~/blog
Articles, tutorials, and posts.
A hands-on walkthrough of NVIDIA's NVFP4 4-bit format — how decimal values become E2M1 bit patterns, block scaling, and a complete PyTorch implementation.
A deep dive into PagedAttention and vLLM — how OS-inspired memory paging eliminates KV cache fragmentation and unlocks 2-4x higher throughput for LLM serving.
Contract-Net Marketplace — Auction Experiment When multiple specialized AI agents share a task queue, matching the right agent to the right job is a coordinatio…
A neural network for images is usually a CNN, and a CNN is usually a stack of small filters sliding over the picture. That is comfortable — each filter sees a s…
Every few months, a new agent framework ships with "self-improving" stamped on the README. Most of the time, nobody asks the obvious follow-up: self-improving h…