No Priors Ep. 33 | With Replit's CEO & Co-Founder Amjad Masad

Replit’s develop-to-deploy platform and new AI tool, Ghostwriter, are breaking down the barriers to entry for beginner programmers. Replit’s CEO, co-founder, and head engineer Amjad Masad joins hosts Sarah Guo and Elad Gil to discuss how AI can change software engineering, the infrastructure we still need, open source foundation models, and what to expect from AI agents. Before co-founding Replit, Amjad Masad worked at Facebook as a software engineer, where he worked on infrastructure tooling. He was a founding engineer at CodeAcademy. Throughout his career, Masad has been an advocate for open-source software. 00:00 Introduction 03:55 - Impact of AI on Code Generation 11:09 - Breaking Down Barriers to Entry in Development with Replit 14:35 - The Impact of Open Source Models, Meta/Llama 20:32 - Bounties, Agents who Make Money 24:26 - The Missing Data Spec-to-Code 32:29 - Building the Future of AI, Money as a Programmable Primitive

Sarah GuohostAmjad MasadguestElad Gilhost

Sep 20, 202329mWatch on YouTube ↗

WHAT IT’S REALLY ABOUT

Replit’s Amjad Masad on AI Coders, Bounties, and Open-Source Futures

Amjad Masad, CEO and co‑founder of Replit, explains how Replit evolved from a simple in‑browser coding prototype into a full-stack development platform with 22 million users and a deeply integrated AI suite called Ghostwriter. He argues that AI coding tools are still primitive and outlines a near-term path to agentic systems that can autonomously build features, manage files, and interact with infrastructure. Masad emphasizes Replit’s advantage as an end‑to‑end platform that captures rich feedback loops from coding through deployment, which can be used to train better models. He also discusses open‑source LLMs, Meta’s role with LLaMA, Replit’s bounty marketplace, and the future of money and payments for both human and AI “developer agents.”

IDEAS WORTH REMEMBERING

5 ideas

AI-assisted coding is already boosting productivity but remains early-stage.

Tools like Ghostwriter and Copilot can deliver roughly 30–50% productivity gains in coding tasks, yet they still touch only a small slice of the overall software lifecycle such as debugging, deployment, and spec translation.

The next leap is agentic AI that can act inside dev environments.

Masad sees near-term (6–18 month) potential for agents that can read/write files, install packages, access the internet, and autonomously implement features like login flows, with multiple models orchestrated via new infrastructure.

Training code models on execution and full workflows can improve semantics.

Current models mostly learn from static code; feeding them execution results, editor interactions, debugging data, and deployment feedback could give them deeper semantic understanding and better performance on real-world tasks.

End‑to‑end platforms create stronger data moats than raw code corpora.

Masad argues Replit’s true edge is not just code data but the integrated journey from first line of code to deployment and production crashes, enabling richer reinforcement and more realistic training signals than a simple editor plugin.

Model training decisions should be driven by product needs, not prestige.

Replit trained a 3B-parameter code model mainly to deliver fast, cheap autocomplete when commercial APIs weren’t suitable, and Masad criticizes open-source one‑upmanship on benchmarks divorced from real user value.

WORDS WORTH SAVING

5 quotes

“I always thought that you can probably apply machine learning to [code tools]… and nothing ever really worked all that well up until GPT‑2… obviously GPT‑3 was like, ‘Okay, it’s here.’”

— Amjad Masad

“AI hasn’t touched yet… everything that we do just around the act of coding itself, and I think it will touch every part of the software development lifecycle.”

— Amjad Masad

“My feeling is this is within reach, even in the current capabilities… we’re within months of actually having really cool basic agentic experiences.”

— Amjad Masad

“The real advantage is the end‑to‑end journey from the first line of code, to the deployment, to getting crashes in production… and the real magic of Replit is putting all of that together.”

— Amjad Masad

“Can you create an AI that can go earn money for you while you sleep? And I think that that would be a really cool, cool idea.”

— Amjad Masad

Origin and evolution of Replit as an in‑browser coding platformDevelopment and capabilities of Ghostwriter and AI-assisted codingFuture of agentic AI in software development and productivity gainsReplit’s data advantage via end‑to‑end developer workflow integrationTradeoffs between training proprietary models vs. using external APIsRole of Meta’s LLaMA and the open‑source AI ecosystemReplit bounties, global developer labor, and programmable money in software

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.