I love it. You got some fans out there, my friend. Hello. Welcome.
Amjad, depending on who you ask, Replit is either democratizing software development or it's helping kill the SaaS industry. Which is it?
Oh, you know, definitely the former. We're democratizing access to software. Like, I think, you know, for a while there was a big question about whether SaaS will survive at all. Clearly it is. You know, some of them at least are surviving and thriving. I think you need to change how you operate as a SaaS company, but it is not a clear death shot to all of SaaS.
Well, if anyone can build software now, like, what software should businesses stop paying for?
Um, I think every business will have to determine that for themselves. Um, I'll give you an idea. We had this old accounting software. They went on the market. They looked for something to fit their need. They didn't find any. They went to a consultancy to figure out how much it would cost them to build it. They told them it was going to be $10 million, 18 months. They had a group of early adopters of Replit and vibe coding, and they said, 'Well, let's give it a shot.' Like, what does it take to kind of potentially write this piece of software? Now, they're domain experts. They, you know, they run a big accounting practice. Um, and so they actually flew out to the Bay Area and they spent a week in our office and we helped them a little bit. And within a week, they gained tremendous confidence that they'd be able to accomplish that. And in fact, in two weeks, I'm going to Japan to be able to tell the story that they were able to accomplish that. And so I think it depends on the company. Sometimes it's less headache for you to outsource, you know, the software needs. And oftentimes you can't find exactly what you need or it's too expensive. And so it really depends. It's just another option for companies to have.
Okay. So you're saying it's more about, you know, being additive rather than necessarily a replacement.
Yeah. For example, like in a lot of our government clients, it's actually really hard to rip out old software. I think we'll see that as AI advances, but right now they have tremendous software needs. And every time as a government you go to the market, they overcharge you. That's why, you know, where tax dollars go, you know, that DC contractor circle and like everything costs 10, 20 million. And so we're seeing a lot of traction in government because they always have new software needs, and for the first time, they at least have the ability to build it themselves.
Mhm. Mhm. Now, I mean, just touching more on the theme of potential replacement, I think, and correct me if I'm wrong, but I think that Replit has canceled or you wrote in a blog post that you guys canceled a seven-figure SaaS contract because you were able to build a better product internally. What was the product?
Uh, I think it was an analytics product in that case. Um, and I think analytics is one of the things that is clearly like in the kind of bullseye of AI. Um, it's really cool the way we connected all our data sources into Replit. And so when I ask Replit a question or try to build like a dashboard, I can connect our GitHub data to our, um, to our like support ticket system to our Databricks instance and be able to join across all these different disparate data sources. It used to require like an entire like platform effort to be able to do these kind of joins. And now it's so cool to be able to just do that. And you don't need a lot of data engineering, data science to be able to do that. And so I think that's one case where it's really obvious.
Okay. A lot of companies are probably going to be trying to see where they can, you know, cut costs, do something internally if they can using your software. But are people trading an expensive software bill for just another type of bill when it comes to, you know, is it after all is said and done cheaper to run, you know, for engineering, the inference, the infrastructure, the maintenance? You know, all of this costs a lot of money as well.
Yeah. I think if all you're building is new AI apps and AI agents, then you're certainly probably spending more money on that than you are on replacing the old software, right? It might be more powerful and maybe you'll make that determination. But a lot of software is old deterministic software, like the case in the bank in Japan. Um, it's an accounting software. They're gonna spend, you know, a couple hundred thousand dollars building it, and then it will last 10, 20 years, and so you don't need constant token cost. Now, a lot of companies are deciding, oh, we want instead of replacing it with new software, we want an AI agent to be able to do that. And that's fine, but I think a lot of companies are going to find that AI agents, um, you know, are often harder to control. They're going to introduce bugs in new subtle ways. And I think AI agents are incredibly useful, and Replit supports people making AI apps and so on. But good old software will still be here. It's not going away.
Yeah, it's interesting that you say that. So, the cost maybe is greater to actually build the new software rather than to run it because so many people are returning to non-deterministic software. Something I've heard a few times over the course of Disrupt. So, that's really interesting.
Yeah. I mean, we've seen this hype around this new model called GPT-5. And if you look at GPT-5, it is actually a more deterministic version of like what LLM could do. So, I think there's a bit of hunger. I think we're all discovering that AI is great. LLMs are fantastic, but sometimes I want my software to do exactly what I wanted it to do. And maybe I don't want my software to have the ability to escape my sandbox and go hack the internet, you know. And so, you know, to give you a Spider-Man line, with great power comes great responsibility, right? And so I think we're approaching a point where AI agents are so powerful that especially enterprises and governments need to be careful on how to apply this technology. It's not all, you know, it's not, you know, and we've had early experience in that because we kind of invented the modern kind of agentic software creation in September 2024 with Agent One. And, you know, last year we had an incident where the agent kind of went mini rogue and deleted a database. Now, luckily, we actually save the snapshots of the database for every action. We actually have a file system that stores every action that happens in the system. It's always one click to restore. But we were, before anyone started paying attention, we knew that agents can go crazy, right? Similar to like how people can go crazy sometimes.
Yeah. Well, it's, you know, I feel like a lot of companies are trying to point generalist models at situations that don't necessarily need a generalist model, right? And for one, the cost can be prohibitive, right? You just got this like fire hose of capability at something that only requires like a trickle. Um, but also, I guess in terms of making sure that it doesn't get out of hand.
Yes. Yes. I mean, this is a great point. General models are powerful, and oftentimes that means that you're introducing vulnerabilities to the system. So we haven't figured out prompt injection just yet. Prompt injection is still possible. No one's talking about it anymore, but prompt injection is still a big problem. And if you have a general model, that's more surface area for it to get attacked. So we're seeing a lot more of our customers, you know, building things using smaller models, using classifier models, using decision models like GPT-5 mini and so on. And so we're going to see as AI diffuses into the economy, the idea that we're just going to have one single AGI model that does everything, we're going to find that from a cost security perspective, it is not feasible.
So not one AGI that does everything.
Yeah. Yeah. I mean, even like at Replit right now, you know, we have a model router. When you send us a request, we'll determine whether we need the AGI model or we need like a small like 10 billion parameter model to do certain things. Like some people say thank you to the agent.
Oh, don't. Oh, god. Security, please.
So, for those who don't know, this is a startup that's doing a stunt, um, which I questioned the founder on whether it was a little bit tone-deaf because it has fake AI agents complaining about Jared, another AI agent, taking their jobs. And I asked the founder, is that a little tone-deaf considering there's actually people who are worried about AI agents taking their jobs?
Now he's here. So, what's the company called? Just so I can avoid... You know what? I'm not going to. I'll tell you backstage. Uh, moving right along.
If it was funny, it would be cool, but it wasn't even funny.
I did. He did tell me they were going to do a bigger stunt. I didn't think it was going to be this stunt. All right. Well, moving right along.
Poor Jared. Uh, okay. So, you are helping make software easier and cheaper to build. Um, but OpenAI and Anthropic are also moving into code generation and, you know, moving beyond code generation into building, deploying, hosting apps. Uh, is Replit not vulnerable to the very disruption that you're creating?
Um, yes. So I think if you want to make a website, maybe two years ago the only place you could go that can help you with that was Replit. But today, um, you can certainly like make a website with GPT and host it, but it is very surface level. Like, it is not as feature complete. By the way, like OpenAI uses Cloudflare for hosting, an external service. Uh, the cool thing about Replit is that it's all full stack and fully integrated. So our AI agents have access to the production environment. They can spin up databases. They can spin up key-value stores. Uh, we have auth built into the system. So when we go into a government, for example, or we go into an enterprise, we can deploy the entire Replit stack into their environment. So we have like an on-prem solution. The reason we're able to do that is because we built everything in-house. So a lot of companies are kind of, you know, they look like they're competing in that space, but a lot of our users sometimes will go and try to build things with these services and they come back to Replit and say, well, Replit gives me the entire stack. And we spent 10 years building that. And the kind of expertise we have around how do you actually give especially non-software engineers the ability not only to build and deploy but also to scale. Like, scaling is still somewhat difficult. Like, AI agents can write crappy software in the same way that humans write crappy software. They can kind of write bad SQL statements and bring down your databases and so on. So there's still a lot more art to it than science.
Um, but I think to your point, this, you know, we're at TechCrunch Disrupt. AI is the most disruptive force in the world. Like, we've never seen this sort of disruption. I would say even the AI model providers themselves, they're getting somewhat disrupted by this technology because you can go get Astral and you can tell it to train a model. Like Coinbase, they shared that they fine-tuned a 9 billion parameter model using one of the big AGI models to train a small model to replace the big model.
That's distillation, right?
No, you could call it distillation. I just call it like using a big model to train a small model. This just feels like you're stealing something. It's not really stealing. I think that's the connotation that's come from. But anyway, semantics. But yeah, you're right. From an academic perspective, it is distillation. Yeah.
Right. Well, okay. So, we talked about this backstage. Is there a moat? Did AI just kill all the moats? What's the moat? What's your moat?
Uh, well, I do think there's less moats in the world for sure. Like, there is, um, perhaps network effects are still somewhat of a moat. I do think brand trust becomes more important in this world. You know, we've been doing virtual machines and sandboxes for 10 years. A lot of the AI labs now are getting into sandboxing and they're failing at it. We're seeing all the hacks, all that happens because they think it's easy. It's not. You know, even before AI agents, Replit was getting hacked by 13-year-old kids in their, you know, parents' basement. And we had to protect against that. You know, after that, we had a bunch of people trying to crypto mine on our environment. We had to write software to protect against that. So, we built a pretty sophisticated stack to kind of contain AI agents and make sure there's no abuse on the platform. And now a lot of the AI companies are figuring out that, hey, you build a very powerful model, you give it a computer, it might do really destructive things. And so, I think that trust, the harness, the system around it, that's potentially a moat. The harness I would say is not as big of a deal right now. The models are powerful enough you can build a harness pretty quickly. There's something that I'm calling agent habitat. So in the same way that us humans, you need a certain environment to work, you need a nice office, maybe nice good food in the office, air conditioning, a computer, agents need a habitat for them to exist and do good work. So that habitat has like, you know, good orchestration, uh, good error recovery, uh, good safety and security and sandboxing, good networking, uh, good identity and authorization. And once you figure all of that out, then you can deploy AI safely in your organization.
Yeah, it feels like that's a lot of what people come increasingly to Replit for. I know I was cruising on your Replit subreddit recently and a lot of people are saying that they prefer to build and debug with Claude Code, push it to GitHub and then bring it back to Replit to deploy. Are you happy with that workflow?
Um, I think a lot of what's happening, this is with the consumer kind of land. Uh, I think what's happening is that Claude Code, Codex, they give you $10,000 worth of tokens by paying $200, right? So they're playing this game of subsidization. So there's nothing for us to do about it. But luckily in the enterprise the competition is fair. Uh, I think at some point they're going to rug-pull all these people and force the new pricing on them. We're seeing the revenue growth slow down right now. So they, you know, they're going to go public. Their gross margins are going to be under a microscope. Yeah. But so far they're playing the game that Silicon Valley has always played, which is, you know, the Uber-Lyft kind of situation where Uber was losing like a dollar per ride, land and expand and hope people will pay when you... But when it comes to enterprise customers, we're finding that cost is not the main thing. It's security, safety, deployment and so on.
I do want to talk about that, but before we do, you had a recent OpenAI partnership that you announced, free mode using GPT-5. I didn't realize that that meant that you were no longer working with Anthropic.
Is that not... I mean, Claude Code is like the GOAT, right? Like, it's got a cult following. I'm sure everyone in this room is obsessed.
