Tag
The tweet highlights a research finding that signals of physical plausibility can be extracted from the geometry of frozen image encoders without video training or physics supervision.
A detailed personal survey of modern deep learning, focusing on foundation models, vision-language models, and their architectural decisions, written for those who want intuition rather than dense math.