Hugging Face's GitHub moment
Microsoft unveiled a suite of developer tools built around AI models, including a direct plug-in with Hugging Face.
Microsoft this week put out a slew of announcements focused on creating AI-powered experiences embedded within its apps as part of its Build conference.
Like Google I/O, Microsoft devoted a considerable amount time to releases for consumer-facing products, such as how ChatGPT will integrate with Windows and search. It’s an example of just how much of a zeitgeist these companies have captured with the rise of AI models starting with ChatGPT last year.
As part of Microsoft’s slew of build announcements with AI Studio, it also unveiled a model catalog, which gives developers a way pull in models to their workflows. That includes OpenAI models and open source models, but perhaps most importantly, it includes the Hugging Face hub.
This may seem like a small, buried entry to a much larger product launch. But it really should not be overlooked in its importance—these sometimes small integrations are often harbingers of much bigger things to come.
Hugging Face has become the de-facto home for machine learning models, particularly with the explosion of development following the leak of Facebook’s more compact LLaMA language model. But even prior to that it hosted the majority of the Stable Diffusion offshoots. And even before that, it was one of the main homes of BERT, one of the first models that you’d actually see fine-tuned in the wild within enterprises.
When it raised at a $2 billion valuation last year in a round led by Lux Capital that included Addition, Sequoia, and Coatue, it wanted to be the GitHub of machine learning. It already had a suite of developer tools with the Transformers and Diffusers packages, which are go-to standards for deploying models in just a few lines of code.
Now it's directly baked into Microsoft’s new AI development interface—right up there with OpenAI. And what was once a popular data science community with big aspirations looks like it’s on its way to being a permanent fixture.
Hugging Face and Microsoft’s blooming partnership
Hugging Face’s closer partnership with Microsoft goes back to last year when it released an endpoint product in May for Azure. I had heard at the time that the product actually came together in a somewhat rushed manner to align with Microsoft’s Build conference in 2022, which surprised a fair number of people.
Now Hugging Face has fully assumed that “GitHub for ML” role, and a $2 billion valuation in 2022 that seemed like it would be pricy at the time is now looking pretty sane in retrospect. At the time of the round the company had an annual recurring revenue in the range of $7 million to $10 million, according to several people I talked to about the deal at the time.
While Hugging Face is the GitHub for ML, Microsoft owns the literal GitHub for everything else. It purchased the platform for $7.5 billion as it sought to become a full-stack developer experience, from the coding and development interface (VS Code) to the actual deployment (Azure). Like Amazon, Microsoft sees Azure as its next cash cow.
What Microsoft (and pretty much everyone else) probably didn’t foresee was the speed of innovation and development in the open source community. After Facebook unintentionally seeded its LLaMA large language model, the community quickly found ways to iterate and build new approaches to those models. Meta, too, continued to iterate with new techniques.
For a brief moment some of the online commentary compared it to OpenAI when it released its own large language model, BLOOM. That OpenAI comparison was mostly a red herring, from folks I talk to, and instead its focus stayed on its goal of being the host for all machine learning models.
The Takeaway
Microsoft is doing its best to become a one-stop shop for AI development. With the growth and evolution of open source AI models, it’s clear that goal doesn’t happen without Hugging Face. And the company has been taking increased steps to ensure that Hugging Face has a place within its products, even if they seem nascent and small for now.
At the moment, the catalog consists of model endpoints for inference. But you could see how this naturally extends as models get more compact and efficient, which is what a considerable amount of the work in the open source community is focused on at the moment.
The emergence of LangChain (and recently Hugging Face’s Agents tool) has also made agent-based workflows increasingly popular, though still experimental. There’s an open-source model for pretty much every very generalized use case today, and the use cases get more specific by the day—and they’re all hosted on Hugging Face.
The timeline where LangChain and agents become dominant parts of the workflow involves a vast, vibrant open source community building and iterating on an endless sea of ML models and chaining them together. Companies then tap those models (and maybe one or two of their own home-cooked ones) to increasingly customized experiences for every possible niche.
While GPT-4 and GPT 3.5 is pretty much the standard for everyone I talk to (most startups, and by extension startups they know, use it), the nascent open source community finds new use cases and niches every day. And inferencing an existing LLM is going to be one part of the workflow. (We haven’t even explored sentence embedding models yet and the opportunities that lie there.)
Now Hugging Face has a spot directly in Microsoft’s hopeful full developer workflow in Azure AI studio at a time when the company has thrown its full weight behind its investment in OpenAI. And Hugging Face is on its way to securing itself as a permanent fixture in AI development—if it hasn’t already.
ICYMI
Experimentally realized in situ backpropagation for deep learning in photonic neural networks (Science): A team at Stanford said it has achieved backprop with a photonic chip, an emerging and highly experimental technology that aims to break the hungry cycle of energy consumption that training takes. Modern data centers that handle training with GPUs consume enormous amounts of energy to develop models.
Audit shows that safetensors is safe and ready to become the default (Hugging Face): Hugging Face ordered an audit on the safety of the safetensors package in model deployment and usage in collaboration with EleutherAI and Stability AI. The results are that safetensors—a successor to traditional Python pickle files—are, well, safe. As AI models continue to grow in usage we’ll see more novel supply chain attacks here, and this was one big step to overcome.
Microsoft announces Windows Copilot, an AI ‘personal assistant’ for Windows 11 (The Verge): Microsoft’s goal seems to be to integrate GPT-4 into every possible part of the Microsoft experience, whether that’s Windows or its Office 365 products. It’s already looking to compete directly on Google’s home turf, but now its aspirations seem even bigger than that.
Google to test ads in generative AI search results (Reuters): Google is starting to insert some ads that are powered by gen AI, which will again continue to radically evolve the search experience across the board. But as an expert on the matter told me once, “you can have a low-margin business or a no-margin business.”
The 2023 State of Data + AI: How Businesses Are Preparing for the New Age of AI (Databricks): Databricks put out a big report on AI usage inside companies. The thing to note here is that 8 out of 10 of the most widely-adopted data and AI products are open source and Hugging Face ranks at number 7, according to the report.
Still on my radar
How has Snowflake structured its machine learning team to respond to the growth of language models?
Which foundation model startups are using Nvidia hardware?
How is Amazon structuring its contracts for A100 GPU usage?
Which foundation model developers are investing in chain-of-thought development?
If you have any tips (or answers to any of the above), you can email me at m@supervised.news or contact me directly on Signal at +1-415-690-7086. As always, please send any and all feedback my way.