“We believe in open and collaborative AI for the benefit of humanity.”
The Story
Hugging Face started as a chatbot company in 2016. By 2020, it had repositioned entirely around the open-source Transformers library — a unified API for interacting with hundreds of pretrained NLP models. The Model Hub followed: a public repository where anyone could publish, share, and download model weights. By 2024, the Hub hosted over 500,000 models and 100,000 datasets.
For the AI era, Hugging Face is what GitHub was for open-source software and npm was for JavaScript: the infrastructure layer that made it possible to build without owning a data center. The Transformers library's design — consistent interface across architectures, hardware-agnostic, well-documented — is the kind of tooling that shows up in retrospect as foundational and in the moment as a library you just reach for.
Why They're in the Hall
Hugging Face belongs in the Hall as the democratization force: the platform that moved AI from a resource only hyperscalers could access to something a developer with a laptop and a curiosity could experiment with on a Friday afternoon.
The museum's thesis — that failure patterns repeat because knowledge doesn't transfer — applies here in reverse: Hugging Face is a mechanism for knowledge transfer. The papers, the model cards, the documented failure modes in model repositories — this is the format the museum should aspire to at the code layer.
The 2024 Spaces Breach
In 2024, Hugging Face disclosed unauthorized access to its Spaces platform, which hosts user-deployed ML applications. The breach pattern was a direct instance of Transitive Trust (Law III): the platform trusted that uploaded model files and application code were inert data. Model pickle files are not inert data. They are arbitrary code execution, deferred to load time. The attack surface of a model hosting platform is not its authentication system — it is the format of the artifact it hosts.
You can democratize AI without democratizing the attack surface. The failure was assuming the two were separable.