Prefill is compute-bound, decode is memory-bound, and running both on one GPU wastes most of it. Splitting them into separate pools, the KV-cache transfer tax that comes with it, and when it pays off.
Retrieval gets worse as the memory pool grows, while the confidence score keeps climbing. Here is what causes RAG memory decay and the four fixes I put into production systems: topic routing, semantic deduplication, relevance based eviction, and lexical reranking.
Prompting stops scaling after a few sessions. Here is how agent skills put your recurring logic into a SKILL.md file instead, with a weekly visualization example and the directory layout I use.
Cowork gets written off as Claude Code for people who don’t code. I use it for architecture work and multi-file debugging, and this is how I keep the context window lean, why I stay in plan mode, and which repeat jobs I turned into skills.
One working directory holds one state of the project, which is why you and your agent keep colliding. Git worktrees give each branch its own folder, and here is the script that pays the setup tax: env files, installs and a unique port offset.
Connecting Claude to a WordPress site over the Model Context Protocol turns your posts and comments into something it can read. Here is the setup, the gap-analysis prompt I use, and how I score the ideas that come back.
Vector databases have become the default answer for AI agent memory, and for a project-scoped agent they are usually the wrong one. Plain Markdown files plus a local SQLite index give you grep, git diff and no server to babysit.
LLM agent memory architecture is what decides whether an agent holds up over a long task. This covers the four temporal scopes, the manage step almost everyone skips in the write-manage-read loop, and how summarization drift creeps in.
Six things I ran into building a GPT-2 clone from scratch in PyTorch: the scaling flaw RsLoRA fixes, why RoPE replaced additive positional embeddings, why weight tying stopped paying off, and what KV-caching costs you in VRAM.
Notes from WordCamp Asia 2026 in Mumbai: plugins and themes landed on every paid WordPress.com plan, the speed build put AI against raw FSE work, and Nirav Mehta’s session covered where LLMs still break. Here is what held up as useful for real builds.