Compositionality in Models
The discussion delves into the intriguing limitations of models like Clip, which lack compositional understanding, and how that might lead to unexpected creative outcomes. The excitement surrounding the Clip plus VQGan moment highlights the exploratory nature of these tools, contrasting with the more uniform outputs seen in language models. There's a sense of anticipation for when language models might break free from their common denominator voice to inspire new ways of thinking.In this clip
From this podcast
The Gradient
Ted Underwood: Machine Learning and the Literary Imagination
Related Questions
How does the understanding of language relate to the understanding of images in the episode Ilya Sutskever: Deep Learning | Lex Fridman Podcast #94 and the clip Merging Language and Vision?
Can language models understand stories as discussed in the episode Ted Underwood: Machine Learning and the Literary Imagination and the clip Literary Prestige Models?