Decentralizing DevRel

🌈 Abstract

The article discusses the work of the DevRel (Developer Relations) team at Hugging Face, a community-centric company known for its open-source ML (Machine Learning) work. It covers topics such as the role of the Chief Llama Officer, the company's community-focused approach, metrics for measuring DevRel success, and the strategies for prioritizing and organizing the various open-source projects.

🙋 Q&A

[01] What does a Chief Llama Officer do at Hugging Face?

• The Chief Llama Officer's role is similar to a "Head of Platform and Community" at another company, with a broad scope.
• There are two main aspects to the role:
  • Fostering the thriving ML ecosystem and community through collaboration and support.
  • Overseeing the platform and community-related work at Hugging Face.

[02] What is under-appreciated about Hugging Face's open-source work?

• Hugging Face is highly community-centric, and the significant behind-the-scenes effort put into fostering the thriving ecosystem is often underappreciated.
• The company's collaborative approach, where they view themselves as the "Switzerland" of the ML community, actively contributing to and supporting the ecosystem, is a key aspect that is sometimes overlooked.

[03] What are some "real" metrics that Hugging Face's DevRel team tracks to measure success?

• The primary goals are to see increased usage of the Hugging Face Hub platform and open-source tools, rather than focusing on revenue.
• Key metrics include the number of repositories on the Hub (2.3 million currently, up from 780k a year ago) and logged-in usage of the Hub platform.
• The team also defines metrics specific to their areas of focus (e.g., Computer Vision, Audio ML, etc.) and prioritizes usage-based metrics over visibility-driven ones.
• However, the team is cautious about overemphasizing metrics, as they can be an imperfect proxy for impact and can be gamed.

[04] How is Hugging Face's DevRel approach different from the ZIRP DevRel article?

• Hugging Face's DevRel team is primarily an engineering function, sitting between the open-source and product organizations, with the goal of increasing usage of the Hub platform and open-source tools.
• This is different from DevRel teams that are part of a marketing or monetization function.
• Hugging Face has fostered a collaborative culture where people are excited to work together, both internally and externally, to amplify the community's work.

[05] How does Hugging Face organize and manage its large open-source work with a relatively small team?

• Key strategies include:
  • Prioritization using an exploration/exploitation framework, balancing new initiatives and maintaining existing projects.
  • Pragmatism in being willing to pause or stop projects that aren't having the expected impact.
  • Decentralized decision-making, where managers and individual contributors (ICs) collaborate on prioritization.

[06] How does Hugging Face decide what areas to invest in, such as the focus on audio ML?

• The decision to invest in audio ML was based on clear community and research validation, as well as growth potential, rather than revenue considerations.
• In contrast, the investment in ML for art (e.g., diffusion models) was more experimental, but showed early promising signs and ended up being a high-impact area.
• The team tries to balance exploration of new areas and exploitation of existing strengths, using a pragmatic approach to pause or stop projects that lack product-market fit.

[07] What are Hugging Face looking for help with, and what questions do they want answered to get to their "next level"?

• Hugging Face is looking for insights and suggestions on how to make their tools even more accessible to non-ML developers, as they've seen an increasing number of developers without an ML background wanting to incorporate ML into their projects.
• The team is also expanding and looking for individuals with strong developer empathy and technical skills based in the Bay Area.

[08] What other areas or tools does Hugging Face wish people would work on to be useful to the ecosystem?

• In research, Hugging Face is interested in areas like few-shot learning, unsupervised learning, and causal ML.
• More broadly, the team is interested in developer-friendly ML tooling that makes it easy for any developer to use ML, not just large language models (LLMs).
• They are also interested in people who can bridge the gap between a specific domain (e.g., biochemistry, chemistry, material sciences, health) and ML, and effectively communicate and work with both audiences.