He turned down a billion dollars for a world with a billion developers

David Cahn has known Amjad Masad for about five years and invested in Replit four years ago. The through-line, he says, has been the same since they met in a San Francisco loft that doubled as office and home: a billion people on the internet becoming developers. AI has made that idea less lonely. Masad dates it to childhood. He first used a computer at six, wrote a math tutor for his younger brother at seven, and did such a good job that the brother now works at Replit. Making software felt like the natural thing to do with a computer. He was surprised it belonged to experts.

The wall, in his account, was piled up by tools. As a teenager building a business he used Visual Basic, put a database in an app, shrink-wrapped it into an `.exe`, and sold it. Then the web arrived, more powerful and more complicated. A newcomer to a JavaScript app today might spend days on the development environment, plus webpack, transpilation, and other “asterisk” knowledge. He did not see an intrinsic reason for that. Open source replaced Microsoft product managers with a decentralized mess, which he still wants as an innovation machine. The old hacker split was cathedral versus bazaar. He calls that a false dichotomy. “We can build cathedrals from bazaars.”

If programming is something more people can do, the divide between developer and user shrinks, which he says was the original vision of computing. The other motive is opportunity. Programming made him money as a kid and later got him an O-1 visa to the United States. He wants that ladder to be less of an accident. Cahn asks him to describe the world a decade out, when the billion have arrived. Masad tries not to sound too utopian, then describes companies and software supply chains that no longer look like factories.

If most knowledge workers can make applications, Masad argues, the industrial-revolution pipeline stops being the design of the firm. Today a product manager writes a PRD, a designer takes it, an engineer builds it, a release engineer ships it, and marketers and researchers sit at the end. Schools work the same way: kindergarten, first grade, second grade. In a world of generalist software creators, a salesperson with an idea for an SDR agent spins it up instead of asking a boss to hire someone. A customer on a call asks whether a feature is possible, and someone prototypes it in Replit Agent live. He stresses that those are real customer stories, not slides.

Construction of software itself becomes a network. Instead of a factory chain of database vendor, host, and front end, an agent with a wallet can buy a database here and an SMS API there, using crypto and stablecoins so machines transact without knowing each other. When the app is monetized, a dollar flows through the whole graph. Ambient services, run by hackers making money on the internet, compose without a central planner. Replit already has users in essentially every country. China is harder, but not empty.

Peter Thiel’s paradox, as Masad retells it, is that the internet was supposed to be the great equalizer and instead concentrated wealth in Silicon Valley. Part of the reason is that the technology is less accessible than the mythology. He tells of a computer-science student in rural India with no laptop, only an Android phone, who learned on Replit’s mobile app and picked up work on the bounties board: people with time matched to people with money and no time. A U.S. founder building a recruiting app hired the student to fix bugs. The kid earned more than his entire family made in a year. That, Masad says, is the wealth machine opening for the first time.

Cahn adds the macro. Social and e-commerce got “techified” first because the barrier was low. Healthcare, education, and industrials are huge slices of GDP that Silicon Valley could not easily sell into. Software-engineer wages kept rising even as supply rose, which he reads as proof that the value they create sits above the wage. A billion developers, versus roughly 100 million GitHub accounts, means more of that value, including for the patient or citizen at the other end. Culture is a constraint too. At Facebook, Masad’s team shipped a vertical-scrolling photos experience after the iPhone; A/B tests tanked. The product was fine. Most desktop users were on netbooks with almost no vertical space. Silicon Valley was on MacBooks. Local builders, he thinks, will make less flattening software for their own communities.

Cahn calls mobile coding one of Replit’s poorly understood secrets. The open-source project that became the company started with a technical stunt: Masad was among the first to compile programming languages to JavaScript and run them in the browser, work that later became WebAssembly. It began as Mozilla’s Emscripten research project. He and a friend compiled CPython, contributed heavily, and had to build a Unix emulation layer in the browser. They posted it on Hacker News. It went viral. Brendan Eich, the inventor of JavaScript, tweeted about kids in Jordan building it.

Phones crashed, especially low-end Androids, because the client downloaded tens of megabytes of JavaScript. He even wanted Nokia Symbian. Client-side execution assumed machines that a lot of the world did not have. At Codecademy, with users in Africa and similar places, he flipped the architecture: the phone became a thin client, code ran on the server, and the JavaScript app had to be tiny with incremental loading. In 2011 that was hard; they invented web tricks that React later made ordinary. The Replit app is still under 100 megabytes, one of the smaller titles in the App Store, while many apps are gigabytes.

Why did IDEs stay local after Sheets and Figma moved to the cloud? Masad reaches for an Arabic saying about the shoemaker walking barefoot: people who make tools for others often neglect themselves. He cannot pinpoint a technical blocker that cannot be mitigated. A lot of it is culture, a feeling of control over the machine, and the old joke that programming progress happens one retirement at a time. The generation that learned on Replit and Codecademy, he says, has no problem coding in the cloud.

Did he foresee LLMs when he started? He had been wrangling compilers, transpilers, and parsers, work that felt like something machines should own. In 2012 he read “On the Naturalness of Software,” a paper he thinks is underrated, almost in the neighborhood of “Attention Is All You Need.” The authors argued that code is statistically close to natural language, built an n-gram completer, ranked it with IDE static methods, and beat classical completion. N-grams were the original language models; Google even trained a trillion-parameter one for simple translation. If that crude setup worked, he figured neural nets would do better. His 2015 seed deck, still online, copied Elon Musk’s master plan: a site for hobbyists and teachers, data on how people learn to code, then models stronger than traditional tools.

He tried machine learning every year. GPT-2 was the first glimpse. GPT-3, for his cohort, was the ChatGPT moment; ChatGPT itself was mostly a UI and some fine-tuning. Replit became the first startup outside Microsoft and OpenAI to train code models, briefly open-sourced 3-billion-parameter weights that were state of the art, and shipped Copilot-like features. In 2022 he posted a viral thread on software agents changing programming and the economy. A TED talk in October 2023 sketched test-time scaling: trade cost, speed, and accuracy, pour compute into a plan, keep the human as the creative engine. GPT-3 agents had been a context-window nightmare, summarizing every function into a sentence. AutoGPT in 2023 showed longer coherence and still did not work well. In March 2024 an engineer named Zen showed him a baby agent. He felt they were there, ran a May layoff to cut burn and drop side projects, and put the company on it.

The unlock was Claude 3.5 Sonnet in June 2024. The property people underplay, he says, is that it could emit up to 32,000 tokens of coherent code and oneshot almost a repository. GPT-4o-style agents wrote a function, opened a browser, wrote another function, the Cognition demo pattern, which was intractable and expensive. OpenAI’s models were lazy: they would insert a comment saying the rest of the code goes here. “Bro, I asked you to make that program, why are you asking me?” They set a ship date of September 5, his birthday. At least one teammate quit, arguing they were not ready, and was partly right. Early adopters glimpsed databases provisioned, migrations run, apps deployed. The mass market found a crappy product. They sprinted until December, left beta, and grew fast.

Agent V2, in A/B tests, is 50 to 500 percent better depending on the metric: about 50 percent better daily retention, conversion, and engagement. New users are twice as likely to deploy an app they made with the agent; power users five or six times. Deployment is the number Replit tracks because people who deploy are ten times more likely to retain. V1 now looks, to him, like something they have to replace. Under 3.5 they had built a complicated multi-agent system with a state machine, isolated manager, editor, and debugger memories, and a custom tool protocol because models hallucinated tools and arguments. Studies showed advertised million-token windows still collapsed on reasoning after about 32,000 tokens. When Anthropic’s computer-use model arrived, they inferred it had been fine-tuned for long-horizon tool loops, and they rewrote toward a simpler single-threaded agent.

Claude 3.7 taught the same lesson from the other side. Cursor’s integration went badly; users hated it until a fix landed, because 3.7 is more agentic and looks worse in a composer-style request-response box that will not let it loop. Over-optimize for present capabilities, Masad says, and you enter a local maximum that is hard to leave. The innovator’s dilemma used to be a story about decades. Kodak had a digital camera and did not launch it for fear of eating film. That cycle now happens on the order of months. Move fast and break things matters more than ever, including delivering crappy experiences on purpose.

The coding-tool landscape exploded in six months. He splits it into incremental and clean-slate. Cursor is a fantastic AI experience on VS Code, and many people at Replit use it, but it is by definition a layer on a decade of Microsoft work. Replit Agent started over: no coding, no DevOps, no IT, no database setup, no migrations. Version control, schema, and deployment config drifting out of lockstep is, in his view, one of the worst parts of building software. Read the Google Cloud or AWS postmortems and most outages are configuration errors, often an environment variable left behind after the software changed. Their agent runs a transactional lockstep system. Hit revert and you roll back code, database changes, and environment variables together.

In Cursor you still hook into APIs a backend engineer already built, or you architect the larger system yourself. Replit is trying to be the final tool you need to go from nothing to a deployed system. He reaches for sushi: everyone else is à la carte, Replit is omakase, the chef chooses. If you want every knob, do not come. Vercel helps with hosting and not the rest; Cursor helps with coding and not the rest. That is the bazaar again. For a billion people, most will never set up a deployment environment, so the product has to include one.

The “toy for students” narrative lasted a long time and, he admits, kept competitors asleep. Paul Graham saw past the mattress-startup version of the company and stayed supportive in hard moments. A Netflix engineer now uses it; a respected friend prototyped a new company on it while the room still called it a toy. They also made it premium on purpose. Cursor Composer is about four cents a request; Replit is about 25 cents, because an agent that iterates and calls tools is expensive. That alienated some users and pulled in more professionals. To reach a billion people they will have to go down-market again, probably with open-source models: Roadster first, then Model 3. He does not like “vibe coding.” It cheapens the possibility, the way “gen AI” did. Karpathy-style vibe coding assumes you are already a coder hitting Enter. On Replit Agent you start from an idea and do not even have the luxury of reading the code. His version: not vibe coding. Just vibe. Don’t code.

Silicon Valley’s popular AGI story is that software engineers disappear and everyone lives on universal basic income. Masad’s is the opposite: a billion more engineers, more powerful, working. Near term, he thinks AI gets really good at two things, work heavily represented in the data and work for which you can build a tight reinforcement-learning environment. Games had AlphaZero and self-play. Math, especially in Lean, is executable. Code execution is the same. GitHub already contains a lot of that distribution. Fundamentally new ideas, new knowledge, creativity, and understanding the world in a complicated, intractable way still look like a human domain. You might approximate some of it with a world model. He is not sure you can RL your way to an idea that changes how things work.

He names the destination functional AGI: systems that do many jobs humans do today because the training data exists and some of those jobs have ground truth, yet they are not general. Drop them in a super-novel environment with no explicit feedback and they will not learn efficiently. Labs define AGI as economically useful work in front of a computer, a remote worker. He thinks we will get there. His reaction is not unemployment. If he has a remote worker, he will create a hundred and implement all his ideas. If you are a code monkey, you are in trouble. If your place in the world is generating ideas and products because you understand what people want, he still thinks that is irreplaceable.

His first memory is his father opening a computer, assembling it, and hunt-and-peck typing DOS commands from a manual thick enough to hurt someone. It felt like a machine you could talk to. A DOS prompt is a REPL, read-eval-print loop, which is where the name Replit comes from. His father, a Palestinian refugee, poured savings into education and into that machine, attended night classes, and came home with notebooks. His mother was a free spirit who taught him Arabic poetry. He still describes himself as that mix: intuition and risk on one side, his father’s rigor on the other. One day his father found the computer in pieces. The boy put it back together, showed him how to actually open applications, and was told the machine was his. The math tutor for his brother was the first program.

As a teenager he was obsessed with Counter-Strike, spent scrap money in LAN cafes, won tournaments, and noticed the businesses ran on pen and paper. He wrote client-server software for accounts, time, and security so customers could not format the machines, sold it across Jordan, and made real money. At college his father pushed electrical engineering because the Jordan Engineers Association, where he was vice president, would not admit computer scientists as engineers. Masad read Paul Graham on Lisp, treated languages as aesthetic artifacts, and built Replit partly so he could try them all in a browser. After the Hacker News breakthrough, MOOC companies in the 2010-2011 wave, Udacity, Coursera, Codecademy, started using it. He took an O-1 visa on the back of that press, landed in New York in early 2011, and was forced at Amman airport to buy a return ticket because no one there had seen an O-1. Five or six hundred of his $700 went to the ticket. He arrived with about $100, on a $70,000 to $80,000 salary, with rent around $2,000.

Cahn asks about the acquisition. They were six people. Numbers thrown around were $500 million to $1 billion. Culture was already souring as they hired beyond the family core. At the same time his mother was diagnosed with cancer in Jordan, and half the team, the family, had to go home. The rational move was take the money and live happily ever after. He did not. He had wanted Silicon Valley since a low-budget movie, *Pirates of Silicon Valley*, and found suburbs instead of flying cars. Paul Graham, Marc Andreessen, and others he respected had invested and were excited. Selling felt like leaving the potential on the table and becoming another rich person doing angel checks, never again putting a life’s force into a company unless you are Elon Musk. They turned it down. Replit now has 40 million users, a valuation above what they would have sold for, and he still thinks the company is underpriced.

Masad describes his management as a deficiency, closer to Jensen Huang skipping traditional reviews than to MBA process, and closer to running an open-source project or a sports team. He keeps a duality of micromanagement and trust: dig into a button, then let people deliver and invent. From the start he kept a text file inside Replit with every name and the one thing he expected that person to ship. In weekly meetings they went around the table: did you do it, yes or no, why not, what next week. Executives still write that. Everyone in the company, every Friday, writes what they got done and what they plan next. He can walk the office and say what each person is on. Week after week of missed promises is, in his words, an obvious reason to let someone go. Attrition in the first few months runs 20 to 30 percent. People leave or are let go because they cannot keep the pace.

Hiring is for weirdos and misfits, a talent he traces to being one. Unused talent feels like waste. Mason arrived at 18 from a Santa Cruz-adjacent runaway story, a bootcamp that used Replit and sent bug reports. Masad asked for their best engineer and got an awkward kid. Co-founder and wife Haya complained that he slammed doors in her face. Masad’s defense was that the kid was thinking about code while walking. Most people reject that communication gap. He builds teams where peaks cover valleys: a cowboy slinging thousands of lines next to someone meticulous about tests. Sources are places others ignore, Replit hackathons, prizes, kids flown to San Francisco at 16 to 21, indie game hackers who never had a traditional job. If he started over he would look first in niche crypto and niche language communities.

He and Haya met in Jordan as employee one and two at a Belgian founder’s dysfunctional consultancy-that-would-become-a-product company, a pattern he says never becomes a product company. Dating was coffee and a 7 p.m. curfew. She designed the Replit logo. They married the week of his U.S. visa; he was 24. YC said get a co-founder, he went on co-founder dates, and she volunteered, then told him three or four months later that she had underestimated the pain. Sharing 12- and 14-hour days beat the Codecademy years when she could not relate to the hours. VCs disliked husband-and-wife founders; Paul Graham and Jessica Livingston were the famous exception, with Canva later. Kids add stress. The unique failure mode is that they go home and keep talking about company problems until the problems feel larger, so they set rules about when work is allowed and then break them.

In the lightning round he says scaling laws will hold, just on a changing set of things. Best advice came from Paul Graham, who asked whether this was his life’s work for ten or twenty years and then told him to stop worrying about daily tribulations: if you are not dying, you can fix it. Favorite new AI app is often something he spun up on Replit that day; he also flags Manus as a demo of models staying coherent for about an hour. A century-old Sears Home Services team, six months after a COBOL migration, with no ERP and no modern SaaS, leapfrogged a generation of enterprise software: operations people, not engineers, built a Replit agent that gives field workers an optimal route each morning. Favorite book is Douglas Hofstadter’s *I Am a Strange Loop*, even though he disagrees with the premise. Best foundation model for agents is Claude 3.7. Most underrated person in AI is Michele Catasta, Replit’s president and an early pioneer of language models for code. Cahn closes the arithmetic: 40 million users is 4 percent of a billion, 960 million to go.