A Beginner Just Shipped a $250K Business. Professionals Got 30 Percent.

Sarah Guo introduces Replit as a develop-to-deploy platform with more than 22 million users, a new Google partnership, and an AI suite called Ghostwriter. Amjad Masad — CEO, co-founder, and head engineer — traces the company not to a Facebook exit but to college. He was obsessed with programming languages and miserable about setting up development environments, and he did not have a laptop. The naive question was: why can't you open a browser and start coding? Nobody had done it, so he tried. The first prototype was a web page with a text box and a button that said Run. His friends loved it. In a lot of ways, he says, that is still Replit.

The prototype became a multi-year project to build an in-browser sandbox. The breakthrough was compiling Python, Ruby, and other languages to JavaScript so the code could run in the browser; he open-sourced that work and it got him a job as a founding engineer at Codecademy. From there he went to Facebook as one of the founding engineers on React Native, working on React, Jest, Babel, and other tools JavaScript developers still live in. The goal was to make mobile development as fun, easy, and fast as the web, and he thinks they succeeded. Toward the end of his time there he looked back at online editors and found plenty of them, none of which felt web-native — no shared URL you could send so someone hops in and codes with you, the way Figma or Google Docs work. They started Replit in 2016 to close that gap.

Guo asks about the AI bet and Ghostwriter. Masad's whole career has been code that handles code — teaching it at Codecademy, compiling it in his own project, building the runtime around React Native. Those tools felt rigid and laborious, classical intelligence systems full of algorithms and heuristics. He always thought machine learning could be applied instead. The paper that marked him was 2012's "On the Naturalness of Software," in which researchers pointed NLP at code and found it could be modeled like any language. Code is repetitive the way language is; even a basic n-gram model could generate compilable programs. That is what "naturalness" meant, and it hit him a decade before ChatGPT.

Every year or two after starting Replit he would check the state of the art on ML for code. Nothing really worked until GPT-2, which you could fine-tune into something "kind of okay." GPT-3 was the moment he decided it was here. Ghostwriter began as tools sprinkled through the IDE — explain this, generate that — then an autocomplete product in the Copilot style, then chat, and the suite got a name. Elad Gil notes how widely accepted contextual autocomplete has become as a workflow change, with Replit and GitHub Copilot among the handful of tools people actually like. Masad is cooler about it. The tool is amazing and still very primitive. The full capabilities of GPT, and of transformers on code in general, have not been explored.

A lot of the next work sits around the model: better context, tools, the ability to read a file, install a package, evaluate code, and write entire features. Then there is work on the models themselves. Today's training diet is a large corpus of static code, so the systems understand structure and syntax and not semantics. You can imagine feeding them the results of executing that code, or training them not just to be good at code but at the editor, the debugger, and everything around it. We are, he says, at the very early beginnings.

For professional programmers, Replit's research on how people use the product puts the gain at 30 to 50 percent more productivity on coding itself. That is a narrower claim than it sounds. Debugging, deployment, meetings, and translating specs into code — everything around the act of coding — AI has not touched yet. Masad thinks it eventually will touch every part of the software-development life cycle. Right now the people getting the most value tend to be beginners, people who typically would not have the skills to start a startup. The company had just profiled an entrepreneur with a quarter-million-dollar ARR business built in a few months, who earlier that year did not know how to code. That, he says, is where most of the impact is today.

Project forward and those same people get more tools; professionals are where the 2x, 5x, and 10x multiples eventually show up. Gil asks for a timeline on massive productivity gains and on the agentic world of generating features and grabbing packages on your behalf. Masad's first answer holds today's models constant: a lot of this is within reach in six to 18 months. "We're in the demo era of GPT models right now." Replit is doing the unglamorous work of building a service inside the container that embodies the AI in the development environment, so it can go to the internet, install a package, read and write files, start another project, git clone. That is not one model. It is probably several. Within months he expects basic agentic experiences — give it "build a login page" and get a decent starting point.

Layer on actual AI progress — longer context, more performance, better prices, even new ways of training code models — and prediction gets hard. Software could look totally different in three to five years. His sketch is that programmers, even individual contributors, start to look like engineering managers whose employees are mainly AI, spending most of their time reviewing code and giving instructions instead of typing. He is confident only about the nearer claim: with current capabilities, there is still a lot of infrastructure to build over the next six to 18 months.

Gil wonders whether 22 million users are a data moat for reinforcement learning, and mentions a Google paper in which human feedback and AI feedback roughly netted out. Masad is measured. It is easy to overstate data moats. The real advantage is the end-to-end platform. A code model without a place to apply it and get feedback is hard to improve; an editor extension does not see deployments or execution, only user feedback, which is the RLHF-style signal and not the one he thinks matters most. The magic of Replit is the full loop from the first line of code to deployment to production crashes to the next edit — a cycle that is usually split across GitHub, VS Code, and AWS. That loop is how you get training data over a large action space.

Why train at all, instead of calling existing APIs? When Microsoft got GPT-3, Masad says, they distilled it, hosted it on their own infrastructure, and layered on caching. Anyone stuck on the OpenAI API could not do that. It was expensive and slow. Even now the commercial lineup has become chat models; they are not going to release completion models going forward, which makes a completion-based product fairly difficult to build on rented APIs. Replit wanted something cheap, fast, and good enough for autocomplete. Three billion parameters was the sweet spot: a 1-billion model felt too dumb, and 3 billion still inferred quickly enough to host.

The timing helped. The Chinchilla paper had just argued for training smaller models longer, Llama had been announced, and the idea of pouring tokens into a compact model was in the air. They trained on open-source tokens, then added more tokens from Replit's own data, which gave a 50 percent improvement over the base model they released. They open-sourced that base because they had trained on open-source data and wanted to give it back. At the time it was better than most open-source code models, including some 20 times its size. The need, he keeps repeating, came from a product, not from a leaderboard.

He is not a fan of the culture that followed. A lot of founders now treat training a model as pride and status. The moment Replit put out the 3-billion-parameter model, three or four copycats appeared, competing on what he calls silly benchmarks. The right order is to start from a user and a customer need. Right now, if they can do the job with Llama they will use Llama; if they have to train, they train. That, he says, is how the company has always been built: start from the customer, then solve whatever technical problem shows up on the way.

Gil asks what happens if Meta stops open-sourcing Llama. Masad rewinds to early 2022, when he realized commercial APIs could not build the product he wanted and felt depressed for a moment: Replit did not have much funding, so maybe only Microsoft would get to ship this class of tool. He got a meeting with Mark Zuckerberg and pitched an analogy. Google and AWS keep data-center secrets because they sell cloud. Facebook did not sell cloud; compute was a complement, not an advantage. So Facebook pulled together other companies in the same position and built the Open Compute Project to share hardware designs, and it was a huge success. Masad's ask was: do that for LLMs. Open-source AI would let Meta build a better product. It is the classical strategy of commoditizing your complement. Zuckerberg nodded. Later, in an interview, he used Open Compute as the analogy for what Meta is doing with LLMs.

What surprised Masad is how little industry participation followed. The difference from Open Compute is the AI-safety angle, which makes the work a little toxic: companies do not want to be associated with a model that spews something harmful. Releasing an open-source language model, especially under the political heat Meta is taking, takes guts. "There isn't a lot of guts in the industry... and Zuck has balls." If Meta stopped, he thinks it would be bad for the ecosystem, and it is not obvious who steps in.

Gil adds the historical pattern: almost every major open-source wave had a corporate sponsor. Linux was sponsored by IBM on the order of a billion dollars a year for a period; WebKit had Google and Apple. The money to train open models exists in the ecosystem; talent is harder. Meta's magic was the intersection of guts, capital, and an amazing talent base, and the replacement is unclear. Masad agrees that on a long enough time scale open-source models will match commercial ones, but the current landscape is hard to game out. Architecture is another wrinkle. Tools are easy to open-source because they are a binary you can run, which is roughly what we have today. If you add mixture-of-experts and stateful components, an open-source alternative gets exceedingly harder.

Gil points at Bounties, the marketplace Replit stuck inside the product so developers can collect dollars for code, and asks whether development agents end up in that loop. The original idea, Masad says, was duty as much as product. Replit is full of people around the world learning to code and becoming strong programmers that nobody knows about — no résumés, few traditional job skills. Getting them paid was a community-building act. People in Africa, India, Thailand, Serbia, and elsewhere are already making money on bounties. The thesis he wanted to test was that AI's near-term impact is largest on beginners, who are getting a 2x, 3x, even 10x multiplication. Learn a little code and you can build a product that makes $200K ARR. A human who knows a little code plus AI can be a really great developer. It turned out to be right: you can go to Replit and get a startup prototype or MVP for as little as $50 or $100.

Bounties were the first step toward something larger. He wants a currency, a way to transact, built into the platform. Open source's currency is GitHub stars, a fake currency with weird side effects, and the most impactful projects tend to need a corporate sponsor. The software economy is missing a native unit of value exchange so contributors can be rewarded and the work can be sustainable. If the development and deployment environment has money built in, you can publish a module, a component, or a service, price it per function, and let another app's wallet pay as it composes those pieces. Money is a coordinating mechanism. Agents make the picture sharper. Today's AIs are copilots, predicting the next autoregressive step. Semi-autonomous agents should be able to spend money and make money, interact with the wider internet, and — the version he actually wants — go hunt bounties fully automated. "Can you create an AI that can go earn money for you while you sleep?" On the data side, bounties also fill a hole: complete product specs paired with finished code, the end-to-end examples code models have been missing.

Guo describes a college hackathon her fund hosted, where 19-year-old MIT students who design their own GPUs were sure they would not need to learn to code the old way. Architecture is prompt guidance; debugging is another agent; they would rather orchestrate bounty-hunter bots. Masad hears a hacker attitude inside the bravado: if it becomes a problem, I will learn it. If the AI cannot do it, I will learn it. Just-in-time learning. He is not going to move himself up the MIT computer-science curriculum; he is going to figure out what he wants to build and work backwards, jumping 20 steps forward because of Ghostwriter.

On the payment rails he is pragmatic. If Stripe gets you there, use Stripe; if it is Bitcoin, use Bitcoin; he does not see them as that separate. Stripe, Square, and Bitcoin started the same year, which prompts a joke that Satoshi was actually Jack Dorsey and Patrick Collison working together. The idea in the air was that the internet was missing money as a primitive. When he designed bounties he borrowed staking from the Ethereum community: instead of a rigid apply-and-accept flow, you stake — say $100 — promising to finish in five hours, and every extra hour slashes your funds. The biggest, most confident bet takes the task, and coordination happens semi-automatically. He still thinks Bitcoin will be the long-term TCP/IP of money, with Stripe and Square as services on those rails.