The research arm of Meta has decided to share some of its internal artificial intelligence projects with the wider research community in order to enhance its AI models. The open science research group, Fundamental AI Research (Fair), consists of 500 to 600 individuals in Europe and North America who are focused on addressing fundamental AI challenges.
Fair has recently introduced several new research tools that they hope will inspire innovation, exploration, and the discovery of new ways to implement AI on a large scale. These tools include Chameleon, a unified platform for text and image input and output; multi-token prediction for training language models to predict multiple future words simultaneously; and AudioSeal, an audio watermarking technique.
While Fair concentrates on fundamental innovation, research outcomes are also shared with Meta’s applied research team. This team collaborates with Meta’s product teams to transform concepts like Chameleon into products. Notable innovations from Fair have influenced products such as Meta’s smart glasses and the Llama models.
Chameleon utilizes tokenization for text and images, providing a simpler approach that is easier to manage and scale. The multi-token prediction model tackles the inefficiency of predicting the next word in large language models, aiming to generate multiple tokens in a structured manner inspired by code generation processes.
Regarding the debate between closed and open AI models, Fair believes in openness as long as appropriate safeguards are in place. For instance, while the Chameleon model has not released its image generation capabilities due to safety concerns, Fair’s voice synthesis model is still in a developmental stage and not publicly available due to existing authentication challenges.
In order to use the new AI models released by Meta, users are encouraged to have a level of proficiency using similar models. Getting started with these models does not require high-end hardware, as smaller models can run on a single GPU. However, larger models may require more advanced knowledge of distributed systems for optimal performance.