Even Paul Graham, who trims 1,000 lines to 980, had to grieve: Replit's Amjad Masad says "don't ration tokens, find the person doing 100x"
Nov 12, 2026Original Y Combinator(폴 그레이엄 × 암자드 마사드) · TechCrunch · Platformer ↗
In 3 lines
- Replit built its agent after watching users copy code between ChatGPT and Replit by hand
- Masad says don't hand out equal token budgets; find the people who deliver 100x
- Graham: don't hold on to a palm tree on the beach, surf the wave
Y Combinator co-founder Paul Graham sat down with Amjad Masad, CEO of the coding-agent company Replit, in a 44-minute conversation YC published on Nov. 7, "How To Build The Future." They talked about how vibe coding came about, how companies should spend on AI tokens, and how to face a world that keeps changing under your feet. Graham was an early backer of Replit. We read the conversation alongside Masad's interviews with TechCrunch and Platformer.
The agent came from users copying and pasting
Masad said that even before vibe coding, Replit's autocomplete was already writing 50 to 60% of the code. The turning point was watching users he invited to the office. People who had never coded kept ChatGPT open in one window and Replit in another, pasted code across, and when an error came back they pasted the error into ChatGPT. Replit automated that loop with an agent that keeps going until the program works and launched it in September 2024; Masad said model companies reached out to ask how they had made it work.
The 2024 harness was 10 times more complex than today's
Models back then hallucinated line numbers and got lazy halfway through a function, leaving a note to finish it later. Masad said both OpenAI and Anthropic models had that bug, so Replit had to build its own edit tool. He said the 2024 harness was about 10 times more complex than today's: as models improve, the engineering needed to turn them into agents shrinks, and so does the barrier for competitors. His conclusion is that you have to accept that what you do keeps getting commoditized and always be doing the next thing.
A $100,000 outsourcing job becomes $1,000
The first customers who paid real money were people who would otherwise hire programmers on outsourcing sites. Masad said an app that would cost $100,000 from an engineer costs about $1,000 on Replit, a two-orders-of-magnitude gap that still holds. He said banks and manufacturers tell him projects scoped at $10 million were done for $250,000. These are customer anecdotes, not independently verified figures. Graham said he had dismissed outsourcing shops as terrible, when he should have seen them as a sign of demand.
"Don't ration tokens": one person can deliver 100x
Masad disagreed with giving every employee, say, $100 of tokens a month. What companies should care about is ROI, he said, because inside a company one person can drive 100x the next; the gap that used to be 10x in software is now 100x with agents. Graham replied that big companies' fundamental problem is that they can't tell the value of work done by individuals, so even a 10% improvement there would be huge. They also noted that people who bring Replit into their organizations often get promoted.
The "self-driving company," with Replit as customer zero
Masad said Replit no longer describes itself as vibe coding. It is a platform for the shift to the "self-driving company," where agents drive much of the work, and Replit is its own customer zero. At TechCrunch Disrupt in October he described a flow in which an agent first triages support tickets, a human reviews the case, and one click sends another agent to apply the fix. He also said "You need human judgment; you need taste" and "We're spending a lot more time figuring out what to work on."
Amjad Masad, CEO of Replit · TechCrunch ↗"That's object code": the grief Graham went through
Graham recalled that Masad seemed like a Bond villain when he demoed the agent in summer 2024, telling him: don't look at the code, that's object code. For someone who spends three years writing a thousand lines and is happy to get it down to 980 the next month, it meant his life's work was now object code. Masad said everyone goes through denial, then depression, then acceptance with AI, and that he had already grieved the craft he loved around 2020. Both placed the moment most people accepted it at the end of 2025, and Masad credited OpenClaw as the trigger.
Banks are still on 2023 ChatGPT, but the numbers move fast
Masad said Silicon Valley underestimates how much the rest of the world has yet to experience. Outside engineering, most banks use something like 2023-era ChatGPT, an LLM barely connected to anything. At a hackathon with a large telecom company, he said, 80% of participants were anxious about AI before using Replit for an hour and 80% were optimistic after. The company itself is growing quickly: per Platformer, Replit raised $400 million at a $9 billion valuation in March and says engineers' code contributions rose 5.8x in the first half of 2026. In the same interview Masad said "Net net, there's going to be more jobs," but companies will get smaller and do layoffs.
Amjad Masad, CEO of Replit · Platformer ↗"Don't hold on to a palm tree; surf the wave"
Asked what to teach children about AI, Masad pointed to flexibility, adaptability and lifelong learning. Graham said that if a tidal wave is approaching the beach, the safest place to be is surfing on it, not standing on the beach holding on to a palm tree. Masad said many people are rigid about "this is my identity, this is what I do" and recommended Graham's 2009 essay "Keep Your Identity Small," which argues that labels constrain what you'll let yourself believe.
Paul Graham · paulgraham.com ↗Living in the future: make the exponential your default
Graham's last question was why Masad always seems to be living in the future. Masad said he actively makes predictions and revises them, and that he trusts Ray Kurzweil's chart: plot the world's compute on a log scale and it's a straight line. Looking at GPT-2 in 2019, he said, he imagined chat, agents and personal assistants, and Replit's 2015-16 pitch deck already described training a model on code written on its platform. Graham summed it up: in a world where everything changes every six months, the person suited to it is the one whose default assumption is dramatic change.
