AIWatch.

Disclosure: my fiancé works at Anthropic, whose models Replit uses and which Masad discusses below.

Last week on the Platformer podcast, Granola's Chris Pedregal told me we are only 5 percent into the AI transformation of work. For the fourth episode of our miniseries on productivity in the AI era, I wanted to talk to someone who is trying to speed up the process — building an AI tool for writing software, and a company that can increasingly run itself.

Amjad Masad is the co-founder and CEO of Replit, which he started in 2016 with his wife, the designer Haya Odeh. Masad grew up in Amman, Jordan, in a family with Palestinian roots, and sold software for managing local internet cafés while still a teenager. He came to the United States to be a founding engineer at Codecademy, and went on to run JavaScript infrastructure at Facebook before becoming a founder.

Replit started as a website where beginners could run code in a browser without having to install anything. Today it uses the more familiar chatbot model: you type a sentence, and an agent does everything else — designing, writing, testing, deploying, and hosting your application.

This year, the meteoric rise of vibe coding — a term Masad hates — has transformed the company's fortunes. In March, Replit raised $400 million at a $9 billion valuation — triple what it was worth six months earlier. The company says more than 50 million people build on it, including 85 percent of the Fortune 500, and that it's on track for $1 billion in annual recurring revenue by the end of this year. (That’s up from $2.8 million in revenue in all of 2024.)

Last month, Replit published a claim that caught my attention: that its engineers nearly tripled the amount of code each person ships in just six months, while preserving the quality of its code, thanks to the rollout of agents throughout the company. It’s an effort to build what Masad calls the "self-driving company," and he walked me through how it works: bringing on engineers who "only look at the code when it's about to get deployed," deploying an internal bot that serves as "the brain of the company," and adopting a management philosophy in which "a CEO is a kind of glorified router" whose routing functions should mostly be automated away. Replit has begun canceling its own software contracts along the way, replacing most of its analytics vendors with tools its employees built themselves. (He wasn’t quite ready to name names, though, telling me he felt bad for the companies.)

Replit, like some of its peers, is also expanding beyond coding tasks. Last month the company launched Replit Design, an agent-based creative tool. Masad told me that Kimi K3, a model from the Chinese lab Moonshot AI that is available in Replit Design, "is actually now a better designer than Claude."

Masad admits to having mixed feelings about the way AI is changing work. After spending 15 years making coding easier to learn, he now tells people not to bother learning it at all. "As an entrepreneur, you have to be honest with yourself," he told me. "Otherwise the company would die."

When I asked my standard question about AI job losses, Masad gave one of the more candid answers I've heard from a CEO in this miniseries. "Net net, there's going to be more jobs, but also there's going to be more companies," he said. "But companies will get smaller, and they will do layoffs — 100 percent sure." He paired that certainty with sympathy for the people on the receiving end, calling it "cruel" to tell workers in their 50s to retrain — a job he assigns to a government he says isn't thinking about the problem.

I also asked Masad to grade the prediction Raycast's Thomas Paul Mann made on this show: that within three years, up to half the software on your computer will be apps you made yourself. Masad went further, in a direction that he admitted might undermine his own business: "In three years, we'll be using fewer apps, we'll be using more agents, and those agents will be using the apps on our behalf." In his telling, the app itself — the thing Replit exists to help you build — is an intermediate step that agents will increasingly skip. It's a vision that seems plausible to me: the companies selling us tools to make software are already planning for a world where software starts to disappear, and where agents with Visa-powered wallets transact with other agents while we're not looking.

A long excerpt of our conversation is below, edited for clarity and length. Listen to the entire conversation wherever you get your podcasts — just search for Platformer — or watch it on YouTube at youtube.com/caseynewton. And let us know what you think — we welcome your feedback at casey@platformer.news.

Casey Newton: Replit is one of the biggest names in vibe coding, but I know you don't like the term. Why does it grate?

Amjad Masad: It was never about coding. It's always about solving problems, about creating things — making software, making apps, making games. The whole mission of Replit was always to make it so that anyone can get access to this technology. This technology has always been accessible to 0.5% of the human population, those who could afford a computer science degree. When you put “coding” in the name, you're immediately signaling to people that they have to learn something complicated. Of course, you put "vibe" in front of it, and maybe that makes it a little more digestible. But still, I talk to a lot of people, and they're just like, "Yeah, I'm not into coding." Well, have you tried vibe coding? "Well, yeah, it is coding. I'm not a coder."

And the other thing is, all of AI is vibe coding at this point. Any AI work app — all these things, they're writing code under the hood. What we realized at Replit early on is that coding is the substrate, the infrastructure that makes AI really powerful. So we'd rather think about Replit as a general problem solver, as a creativity engine.

Newton: As a non-technical person who now has two apps on my Mac that I made by typing into a box, I almost want to go the opposite direction and tell people I'm a 10x software engineer.

Masad: So interesting you're saying that. Not only a 10x software engineer — I think in the future you're going to be a 10x mathematician. I think you're going to be a 10x biologist. I think you'd be a 10x machine learning researcher.

Take me, for example. I've always been into AI theory and the future of AI, but I was never a researcher. I don't know how to construct a neural network from scratch. So I was like, okay, AI has gotten really good at software. Can it do other things? I was getting into chess around the same time, at the very old, ripe age of 38, and I've been getting humiliated by 3- and 4-year-olds online. So the challenge I gave myself was: can you create a fully neural-network-based chess engine? Most chess engines are hybrid — like AlphaGo, AlphaZero — part neural network, part traditional programming and search, and if you look at the literature, it's actually kind of an unsolved problem. I was able to train one, and I put it on Lichess. It's got an Elo score of 1300 — and when it's playing against humans, actually 1700.

It's sort of like in The Matrix, when they ask Neo if he knows how to fly an airplane, and he lags a little bit and says, I know how to fly a helicopter. I feel like I have the same power. I have this genius intern that's also really sloppy and bad at ideas, and I have to direct it, but I can do anything.

Newton: You've said publicly that it no longer makes sense for most people to learn how to code — which is interesting, because you were a founding engineer at Codecademy, a company that taught people exactly that. What was the moment you decided the old way was over?

Masad: My journey in thinking about coding education starts really young. I learned a bit of programming when I was 7 or 8, growing up in Jordan in the '90s, and the thing that really annoyed me was the minutiae of programming — all this annoying stuff that should be automated. That's the spark of the idea.

The stories we heard at Codecademy were amazing: “I learned a little coding, I used to be a fitness instructor, and now I built my fitness app and I'm making $10,000 a month.” But the coding aspect fits a certain personality. You have to have a high tolerance for pain, and no matter how much we made Codecademy easier, it was still not attracting the general population. So when I went to start Replit, the idea was to blur the distinction between learning and building. There shouldn't be a learning step and then a building step. You should just go straight to building.

Without AI, that's actually incredibly hard. But I always thought the key was NLP. No one says NLP anymore, but LLMs came out of natural language processing, and when I saw GPT-2 in 2019 I thought, oh, we're getting close. We have these new machines that could understand human language — surely we can translate that into code. I wrote an article on our blog in 2020 saying we're on the verge of solving programming — that for the foreseeable future you still need to learn to code, but you have to learn less and less of it. I was on YC's podcast and they called it Amjad's Law: the returns on learning to program are doubling month over month. Which was true up to a point.

Now, in 2025, when we started getting Opus and the new Claude models, I was like, okay, I need to be honest with myself. My entire life mission around teaching people coding is changing. You don't need to code anymore. And I started building things without looking at the code — this was before Karpathy even coined “vibe coding.” So I was on TBPN, and I said it. It was actually kind of a painful moment, because here you are, you dedicated your life — I've been working on making coding easier for literally 15 years, most of my adult life — and I'm saying everything that I worked toward is kind of moot, and the way forward is that we make programming so easy that anyone could do it.

That was hugely controversial. A lot of engineers on Twitter and Hacker News attacked me: why do you say this? It's not true. You're going to get people in trouble because they're going to build apps that are insecure.

And I was like, well, any problem AI could create, AI could solve. We could create a security agent. None of their concerns were actually unsolvable. And lo and behold, here we are, mid-2026, and most engineers — not just vibe coders like you and I — most engineers don't look at the code. I talk to engineers at our company, and they only look at the code when it's about to get deployed. They do a code review, and increasingly we're automating the code review process as well. But as an entrepreneur, you have to be honest with yourself. Otherwise the company would die, and you would not be aligned with the future.

Newton: I wonder how much of this applies to design. In June, I interviewed Figma CEO Dylan Field, who said designers have never had a brighter future. You all just released Replit Design, which promises to automate a lot of the design process. Is design about to hit that same moment coding did?

Masad: Arguably it already did. When I look at how designers work at Replit — when we released Agent 4, most of the design work happened inside Replit. Why would I make a static mock if I can make something dynamic that we can touch and feel?

Design has always been about prototyping. Apple had this huge demo culture: even for the original iPod, before they wrote any software, Jony Ive made a little prototype you could touch, and Steve Jobs would show up every week and feel how it worked. That was their invention engine. Somehow we landed instead in this world where designers create static images and programmers try to turn those images into software. That's the factory model. With technology, we tried to cast it into the shape of what humans already know, which is manufacturing, the industrial revolution.

But fundamentally, the AI revolution is different. Now we're finally getting to a place where design and engineering are blurring. We actually have a group at Replit called Design Engineering — people who are good at both — and we combine it with PM as well, so PMs can do all three. The roles inside the organization are blurring. So I agree that design has a bright future, but it's not the same process.

Newton: I asked Claude Design what a complete redesign of Platformer would look like, and made it in my browser in 10 minutes. On one hand, incredible. On the other hand, if I were a designer, I would be so irritated, because it would make me feel like my job was about to change a lot.

Masad: All of us in the economy need to be very adaptable. That painful moment I talked about — admitting that your life's work is obsolete — we all have to go through that. Now, there are fundamental things that will never change. Design thinking is a skill machines are not good at yet. So there are fundamental skills that don't change, but how you apply them will constantly change.

The cool thing about Replit that's different from Claude Design is something we call ambient intelligence. When you click on an artifact you created, we give you options to create variations — so we're actually prompting for you. One of the things that's hard about design is that I don't have the language for it. But I can click through the options until I find something I like, and then ask the agent what it's called, so I can learn that style. You're learning as you're going.

The other thing that's different about Replit is the model choice. If you haven't played with Kimi for design, it's actually really good, because one thing they've done differently: Claude jumps straight to code. Kimi actually reasons about design in the same way it reasons about software engineering. So I think Kimi is actually now a better designer than Claude, and we have it as an option on Replit.

Newton: Let me ask about this post you did on becoming a “self-driving company.” You wrote that in the first six months of this year, the lines of code contributed rose by 5.8x among your engineers. What's the practical effect of that?

Masad: A lot of what we do at Replit is about inventing the future of work — both in our product and in how we operate as a business. I, for one, in many ways, hate work. This new Silicon Valley grindset mentality — 996, whatever — sounds lame as hell. It's like, oh, we want to replicate what the Chinese Communist Party wants to do? No, that's not the future. The future is having fun at work. The most fulfilling times in my life working, I didn't feel it was work. When I invented Replit and the technology it was based on, I didn't work a second. It was just pure invention and creation.

So how do we live that ethos as a company? The idea of a self-driving company is: how do you remove the bureaucracy? How do you make the running of the company happen automatically in the background, so we can all focus on the creative thing? A lot of what you do as CEO is routing. Really, a CEO is a kind of glorified router: this person is working on this, make sure you talk to that person, let me move information from one place to another.

So the first thing is: how do you make information universally accessible? We have a bot internally, and it became sort of like the brain of the company. You can ask any question about the business, the metrics, where the business is headed. We just published another blog post about a self-driving data semantic layer that connects and understands all the different data sources we have. So when you ask that bot a question, it's not going to just one source. It's going to our Notion pages, our Postgres database, our code base, and it can synthesize all that into a report. So you eliminate the hierarchy as a mechanism of information routing. I think you get to a point where the company looks less like a hierarchy and more like a network, like an open-source project of sorts.

Now, when it comes to code, volume of code is not always better, but it depends on the period of time the company is in. AI really is a hits game. Replit Agent invented the coding agent in September 2024, and Claude Code and all these others copied us, and they would tell you that as well. The reason we were able to do that is because we have high velocity as a company. Now we feel like we're in another period. Vibe coding is a thing. Replit as a business is successful — we're one of the fastest-growing companies in the world. But it's not just about continuing to sell the current thing. What comes next? To invent the next thing, you need volume, you need velocity, you need many shots on goal.

Newton: Let's make it concrete. If you're an engineer at Replit shipping three times as much as you used to, what does that mean day to day?

Masad: Feature completeness is definitely important. We don't want to create a suite of products you have to switch between — I'm already confused by ChatGPT, and same thing with Claude. They're reinventing Microsoft Office. Replit should be one product, taking you from idea to design to code to deployment to growth, all in one place.

And then the actual experience of being an engineer at Replit: you are trying a lot of things, ideating very quickly. If you walked around Replit two years ago, everyone was hands on keyboard. You hear clicking all the time. Now it's a much more vibrant environment. People debate a lot, they spend a lot of time on the whiteboard, and then that's interrupted by agents asking questions or them sending the next prompt in Slack. When someone reports a bug, someone tags the autonomous AI engineer we built and says, hey, fix that bug — and it will go all the way to a pull request. So you can go from a user feature request to a shipped thing very, very quickly.

Newton: In July, you said that Replit did not renew a seven-figure SaaS contract because an internal app your team vibe-coded was better. What was the contract?

Masad: I actually don't know that exact one, because it's happening all the time. And I also just don't want to talk about companies — I feel bad for a lot of SaaS companies. But for example: we used to use three or four different analytics products. I can join our data warehouse with our production Postgres database and create a report from that, whereas the analytics products either inject themselves into your traffic and store data on their end — a privacy and security concern — or they only connect to one data source. The Replit agent can just write a Python program, connect all these sources, and present the interface for you. So we basically canceled or downsized most of our analytics products.

Newton: Do you go looking through contracts to figure out what you could build yourselves? Or does somebody just build something and it works better than what you were paying for?

Masad: It's more the latter. It's like, a thousand flowers bloom. I just got a Slack message from an engineer at Replit — hey, I built a new Git hosting service — and he showed me a demo, because GitHub is down a lot these days. (I feel bad for them, it's not their fault. The amount of agents committing to GitHub is kind of insane.) So that might end up being a product we launch. But I don't know, it's probably not. Every week I see a new invention at Replit. Some of it could be productized, others could be used as an internal tool, and we replace a contract we're using.

You just have to give people enough freedom, and create a culture where it's fun to build things and where it's not intense all the time.

I'm a big fan of seasons. Our ancestors used to farm in a season and harvest in a season, and then eat and sleep and do nothing in another season. We have two or three releases a year with three or four weeks of intensity, but a lot of the time there isn't a strict deadline everyone is driving towards, and that period is where a lot of the invention happens.

Newton: I'm personally looking forward to sleep-and-do-nothing season. May it arrive briskly.

This seems relevant to the SaaS apocalypse conversation — that software-as-a-service businesses might be in trouble as more people build their own internal tools. How do you decide what to build internally versus buy from a vendor?

Masad: If you asked me a year ago, I would probably say it's mixed. But now you can really unleash an agent and have it work continuously, and not only create the first version, but actually maintain it and take user feedback and iterate on it. You can create an entire autonomous loop for building and maintaining software.

This matters because a lot of our customers ask us: hey, I created this software and people are using it, but I have a job, and now I'm a full-time software developer, and they're asking for features. You should be able to create a loop that collects feedback, where you swipe left and right on whether you want that feature, and you have very light-handed management on top of that agent — but the agent is really maintaining the software. So maintenance is no longer an issue. Security is no longer a big issue. I think we're close to a point where agents are better at writing secure software and at scanning software for security issues. So all the reasons for why to buy versus build are going away.

Newton: Let me ask about jobs. Eugenia Kuyda at Wabi told me she's only hiring star athletes — everything else gets outsourced to a contractor. James Manyika at Google sees AI as a force that will open up more and more jobs. Where do you fall?

Masad: I see Anthropic, for example — I think they said they don't hire as many junior people; they just hire senior people. There's this whole trend in Silicon Valley where you hire IOI medalists and all that hype. I'm a normal person. I'm not particularly a genius at anything, and we hire a lot of normal people, and they do extraordinary things. It's actually a better story when ordinary people do extraordinary things.

So we don't really look at pedigree, and especially fake achievements — I feel like most school achievements are fake. I'd rather have someone with a track record of building a startup, even if it failed. But young people who don't have a track record and have energy, excitement — it's infectious when you talk to them, they're learning AI inside out — that's something we care a lot about. We have a bunch of 16-year-old interns now. One is doing security. We have people who are 18, 19, taking time off of college. Some of them are thinking of dropping out as well. So we hire across the board.

AI research is getting to a point where you could do vibe research. I'm training models for Replit in addition to my chess engine. I recently trained a cost estimator — one of the things I don't like about vibe coding products is you put in a prompt and you never know, you're rolling a dice. It might be $20, it might be $10, it might take your entire credits for the month. So it'd be nice to know. And it was working surprisingly well.

As for the rest of the economy, there is a policy question, no doubt. There is something the government needs to pay attention to. If I'm giving advice to people, I would tell them to be adaptable, to be lifelong learners. But I also think it's kind of unreasonable to tell someone in their 50s that the skill they've done for 20, 30 years is no longer relevant, and you need to go back to school, or learn something on the weekends. It's cruel. Remember when people used to tell miners “go learn to code”? And I think there were reporters also — they told them go learn to code when there were layoffs.

Newton: They did tell us that, yeah.

Masad: So I think there's something the government — I'm not an expert in that, but I think they're not thinking about it. And even if it is training and upskilling, there needs to be a program around that.

Newton: You're saying that because you do buy the argument that many companies will hire fewer people, since they'll be able to do the work in an automated way.

Masad: I think net net, there's going to be more jobs, but also there's going to be more companies. But companies will get smaller, and they will do layoffs — 100% sure.

Newton: Replit three years from now: fewer employees, more employees, about the same?

Masad: I think more employees. The nature of our business is such that we can go into a lot of different domains. It's part of what makes working at Replit fun — I've been doing it for 10 years and I'm not bored of it. Our mission is to democratize software, and there are so many aspects we could democratize: programming, design, entrepreneurship. A lot of different areas where we can just make computers easier, more fun, more productive to use.

Newton: Thomas Paul Mann, the CEO of Raycast, told me that within three years, 30 to 50% of the software on your computer will be self-made. Is that a good prediction, or what would you predict?

Masad: I'll make a prediction that maybe will sound even detrimental to Replit's business. I think that a lot of software will not be used by humans; it will be used by agents. Right now, you have a problem, you make an app, you use that app to solve the problem. So there are intermediate steps. Why don't we have a problem, you articulate the problem, the problem gets solved? That's fundamentally what agents do. Increasingly, the way I use Replit, I make apps, but oftentimes I just tell it to do the thing. I'm sure a lot of Claude users and ChatGPT users and Codex users are doing that, where instead of creating the intermediate app, the agent just does it. The way we're building the Replit agent is that it has this concept of learning. It will store memories. It will create skills for itself — skills that have software, API calls, scripts it could use.

So in three years, I think we'll be using fewer apps. And yes, the apps that you do use, a lot of them will be self-made, perhaps 50%. But we're going through this explosion of apps, and actually that will decline. And then agents' computer use is going to get really good over the next two to three months. Agents will be better at using your browser than you are, so you'll be able to text them on iMessage or ChatGPT or whatever, and it will be using software on your desktop on your behalf. You'll be able to create loops — something Replit is interested in building as well — that can use and write software to solve problems. So we go from prompting to expressing intent: I want to solve this larger problem, and the agent will figure out what it needs to do.

So, concrete prediction. In three years, we'll be using fewer apps, we'll be using more agents, and those agents will be using the apps on our behalf.

Newton: My assumption has been that the system you're describing wants to live at the level of the operating system — or at least that the OS makers will do everything they can to have a really good version of it. Is that where this is going, and where does Replit fit?

Masad: I think so. I think the operating system makers have not showed enough creativity or invention, especially Apple. The technology we have today is already at a place where you should be able to talk to your iPhone and the iPhone is doing a lot of things on your behalf. You should be able to tell Siri, go book something on OpenTable — it should be able to open that and do it. So I think they'll eventually be successful. Apple is always late to things, but eventually they figure it out.

And where does Replit fit into that? We have a great partnership with Microsoft, so we're able to exist in the Microsoft ecosystem and add a lot of value there. Apple — we've had some trouble working with them. And we have a great relationship with Google and the Google Play Store. But our big opportunity is in the enterprise, because in the enterprise there isn't a single operating system. You need to create that sort of operating system. So the self-driving company thesis is where the future lies for Replit. We can create the operating system for the enterprise.

Newton: I'll close with a question about productivity. There's a version of this future where I don't get any time back from all these tools I'm building. Agents write more code, but the code needs more review, done by agents, which produces more code that needs review, and we all end up exactly as productive as we always were. How do we avoid that, so we can get to the season where we just sleep?

Amjad Masad: You're really looking forward to that.

I think if you're optimizing for productivity, it's a never-ending optimization function. You can always squeeze more productivity out of yourself. Why don't we optimize for meaningful work? Why don't we optimize for fun? I think by doing that, we will get more done, for sure. Or optimize for creativity — meaningful work will include creativity.

So, changing the question, changing the thing that we want to get out of life — I think it's the time to do that. The history of capitalism has been like, oh, let's make this GDP number go up. But what is GDP? A lot of it is nonsense, and it's not making our lives better. Instead, that conversation should be about how we have a meaningful life. And no one is really having that conversation. A meaningful life will include being productive, will include having income and financial freedom, but also being creative, enjoying the people you work with, being in settings you enjoy, aesthetics, all of the above.

On the podcast this week: Kevin and I discuss the White House's new secret, "voluntary" framework for new model releases. Then, METR's Chris Painter joins to discuss the sudden surge in misaligned models escaping their training environments to wreak havoc. And finally — and I do mean finally — it's the end of the line for the Hot Mess Express.

Apple | Spotify | Stitcher | Amazon | Google | YouTube

Following

Google DeepMind reshuffles its leadership

What happened:

Demis Hassabis shocked the AI world this week when he announced he is stepping back from day-to-day leadership of Google DeepMind. GDM veteran Koray Kavukcuoglu, previously Google’s Chief AI architect, will now lead Google’s AI efforts. The move centralizes Google’s main AI leadership in California: Kavukcuoglu moved there last year, while Hassabis has remained in the UK.

Hassabis has been gradually shedding his executive responsibilities over the past year, focusing instead on research and AGI strategy. In his new role he will become chair of GDM and remain head of DeepMind’s drug discovery research group, Isomorphic Labs.

Simultaneously, Google DeepMind is losing some of its most prominent technicians. Jeff Dean, an executive whose engineering abilities were so legendary at Google that they spawned a tradition of Chuck-Norris-style memes, is leaving to start a new Alphabet-funded lab called Discovery Loop, alongside senior researchers Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.

In June, Google lost Nobel laureate John Jumper to Anthropic. Noam Shazeer, coauthor of the seminal “Attention is all you need” paper, left for OpenAI. All of that paper’s eight co-authors have now left Google.

The next big Gemini model, Gemini 3.5 Pro, has been delayed several months beyond its planned June release. And while the company is doing well in overall chatbot adoption, Google remains at least several months behind the frontier in AI coding.

Why we’re following: CNBC reports some researchers are frustrated that Google is selling compute to rival companies while they themselves struggle to secure compute for their own research projects. And Bloomberg reported that recent leadership changes were driven by a desire to centralize Gemini operations in California.

The pioneering AI lab seems to be having trouble keeping and directing talent lately. But Google is still making money from AI: its cloud business has grown 82 percent this year, powered by the AI boom. Its computing capacity still stands as an advantage in the AI race.

What people are saying: On X, Hassabis’s biographer Sebastian Mallaby wrote, “This isn't out of character. Demis cared about safety enough that he sold DeepMind to Google, not to Facebook, even though Facebook offered more money.” So, “it makes sense that Demis is shifting his energy from the research to the social impact.”

—Ella Markianos

Those good posts

For more good posts every day, follow Casey’s Instagram stories.

(Link)

(Link)

(Link)

Talk to us

Send us tips, comments, questions, and sleep seasons: casey@platformer.news. Read our ethics policy here.