How Hugging Face Transformed a $4.5B AI Powerhouse [Pivot Case Study]

iOS Chatbot → GitHub for ML

In the fast-paced world of AI startups, Hugging Face stands out as a remarkable success story of pivoting and finding product-market fit. Founded in 2016 as an AI-powered chatbot for teenagers, the company has transformed into a $4.5 billion open-source powerhouse that's reshaping the machine learning landscape.

This case study explores Hugging Face's journey from its initial premise to its game-changing pivot. We'll examine how the founders recognized the need for change, validated their new direction, and ultimately built a thriving ecosystem that serves developers and researchers worldwide.

Hugging Face's evolution is a testament to the power of adaptability in tech entrepreneurship. By shifting focus from consumer-facing chatbots to developer tools and open-source libraries, the company not only survived but thrived, becoming a cornerstone of the AI community.

1. Original premise?

What problem were they solving? For who? What was the intended solution?

In 2016, Clement Delangue, Julien Chamound, and Thomas Wolf launched an AI-powered chatbot for teenagers called Hugging Face. Their vision for the product was a digital friend that was entertaining enough for people to have fun talking to it. The co-founders chose to build in the space of open-domain conversation AI—in simple words, AI that could understand any conversation—because they perceived it to be the most difficult problem they had the technical expertise to tackle.

In the process of training their chatbot, the Hugging Face team built a foundational library that included various machine-learning models, including a model capable of detecting emotions behind a text message, and another that could generate coherent responses. The library also included many datasets to help the chatbot understand a wide-range of conversational topics, from sports to high-school gossip. The Hugging Face team released free pieces of this library as an open-source project on GitHub. This commitment to knowledge-sharing was perhaps an early sign of what was to come!

2. Warning Signs

What made them think they needed to pivot from their original premise?

The Hugging Face chatbot was an early success with around 100,000 daily active users at its peak and modest retention numbers. However, the team, which was deeply passionate about NLP technology, became increasingly frustrated. They found that significant technological advancements, such as enhancing the accuracy of the bot's responses, didn't translate into corresponding improvements in user growth or retention metrics.

3. Time to Pivot

How long after they started the company did they think about pivoting?

Hugging Face pivoted around 2 years after it was launched.

4. Initial Funding / Team

Had they raised money before their pivot? How much? How big was the team?

Hugging Face raised an angel round of $1.2 million from Betaworks, SV Angel and NBA star Kevin Durant and others in 2017. The objective was to fund applications of AI for entertainment purposes. 

5. New Opportunity

What were the key insights that drove their pivot?

2017 was a banner year for NLP. Researchers from Google and the University of Toronto authored the now infamous paper called Attention is All You Need which conceptualized the architecture of “transformers” which, to put it plainly, enabled machines to understand a word in the context of the words around it. 

Even though this was a huge breakthrough, it remained inaccessible to the average developer. Hugging Face, which had already open sourced parts of its library on GitHub, developed an open-source transformers library to solve this problem.

6. Pivot Validation

How did they validate those insights?

The watershed moment for Hugging Face’s pivot had to do with the release of Google’s language model BERT in October, 2018. BERT was too complex for most users and was only available on Google’s own TensorFlow platform. Within a week, the Hugging Face team had created a simplified version of BERT on the machine learning framework PyTorch—and open-sourced it for anyone to use or modify further. This made Hugging Face wildly popular among the machine-learning community. Looking back on this Delangue said that they released open-source versions simply as a benefit to the community “without thinking about it too much” and the team was pleasantly surprised by the reaction they received.

7. Length of Pivot

How long did it take to get clear signal the pivot was working?

The machine learning community’s enthusiastic reaction to Hugging Face’s custom version of BERT reaffirmed the team’s commitment to knowledge sharing. The mission of the company fundamentally changed and the co founders decided to share everything they had learned about in the process of building their chatbot — and soon enough, Hugging Face became a favorite for developers in the machine learning space. 

Fun fact: While Hugging Face officially adopted a mission focused on knowledge sharing in 2018, this ethos was already deeply rooted in one of its co-founders, Delangue. When he was at university, Delangue had developed an open-source tool called Unishared for college students to exchange knowledge with each other. 

8. Pivot Funding

Did they raise money during or after their pivot?

Yes. Hugging Face raised a $4 million seed round in May, 2018 led by Ronny Conway, with participation from pre-seed investors. The company went on to raise four more rounds—the latest being a $235 million Series D  in August of 2023 at a valuation of $4.5 billion.

9. Pivot Outcomes

What happened after the pivot? Did they ever get to product market fit? Where is the company today?

Apart from being one of the most-liked players in the AI scene, Hugging Face is a powerhouse in the machine learning ecosystem. The company has strategically positioned itself at the crossroads of open-source idealism and practical business solutions, creating a value proposition that appeals to a wide audience. The company’s 2023 stats included hosting 500,000 models, 250,000 data sets, and 250,000 apps; numbers that it aimed to triple in 2024. 

In early 2024, Hugging Face put out feelers for robotic engineers, sparking discussion about the company’s first shift into the paradigm of hardware.

10. Lessons Learned

What are 2-3 key takeaways from their experience?

1. Experimentation trumps long-term strategies. Looking back on the evolution of Hugging Face, Delangue remarked that something he learned while building the company was to experiment and respond to user feedback, instead of drawing up mighty business plans that stretched far out into the future.
2. Build a technical moat around your business. Delangue reminds startup founders that if they want to build a strong technical moat around their business they should build with AI, as opposed to simply using it. He thinks of an AI-native startup as one which is building and training their own models.

• Share
• Tweet