Meta’s New AI in Virtual Agents and Security: Explained

In a time of concerns about AI safety, transparency and ethical development, technological developments to tackle these challenges are urgent.

Especially in regions like the EU which has recently implemented AI regulations, the tech sector is grappling with questions about AI's role in society, its potential impact on employment and its implications for privacy and security.

With major players like OpenAI, Google DeepMind and Anthropic making significant strides, Meta's latest announcements are another strategic push to advance the field whilst maintaining a commitment to open-source development and collaborative research.

It has advanced its position with AI development with its research through its Fundamental AI Research (FAIR) division, unveiling a suite of groundbreaking developments that span multiple domains of AI.

The research initiatives address several critical challenges facing the AI community: the development of more sophisticated embodied AI agents, the pressing need for content authentication in an era of synthetic media and the ongoing quest for more efficient and capable language models.

Off the back of this research, Meta has introduced several significant innovations: Meta Motivo, a sophisticated model for controlling virtual agents; Video Seal, a robust video watermarking system and advancements in language model architecture through their Large Concept Model (LCM).

Meta Motivo: virtual agents and embodied AI

One of Meta’s key announcements is Meta Motivo, a foundation model designed to control the behaviour of virtual embodied agents.

This model is an advancement in unsupervised reinforcement learning, allowing virtual humanoid agents to perform complex tasks without specific training for each action.

Something unique about Meta Motivo, is it’s trained using a novel algorithm that leverages unlabeled motion datasets to create human-like behaviours while maintaining zero-shot inference capabilities, which is the ability of an AI model to perform tasks it has not been explicitly trained on.

This research could lead to more lifelike non-player characters in virtual environments and new immersive experiences in the future and has potential applications in gaming, virtual reality and simulation training.

Video Seal: video watermarking and content security

In addition to Motivo, Meta has also introduced Video Seal, an open-source model for video watermarking.

This technology adds imperceptible watermarks to videos, a technique used to embed hidden information within video content, which can be used to verify authenticity or track the source of the video and in turn can later be uncovered to determine the content's origin.

This is particularly relevant in an era where AI-generated content is becoming increasingly sophisticated and difficult to distinguish from human-created content.

An element that makes Video Seal stand out is its resilience against common video editing techniques and compression algorithms, making it a potentially valuable tool for content traceability and mitigating the risks of AI-generated content misuse.

Meta Omni Seal Bench

Furthermore, Meta is also releasing Meta Omni Seal Bench, a leaderboard dedicated to neural watermarking covering several modalities, which aims to enable the research community to easily test and add their own work in the field of digital watermarking.

LCM: language models and memory scaling

Meta has made big strides in language model development particularly this year and now even further with the introduction of the Large Concept Model (LCM).

This approach is a departure from traditional token-level language modeling, instead focusing on predicting high-level concepts or ideas.

Whilst traditional models typically operate at the level of individual words or subwords (tokens), the LCM aims to work with larger units of meaning.

What makes LCM stand out is that it outperforms or matches recent large language models in tasks such as summarisation and offers strong zero-shot generalisation to unseen languages.

This could potentially lead to more efficient and effective language models capable of operating across multiple modalities and languages.

Meta Memory Layers at Scale

Another development is Meta Memory Layers at Scale, a method for scaling memory layers in AI models.

According to the company, this technique enables an increase in factual information retention without significantly increasing computational requirements.

Meta claims that models augmented with these improved memory layers outperform denser models with more than twice the computation budget.

More broadly, this advancement could lead to more efficient AI systems that can store and retrieve information more effectively.

Crucially, Meta is showing interest in the developments of these innovations as well as the results - a company blogpost stated: “By publicly sharing our early research work, we hope to inspire iterations and ultimately help advance AI in a responsible way.

“As always, we look forward to seeing what the community will build using these new releases and continuing the dialogue about how we can all advance AI together responsibly and build for the greater good.”

Explore the latest edition of Technology Magazine and be part of the conversation at our global conference series, Tech & AI LIVE.

Discover all our upcoming events and secure your tickets today.

Technology Magazine is a BizClik brand