Meta has recently announced the launch of four new AI models with multi-modal, multi-media capabilities. The company's Fundamental AI Research (FAIR) team introduced these new AI models as part of the Meta Chameleon family.
Public Release of Key Components
Meta unveiled key components of its Chameleon 7B and 34B language models that are designed to support mixed-modal input and text-only outputs. These models were safety tuned and are intended for research purposes. Although Meta has taken responsible steps in their development, they acknowledge the existing risks and have chosen not to release the Chameleon image generation model at this time.
Future Developments
Additionally, Meta hinted at the potential release of more advanced large language models that focus on multi-token prediction. The company also made pre-trained models available for code completion under a non-commercial/research-only license.
Meta JASCO and AudioSeal
A paper was released for Meta JASCO, the generative text-to-music models, with the pre-trained model expected to be launched soon. Meta also introduced AudioSeal, which is described as the "first audio watermarking technique." This technique is designed to detect AI-generated speech and AI-generated segments within longer audio tracks.
Collaboration and Community Building
Meta emphasized its belief that access to cutting-edge AI technology should be inclusive and not limited to a select few tech giants. The company expressed excitement about sharing these advancements with the community and looks forward to witnessing how individuals and organizations will leverage this technology for learning, iteration, and innovation.
Despite the public availability of these models, Meta has implemented restrictions such as research-only licenses and safety limitations. The company is dedicated to maintaining an open AI system while ensuring responsible usage of their AI models.
For more information on this announcement, you can read the full article here.
