Learn LLM internals step by step - from tokenization to attention to inference optimization.

We read all 512K lines of Claude Code's accidentally exposed source. 82 docs, 15 diagrams, every subsystem mapped — from the hidden YOLO safety classifier to multi-agent swarms.

Does post-training give the model a self? Testing whether RLHF instills a privileged self-emotion direction in the residual stream, and whether surgical steering on emotion sub-directions can decouple sycophancy from harshness. Open-weight replication on Qwen2.5-32B.

A reference point for phenomena that have been reported to occur inside AI systems but have no direct mapping into natural language.

Mechanistic interpretability of transformer hallucinations via attention flow, residual stream geometry, and head-level attribution analysis.

Analyze Claude Code source leaks with static reviews, architecture notes, security findings, memory flow, and tool integration details

Analyze Claude Code’s architecture, modules, and design patterns to understand its internals and build better tools and integrations