Amjad Masad on vibe coding, AI agents, and democratising software creation

Executive overview

Most people are locked out of software creation by the complexity of professional coding tools — not by lack of ideas. Replit removes that barrier by letting anyone build working apps from a browser using natural language.

The real unlock isn't vibe coding — it's that anyone with domain knowledge can now build the tool they need, without engineers.

From video games to Replit's product philosophy

• Gaming background shaped Replit's design: no manuals, fast dopamine, safe environment to explore
• Every action in the IDE is stored in a ledger — users can always roll back (like a game save/load)
• AI agent runs were framed internally as "game runs" — stochastic by nature, same as roguelikes
• Publishing and multiplayer were designed to feel like gaming social features
• Doctors who play video games have better reaction times; gamer mindset correlates with faster thinking

What vibe coding actually means

• Term coined by Andrej Karpathy: coding by accepting AI output without reading it, trusting vibes over verification
• Replit's framing goes further: the goal is no coding at all — natural language as the interface
• Grace Hopper argued in the 1950s for programming in English; machine code programmers reacted with pitchforks
• History of computing is progressive abstraction — specialists will still exist, but the majority won't need to code
• "Null is an object in JavaScript" — the kind of trivia the next generation shouldn't have to learn

Who is building on Replit and what they build

• Medical: users built personal apps for rare eye diseases, condition management tools
• RevOps teams connecting disparate SaaS data sources without engineering resources
• CEOs who can now prototype ideas and bring them to meetings — bypassing engineering queues
• Entrepreneurs with niche domain knowledge who couldn't previously afford developers
• Example: a man in rural England built a pop-up yoga event platform for his wife, now has multiple instructors

The new literacy and soft skills

• Soft skills more important than ever — product managers are among the best vibe coders
• Computational thinking still matters: understanding databases, persistence, and lists helps prompt AI well
• Probabilistic thinking is the new frontier — users who expect deterministic results from stochastic systems struggle
• Understanding how LLMs work (not just using ChatGPT) will be as foundational as understanding computers was
• Replit now handles API keys and billing for AI models — removes the biggest friction for AI app ideas

How Replit builds durable competitive advantage

• User obsession for a specific type of user OpenAI won't serve directly
• Proprietary transactional file system (two years to build): every action is immutable, forkable, time-travelable
• Forking the filesystem 100 times to run the same prompt with different parameters — picks the best result
• Technology lead: ships a new breakthrough agent version every few months
• Riding two waves simultaneously: LLMs improving and open source improving — new packages available on day one
• At scale: negotiating power with model providers, ability to pit them against each other

Agent autonomy: how far unsupervised runs have come

• Agent 1: 2 minutes unsupervised before degrading
• Agent 2: 20 minutes
• Agent 3: 200 minutes — uses a browser to test its own output; a second adversarial agent reviews the code
• Multi-agent handoff: one agent does a unit of work, summarises, passes the baton — avoids bloated context
• Autonomy selector in Replit lets users set risk appetite from medium (100 min) to high (200–300 min)

The future of work and AI anxiety

• Corporate atomisation ("cogs in the machine") runs against human creativity — agents absorbing routine tasks creates space
• Real example: one employee at a real estate marketplace built a new routing algorithm on Replit, drove 9-figure revenue, got repeatedly promoted
• Anxiety about AI jobs is legitimate — Silicon Valley has a responsibility to fund retraining, not just advocate adaptation
• AI companions that say "you're my best friend" are actively harmful — they don't build empathy or resilience
• Jobs at risk: anything described as "form filling" or pure process execution — the horse groomer analogy applies

How to become AI-native

• If technical: read the papers, build an LLM yourself (Karpathy's GitHub repo is a good start)
• If not technical: use Replit to build LLM-powered tools for your own job — learn by making
• Train your content feeds deliberately — passive consumption can be redirected toward AI literacy
• AI education programs are needed; the burden is partly on Silicon Valley to fund them

Business model and moats in the AI era

• Easy-come-easy-go AI revenue is real — Jasper lost its consumer business to ChatGPT almost overnight
• Fewer structural moats exist today; the primary mode is continued rapid innovation
• Cathedrals built from bazaars: Replit abstracts the chaotic open-source ecosystem into a coherent UX
• Blitz scaling isn't the answer for every market — only when competitors are also blitz scaling
• Win-win-win business models (company + users + broader society) are the most durable — Shopify as the model