Meta’s Fundamental AI Research (FAIR) team has launched several innovative initiatives aimed at enhancing AI development, with a particular emphasis on large language models, multi-modal capabilities, and tools designed to identify AI-generated content.
Key Highlights
Meta FAIR is unveiling a series of new research contributions that showcase our recent advancements in the development of agents, as well as improvements in robustness, safety, and machine learning architectures.
The research we are presenting supports our ambition to achieve advanced machine intelligence and includes Meta Motivo, a foundational model for managing the behavior of virtual embodied agents, and Meta Video Seal, an open-source model for video watermarking.
Our objective is to democratize access to cutting-edge technologies that reshape our interactions with the physical environment. To this end, we are dedicated to nurturing a collaborative and open ecosystem that promotes progress and discovery.
As we strive towards the realization of advanced machine intelligence, we are eager to share our advancements with the research community, enabling them to build upon our findings. Today, we are pleased to present some of the latest research, code, models, and datasets from Meta Fundamental AI Research (FAIR). The materials we are sharing today concentrate on developing more capable agents, enhancing robustness and safety, and innovating architectures that allow models to learn new information more efficiently and expand beyond existing limitations.
In this release, we are glad to provide a demonstration and code for Meta Video Seal, an open source model to video watermarking considering at the scene of the success of Meta Audio Seal which we introduced year before. Furthermore, we are including an array of other resources such as a basic reference model for the control of behavior of virtual embodied agents, a technique for boosting layers of memory to accommodate increased factual data, and code for increasing the social component of intelligence of models. You can get plenty of information from this post as this post includes nine projects and artifacts, and all of them are ready for downloading and no time is needed.
This effort speaks to our continued dedication towards increasing openness and replicability of work we do within the community. Our aim while sharing the findings of our first-phase studies is that this will be of help to interested researchers to build on it to contribute to the positive growth of AI. We are already looking forward to the use that the community will make of these new resources and also the continuing debate about how AI can be responsibly developed and used to the benefit of everyone.
Meta Motivo
Meta Motivo is the name of the developed behavioral foundation model that was created to regulate the actions of a virtual humanoid agent to perform complex tasks.
This model is built using an algorithm that employs an unlabeled motion dataset to accomplish unsupervised reinforcement learning to enable the learning of human-like behaviors while keeping the model’s ability to perform zero-shot inference intact. One of the most important technical enhancements of our algorithm for learning a representation is that it can simultaneously encode state, motion, and reward into a shared latent space. As a result, Meta Motivo can provide efficient whole-body control solutions for a broad range of problems related to motion tracking, reaching for target positions, and maximizing reward, without the need for additional training or planning.
Meta Video Seal
AI tools have immense potential to connect the world, but implementing safeguards is essential to prevent misuse, such as imitation and manipulation. Post-hoc watermarking plays a vital role in ensuring content traceability and accountability.
Meta’s Video Seal introduces a watermarking technique that embeds an imperceptible watermark (optionally containing a hidden message) within videos. This watermark can later be extracted to verify the video’s origin. The technique is robust against common video alterations, including blurring, cropping, and compression algorithms widely used in online content sharing.
To encourage transparency and collaboration, Meta is releasing the Video Seal model under an open license, along with accompanying resources like a research paper, training and inference code, and an interactive demo for hands-on exploration of the model.
Chatbot ‘encouraged teen to kill parents over screen time limit’ 2024