Mastering a World of Content
Long before generative AI became the talk of Silicon Valley, Facebook faced a monumental task: understanding the tidal wave of human expression being uploaded to its platform every second. This challenge gave rise to some of its first major forays into
deep learning. One of the most significant was 'DeepFace', a research project revealed around 2014 that could identify human faces in photos with an accuracy that rivaled humans. Using a neural network trained on millions of user images, DeepFace wasn't just about suggesting who to tag; it was about teaching a machine to see and recognize patterns in unstructured visual data. This work in computer vision laid the groundwork for countless features, from content moderation to augmented reality filters. On the text side, a similar revolution was happening with 'DeepText'. Announced in 2016, this AI engine was designed to understand the meaning and sentiment of thousands of posts per second across more than 20 languages. It could differentiate between 'I need a ride' and 'I just got out of a taxi,' enabling features like Messenger suggesting you call a car. These projects, focused on comprehending text and images, were the foundational building blocks for any future AI aspirations.
The Ghost of Assistants Past
Perhaps no single 'failed' project was more influential than 'M', a personal assistant built into Facebook Messenger in 2015. M was designed to be a competitor to Siri and Cortana, but with a twist: it was a hybrid system powered by both AI and a team of human supervisors. Users could ask M to do almost anything—book a restaurant, buy a gift, or even call their cable company and wait on hold for them. While the AI attempted to handle requests, humans would seamlessly take over for complex tasks. The project never expanded beyond a small test group and was officially shut down in 2018. But to call it a failure is to miss the point. Facebook acknowledged that the goal of M was to learn what people wanted from an assistant. The massive, real-world conversational data gathered from M was an invaluable treasure trove. It provided critical insights into natural language, user intent, and the sheer unpredictability of human requests—lessons that are directly applicable to building the large language models that power today's chatbots and AI agents.
Building a Digital Babel Fish
For a platform connecting billions of people globally, language is the ultimate barrier. Initially, Facebook relied on thirdrd party services like Bing for translation. But the stilted, literal translations couldn't capture the nuance of social media slang, idioms, and cultural context. This led to a massive internal push to build a world-class machine translation system. The Facebook AI Research (FAIR) lab, founded in 2013, was instrumental in this effort. Researchers pioneered the use of Convolutional Neural Networks (CNNs) for translation, which proved to be significantly faster than the recurrent neural networks (RNNs) used by other systems at the time. This focus culminated in projects like M2M-100, the first model capable of translating directly between 100 different languages without having to pass through English first, a breakthrough that dramatically improved accuracy for non-English pairs. By 2020, Facebook's systems were powering 20 billion translations per day. This deep investment in neural machine translation didn't just improve the user experience; it built the institutional muscle and expertise in language modeling that made the rapid development of Llama and other generative AI possible.
The Tools That Built the Tools
Meta's current AI dominance also rests on another category of forgotten projects: the open source tools it built for its own researchers and then shared with the world. In its early days, the FAIR lab prioritized creating and releasing frameworks that would accelerate AI development for everyone. They open sourced key components of Torch, a deep learning library, in 2015. This work evolved into PyTorch in 2017,
a framework that made it dramatically easier for researchers to go from idea to execution. PyTorch quickly became one of the most popular AI development platforms in the world,
used by academics and competitors alike.
Similarly,
FAISS (Facebook AI Similarity Search) was a library created to solve an internal problem:
searching for similar items within billions of data points,
a key function for recommendation engines.
By open sourcing these fundamental tools,
Facebook not only attracted top AI talent but also helped create the very ecosystem in which its own future ambitions would be built.
