Kyle Chung
← Writing
EN繁中

How a Teen Chatbot App Became Hugging Face, the GitHub of AI

5 min readBy
aientrepreneurship

Most people using Hugging Face today have no idea it started as a chatbot for teenagers.

Not a metaphor, not an early prototype that got scrapped in week one — an actual consumer app, with actual users, that ran for two years before the company became one of the most important pieces of infrastructure in AI.

The Chatbot Nobody Remembers

Hugging Face was founded in 2016 in New York City by three French engineers: Clément Delangue (CEO), Julien Chaumond (CTO), and Thomas Wolf (Chief Science Officer). The name comes from the 🤗 emoji — a deliberate choice, because the product was supposed to feel warm and approachable.

That product was an AI companion app aimed at teenagers: something to talk to, powered by in-house NLP models. It found a real audience — over a million users at its peak — which is more traction than most startups ever see. By any normal startup scorecard, it was working.

It just wasn't the thing that mattered.

The three founders had met through an online engineering course, not a shared thesis on the future of AI infrastructure. None of them set out to build developer tooling. They set out to build something teenagers would enjoy talking to, and the NLP stack underneath was just the plumbing that made that possible — until the plumbing turned out to be more interesting than the app it was built for.

The Week Everything Changed

In late 2018, Google released BERT, a new approach to language models that outperformed nearly everything before it. Instead of just using it, the Hugging Face team ported BERT from TensorFlow to PyTorch and open-sourced the implementation — reportedly within about a week of the paper landing.

That single decision did more for the company than two years of chatbot iteration.

Researchers and engineers who wanted BERT in PyTorch had exactly one clean place to get it. The library kept growing to cover more models, and the open-source community started treating it as the way to use modern NLP. The founders were watching a side project outgrow the product it was built to support — and they made the call to follow the traction, not the plan.

That side project became transformers. The chatbot was retired. The infrastructure became the company — formalized after the team went through an accelerator in 2019 and committed to open-source ML tooling as the actual business, not a side project that happened to support one.

It's worth sitting with how counterintuitive that call was at the time. The chatbot had users, a name, momentum. The PyTorch port had no business model, no pricing page, nothing an investor could point to as "traction" in the conventional sense — just a library that kept getting cloned, starred, and built on by people the founders had never met. Betting the company on the thing with no revenue attached, over the thing with a million users, is exactly the decision most founders talk themselves out of.

From Side Library to the Hub Everyone Uses

Once the pivot was formalized, growth compounded fast:

  • Series A (2019): $15M
  • Series B (2021): $40M
  • Series C: $100M
  • Series D (2023): $235M, at a reported $4.5B valuation

By 2025, the platform hosted over 2 million public models, 500,000+ public datasets, and more than 13 million users — spanning the Hub, the transformers library, Spaces (hosted model demos), and inference infrastructure.

None of that came from a single flagship product the way a chatbot company scales. It came from the same pattern repeating: ship a piece of infrastructure other people need, let them build on top of it in the open, and the surface area grows without Hugging Face having to build every layer itself. transformers proved the pattern once. The Hub, Datasets, and Spaces are the same bet, made three more times.

Why "the GitHub of AI" Undersells It

The comparison gets made constantly, and it's directionally right: both are git-based, both use the same username/repo namespace pattern, both are where a technical community stores and shares its work in the open. But GitHub hosts code; Hugging Face hosts artifacts — trained model weights that run into tens of gigabytes, datasets, and running demos — which is a different kind of asset with different problems. Plain git chokes on a multi-gigabyte binary file; Hugging Face had to build its own large-file storage and versioning on top of git to make model weights behave like normal repos at all. The nickname captures the cultural role. It undersells the actual infrastructure problem the company ended up solving.

The Lesson for Builders

The interesting part of this story isn't the exit multiple. It's the discipline behind the decision.

The founders had a working consumer product with real users — the thing every early-stage founder is told to protect and iterate on. What they noticed instead was that the infrastructure underneath the product was generating more organic pull than the product itself. That's an easy signal to miss, because it doesn't look like failure. It looks like a side effect.

The pivot wasn't abandoning a bad idea. It was noticing that the thing they'd built to support their idea was the actual idea — the same instinct behind knowing when to build infrastructure instead of buying it.

The Short Version

A chatbot for teenagers, a fast PyTorch port of BERT, and the willingness to kill the product that got them users — that's how Hugging Face became the place open-source AI actually lives. The tool underneath the tool turned out to be the business.

Sources

Share this postXLinkedInThreads