It’s been a turbulent year for Replit, but CEO Amjad Massad spoke to me Monday in an airy conference room over a glass of coconut water as the sun set over Foster City, California. It was like Zen.
The AI coding company moved its headquarters from San Francisco in April and underwent layoffs in May, cutting its headcount in half to about 65 people.
But Massad said the company’s five-fold increase in revenue in the past six months was thanks to breakthrough advances in artificial intelligence capabilities that enabled a new product called Agent. Natural language prompts.
“It was a big hit,” Massad said. “We launched this in September, and it’s basically the first software agent that works at scale that you can try in the world today. And I would say it’s the only one. Masu.”
Replit, which Massad co-founded in 2016, has embraced AI from the beginning, and in recent years has launched products that automate various aspects of the coding process.
But if you’ve listened to Massad in recent years, Agent should still be impossible. He said at one point that this decade might not be possible. Even when they formed an “Agent Task Force” to develop the product last year, they weren’t sure it would work. That’s changed with Anthropic’s new model, the Claude 3.5 Sonnet, which achieved a record score in October on a coding assessment called SWE Bench.
Replit was building its own model and was hoping that its proprietary data, which included all aspects of the coding process from ideation to deployment, would give it an advantage. Suddenly, that wasn’t the case anymore.
“I knew something like this was going to happen. I didn’t think it would happen this quickly,” he said.
This acceleration will impact every industry, not just Replit. Writing code is the first thing that makes so-called generative AI models, like OpenAI’s GPT, work well, and provides a glimpse of what other sectors of the economy might look like as AI capabilities improve. Masu.
And this huge improvement is a double-edged sword for Replit. The agent is a huge success. At the same time, Replit dropped the idea of developing its own model. And the Anthropic model that made it possible is now available to a growing number of competing startups.
“The mere fact that we were able to get here without using data raises a lot of questions for the industry,” Massad said. “As long as we maintain the rate of innovation and the rate of progress and continue to deepen it, I think we’ll continue to stay ahead of the curve. But the business question is, ‘Durable What is a moat?