Amjad Masad: Replit demonstrated AI running to completion before it had a name

Replit's agent proved AI could work until the software actually ran

Amjad Masad tells Paul Graham that before Replit built its agent, users were manually copy-pasting code between ChatGPT and Replit, hitting errors, pasting the error back, and repeating the cycle. Watching this convinced him the entire loop could be automated, but he says it wasn't obvious at the time because "AI models would like do a couple of tool calls" and stop rather than looping until the program worked.

Masad says Replit demonstrated two things no one had shown before: that an LLM could be run to completion with the terminal condition being a working program, and that AI could provision its own infrastructure — spinning up databases and Redis instances via tool calls. He notes Andrej Karpathy called the July 2024 demo an "AGI moment," and that model companies reached out asking how Replit had pulled it off.

"The harness circa 24 is 10 times more complex than the harness today... as models get more capable, the engineering you have to do to make them become agents is less and less."

A decade of thinking about accessible programming set up the moment

Masad says he started thinking about easier programming around 2008, and that Replit's original 2015-16 pitch deck described training a model on millions of coders' data to predict code — before the Transformer paper. Graham suggests this long gestation, not sudden insight, is why Masad could extrapolate correctly when GPT-2 appeared in 2019, imagining chat, agents and personal assistants from it.

Masad credits Ray Kurzweil's compute-growth charts for shaping his thinking, noting the log-linear line of global compute "AI didn't bend... it's still a straight line." He argues most people treat the world as static, while his default assumption is exponential change, which he says is the real differentiator.

Replit's systems-programming depth became its enterprise moat

Masad describes Replit's infrastructure as an iceberg: a simple English-to-code surface sitting atop years of work on sandboxing, a custom transactional file system, and collaborative editing primitives built for human co-authors that turned out to plug directly into agent collaboration. He says this depth is winning enterprise customers who are "a lot more sensitive to reliability, security," unlike many competitors he describes as reckless about building unsecured apps.

He adds that early agents needed native IDE-style edit tools because they would hallucinate line numbers or leave functions unfinished — bugs he says both OpenAI and Anthropic shared, likely from chat-model training that discouraged long outputs. Today, he says, agents can just use a shell with basic commands.

Enterprises are swapping $10 million projects for $250,000 of Replit

Masad says Replit's growth inside large, non-tech companies — banks, manufacturers — comes from individual "scrappy" employees adopting the tool internally before it spreads. He says these companies tell Replit they had scoped a project at $10 million and instead spent $250,000, and argues the outsourcing industry is effectively being replaced by this dynamic.

"They've actually scoped this project out and it would have been $10 million, we spent $250,000 on Replit." — Amjad Masad

He also describes an internal "schlep" problem: making the product enterprise-ready requires giving IT admins controls, budgets and rules, which can slow or degrade the product for end users — a tradeoff he says is still being worked out.

Masad wants to measure ROI per employee, not ration tokens

Masad says he disagrees with the common corporate approach of giving every employee a flat token allowance, arguing companies should instead track return on investment, since some employees drive "100x" the value of others with the same tools. He frames this as solving a long-standing problem for large companies: the inability to measure individual contribution.

He says the near-term approach is to give customers what they ask for to gain adoption, then build better tools to surface high-value usage patterns — something he positions as more valuable than vibe coding itself and central to Replit's "self-driving company" vision, with Replit treating itself as "customer zero" for the idea.

Masad sees a widening gap between Silicon Valley and most companies

Masad says most banks are still using something close to ChatGPT-3 level tools outside engineering, barely connected to other systems, while Silicon Valley has already moved to autonomous agents. He says this two-year lag is significant because vibe coding itself did not exist two years ago.

He predicts usage will scale from one agent per person to 10, 100, and eventually 10,000 agents, which he says will push work away from local machines entirely. He also notes AI is finally displacing COBOL, something companies previously asked Replit to rewrite directly before Replit chose instead to teach customers to use the tool themselves.