Amjad Masad, Replit CEO: The Software Era is Dead | End of Limits Podcast by HUMAIN

Just because you're a Silicon Valley company that can make software is no longer an advantage. Competitive dynamics have never been that cutthroat. And for you to actually survive, you need to be very, very competitive. So I say don't learn to code. Just go build it. The market is going to get disrupted and then something else will emerge.

Thank you for having me.

I saw a tweet recently from one of our customers, one of our users, and he said that as a non-technical founder, before ChatGPT I had to call a friend to make software for me. After ChatGPT I could like copy-paste software from ChatGPT to Replit or somewhere else. And then Agent V1, I had a crappy intern. Agent V2, I had a good junior developer. Agent V3, I have a founding engineer.

I have a partner to work with to build my startup. And that's how a lot of our customers feel. And so, I think it'll just keep going. So, I think Agent 4 is perhaps—I don't want to overpromise—but maybe it'll start feeling more like a staff software engineer. Maybe Agent 5 is your CTO. And maybe Agent 6 will sort of feel like a team of developers.

So I think it'll just keep going. So in many ways you can say 2025 is the year of the agents because we made so much progress and it finally became possible. Like Agent V1 came out September 2024 or like in beta. And that was just a glimpse of a possibility. And now we have a real product that's adding a lot of value, a lot of people, a lot of enterprises. And it'll just keep going. So, in many ways, 2025 is the year. I think maybe it's not going to take a decade to fully mature, but I think it's definitely over the next 3 to 5 years, it's going to continue to get better.

No, I think I haven't seen a good non-software engineering agent yet. Have you?

Yeah. I don't think you talk to anyone outside of software engineering. That's the crazy thing because people do not know the craziness that's happening in software. Although anyone can use Replit now and so the more users we get, people are being surprised, like shocked, like 'Oh, I have this virtual employee in front of me.' But we haven't seen an accounting agent that's really good or a sales agent that's really good. One thing we did with Agent 3 is now Agent 3 can make other agents.

It's closed-ish. It's dependent on a training AI training method called reinforcement learning. The way reinforcement learning works is you give a reward as part of the training algorithm. So it's sort of like training a dog. You give it a reward whenever it does something.

So in software engineering you can give a reward on, for example, if you pass a unit test, if the application is running. And so there's easily definable rewards whereas a lot of other aspects in business are soft.

Exactly. So, it'll take a little longer to mature these areas.

That's all our IP. Replit uses a lot of the commercial foundation models. A lot of what we build is the infrastructure, so that's the development environment, package ecosystem, the storage, the database, the deployment, all of that, all the development infrastructure. But also all of the context engineering around the agent and the tools infrastructure and their liability and things like that. So that's where we spend, you know, 50% of our time is basically just like making it more reliable. So often times it can... because the way reinforcement learning works, it is motivated to try to finish the job at any cost. Sometimes it will like change the unit test to make them pass, you know, do certain scheming behavior. So for example, one thing we did is we have a different model judge its work. So as you watch your applet work, it's actually multiple agents. So you have an agent writing the code and then it sends the code to something called the architect. The architect reviews the code and sends it feedback. And once it's done with that loop, it goes to another loop which is the testing agent. So it sends it to a testing agent, boots up a browser, tests the app and then gets feedback to the coding agent. And by doing that you create this adversarial system between the agents and they get better because of that because every agent is trained differently. So there's a lot of techniques, you know, I can spend the next hour talking about those but this is what we do.

You know, I made an app two days ago. One of my favorite apps on my phone is getting sunsetted. It's called Pocket. It's a reader app, so you can put links into it and you can read at any time. And it was all about the company and they're shutting it down. And I was really disappointed because I really like it and I have links going back to 12 years ago that I saved in this app. And so they allowed me to export the data. I'm like, what am I going to do with the data? So I went to Replit and I said like, build me an alternative for this app. In 20 minutes I got a working version and spent another hour kind of like fixing it and buying the domain, linking it and kind of doing final touches. But within an hour or two, I had a replacement for that app. Quite simpler, but it was fully functional and I launched it on Twitter and now I have hundreds of users. And I didn't look at the code once and I didn't micromanage the agent at all. So I just gave it the high-level objective. Even before I launched, I was like a little nervous about the security. So I asked the agent, hey, do a thorough review of all the code and give me a list of security issues. And it gave me a list of high priority issues. Security is never final, right? You can always work on security. High priority issues, lower priority. And I was like, okay, this doesn't matter for us, whatever, you know, this one thing, just go fix that. And it went and fixed that. I didn't even feel like I needed to look at the code. And so I think we're at a point where we can, in software at least in Replit, you can delegate complete tasks to agents. I think when it comes to software it's a little easier because you can use the app yourself, do your own quality assurance. When it comes to information with hallucinations, like you can be totally misled and you can believe something that's totally wrong and things like that. So I feel like hallucinations are not as big of an issue for software engineering agents right now as much as it is for others. The way I see collaborations between AI and humans, I think there's like three different modalities and all three are useful. There's the co-pilot modality which is basically, you know, Microsoft Copilot or GitHub Copilot, you know, it's autocompleting your work, it is providing some assistance. There's the agent model where you can just delegate entire work to it. And then there's the standard chat consumer chatbot. I think these are really the three modalities for the foreseeable future. And all of them will be useful but different jobs will... you know, we don't have a consumer agent yet that's really good for example but we have a consumer chatbot.

That's right, yeah. A lot of people when they talk about agents they're like, 'Oh, agents, they can book your tickets.' And why hasn't it happened yet? Because well, it needs to be really good at like Booking.com and also whatever other sites that people use and it needs to handle all these different complexities and those sites don't like agents to be using them and there's an adversarial aspect of it.

Well, depending on the incentives, right? Like the question is for those websites... I heard Brian Chesky get asked on an interview, Brian Chesky, CEO of Airbnb, and they asked him why aren't you on the ChatGPT new app experience and he basically said they don't see the incentive for them to be there because ChatGPT is like taking over the experience and they want to own the consumer experience. And so the thing that made Google really powerful is that they aligned the incentives between publishers and consumers and advertisers. It was a really aligned system. Right now, ChatGPT actually is sort of a misaligned system. Publishers hate it because they feel it's stealing content and they're suing them all the time.

It's hard.

Trillion, kajillion. Yeah. I mean I don't think it's an existential question because all these companies are locked into a race with each other. So they're all going to build the chatbots and the publishers might get screwed and maybe some new model will have to emerge. So it's disruption technology. So the market is going to get disrupted and then something else will emerge. But right now I don't think anyone has a good idea for what the business model for AI chatbots and agents could be. A lot of people are starting to talk about e-commerce agents, like agents that could buy things on your behalf. They could be very aligned because Amazon doesn't care if a human is buying or an agent is buying.

But people like shopping. I don't think my wife would want to outsource shopping. That's like the one thing you don't want to outsource.

But maybe the nature of software engineering itself changes, right? So, I mean compare the job 10 years ago to how it is now.

Yeah. So even going back, let's say I'm going to go back 30 years. We were in like low-level languages, right? People writing C and C++ and you had to be very meticulous, very detailed in everything that you're writing. And then we had the emergence of dynamic languages like Python and JavaScript and things like that. And then we had the emergence of code repositories like GitHub and open source packages and every one of those movements created a lot more developer efficiency and made it a lot easier to make software and a lot more developers came online because it's easier. And I think we're in the latest instantiation of it. I mean it does feel like it's different but in many ways it's not. It's like the same thing. It's just become this higher level abstraction. And now I think we're... and AI went through a rapid cycle where we had the co-pilot era and now we're in the AI agent era. And I think software engineering is becoming increasingly about managing agents. I think there are aspects of very specialized software engineering that is not like that yet but if you're building just web applications and mobile applications, your job is basically managing AI agents.

The value of software becomes the value of the ideas, the value of the distribution, the value of the marketing. The value stops being around like how fast you can make your website or how nice you can make your website. It becomes more like how solid is your business idea. I think still like how hard you work and the culture around your company and all these social behavioral things and companies will maintain even if you're using AI agents. You still want to have a culture that produces the best ideas, that attracts the best talent. So, a lot of those things, those really human things, I don't think will change. But you can no longer just because you're a Silicon Valley company that can make software is no longer an advantage because anyone else can.

Yep. We call it in Silicon Valley, we call it the ZIRP era, the zero interest rate era. And so during 2013 to 2023, you had an era of zero interest rate. And that made it so that money is free and companies can spend as much money as they want. They can spend it on all the employee perks. They can hire infinitely. And suddenly we got into a period of austerity. So that's one big factor, the economy. The other factor is, you know, for a while the winners were established during the SaaS, the cloud and mobile era, the winners are established and people thought that crypto might be the disruptive technology, it wasn't. But now we have a disruptive technology and even a company like Google, it's existential for them. And so the entire culture around Silicon Valley work started to change because we're all locked into these races. Like everyone's competing with everyone. Everyone's trying to do everything. Like every company, especially all the big companies are trying to compete with us, right? And even smaller companies move to different spaces and different markets very, very quickly. The competitive dynamics have never been that cutthroat. And for you to actually survive, you need to be very, very competitive. So people are working harder, they're working longer. The pay is still very high and higher than ever actually, especially in AI companies. So that didn't change, that continued to grow but people are expected to deliver.

It's a very important problem. Yeah, I don't think that many people are thinking about it and I don't have a macro solution to it, but I can say as advice to young people, by the time you're graduating college and going into the job market, you need not to be a junior software engineer. In other words, you need to be so good at managing these AI agents. You need to have gone through a lot of AI, a lot of internships. You need to be entrepreneurial, maybe try to start a startup or nonprofit or something like that. So there are ways for young people to differentiate themselves on the job market by being early adopters of these tools and really understand. So they have an advantage in many ways in that they don't have priors. Like I know a lot of software engineers who still struggle with adopting AI. A lot of CEOs that I talk to at big companies, there's antibodies internally, like they don't want to adopt AI, they're just used to the way they're doing things, they don't like change. So there is a disadvantage, yes, but there's an advantage for people who seek it. It is there.

So, I have a nuanced take on coding. When an entrepreneur comes to me and says, 'Hey, I have an idea. I want to build, like where should I go to learn to code?' I say, 'Don't learn to code. Just go build it. Just go build it.' Like, you can build it and maybe learn some things along the way you need to. But if you're really into computer science and you want to be a specialist and you want to be like a low-level engineer or you're going to work at NASA or you're going to work on some really hard software things or you're going to work on AI, yes, I think you should still do CS and learn to code. In terms of soft skills, I think those are going to become higher in demand. The thing about communicating with an AI agent, it is very much like communicating with a human. It is easier for say product managers to write product descriptions for an AI agent than it is for someone who is otherwise a let's call it somewhat left-brained, you know, quoter personality. They're typically not very good at writing natural language. So they need to learn that now. So, whereas previously you could get away by being a brilliant programmer but not knowing how to talk to people, you can't anymore because you're computer people now, which is such a weird thing. And so, it is giving an advantage to people who have soft skills and we're seeing it in our customer base, product managers. I mean, here even at Humane, I was just outside the moment I got out of the car, someone on your marketing team was like, 'I'm spending way too much money on Replit.' That's a marketer, right? It's not a software engineer. And so it's a very weird world where computers are suddenly becoming more human and those people who are able to communicate best with humans can communicate with agents.

So there are few answers to this. I think on an education system level, I think every kid should get an AI tutor and it should be a really good AI tutor that understands the curriculum, that allows them to look things up really quickly and I think it should be better than ChatGPT. Like the Saudi curriculum is very different than the American curriculum. Different values, different things like that. There should be an AI tutor that has all that knowledge and context. In Jordan, they just rolled out an AI tutor called Siraj. They were able to build it on Replit in record time. So the crown prince put together this Jordan Technology Council and they're tasked with modernizing the country's infrastructure with AI and things like that. So they had this idea of an AI tutor and typically it would take years in the bureaucracy but they were able to build it and pilot it very quickly within a month and then they went to the Ministry of Education with hard data like this is benefiting students and now they're rolling it out to 1.6 million students. So it's amazing how much even the software engineering agents can benefit governments being able to move really fast at modernizing education and that's the kind of mindset you want to see, like even ministries of education need to be nimble now, need to be open to disruption. And we can only learn what works by trying a lot of different things. Like I don't think you can get researchers to just spend the next 10 years studying how does ChatGPT affect my students? Well, by that time, your students are going to be totally different because ChatGPT is already affecting... ChatGPT just disrupted homework entirely. Like homework just doesn't work anymore. So, you need to come up with alternatives. And so I don't claim to have all the answers, but I feel like personal tutors are very important. I would like to see a lot more entrepreneurs building personal tutors. Like I'm looking for my son. There's one company called Synthesis that has like a really cool AI tutor.

You know, Musk was... so Synthesis was a spin-off from an Ad Astra school.