The discussion centers around the newly released work titled 'Liquid' which presents a significant advancement in the scaling of language models (LLMs) that incorporate multi-modal capabilities. It reveals that despite prior works not delving deeply into this, Liquid demonstrates a performance drop associated with unified training across visual and language tasks that diminishes as the model scales in size. The findings suggest that LLMs can indeed learn visual tasks effectively as a form of language, supporting a cohesive framework for multi-modal generation. Users praise the presentation of the paper, highlighting its clarity in discussing complex behaviors of multimodal models while acknowledging the challenges posed by naming conflicts with existing entities like Liquid AI.