In August 2022, Ben Firshman woke with a splitting headache. He stumbled through his San Francisco apartment wracked by achiness and chills and grabbed a COVID test. That’s when he noticed something strange on his computer screen. It was a massive spike in traffic to the site he and fellow software engineer Andreas Jansson had launched in 2021, called Replicate. 

Replicate is a platform that hosts open-source machine learning models that anyone can run in the cloud with a single line of code. While the company had a small subset of dedicated users since its founding, it had struggled to gain meaningful traction—until now. 

Sitting at his computer, waiting for the results of his COVID test, Firshman studied the uptick in visitors. Stable Diffusion, the most advanced text-to-image model to date, had just been released. Thousands of developers wanted to use it, but most were not AI experts able to navigate the complexity and cost of running the model on their own. And so they were turning to Replicate, which had packaged, optimized and deployed Stable Diffusion, making the model available for anyone—AI expert or not—to use. 

Terrified that the onslaught of API calls would crash the site, Firshman phoned Jansson. They agreed they needed to work fast to ensure Replicate could handle the surge of developers coming to the platform to build and collaborate on AI-based tools and products. This was why Firshman and Jansson had started Replicate in the first place: to make machine learning accessible to software engineers.

Just then, Firshman’s timer went off. He checked his test: positive. He wouldn’t be leaving the house, but he would be working day and night to capitalize on this moment. Firshman ordered some Paxlovid and dove in. 

In the 1950s, if you wanted to use a computer you most likely needed a Ph.D. in electrical engineering—that’s how complicated they were. But over time, software was introduced that made using computers easy, no matter a person’s background or experience level. 

For the greater part of the 2000s and 2010s, machine learning’s useability was akin to mid-century computers, despite its growing ubiquity. Machine learning models were bulky, complex, on-prem systems, inaccessible to most engineers. To run them required expertise in the field and resources only available to larger institutions that could afford expensive servers and GPUs.   

Jansson, who had been a machine learning engineer at Spotify prior to founding Replicate, understood firsthand the barriers developers faced in using AI models. “I could read all about new developments in AI, but there were no tools readily available to actually apply any of it,” Jansson says. When he encountered an issue at Spotify, he would delve into relevant research, which would be in the form of scientific papers—not usable code. “Experts would write software, turn it into prose for the sake of academia, and then engineers like Andreas would have to turn their research back into software in order to use it,” Firshman says. “The whole process was bananas.”

“I could read all about new developments in AI, but there were no tools readily available to actually apply any of it.”

Andreas Jansson

Once the prose had been translated back into code, there came another issue: deployment. There was no guarantee that the software would run properly when it was transferred between computer systems. Researchers and engineers would have to work together for weeks to ensure other engineers could run the models on their own computers. 

Jansson and Firshman didn’t understand why AI models couldn’t just be distributed as software ready for production. This question connected directly to the work Firshman was doing at the time at Docker. Docker allows developers to package their code into containers (think shipping containers), which ensures that software runs the way it is supposed to on any computer. 

Firshman and Jansson wanted to build such a container, but for AI. “If only we could define a standard box for machine learning models, then researchers could put their models inside it,” said Firshman. “They could be shared with other people. They could be deployed to production. They could run anywhere, and always keep on running.” 

In October 2019, Firshman and Jansson began working on Replicate. As Firshman explains, “The heart of Replicate is taking all of these incredible new advances in AI and making them accessible to software engineers who don’t understand those things, but know how to build great products.” 

Replicate would be a platform where machine learning researchers could host and share AI models, and where engineers of all backgrounds could use those models to build applications. 

While Firshman and Jansson believed in Replicate’s potential, they knew they were taking a calculated risk. “Generative AI is everywhere now, but back then, it was not a term at all,” says Stephanie Zhan, a partner at Sequoia. “Ben and Andreas’ bet was that there would be a flourishing open-source model ecosystem, and they could become an important enabler of it by providing the easiest-to-use tools and infrastructure for developers to use these models.” 

To confirm Replicate’s viability, Firshman and Jansson spent weeks talking with machine learning experts and software developers, assessing their needs and experiences and comparing them with their own. Open-source machine learning was a nascent field and Replicate was a gamble on its future. Ultimately, Firshman and Jansson determined it was a gamble they were ready to take. 

“Generative AI is everywhere now, but back then, it was not a term at all.”

Stephanie Zhan

Firshman and Jansson first met in London in 2012 while working as developers for a music platform called This Is My Jam. On a team of just four people, Firshman and Jansson quickly formed a friendship, bonding over favorite bands and a shared attitude towards their work. “In a field where it’s easy to be cynical, we both had a childish enthusiasm for programming,” Jansson says. “We were both excited about building things.”  

Firshman’s interest in building came from his father, who he describes as a “nerd who was always tinkering with things.” Firshman’s mother was a classical guitar teacher, and during the hours she was with students, Firshman and his father would build and take apart anything they could get their hands on, from model airplanes to the family’s car. Firshman recalled being five years old and attempting to turn a landline into a wireless phone using a soldering iron. His father encouraged him to explore how the world worked, and Firshman seized every opportunity to do so.

As a teenager, Firshman applied his curiosity to software. He became involved with an online developer community who attached cell phones and cameras to weather balloons and sent them to the edge of space, where the cameras would take pictures and the phones would text their location. Firshman helped write the code for these explorations, collaborating and learning from other engineers. 

Meanwhile, deep in the forest outside Habo, Sweden, twelve-year-old Jansson was also learning to code. Armed with only a book about BASIC and an old Amiga 500, he began to recreate the games his friends had on their more advanced computers, like Brick Breaker. He was proud when he finally got the physics of the bouncing ball to work properly, and soon started building games of his own. 

When he was 15, Jansson left home to pursue his other passion, music, enrolling at a high school in Rättvik specializing in Swedish folk music. Following graduation he completed his military service, worked a stint as a forklift driver, and then moved to England to study sound engineering and, later, computer science.  “Maybe there’s a way to combine music and engineering,” Jansson thought. He was right. Soon after graduating, Jansson started working at This Is My Jam and met Firshman.

In 2014, the two friends parted ways. Jansson accepted a job as a software engineer at Spotify and moved to New York. Firshman began working at Docker, but over time began to question his purpose. “I felt like I always stumbled into the next opportunity,” Firshman says. “I really wanted to figure out what I wanted to do with my life and what was important to me.” 

In 2017, Firshman quit his job and decided to travel. He cycled across India and drove a van through China. He ate homemade pasta for breakfast at a stranger’s house in Italy. Eventually, he made his way to Greece, where he met up with his old friend, Jansson, on a small island called Naxos.

One afternoon, sitting by a pool, Firshman watched as Jansson struggled to read a research paper about AI on his phone. The paper, only available as a PDF, wasn’t designed to be read on a smartphone, so Jansson had to zoom in on every sentence. Firshman understood his exasperation. In the early days of their friendship, they’d talked about how the internet had originally been designed to share academic work. But decades after the internet’s founding, so much scientific research remained needlessly difficult to consume, like these mobile-unfriendly PDFs. Lifting his legs out of the water, Firshman proposed they work together to find a solution. Over the next few days, Firshman and Jansson built a program that converted research papers to web pages. Using the program, papers could be scaled appropriately to phone screens, and the hyperlinks, which previously would have to be copied and pasted into a new browser, were clickable. Now anyone anywhere, and on any platform, could easily read and digest a paper. 

Although Firshman didn’t know it at the time, he’d found what was important to him: he had a passion for making the inaccessible accessible, and he loved working with his friend. Whatever project came next, he wanted it to involve Jansson. 

Back in his apartment in the summer of 2022, the positive COVID test beside him, Firshman realized that for Replicate to handle the incoming surge of traffic, he, Jansson and their two other employees didn’t just need to patch the site, “we needed to rewrite our entire infrastructure,” he says. “It wasn’t working like we needed it to. It wasn’t up to par.” 

Firshman recalls the next few days as “24/7, all hands on deck.” One of their users informed them that an application built using Replicate was going to be featured on Japanese national television. “The show’s airdate was our deadline,” Firshman says. “And we managed to meet it. We rebuilt the whole Replicate infrastructure in a couple of weeks.” 

“The show’s airdate was our deadline, and we managed to meet it. We rebuilt the whole Replicate infrastructure in a couple of weeks.”

Ben Firshman

Their work paid off. Following the launch of Stable Diffusion, software engineers and AI enthusiasts across industries flocked to Replicate to create new products. One popular app allowed users to upload pictures of their homes, describe an interior decorating style, and watch as the program generated images of their homes in that style. Another converted blurry, old photos into crisp, high-definition images. More than growing Replicate’s user base, Stable Diffusion’s release brought attention and legitimacy to the idea of open-source AI models. And it underscored the value of Replicate, a platform where they could live and be deployed. 

In July 2023, Meta and Microsoft released Llama 2, a large language model whose addition to Replicate’s library resulted in the platform’s biggest week of growth to date. For Replicate, this represented yet another vote of confidence—this time from two of the world’s most significant tech companies—in open-source AI models. 

If Replicate’s initial challenge was getting people to buy into open-source AI models and believe they could be easy to use, today its biggest challenge is competitors crowding into the space. “Developers care about cost, performance, and the developer experience of running these models,” Zhan says. “The ecosystem of models is flourishing and ever-evolving, so pushing the envelope on these dimensions is a continuous journey.”

Cost and performance aside, according to Zhan, what sets Replicate apart from the host of other AI inference platforms is empathy. “They understand what developers actually want,” Zhan says. “They didn’t just build something that works, they built something that is delightful—a few lines of code to run any model in the cloud directly via their API.” Replicate’s website features a neon background gradation and large, stylized image tiles for each model in its library. Detailed instructions, replete with colorful visual aids, guide users through how to run, customize and deploy models. It’s a platform that front to back inspires the imagination.

Scaling a startup is almost impossibly complex, but similar to how AI models learn, Firshman and Jansson have gotten better at it over time. “At first, we tackled every responsibility together,” Jansson says. They soon realized this was a quick path to burning out and decided to delineate duties. Firshman focuses more on products and the business side of the platform, while Jansson keeps up with developments in machine learning. A decade into their friendship, the two maintain the highest admiration for each other. Jansson says he got an MBA by osmosis just by working with Firshman, who understands business and people “in a magical way.” Firshman describes Jansson as a “ridiculous, multi-talented person.” They’re certain that at the core of what makes Replicate successful are their friendship, respect and shared ethos. 

Although Firshman lives in San Francisco and Jansson in rural Sweden, they make it a point to spend time together. Firshman visits Jansson about four times a year, and the two take walks through the forest, discussing the future of Replicate and AI in general. They continue to share the youthful passion about software they had when they first met, and they still can’t help but build together. When Jansson was getting married in July 2022, he constructed a dance floor in the woods near his home for Swedish folk dancing. A week before the wedding, he realized it needed a roof. Firshman, who had flown out early, spent days with Jansson building the roof for the dance floor. The last plank went in at dawn, hours before Jansson’s wedding. That night, the two celebrated with friends and family, dancing beneath their creation.