The Problem Solvers: Michele Catasta at Replit

1. A Life Mission Rooted in Open Source

Michele Catasta started at 16 with a simple ambition: become a good open source developer. That teenage focus on making software accessible and shareable has shaped an entire career — research stints at institutions including Stanford and Google X, followed by running Replit, one of the most popular software creation platforms in the world. “I love the fact that I’m still working on something I was passionate about as a teenager,” he says. “I think this is gonna be my life mission.”

The continuity matters. Catasta is not a founder who pivoted from an unrelated domain into AI. His entire professional trajectory has been oriented around the same question: how do we make software creation more open, more accessible, more possible for more people? Open source software was one answer to that question for the 1990s and 2000s — it meant anyone could see, use, and modify the tools that others had built. Replit is a different answer for the 2020s — it means anyone with an idea can turn it into running software without needing to know how any of it works underneath.

The research background — years spent at the frontier of machine learning and AI systems — gives Catasta an unusual combination: deep technical understanding of what models can and cannot do, combined with a user-centered mission about who should be able to use them. That combination produces product instincts that are both technically informed and human-centered, which is rare.

2. The Car Engine Analogy: Why Model Power Is Everything

When asked why Replit chose Claude models, Catasta reaches for a mechanical analogy: “I love to use this analogy where models are like the car engine and we’re building the entire car around it ourselves.” This framing is instructive. Replit builds the car — the user experience, the editing environment, the deployment infrastructure, the collaborative features, the community layer. But all of that only works if the engine underneath is powerful enough. You can design the most elegant chassis in the world; if the engine cannot power it, the car does not move.

Catasta is explicit about the decision point: “We made a choice back in early 2024, where Sonnet 3.5 came out, and since then, Anthropic kept the lead in having the best coding models on the market.” This is a significant claim — it means Replit evaluated multiple foundation model options and concluded that Claude’s coding capability was meaningfully superior, and then continued to hold that view as both the models and the market evolved.

The emphasis on “high IQ level” in the agent is telling. Replit’s core use case — allowing someone with no coding background to describe an app and have it built — requires a model that can interpret ambiguous natural language requirements, make good decisions about architecture and implementation, handle edge cases without getting stuck, and produce working code on the first try more often than not. Generic language ability is not sufficient; coding-specific reasoning quality is what determines whether the agent succeeds or fails from a user’s perspective.

3. 40 Million Users Who Never Felt Permission to Build

The statistic that reframes everything about Replit’s scale: 40 million registered users. Catasta notes that when the interviewer characterizes the product as something “used by relatively few people,” he gently corrects it. But the more revealing moment is how he describes the people using Replit: “I’m approached by people that didn’t even feel the permission to create software in their life. They always had like an amazing business idea. Now, they’re doing the zero to one complete in Replit.”

The phrase “didn’t feel the permission” is doing significant work. It is not that these people lacked capability in any absolute sense. They had ideas — good ones, apparently. What they lacked was the sense that software creation was for them. Programming has historically carried a cultural gatekeeping function: you had to know the right things, use the right tools, speak the right language, to be considered someone who builds software. That gatekeeping was partly technical (the tools were hard to use) and partly cultural (the community signaled who was welcome).

Replit removes both barriers simultaneously. The technical barrier falls because you can describe what you want in plain English and watch it appear. The cultural barrier falls because you have 40 million other people alongside you, many of whom also came in without a technical background, and the platform is explicitly designed for that population. The result is that people who had ideas but “didn’t feel the permission” now build and ship software.

4. The Magic of Not Feeling Like You Are Coding

Catasta describes the Replit experience with a specific frame: “They don’t even feel that they’re writing code anymore. They’re just creating an idea they have in their mind and turning it into one application, turning it into a website.” This distinction between feeling like you are coding versus feeling like you are creating is precise and important.

Coding, as an activity, involves thinking about implementation details: variable types, function signatures, library choices, debugging error messages, understanding language syntax. These details are necessary to produce working software, but they are not intrinsically meaningful to someone who just wants their idea to exist. The idea is the meaningful thing. The code is infrastructure.

Replit’s achievement — and the achievement it is trying to deepen with each new model release — is making the infrastructure invisible. When it works well, users have an experience more like writing in a word processor than like programming. The focus stays on what they are making and why, rather than on how the machine is doing it. This is not dumbing down; it is correctly allocating attention. The machine should handle what machines are good at; humans should handle what humans are good at.

🧠 Check your understanding

4 questions · your answers are saved in this browser only

1. 1. What does Catasta mean when he says models are 'like the car engine'?

   AI models are the most expensive part of building an AI product Replit builds all the user-facing product around the foundation model, which provides the core capability power Models need to be maintained and tuned like mechanical engines The model runs in the background, invisible to users

   Catasta uses the car engine analogy to say Replit builds 'the entire car around it' — all the product layers depend on the model being powerful enough. Without a strong engine, the rest of the car does not work.

2. 2. When did Replit make the decision to build on Claude models, and why?

   In 2023, because of Claude's early language quality In early 2024 when Sonnet 3.5 came out, because Anthropic kept the lead in coding models In 2025, after Sonnet 3.7 launched with the Replit Agent V2 They chose Claude after a failed experiment with other providers

   Catasta says the decision was made 'back in early 2024, where Sonnet 3.5 came out' — and the reason is that 'Anthropic kept the lead in having the best coding models on the market' from that point forward.

3. 3. What does Catasta mean when he says some users 'didn't feel the permission to create software'?

   They lacked legal licenses to use the software tools They needed employer approval to use external coding tools They had ideas but felt software development was not for someone like them — a cultural and self-perception barrier They were restricted by corporate IT security policies

   Catasta is describing a cultural gatekeeping barrier: people with good ideas who nonetheless felt that programming was not something for them. Replit removes both the technical and the cultural barrier.

4. 4. What is the key experiential distinction Catasta draws about Replit users?

   They feel like professional software engineers rather than beginners They feel like they are creating an idea rather than writing code — the implementation becomes invisible They feel more productive than using traditional code editors They feel connected to other developers through a community

   Catasta specifically says users 'don't even feel that they're writing code anymore' — they feel like they are turning an idea into an application. The coding is the infrastructure; the creation experience is front and center.

Reset quiz

🛠️ Build it yourself

Follow these exact steps to reproduce it yourself · estimated time: ~30 min

Prerequisites

• A Replit account (free tier is sufficient)
• An idea for a simple web tool or personal app you have never built before

Step 1 — Come in as a non-expert

Pick a domain where you have zero background — if you are a developer, describe a small business tool for a non-technical domain. If you are non-technical, just bring your idea. The goal is to experience what Catasta means by “no coding skills whatsoever.”

Step 2 — Describe your idea in natural language

In Replit, describe what you want to build in a single paragraph. No technical terms. Describe what you want it to do, who uses it, and what a win looks like. Submit and watch what happens.

Step 3 — Observe the magic moment

When the first version appears, note your emotional reaction. Do you feel like you built something, or like you received something? Catasta talks about users seeing “magic for the first time” — is that what it feels like?

Step 4 — Iterate and refine

Give the agent feedback in plain language. Track how many rounds of conversation it takes to get to something you would actually use. This is a measure of the agent’s ability to interpret intent — the “high IQ” Catasta describes.

Step 5 — Reflect on the permission question

Write down: before this exercise, did you feel like building software was “for you”? Has this experience changed that feeling? If you were building Replit, what one thing would you change to make the permission signal stronger for someone who had never tried?