Hugging Face Democratizing GenAI for Everyone

Transcript

Good morning, everybody. I wish I could be here in person, but I have a very full schedule, so at least we get to meet virtually. I hope to be in person next time. My name is Julien. I am the chief evangelist for machine learning at Arcee. I've been with Hugging Face for about two years. The mission of Hugging Face is to promote open-source, state-of-the-art AI. We build a website called the Hugging Face Hub, which is the focal point of open-source AI. People call us the GitHub of machine learning, which I think is a fair analogy. Just like you go to GitHub to find and share models and code, you go to the Hugging Face Hub to find and share models and datasets. We also build libraries, open-source libraries that are extremely popular. If you want to think about Hugging Face in three words, open, simple, and state-of-the-art. We insist on those three things. We help you use and work with the latest and greatest models coming from the best tech organizations in just a few lines of code, keeping everything open-source and transparent. Before working for Hugging Face, I worked for AWS for a few years. Hugging Face is also partnering with cloud companies such as AWS and others, as well as hardware companies like Intel, to help accelerate those models on different hardware. It's a very lively platform, very busy. We have a huge community, and we're really excited to help customers build with open-source AI. Let me explain what we do on the open-source side and the commercial side. On the open-source side, Hugging Face is a technology company. I would say 90% of the team is engineers and PhDs. We contribute models, datasets, open-source libraries, together with the community, which is really the core of the company. We insist on keeping everything open so that if you decide to work only with open-source, you can build everything you want. For commercial organizations, there's always the need to go faster and focus on solving business problems. For this purpose, we also build commercial services for application hosting, model deployment, etc. We partner with cloud companies so that you can take your existing state-of-the-art model, released by Meta or Google, and deploy it in the cloud and train it in the cloud if you want to fine-tune it for your own purpose. The vision is to stay open-source but also make it easy to run the open-source ecosystem in the cloud across different hardware. I meet with a lot of customers, including in Dubai, and have amazing conversations. Many generative AI scenarios deal with increasing productivity, replacing manual processes, or legacy IT. A lot of this involves document management, classifying, summarizing, validating documents, and ensuring customer signatures and correct form completions. These are reasonably easy to automate. Code generation helps developers write better code quicker, which is very important. Chatbots, knowledge bases, finding the right information at the right time, content generation, answering customer emails automatically, generating marketing content, email campaigns, personalized in natural language, generating images, branding, product placement—the list goes on. GenerateAI brings many new capabilities for enterprise and public sector customers to innovate. How can we help startups? Startups need to be very nimble. Finding product-market fit, going quickly, and getting user feedback is critical. You have a small team, limited time, and limited capital. You need to pick your battles. What's the core value proposition of your startup? What's the core product you're trying to build? Everything around it is needed, but you can't build everything yourself. If you need voice capabilities, such as speech-to-text, there are plenty of good models out there that you can integrate easily into your app with a few lines of code. If you serve international customers, you will need translation models, and there are plenty of good ones available. These tasks are now commodities that can be solved with pre-trained models and a few lines of code. For the core business stuff, there's an opportunity to create value by training or fine-tuning a model on your own data. This could be a chatbot, a customer support application, or an online banking application. The unique experience comes from a pre-trained model that you fine-tune on your own data, creating defensible intellectual property and adding value to the startup. For large companies, innovation is key. They may have a stable business, but it's a competitive market. Competitors could create new models or user experiences based on data, and startups could be doing the same. Large companies have a ton of data, which is their main asset. If you're a manufacturing company, you have mountains of data. You need to unlock this data, which could be stored somewhere in the company. Identifying what you have is already a challenge. If you do the data work, you realize you have unique data going back decades. Financial data from 30 years ago is still valuable. We worked with Bloomberg to help them train their own model for Bloomberg GPT. Bloomberg is known for having tons of long-lasting, quality data. Other companies can do the same. If you have healthcare data from 50 years ago, it's still good and will be good in 50 years. Investing in preparing and curating your data can radically improve and invent business processes that give you an edge on the competition and help serve your customers better. How does Hugging Face differentiate? We are open-source, open-source, and open-source. Some open-source companies have gone the community edition versus enterprise edition route as they grew, but we're not doing that. Everything we build, everything the community builds, is open and will stay open. If you're a student or a small company and want to tinker and experiment, you can do everything the larger players do with open-source. Open-source models level the playing field because you get access to the latest and greatest models, like the Falcon model trained in Abu Dhabi, which is currently the best open-source large language model available. You can use it in a few lines of code. You get full transparency: you know what the models are, their architecture, and the data they've been trained on. The datasets are available on Hugging Face, and there's probably a research paper explaining the model. We integrate these models into our libraries, write the code, and you can read and contribute to that code. Full transparency is critical, especially in heavily regulated industries and for engineering teams. Open-source AI brings maximum transparency, compliance, privacy, and flexibility. The best business value is realized when you take off-the-shelf models and specialize them on your own data, which is more cost-effective than working with a general-purpose model. We partner with a wide community, from individual developers, students, and enthusiasts contributing data, models, and to the open-source libraries, to the largest tech organizations. We recently had a Series D funding round, and now our investors include AWS, Google, IBM, Microsoft, Salesforce, Intel, AMD, Nvidia, and Qualcomm. These companies are investors in Hugging Face, validating the open-source AI philosophy and making it easy to partner with them. You will find models from Salesforce, Google, Microsoft, and Meta, among others, on our platform. It's not open-source versus the rest of the world; the largest tech companies share models and work on their closed models. There's room for both, and we ensure that state-of-the-art models remain open and that there is an alternative. We release models that are unique and alternatives to closed models. It's a collaboration and competition, but open-source has made significant progress and is now extremely competitive with the best models. Where is Gen AI going? We've only scratched the surface. The launch of ChatGPT and its followers was a breakthrough, both technically and in business, as it allowed non-engineers to experiment with AI. We've seen a ton of progress, but the pace of innovation is still crazy. New models are released overnight that improve cost-performance by 2x. What used to require a 13 billion parameter model can now be done better with a 7 billion parameter model. We will see more technology breakthroughs, smaller, more nimble models outperforming larger ones, and amazing performance through hardware acceleration. We will bring these models to edge devices like smartphones and integrate different types of data—text, images, and voice. A lot of the work has been text-centric, but integrating all three modalities will lead to much more. We are three steps into a 5,000-kilometer trek. There's still time to get started, and it's never been easier to experiment. I encourage everyone to do it. Thank you for having me. Next time, I will be there in person.