The Scene
Replit has had a turbulent year, but CEO Amjad Masad’s sonorous voice was almost zen-like as he spoke to me on Monday in an airy conference room, sipping coconut water with a view of the sun setting over Foster City, California.
The AI coding company had moved its headquarters out of San Francisco in April, went through layoffs in May, and has seen its headcount cut in half, to about 65 people.
Yet it has grown its revenue five-fold over the past six months, Masad said, thanks to a breakthrough in artificial-intelligence capabilities that enabled a new product called “Agent,” a tool that can write a working software application with nothing but a natural language prompt.
“It was a huge hit,” Masad said. “We launched it in September, and it’s basically the first at-scale working software agent you can try in the world today. And it’s the only one, I would say.”
Replit, which Masad co-founded in 2016, has embraced AI since the beginning, and in recent years it has launched products that automate various aspects of the coding process.
But if you had listened to Masad in recent years, Agent shouldn’t be possible yet. He said at one point it might not be possible this decade. Even as he set up an “agent task force” to develop the product last year, he wasn’t sure if it would work. What changed was a new model from Anthropic, Claude 3.5 Sonnet, which achieved a record score on a coding evaluation called SWE-bench in October.
Replit had been building its own models and had been hoping that its proprietary data — which includes every aspect of the coding process, from conception to deployment — might give it an advantage. Suddenly, that was no longer the case.
“I knew all this stuff was coming. I just didn’t think it was going to come this fast,” he said.
That acceleration has implications not just for Replit, but for every industry. Writing code was the first thing that so-called generative AI models like OpenAI’s GPT could do well, and they offer a glimpse into what other sectors of the economy will look like as AI capabilities increase.
And the massive improvement is a double-edged sword for Replit. Agent is a runaway success. At the same time, Replit has dropped the idea of developing a proprietary model — and the Anthropic one that made it possible is available to competing startups, of which there are a growing number.
“Just the fact that we’re able to get here without using our data poses a lot of questions for the industry,“ Masad said. “As long as we keep the rate of innovation and the rate of progress, and we keep deepening that, I think we can continue to be ahead. But the business question is, ‘what is the durable moat?’”
In this article:
Know More
Replit’s big advantage today isn’t AI capability; it’s software that makes every other aspect of creating software easier. Replit has been honing that technical foundation for eight years. The new Anthropic model that Replit Agent is based on just removes the final headache: Actually having to write the software. Masad said the company still plans to use its data to fine-tune larger foundation models.
Customers could, in theory, use Claude directly to create software, but then they’d have to handle everything else that goes along with it. “What you’d have to do is pay for Claude, go to AWS to start an EC2 machine, go into that, install Git and Python. Already, most people are just gone at this point,” he said.
In essence, Replit’s latest customer base is a new breed of coder: The ones who don’t know the first thing about code.
“We don’t care about professional coders anymore,” Masad said.
Instead, he says it’s time for non-coders to begin learning how to use AI tools to build software themselves. He is credited with a concept known as “Amjad’s Law” that says the return on learning some code doubles every six months.
Essentially, a very basic understanding of the way software works enables people to take advantage of AI tools that are growing ever more capable.
Masad says it really amounts to a return to the roots of the way computers were supposed to work. Before Windows, people would type in complex, arcane prompts into MS DOS. But only a small percentage of the population was going to adopt a technology that required learning an entirely new language.
DOS gave way to Windows, and mouse-clicking replaced prompts. AI allows computers to be prompted with natural language, instead of arcane commands. “I think we’re going to see this era of Windows and mice and desktops as totally cringe.”
Reed’s view
After meeting with Masad in Foster City, I went home, put the kids to bed, and then proceeded to stay up most of the night trying to build apps that could make my life more efficient.
Let’s just say it’s still not easy to build good software. At the same time, what I was able to build in one night blew my mind. I came away thinking that I could probably do just about anything with enough time.
Steve Jobs said the computer is “like a bicycle for our minds.” Until this AI moment, our bikes had been confined to specific lanes determined by the tiny percentage of people who write the software and manage restrictive walled gardens.
Masad said something that is going to stick with me for a while: “Finding software problems in your life is also a skill, looking at a problem and saying, ‘a piece of software could solve that’ is a skill.”
Since I can remember, that is a skill that only software developers could capitalize on. Now, pretty much anyone can.
Creating custom software from scratch is going to become a normal part of daily life for most professionals. People who are good at it will have an advantage, in the same way people who are naturally organized and efficient have an advantage today.
Room for Disagreement
A study by Uplevel found that the productivity of developers hasn’t improved because of AI coding assistants, and their code has become more buggy.
“Developers could indeed be seeing positive results, given that a report from Copilot’s early days showed nearly 30% of new code involved AI assistance — a number that has likely grown. However, another possibility behind the increased usage is coders developing a dependency and turning lazy.”
Notable
- The buzziest AI coding company is probably Cognition, founded by former “sport coders” who aim to build a virtual coder named Devin that can take on a coding project from start to finish without any human intervention.