The Llama Trail
From Big Tech to Open Source Startup, the first year.
Disclaimer: This post is personal and is not an official Hugging Face post.
Today marks my one-year anniversary working at Hugging Face...and what an adventure it has been 🤗! If you don’t know what Hugging Face is, apart from an emoji, it’s an Open Source company that has created some of the most popular ML Libraries such as 🤗 transformers and 🤗 datasets, as well as the Hub, a central platform through which anyone can share their ML models, datasets and demos.
Background (From Google to an Emoji Company)
Deciding to join Hugging Face was a very challenging decision. At the time, I had been working at Google for two years and I was doing well. I had a somewhat fast promotion, a clear growth track ahead of me, and a team of skilled people from which I was learning every day. I shipped real products being used by many people, was contributing to DEI efforts, and got 14 bonuses (between spot and peer bonuses). In a short span of time, I got to interact with amazing people, including unique opportunities with Jeff Dean, Geoffrey Hinton, Paige Bailey, and many others. So leaving Google was both challenging and scary.
So what went wrong?
Going in-depth on that would make for a much longer post. In summary, I was not fulfilled. Before working at Google, I was engaged with ML Latinamerican communities giving open, free workshops in Spanish. My partner and I co-founded one of the first ML communities in Mexico, AI Learners, in which we organized free events. And after joining my position at Google I didn’t have time to keep going with many of my community-related passions. It’s a weird ( and also privileged) position to be in. Having an extremely good compensation, nice team, growth opportunities, and more, but still, be frustrated/unfulfilled and let you search for other opportunities. This is a feeling that returns every night and returns for more time than I should have allowed.
Back then I started to explore opportunities within Google that would get me closer to working with communities, but I didn’t explore things outside the big G. My partner suggested I check out a cool startup that is doing very cool work in the NLP space called Hugging Face. To be honest, I hadn’t heard much about it, in big part due to the bubble effect of working in a huge company using mostly internal tools. I still recall when I told other teammates I was leaving to HF, nobody knew about it, and usually, they asked about why it was called like that 🤗.
After talking with some awesome people from Hugging Face (Clem, Thom, Lysandre, and Sylvain), originally doing a take-home challenge for a TensorFlow Engineer role 😅, and proposing a two-pager with strategy ideas and opportunity areas, I signed a contract to join HF as Machine Learning Engineer.
Funnily enough, many of the ideas came to life, many of these were already work in progress and amazing people in the team had them in mind.
NLP Course: Unknown to me, Lysandre and Sylvain were already doing some cool work in https://huggingface.co/course/chapter1/1
Demos: Funnily enough, with Spaces and Gradio this is extremely common nowadays https://huggingface.co/spaces. The demos now extend the documentation of different Open Source libraries which was the original goal I had in mind.
Distill.pb article...well...this didn’t age well.
University Outreach: Violette and many members at Hugging Face are now extremely involved in education initiatives, including a demo.cratization tour giving talks to universities.
Blog: Mishig has done an excellent work improving the blog. Now you can filter by different tags, and discover content in specific modalities. https://huggingface.co/blog
Discord: We launched a Discord HF Community Server in October. Nowadays we have 8000 members 🤯
I think the email thread had more emojis than words…
My first 6 months
After taking a couple of weeks on vacation which I took as an opportunity to refresh in transformers and my German, my first day quickly came. The first month was quite active. I worked on integrating two Open Source libraries, Sentence Transformers and spaCy, into the Hub. I also collaborated with an awesome group of people in creating the first part of the Hugging Face Course and slightly contributed to a community event training language models over the internet in a collaborative fashion.
Something I really like about Hugging Face is that the opportunities are there for you to take. If you’re passionate about something and it has impact potential, you have the support to do it. There are many many things one can do and lots of freedom for each individual to decide how to focus their day today, so it’s important to prioritize based on the impact projects can have. Here are some of the things I worked on in my first 6 months that I enjoyed quite a bit:
Co-organized the Flax sprint event in which free TPUs were provided to the community to train their own models and then build demos (such as Chef Transformer, DALLE mini, and CLIP RSCID).
Collaborated with many OS libraries for integrations: Adapter Transformers, TensorFlowTTS, Asteroid, Keras, AllenNLP, ESPnet, SpeechBrain, Stanza, and more.
Built some fun demos, such as using Google Sheets with a TAPAS model and creating a chrome extension for question generation.
Launched Spaces with the fantastic infra and product teams. Spaces has allowed thousands of people to craft and share amazing ML apps and demos to showcase their work and build their ML portfolio.
Launched a Hugging Face Community Discord server with the amazing DA team. The server is used by the community while participating during events; people also join to discuss NLP, Computer Vision, Reinforcement Learning, and Llamas, from time to time. There are over 8000 members now!
Contributed to the Somos NLP community “NLP for Spanish” course
Gave 10 workshops about NLP, Transfer Learning, building ML demos with Python, and contributing to Open Source.
In September of last year, we ended up forming a Developer Advocacy Engineering (Avocado) team with the fantastic Nate, our first Developer Advocate Engineering 🥑 Merve, and me.
The team has grown since then!
I’m super happy and proud of the team we’ve built. A year ago there were 6000 public, open-source models on Hugging Face. Today we have over 42000. Now there are models for Audio, Computer Vision, Reinforcement Learning and other domains! The Hub went from 1K datasets shared by the community to over 4K. And we launched Spaces which is already changing the landscape of how people share ML models and research: by now there are over 3000 public open-source demos.
And so many cool things going on! Here is a non-exhaustive list of what the team has accomplished or contributed.
We’re organizing a free course about Deep Reinforcement Learning. This course blends theory with practical things. All units have a practical part that shows Open Source and free tools to do RL yourself!
Keras models can now be pushed to the Hugging Face Hub with a single line of code, getting automated model cards and many other things. The community even reproduced the official Keras examples and created demos for them (check them here).
The Somos NLP community organized the largest hackathon in Spanish, with hundreds of participants, and dozens of demos, models and datasets for very interesting problems.
We launched Task Pages, an open-source initiative with use cases, tutorials, curated models and datasets, and more, for common ML tasks.
This is really not a comprehensive list of things that have happened, there are dozens of things going on (really), with the team organizing events, contributing technically, to Open Source libraries, contributing to the product roadmap, and being involved in educational initiatives, and more 🔥.
People
The best thing about Hugging Face is the people around it. Both internally and externally, people are incredibly awesome. You get to work with skilled and extremely nice people with lots of positive energy. The HF community is one of the nicest I’ve seen, with an attitude of encouraging each other and incredible teamwork.
And we get to collaborate with amazing people all the time. We got the opportunity to get breakfast with Sebastian Ramirez (from FastAPI) and his partner, get lunch with Ines Montani and Matthew Honnibal (from the amazing Explosion/spaCy), and travel to Paris, New York, and London where I met some really amazing people! I’m super lucky to work with a team of great individuals passionate about building the future of Machine Learning with a collaborative and community-driven mindset.
What’s next?
I’m super excited to share that we’ve raised a Series C and we’re becoming a Llama (🦙== 🦄🦄)!
We want to democratize Machine Learning. This means creating tools, content, and features that help people share their work and collaborate in their ML workflows. This means that we want not just ML Engineers, but everyone, to be able to learn and understand ML. We’ll keep building collaborations with Open Source tools, organizations, and groups.
We have a transparent (to the point I can share roadmaps and tweet metrics all the time) and collaborative culture. We don’t have a competitor-driven mindset. We aim to craft things really useful for the community and maximize the impact of our work. Sometimes this means we do lots of explorations in order to find what has the most impact. If you are a person that would thrive in an environment, I would encourage you to apply to Hugging Face. We are hiring for all kinds of roles you could imagine, Senior Developer Advocates, Python Engineers ready to contribute to Gradio, ML Engineers for Speech Recognition and for Computer Vision, and more. If you don’t see a role of interest at hf.co/jobs, you can apply to Wild Card. If you’re a right fit, no matter where you live, you should apply!
Gran blog post!
Best part: we’re becoming a Llama (🦙== 🦄🦄)!