Meta, the advanced AI research division of the company, has recently released new research assets to further their goal of achieving autonomous machine intelligence (AMI) while promoting open science and reproducibility. These releases include an updated image and video segmentation model, as well as research on improving large language models, post-quantum cryptography, model training, and inorganic materials discovery. One of their latest models, SPIRIT-LM, is a speech and text-generative language model that can generate both speech and text in a cross-modal manner. This model has been evaluated on various metrics and the company plans to continue improving its capabilities while also prioritizing transparency and safety.
While Meta is not the only player in this space, other companies like Anthropic have also released their own AI models, such as Claude 3.5 Sonnet. However, these models do not have features specifically designed for podcast creation or expressive voice functionalities like Meta’s Spirit LM. Meta’s recent paper also detailed their use of the ‘chain of thought’ mechanism, which has been utilized by other companies like OpenAI for their o1 models. Google and Anthropic have also published research on reinforcement learning from AI feedback, but these are not yet available for public use.
Meta’s FAIR team stated that these new releases align with their goal of achieving advanced machine intelligence while also promoting open science and reproducibility. The newly released models include Segment Anything Model 2 for images and videos, Meta Spirit LM, Layer Skip, SALSA, Meta Lingua, OMat24, MEXMA, and Self Taught Evaluator. One of the most notable models is the “strong generative reward model with synthetic data,” which is a new method for training reward models without relying on human annotations. This approach generates contrasting model outputs and uses an LLM-as-a-Judge to evaluate and improve those outcomes, without the need for human labeling. Overall, these new releases showcase Meta’s commitment to advancing AI research while also prioritizing transparency and reproducibility.
