Most of the commentary around big tech’s strides in AI has centered around Google and Microsoft — the former has a long history of AI research, and the latter has acquired OpenAI and integrated its products in its suite. But there’s a third FAANG name that’s quietly making big strides in AI.
Meta has the highest number of public models available on Huggingface, data shared by Huggingface CEO Clement Delangue shows. Meta had 689 models on Huggingface including MusicGen, Galactica, Wav2Vec and RoBERTa. Google followed with the next highest number of models with 591 models. Some of Google’s prominent models are BERT, Flan, T5 and mobilnet. Third was Microsoft with 255 models including DialoGPT, BioGPT, layoutLM, uniML. Salesforce was next with 88 models including CodeGen and Blip, and Nvidia was fifth with 86 models including Megatron and Segformer.
Meta has been quietly working on its AI capabilities for a while. In 2013, the company had set up the FAIR (Facebook Artificial Intelligence Research) group to “advance the state of the art of AI through open research for the benefit of all”. The group now boasts of some big names, including Yann LeCun, who is one of the pioneers of deep learning, and is the Chief AI Scientist at Meta. Meta has also been the birthplace of Pytorch, which is the most popular language in which AI programs are now written.
And Meta has released several interesting models in the recent past, many of which it has also open-sourced. The Segment-Anything model allows users to split an image automatically into its components, which has many applications across image recognition. It had also released the MIMS model, which allows for conversion of speech into text for 1,100 languages. Meta had also released its large language model named LLaMa, which was initially meant for research, but its weights were leaked, and it ended up forming the basis of several popular open-source models.
Meta has now pipped both Google and Microsoft in terms of public models uploaded on Huggingface. These models not only demonstrate the company’s research capabilities, but also show how it’s managing to productize that research into useful models. And having all these models in its kitty — and their underlying research — could help Meta hold its own against the more vaunted names in AI.