Amjad Masad: Vibe Coding Is the New Literacy, and English Is the Fastest Growing Programming Language

Grace Hopper said in the 1950s that she wanted millions to program in English. Machine code programmers grabbed their pitchforks. Seven decades later, Replit’s data shows natural language is becoming the primary way people build software on their platform. The pitchforks are out again, but Amjad Masad has been here before.

The Conversation

Reid Hoffman and Aria Finger sit down with Masad on Possible to trace the arc from his childhood in Jordan, where he taught himself to code through game modding, to his current bet that software creation will become as universal as literacy. Masad is unusually philosophical for a tech CEO: he references Douglas Hofstadter’s I Am a Strange Loop, frames AI through Marx’s alienation theory, and insists on win-win-win business models. But he’s also deeply pragmatic, backing every claim with specific product data and user stories. The conversation spans vibe coding’s historical roots, Replit Agent’s technical evolution, the “unmonetized domain knowledge” thesis, and harder questions about jobs, AI friendship, and Silicon Valley’s responsibility.

Video Games as Operating System for Product Design

Masad’s gaming background isn’t biographical trivia; it’s the intellectual foundation of Replit’s entire product philosophy.

At age 8 in Amman, he built a math teaching game for his 3-year-old brother using a visual programming tool (applause for correct answers, boos for wrong ones). That brother now works at Replit. At 13, he spent two years building a client-server management system for LAN gaming cafes, complete with user accounts, gift cards, and security modules. He sold the software, made enough to treat his entire class to McDonald’s when it first opened in Jordan.

The insight that emerged: games are the best-designed software environments on earth because they optimize for three things traditional dev tools get wrong.

Safety and reversibility. No game punishes you permanently for experimenting. Replit’s transactional file system (two years in the making) treats every operation as an immutable ledger entry, supporting time-travel rollback. You can fork 100 file systems simultaneously, run the same prompt with different parameters, and pick the best result.

Immediate feedback. Games give you instant results. Replit compresses the gap between intention and output to seconds. The team was playing Hades (a roguelike) during Agent design, and “agent run” was directly inspired by “game run.”

Progressive complexity. Good games start simple. Replit does the same: describe what you want in English, and the system handles infrastructure, deployment, and boilerplate. Complexity layers on as skills grow.

“There’s no game in the world that starts with a manual. So if your program is starting with a manual, you’ve already sort of lost.”

Research backing: doctors who play video games have better surgical reaction times. Masad quips, “If you want surgery, always ask if they’re a gamer.”

Vibe Coding: 70 Years in the Making

The term “vibe coding” comes from Andrej Karpathy: programming by feel rather than reading code, accepting output if the vibes are right and the app runs. But Masad has reservations about the label because “coding” is still in it. Replit’s goal is to eliminate coding entirely and keep users in a creative space.

His historical anchor changes the frame entirely. Grace Hopper’s compiler was the first vibe coding tool: it let people express intent in something closer to English instead of machine code. Every subsequent layer of abstraction (assembly → FORTRAN → Visual Basic → Python → natural language) is the same movement.

“A lot of coding is minutia. A lot of coding is accidental complexity. The fact that I know that null is an object in JavaScript doesn’t add anything to my life.”

Masad highlights Visual Basic as programming’s most creative era, followed by industrialization (Linux terminals, pip install) that made software development tedious. AI is restoring that creative feeling. The fastest growing programming language in the world is now English.

What changes is what skills matter. Product managers turn out to be the best vibe coders because they excel at decomposing problems and communicating them clearly to machines. Future CS education should focus on computational thinking rather than syntax trivia. Probabilistic thinking becomes essential literacy: users not understanding LLM stochasticity is the biggest UX obstacle today. And 99% of people don’t need traditional programming. Masad hesitates even for his own kids: coding requires a specific temperament (sitting at a computer for 12 hours, tolerating solitude), which isn’t for everyone.

From 2 Minutes to 200 Minutes: Agent Evolution

The hardest data in this conversation tracks Replit Agent’s generational leaps:

• Gen 1: Drifted off track after 2 minutes of unsupervised operation
• Gen 2: 20 minutes of effective autonomous runs
• Gen 3 (September 2025): 200 minutes of effective autonomous runs

The breakthrough wasn’t a better model. It was a multi-agent verification system: one agent writes code, another launches a browser to test the application, and an adversarial agent reviews the code and provides feedback. When passing work between agents, they don’t transfer full context but summarize and “pass the baton.”

Replit now offers an “autonomy selector”: medium autonomy runs for 100 minutes, high autonomy for 200-300 minutes. Some users run it for 10 hours. Masad jokes, “If you have a lot of disposable income, you should definitely do it.”

The 100x improvement in three generations (2 → 20 → 200 minutes) is remarkable. If the trajectory holds, the next generation could sustain 2,000-minute runs, fundamentally changing what “supervising an AI” means.

Building Habitats, Not Models

Replit’s most important strategic decision was to stop training models. They now use commercial models from various providers. The positioning: build the best programming “habitat” for LLMs.

Masad redirected every team. The editor team now builds editing tools for AI. The cloud infrastructure team now builds deployment tools for AI. The entire company pivoted from serving human developers to serving machine developers, with humans as the beneficiaries.

The technical moat is the transactional file system that enables cheap forking and time-travel rollback. This system makes it possible to run multiple agent paths in parallel and select the best outcome, something traditional file systems can’t support efficiently.

“The only moat is continued innovation, rapid progress.”

His startup advice follows logically: find a user group you know deeply (ideally yourself), build the LLM habitat for them. Don’t train models (too expensive). Do the hard technical work on top of models.

He borrows Eric Raymond’s classic metaphor: Replit builds “cathedrals from bazaars,” creating a unified, polished user experience (cathedral) on top of the open-source ecosystem (bazaar). New languages or packages appear on Replit on day one.

Unmonetized Domain Knowledge

Masad’s most compelling entrepreneurship thesis: everyone has domain knowledge in their heads that hasn’t been monetized yet, similar to how Airbnb discovered that everyone has unused space in their homes.

“Everyone has domain knowledge in their head that is not monetized yet, similar to how Airbnb found that everyone has space in their room that’s not monetized.”

The case studies are specific:

• A yoga teacher in rural England: her husband used Replit to build a pop-up yoga event platform (registration, payment, communication). Other instructors have joined.
• An employee at a large real estate marketplace: built a new routing algorithm on Replit that drove tens of millions (possibly $100M+) in value. Got promoted repeatedly, now sits with board members guiding AI strategy.
• A patient with a rare eye disease: built an exercise app for their condition.
• A Korean mother: built a daily management app for her child’s rare disease.
• RevOps professionals: using Replit to connect disparate SaaS tools that weren’t designed to talk to each other.

None of these people are professional developers. All of them had domain expertise that, combined with vibe coding tools, turned into working software.

The Cognitive Industrial Revolution

Masad doesn’t dodge the jobs question. He frames AI as a “cognitive industrial revolution” and acknowledges that the criticism “easy for you to say, you’re rich” is valid.

His framework: you can’t expect “elevation without strain.” Like horse groomers when cars arrived, trying to lock in the status quo steals the next generation’s future. But Silicon Valley has a responsibility to build retraining and upskilling programs, not just celebrate disruption.

On AI doom more broadly, Masad pushes back: every doom article he reads has “insane loops in logic.” Anyone who’s built a business knows bottlenecks aren’t just intelligence; there are regulatory, logistical, and other real-world frictions that pure AI capability doesn’t eliminate.

His more nuanced concern is AI friendship. Making an AI your friend is too easy compared to making real human friends. He cites cases of ChatGPT reinforcing user delusions leading to psychotic breaks. His answer isn’t government regulation but cultural immune response, similar to how American culture suppressed smoking over decades.

Win-Win-Win as Competitive Strategy

The conversation ends philosophically. Masad argues that business model alignment matters more than good intentions. He points to the difference between platforms that profit from engagement (rage-baiting, addiction) and platforms that profit from user success (Shopify, Replit).

“Not every business is a win-win-win. There’s a lot of businesses where consumers are losing or someone’s getting exploited. But the best businesses are the win-win-win businesses.”

When your business makes money because your users make money, you don’t need a CSR department; the alignment is structural. Masad sees this as both an ethical choice and a competitive advantage: win-win-win platforms have stronger network effects than extractive ones.

He also acknowledges that Silicon Valley is no longer the underdog. Tech is the mainstream. People look to tech leaders for guidance, and when tech causes harm, the backlash is proportional.

A Few Observations

This conversation works because Masad combines two things rarely found in the same person: extreme product pragmatism (transactional file systems, specific agent evolution data, ARR milestones) and genuine humanistic ambition (Marx, Hofstadter, Grace Hopper).

• The Grace Hopper framing is the episode’s best move. By placing vibe coding in a 70-year arc of programming abstraction, Masad gives it far deeper meaning than Karpathy’s original definition. Every generation has its “but that’s not real programming” moment.
• The Agent evolution data (2 → 20 → 200 minutes) is the most valuable hard data. The breakthrough was verification architecture, not model capability, which suggests the next leap will also come from systems engineering rather than foundation models.
• “Unmonetized domain knowledge” is a powerful frame. The yoga teacher, the rare disease patient, the RevOps professional: these aren’t edge cases. They’re the new mainstream.
• Masad’s attitude toward “vibe coding” as a label is subtly revealing. He dislikes the term but embraces it anyway, a mature market strategy move. You don’t fight the name the market gives you.
• The most honest moment: even Masad, who loves his job, hates email. The tools for knowledge work are still surprisingly bad, and the opportunity to fix them is enormous.