Kimi K2.5 Deep Outline
A detailed reading note on Kimi K2.5, native multimodal training, and parallel agent orchestration.
Research notes about agi
Essays and field notes about engineering decisions, debugging, product tradeoffs, and the systems behind shipping software.
Understanding Cross-Attention in Multi-Modal Architectures
A short framework for reading cross-attention design choices in modern vision-language model papers.
Scaling Multi-Modal Understanding with Sparse Visual Tokens
Notes on reducing visual token load while keeping cross-modal reasoning stable in large VLMs.
Training Recipes for Stable Vision-Language Models
A working summary of the training choices that most affect stability, convergence, and downstream transfer.
Controllable Multi-Modal Generation with Unified Latent Spaces
A concise look at why unified latent representations keep appearing in modern image, video, and audio generation systems.