a year ago? What did you do and then what happened?

Michele Catasta

>> It didn't require any external force, I would say. In general, especially if you're a very senior developer, you tend to be so opinionated that, regardless of what you hear from the outside, you're not going to change your mind. So what happened in our case is by them seeing how faster the other developers were becoming, especially the younger ones, they realized that there was definitely some alpha in using AI in the loop for the development work. And that's how they gradually started everyone to embrace it.

>> So at this point, what is the productivity gain? Do you have over, let's say, a year or a year and a half ago? What are we talking about?

Michele Catasta

>> I would say this is just a rough estimate. We don't keep track as carefully as other larger companies do, but I would say at least it's a 2X compared to when we were not using AI very much. And I'm not exclusively talking about writing lines of code. If anything, we all know that the more complex is the project they're working on, the less time you spend actually writing code and you rather spend it doing meetings, writing design docs and debating how something should be built.

>> So when you say 2X, what does that mean? We all know software engineers spend only a fraction of the time writing code, that the rest of the time having meetings or whatnot. So 2X, are you talking about 2X productivity when it comes to writing code or generating code? Are you talking about 2X for a software engineer overall?

Michele Catasta

>> I'm talking about shortening timelines, like shortening timelines for deadlines.

>> You're talking about throughput?

Michele Catasta

>> Throughput, yes.

>> Meaning that for sprint you used to be able to do this much. Now you're able to get 2X out of the same number of people.

Michele Catasta

>> Correct. I think I recall when I joined Replit, we were sometimes thinking in terms of quarters, which is very common in a company. Right now, usually a six month sprint is considered a long sprint at Replit.

>> Six weeks?

Michele Catasta

>> Six weeks, sorry. Yeah, six weeks. Sprint is considered relatively long one. We have a lot of two, three week sprints that are happening as well.

>> Right. And then before you had to do six weeks sprint?

Michele Catasta

>> Six weeks.

>> But now because of AI, you are thinking about two or three weeks sprints.

Michele Catasta

>> Correct.

>> And the overall throughput is 2X. Do you think it would be 5X sometime soon or what's your prediction?

Michele Catasta

>> I think so, 100%. It's going to happen. It's going to happen soon and the key ...

>> What timeframe?

Michele Catasta

>> Maybe end of the year, early next year.

>> Maybe end year of the year.

Michele Catasta

>> Yes. I think one or two model generations from today, so the next version.

>> Well, allegedly GPT-5 is going to be this summer.

Michele Catasta

>> Correct? Yeah. So imagine GPT-5 being much more powerful and at the same time being capable of running agency in parallel. So what you're seeing being done today by Codex, or you can do that with code, like running multiple agents in parallel, that is going to allow developers not only to supervise a single AI, but rather having multiple AIs creating the same PR and then cherry picking the best one or maybe allowing multiple ...

>> How far away is 10X throughput improvement?

Michele Catasta

>> I would say sometime next year if we see people ...

>> You don't think it's even two or three years? You are seeing one year?

Michele Catasta

>> Yes.

>> So basically one year from now the throughput will be 10X of what you had a year ago, roughly speaking?

Michele Catasta

>> Especially for companies that are doing the zero-to-one are relatively early. I wouldn't expect that to happen, say, at Google or Microsoft with giant code bases and monorepos, but I would say if you have a small team of two, three very strong engineers that are leveraging AI correctly, we're going to see them moving much faster.

>> So we have ...

Michele Catasta

>> We see already today a lot of startups that are doing these crazy explosions in terms revenue. That happens not only because AI is allowing people to create more powerful products, but also because from zero to actually creating revenue, that timeline has shrunk.

>> Yes. And then speaking of that, there's a zero sort of one person unicorn thesis from Sam Altman. Do you think one person unicorn will be powered by Replit's and what timeframe?

Michele Catasta

>> I believe so. I don't expect that maybe to happen next year. I think Dario Amodei claimed that it's going to be happening in 2026. I can be that optimistic because I think there are other factors at play to allow you to create a unicorn. It's not purely how fast you are creating software. There is marketing, there is sales, a lot of ...

>> Yeah, exactly.

Michele Catasta

>> Yeah, but I think it's something we're going to be seeing really soon.

>> Yeah, you're pretty bullish about that. So I was going to also follow up on this, right? Within two years, a year ago to a year from now you are saying that engineering or product throughput will increase 10X. That's a pretty big improvement. But also, you are not just VP engineer or VP of AI, you're the president of the company. So you think about business holistically. So what do you think a 10X improvement for engineering, for products, the production part, has impact on the overall business, right? There's a distribution, there's a monetization, there's a marketing. So me as a, let's say, I'm a CEO of a company. If my VP engineer tells me I'm getting 10X improvement for the productivity, what does that mean to my company? What can I do? What should I do?

Michele Catasta

>> First of all, making sure that everyone internally is embracing AI from today. If they haven't started, it's very important that they catch up as soon as they can and they stay up to date with the latest. And they try all the tools that can make them more productive. Even if they don't appear to boost your productivity very much today, it's going to be happening in a few months.

>> Just get hands dirty.

Michele Catasta

>> Yeah, get your hands dirty and start now rather than waiting for when things are perfect. It probably will never appear to be perfect, but they're so damn powerful that it doesn't make sense to wait. At the same time, if I was like a VP of engineering, I will be putting even more effort in hiring the best possible small team that I can. Talent density will matter even more in the future than it does today. If the goal is to optimize the company and making sure that the head count doesn't grow unbounded, so you want to have the right team at the right size that allows you to have quick communication, that everyone is extremely capable. And you want to attract them early in the company because they want to have skin in the game and they want to be there for the long term.

>> So my specific question is when you have the production side 10X, the distribution side needs to go 10X in order to accommodate it. You must have some lesson learned for that. Is there any lesson you can share that to make sure that entire company is growing 10X in terms of the productivity efficiency?

Michele Catasta

>> I think it's very important to communicate within the company where the puck is added. There are teams like sales or marketing that might not be in tune as much as we are on how fast AI is making progress. And they need to know today what we expect to happen in one year. So if you expect your product to be amazing in a couple of months and get incredible adoption within the enterprise, you can't afford to have a team, a sales team of one person today because then it's going to be very hard to scale it fast enough and teach all your account executives basically in a few weeks. So everyone has to start to think about where the company is going to be in a few months. And that level of communication is very important to have before you actually see something working perfectly.

>> So last question. We talked about Replit being a very good tool for knowledge worker, people who have never taken computer science and then be able to produce SaaS in a way. So that's very good. And we also talk about for the executives in the company, how do they think about it? What about software engineers? You came from Google, Stanford. What do you tell people who are graduating from Stanford now who are sort of Google engineers? There are tens of thousands engineers there or even smaller companies. What do you tell software engineers?

Michele Catasta

>> It's as important for them to learn the basics than it is to also learn how to use agents. And on one end, knowing the basics allow you to be not only more productive, but also capable of doing the last mile. Where agent fails today, that's where the human

>> Do you think they should regret taking computer science?

Michele Catasta

>> No. Even myself today, I don't regret it. I do think that the future is changing and it's going to change fast. It doesn't mean that we can't adapt ourselves. And if anything, I see younger people to be much better at embracing those tools.

>> 10X engineer?

Michele Catasta

>> They're native. I wouldn't call myself an AI native. I've done research on that. I learned it for my professional career, but I'm not going to be a person who was basically a teenager when ChatGPT came out.

>> So for those engineers, do more, be more AI, adapt to the AI world. Anything else you would've recommend? For instance, broaden their skill set to do more than just computer science or how do you think about it?

Michele Catasta

>> It really depends if they themselves as future entrepreneurs or if they want to be just exceptional on the technical space. If the former, I don't think there is a better time to become an entrepreneur. Even myself, I'm broadening my skill set mostly by means of AI. I'm not formally trained to do the job that I'm doing today, but I always feel that I have not only a lot of people in my network that can help me and can take a phone call from me, but I can also call ChatGPT and ask how I'll be doing something. So there is that aspect for my job today. And at the same time, if they want to be exceptionally good on the technical side, even here, being in front of an agent really forces you to speed run your learning process, put in front of so much complexity way faster than you were if you were typing the code yourself, if you think about it.

>> Sounds like this is the best time for software engineer if you want it to be 10X, 100X. And then also from a Replit point of view, this is the best time to be a knowledge worker, right? You are not just having the spreadsheet or the Excel. You have Replit, right? Of course I think this is the best time to be entrepreneur, to be the CEO now because you get 10X out of just using the tool.

Michele Catasta

>> Yeah.

>> Thank you very much, Michele. This is a wonderful conversation. Thank you.

Michele Catasta

>> Thank you, Howie. It's been a pleasure. Thank you.