Under 100 People, 24 Million Users: Replit's Bet on Artificial Developer Intelligence

Nathan Labenz, host of *The Cognitive Revolution*, already treats Replit as one of only 15 to 20 "live players" in AI worldwide. In the previous episode he spent ten minutes arguing that the cloud IDE could become one of the most important companies on earth, and that if he had to pick a single platform where a human–AI collaboration economy would take shape, he would pick Replit. This time he goes deeper with Michele Catasta, the company's new VP of AI, who had just left Google and published a manifesto aimed at the next billion software developers. The slogan in that document is Artificial Developer Intelligence, ADI, a deliberate play on AGI.

Catasta calls ADI a mid-term roadmap, not a project that should take decades. He split the name from AGI for a blunt reason: he does not want to go on record predicting when general intelligence arrives. Most people agree it will happen, he says; Replit is interested in what it can deliver to developers in the next few years, not in a twenty-year forecast. GitHub Copilot already proved that completion helps. Replit's Ghostwriter completes code and chats as a pair programmer. The manifesto is about taking that up a notch for people with no computer science background who land on the site and want to go from idea to prototype in the shortest possible time.

Today's models generate code that is usually aligned with what the user wants and periodically wrong. Experienced developers, Catasta included, know how to climb out of a GPT or PaLM bug. A person who has never written much code gets stuck on the first exception. ADI is a picture of how several iterations close that gap. An AI generates code; Replit's execution environment runs it; user feedback and runtime errors are folded back in; the model generates again until the program actually runs and matches the request. That loop, he says, is the company's ambitious goal, and he has a feeling that a couple of years might be enough to crack something powerful. In the manifesto he even unpacks REPL, the old read-eval-print loop, as reflect, evaluate, percolate, and learn.

Labenz is seeing the same wave from another angle. He advises Athena, an executive-assistant company, and has been teaching people who have never coded to manipulate applications on Replit. They skip for-loops and syntax. For the top-level skeleton they often bounce out to GPT-4, still "the boss model," then come back into Replit to iterate locally with the built-in tools. It is only 2023, he says, and people who did not expect to be asked to code in this job are already getting somewhere. Catasta is not surprised. Since the first transformer generators, the timeline has been more compressed than he expected.

Replit hosts monthly hackathons, most of them AI-themed, and a lot of the people who show up are not software engineers by trade. Product managers do especially well. Their day job is putting requirements in English, so in front of GPT-4 they describe exactly what they want, get an output, and let Replit stitch pieces together and debug. Ghostwriter's debugger can suggest a fix the moment an exception fires. Catasta has watched PMs win hackathons on their first experience writing code. His dream is that this path — prompt first, fill in the lower-level details later — becomes the default way people learn.

The obstacle is cost. AI is expensive to train and especially expensive to serve, and Replit had just passed 24 million users. Unit economics matter the moment you put a powerful model in front of that many people. Catasta still sees a future where at least some basic AI features are free, because scale keeps improving. He does not regret learning to code the old way, but he would have loved Replit as a beginner: more time on interesting problems, less time on unclosed parentheses. Andrej Karpathy has said he would waste half his productivity without Copilot. Catasta expects that to become consensus for anyone who writes software, experts included.

Catasta came from Google, where Replit now has a partnership that covers investment, cloud, and product. He was an internal champion, not the person who called the shots. He was already a Replit fan, had demoed PaLM — the coding work was called PaLM Coder — to CEO Amjad Masad and the team, then stayed in touch as an unofficial advisor, then a formal one, then joined full time. "It was a slow cook." He is glad the partnership closed after he left. Replit gives compute to millions of users; that is not something the company could have carried alone.

Labenz notes the trade-off: Catasta left perhaps the densest AI lab in the world for a company still under 100 people, the smallest name on his live-players list. Some things would be easier at Google by sheer resources. The other side is less process, faster iteration, and the freedom to reverse a decision instead of treating it as sunk cost. Catasta does not think he belongs in big tech long term. Replit's focus is narrow: the best possible AI for its developers. Incumbents changed the landscape with APIs; startups still have room if they build products an incumbent would not. He would not try to out-Google Docs Google. He would try a different way to edit, with generative AI as a first-class citizen. Replit, in his telling, is that kind of product.

Labenz offers a twist: in headcount Replit is tiny, but as a platform it is already an incumbent. The assistants he trains never feel the old pain of local environments, Python version hell, and settings sync. The containers, multiplayer, and history are already there; an AI layer can sit on top. Catasta agrees that this pre-AI advantage is why he joined, and he likes that nobody inside talks as if they have made it. The company still sounds like a startup that needs three years for this and five for that. Asked whether Replit identifies as an e/acc company, he says he is more of a Twitter spectator than a partisan. The new-hire page says the company is apolitical. Progress is the mission; labels are not.

On AI, he draws a line between political and responsible, and he is not a fan of politics bleeding into the field. There is no playbook yet. Social media has had fifteen-plus years and, he argues, still has not been regulated correctly; AI is newer and potentially more powerful. Replit also sits in a relatively lucky spot: it generates code, not text or images, so it dodges fake news, copyrighted stock photos, and face detection. Internally the discourse is how to make users aware of what the company does — hence an open-sourced LLM and an open dataset — not AGI. Transparency, he says, beats working in a vacuum. The political argument is one they do not have cycles for.

Replit's State of AI Development report had flagged the rise of LangChain, a library that by mid-2023 was catching public heat after a brief period of being fashionable. Catasta would not guide anyone away from it. At minimum it is a treasure trove: the team, led by Harrison Chase, often has an implementation of a new paper online within 24 hours. Frameworks go through a life cycle. Django and Ruby on Rails took the same beating when they told developers to work a certain way. Debugging a chain is harder than debugging your own code; that is a real cost, not a vibe. Give the project time for a next release, he says. Even if LangChain never dominates production, a one-day reproduction test is how the field finds out whether a paper's claims are real.

On productivity, Catasta is unusually numeric. The 2x threshold is already in hand from code completion. GitHub Copilot has published numbers of developers getting up to 55 percent faster; Replit sees comparable unpublished acceptance rates on Ghostwriter. Plenty of people now tweet that they would waste half their time without modern completion. Ten-x, he thinks, is attainable in about a year, not by generating larger blocks of code but by adding better debugging and more agent-like behavior. The model proposes an architecture, builds the scaffolding, writes a basic implementation; the developer adds glue and kills the remaining bugs.

A thousand-x developer is Amjad Masad's North Star, and Catasta says that one will take a while. The picture is not a better autocomplete. It is a model that can orchestrate tools and services that already exist, so the human becomes a tech lead with a bench of AI interns. Need a data pipeline? You should not have to write it if Replit already has a tool that fetches, processes, and bubbles up the result. That kind of independence is far from easy. It is also, he will argue next, not the same multiplier that amateurs will get.

Labenz feels cognitive dissonance. Add a billion developers *and* make them ten times more productive, and you have something like ten billion current-developer-equivalents — more than one active programmer's output per person on earth. Where is the demand? Catasta splits the curves. Horizontal growth is the next billion. That does not mean ninety percent of them will write the next LangChain or the next Google. It means programming becomes a basic skill, like Excel macros, iOS Shortcuts, or the variables now creeping into Google Docs. Tech-savvy people already want to automate tedious pieces of their lives. At Replit hackathons he has seen WhatsApp bots that order pizza for a group of friends. Those products will never have a business model. People will build them anyway, for themselves and for their friends.

The vertical multiplier, especially 1000x, does not apply to that amateur. It applies to the John Carmacks and Jeff Deans of the world: people who have run very large teams and might replace part of that org with agents and APIs. The goal, Catasta says, is not to displace developers. It is to make more people capable of building serious companies — fewer humongous firms, more mid-sized startups on hard problems. Labenz's own mantra at Athena is "copy and customize": fork a template, wire it to *this* messenger, *that* task tracker, *this* spreadsheet. Flo Crivello's analogy, from an earlier episode, is an aluminum can. In Napoleon's day only he could drink from aluminum; now we throw the cans away. Catasta says that is exactly the future Replit wants: single-use software spinning up all around us.

Large software will still need process, management, and a lot of human effort. What changes is the size of the pool of good ideas that get explored. Labenz presses on the implicit assumption: today's models already beat a random person at coding, they are closing in on expert tasks with a settled protocol, and they still do not supply the spark of a new scientific hypothesis or product concept. Catasta agrees. That is why the company named the bet ADI rather than AGI. The creative leap — designing the next LangChain — is still human intellect. If a system that smart arrives next year, it will remake every company, Replit included. Until then, the interim plan is to level out the basic skills and leave the creative process to people.

When students ask Catasta whether they should drop out of a computer science degree, his answer is no. Someone still has to design the next library and the next language. The last sixty-plus years of the field, as he tells it, have been a sequence of new abstraction layers: punch cards, then assembly, then C, then kernels, libraries, and web frameworks. Each shift came from humans working through anti-patterns and pain. An AI can learn to do things the way we write code today. It cannot think of how we will be writing code in two years. That remains human work, and it is more interesting if the tedious tasks go away.

He is not ruling out a sudden jump in which models start driving the creative process. He is refusing to plan the company around it. AI has a history of hype cycles and winters. In the early 1960s, he recalls, Marvin Minsky thought a summer intern could solve computer vision; it took about sixty years to get models that match that ambition. Putting all of Replit's chips on AGI in two years would be a bias, not a strategy. His probability mass sits on ADI. If the long-shot arrives first, the company is deep enough in the research community to hear about it — and the rest of the world will not be talking about anything else.

Labenz would rather not keep hyper-scaling training compute by another thousand-x, on the chance that something truly disruptive pops out. Catasta, who admits he has not studied the safety literature in depth, thinks the scaling will happen regardless. Letters from influential people have not slowed progress. Even startups now have the capital for large H100 clusters; he names Inflection. Talking about how to stop that, he says, is pointless. The useful work is getting ready so that nothing goes in the worst possible way. He is "macro optimistic" about mankind. Doomsday talk on Twitter and in mass media, including rhetoric about bombing data centers, strikes him as a poor way to start a discourse — especially for a public that has just come out of two-plus years of pandemic fear. Extremes, in his view, have never helped.

Labenz lists candidate moats: the existing infrastructure, including multiplayer; user feedback at inline frequency; community and bounties. On feedback he names two companies, Microsoft and Replit. Catasta cuts in. "We get even more than Microsoft." VS Code is a local, client-side experience; Replit runs in the cloud, so the backend sees everything. The two places with this kind of data, as he counts them, are Google's internal IDE and Replit — and Replit has the much larger user base. Microsoft does log Copilot interactions, which is how it publishes metrics. What it does not log, Catasta says, are keystrokes and a complete edit history. Doing that in VS Code would be a legal and PR nightmare. Replit has to keep that history anyway, for rollback and for reconciling multiplayer sessions. He is eager to train models on it.

The community terms are being written in public. An opt-out is coming in the docs. Replit trains on permissively licensed code and on public Repls only; private Repls are not touched, a reaction to what Catasta calls Copilot's original mistake. Fair use may be a legal gray area in the U.S., but he would rather be "correct towards our users." Longer term he wants attribution — Bing Chat already links to GitHub snippets that contributed to an answer — and even a revenue-share if a user's public code drives a lot of completions. He admits that last idea is close to interpretability-complete: common patterns appear everywhere, and it is unclear whom to pay. Until the lawsuits and the industry standard settle, the company will follow what it considers good practice and adapt.

Custom models are not, in his view, secret sauce in the weights. They are proof that the company can do the work, a recruiting signal, and a way to give back. Replit was built on open source; releasing a baseline trained on BigCode's The Stack was "almost a no-brainer." The production model is different: further pre-training on Replit's own data, which will not be released, with gains of up to 50 percent depending on language. The design constraint was latency. Most completions should return in 100 to 150 milliseconds, the threshold that feels instantaneous. That forced a roughly 3B-parameter model on the GPUs they had. Copilot often takes at least a second. Replit users, Catasta says, prefer shorter completions that arrive now.

The training run was a ten-day sprint by about two and a half people, after a couple of weeks of ablations, on MosaicML's cluster. Catasta's side owned data, smaller-model studies, a custom vocabulary, architecture, and a training schedule; Mosaic orchestrated the nodes and sat on call. They stared at Weights & Biases around the clock. The vocabulary is 32k, the size of the first Llama release, not the ~100k of GPT-3.5 and GPT-4. Compression is a bit worse — about three characters per token instead of four — in exchange for faster inference, and because the tokenizer was trained on code it actually compresses code better. The small model even looked competitive with Llama 7B on some self-contained reasoning benchmarks, a side effect of training on code. World knowledge still wants scale and retrieval. Catasta's point is that not every effort has to be a mega-model.

Labenz has been casually red-teaming code models. In May, Replit wrote him a denial-of-service script without hesitation. In August it refused. Catasta's explanation is unromantic. The team is too small for its own red team, so it leans on third-party providers. Ghostwriter Chat is Google and OpenAI models plus Replit's prompting and retrieval; the custom model is mostly used for completion. What changed between March and June GPT-3.5 releases, in his telling, is behavior under RLHF, not base skills. Llama 2 will be jailbroken too. "That's the nature of the beast." He asks users to be patient. Nobody in the community has cracked the problem, and Replit is unlikely to be first. The fact that a coding platform still rides commercial models for chat is, Labenz suggests, a reason for other companies not to assume the latest open-source weights are production-ready.

The cost picture is a moving target. AI feature usage is growing the way the State of AI Development report described: once Ghostwriter becomes a favorite pair programmer, people use it far more than expected. Third-party models are billed on traffic, not on a server staying up. A detailed unit-cost answer today, Catasta says, would be stale in two weeks. OpenAI has cut API prices by about two orders of magnitude in 18 months, so he expects the economics to keep improving. For now the job is to learn how to make AI useful for Replit users and "try not to bankrupt." He rejects the idea that paying customers are automatically fine. Expose GPT-4 to a power user and you can start losing money on that person in a few days — which is why even ChatGPT Pro limits GPT-4 messages. The hard problem is the 20 million-plus non-paying users. Completion, with a model they own and already know how to deploy, might go free first. GPT-4- or Gemini-class chat for the whole free tier will take a while, even if inference gets cheaper.

Edge inference is already on the whiteboard, either on-device or inside each Repl's limited container. The community has pointed the way: people wrap Replit's open 3B model in ggml, drop it into a VS Code extension, and get completion speed on a MacBook on par with the website. Near-term product teasers are more immediate. Ghostwriter is getting a new internal effort to live in more corners of the IDE, not only as a chat pane. Open-source models will keep coming on a semi-regular schedule; the first release pulled in community instruction-tuning data Replit did not expect. Zooming out, Labenz quotes Amjad Masad's tweet that Replit is "the perfect substrate for AGI." Catasta's reading is not sci-fi. A future AGI should behave like a good developer: it should know which services, APIs, and compute to call rather than generating everything from scratch. AutoGPT-style experiments have not become AGI, in his view, because GPT-4 cannot materialize tools and hardware. Replit is a place that can.

Asked which other company he would add to a short list of world-shaping AI labs, Catasta names Perplexity. It feels like early Replit in the velocity of new features, and it is attacking search, a field one player has dominated for twenty-plus years. He talks from time to time with CEO Aravind Srinivas "to get some inspiration." A VP of AI at a sub-100-person company, watching another small team climb, is a fair snapshot of the 2023 table: not many seats, and everyone is moving fast.