Notes on research, learning, and ideas.
An essay on representational drift in neuroscience and its connection to cross-modal representation shift in point-supervised video moment retrieval, with reflections inspired by DRONE.
A technical essay on the newly proposed Neural Computer (NC) concept: what problem it is trying to address, how it differs from agents and world models, what current prototypes actually show, and where the hard limits still are.
A technical deep-dive into the architecture evolution of Multimodal LLMs, from early dual-encoder designs to unified native multimodal models, covering visual encoders, alignment strategies, training paradigms, and efficient inference.
A technical survey of modern video understanding, covering spatiotemporal attention mechanisms, long-video comprehension, temporal grounding methods, and the emerging Video-LLM paradigm.
An in-depth analysis of video generation architectures, from Latent Diffusion to Video DiT, covering Sora, CogVideoX, and Open-Sora, with focus on spatiotemporal consistency, scaling strategies, and open challenges.