Ben Firshman

Ben Firshman co-founded Replicate with Andreas Jansson and co-created Cog, an open-source tool for packaging machine learning models in Docker containers. Their work connects model builders with software developers who want to use AI without managing its underlying GPU infrastructure. Firshman previously co-created Fig, which became Docker Compose, with Aanand Prasad.[1][3][8][6]

Replicate became part of Cloudflare on December 1, 2025. In their joint announcement, Firshman and Jansson explained the decision through the changing needs of AI application builders: model inference had become one component alongside storage, databases, caching and microservices. They wanted to combine Replicate’s model infrastructure with Cloudflare’s broader application primitives. Edge inference, Workers-based model pipelines and WebRTC streaming were possibilities they proposed for that combination.[6]

The model infrastructure behind that decision grew out of earlier work on containers and scientific publishing. Firshman’s Docker experience supplied the packaging analogy; his collaboration with Jansson supplied practical problems in reading, reproducing and deploying machine learning research.[1][4]

Firshman and Aanand Prasad built Orchard Laboratories to simplify software deployment. After Docker appeared in early 2013, they discarded their earlier work and rebuilt the company around Linux containers. Their open-source tool Fig let developers define an application’s services and start them together in isolated development environments. Docker acquired Orchard on July 23, 2014, bringing Firshman, Orchard’s CEO, and Prasad into the company. Fig continued as a Docker project and later became Docker Compose; Orchard’s hosting service was discontinued.[7][1][8]

After leaving Docker, Firshman became interested in building tools for scientists. He recalled creating arXiv Vanity in 2017 with Jansson during a vacation in Greece, after Jansson struggled to read a research PDF on his phone. They used papers’ original LaTeX source to produce responsive HTML pages that were easier to read on small screens.[1][9]

Their project history describes an initial attempt with Pandoc that failed to handle TeX adequately, followed by a switch to LaTeXML. Mark Hurrell designed the paper layouts; Firshman and Jansson added CSS and JavaScript and wrapped the conversion process in Engrafo. They eventually retired Vanity because arXiv had added HTML papers. They participated in early integration efforts but credited Deyan Ginev and the LaTeXML/ar5iv team with completing the official HTML work.[9]

Firshman and Jansson started Replicate in 2019. Jansson’s experience at Spotify exposed a gap between a published machine learning paper and usable software: code could be incomplete, dependencies difficult to reconstruct and model weights unavailable. Firshman connected these problems to Docker’s portable containers. Their early proposal was a CI-like system in which researchers would package models and run shared benchmarks, allowing comparisons on the same evaluations.[6][1][4]

In his February 2024 interview, Firshman recalled trying unsuccessfully to sell related products before building them during Y Combinator. The pandemic interrupted the batch in March 2020. The pair then explored COVID tools, considered and rejected becoming a government application contracting business, and built an open-source experiment-tracking tool later renamed Keepsake. Firshman described himself as handling front-end and product work, with Jansson handling the back end and a designer named Mark contributing Replicate’s early designs.[1]

They returned to model sharing and built Cog and Replicate, dropping the shared benchmarking component because they considered it too complicated. The initial Replicate product offered web pages where people could try models and download their Docker containers. Firshman recalled that researchers often published a demo and then stopped using the service. The early-2021 community combining CLIP and GANs behaved differently: people continually tinkered, forked and shared models. Replicate began building for that community, including PixRay.[1]

The hosted API business emerged when a user reverse engineered the internal API behind Replicate’s web forms to generate images from a script. Firshman recalled that the team responded by documenting the API in Notion and offering paid access. He identified that first customer as PixRay’s creator and put the initial fee on the order of $1,000 a month. Replicate subsequently added an official API and billing. For Firshman, the lesson was practical: a downloadable container still required difficult GPU setup, while a hosted API made the model accessible to software engineers.[1]

Firshman said he and Jansson considered making models or a consumer image-generation product, but stayed with developer tooling because that was their expertise. He recalled Stable Diffusion’s 2022 release bringing both new model variants and developers ready to build products with them. In February 2024, he reported two million Replicate users, explicitly correcting the suggestion that all of them were developers.[1]

Firshman’s central aim for Cog was a shared interface for runnable models. In 2024, he described it as a Docker container with an OpenAPI specification attached as an image label, defining model inputs and outputs and using a CUDA device when needed. He favored a general-purpose interface using human-readable JSON over an abstraction at the tensor level, so the container could hold arbitrary software and accommodate future model implementations.[1][3]

The configuration format, cog.yaml, reflected what researchers already understood: Python versions, Python packages and necessary system packages. Firshman wanted users to avoid having to learn Dockerfile authoring and Linux administration simply to publish a model. Cog also absorbed recurring deployment work such as choosing compatible CUDA dependencies, connecting GPUs, defining schemas and serving inference. A May 2022 product introduction documented deployment either to Replicate or to users’ own infrastructure.[1][5]

Firshman regarded Cog as complementary to libraries such as Transformers and Diffusers, whose models could run inside its containers. He distinguished its server and NVIDIA GPU focus from tools aimed at local laptop inference. In February 2024, he advocated interoperability and fewer competing standards, while acknowledging that the boundaries between formats and systems were still being worked out.[1]

Firshman co-authored the Command Line Interface Guidelines with Aanand Prasad, Carl Tashian and Eva Parish. The guide combines human usability with script composability and recommends printing feedback within 100 milliseconds. Firshman described his own preference for software that feels physically immediate—like pulling a lever—and criticized Fig’s Python startup overhead. He also argued that a CLI can guide users through a conversation rather than merely succeed silently or reject an incomplete command.[8][1]

In February 2024, Firshman supported model creators receiving compensation when others made money using their work. At the same time, he argued that licenses should preserve the ability to customize and fine-tune models: restricting those activities could remove much of the value of making a model open. He favored sustainable API pricing that developers could afford, and described Replicate’s cloud GPU strategy as aggregating customer demand into larger provider commitments while allowing individual users to make short inference requests.[1]

In his February 2024 interview, Firshman advised software engineers to experiment with language models, diffusion models, prompting and fine-tuning without assuming they first needed to master low-level framework internals. He compared this to building websites without understanding every layer of a network stack. His appeal to fellow tool builders was that accessibility remained a major constraint: “The limiting factor is just like making that accessible to people who build products, because it's really hard to use this stuff right now.”[1]