Karpathy's LLM Wiki structure cuts token costs up to 95% by avoiding full reloads

Archived signal · AI Coding & Agents

東大ClaudeCode研究所@ClaudeCode_UTx

Karpathy's LLM Wiki structure cuts token costs up to 95% by avoiding full reloads

Most AI implementations reload the entire database from scratch on every query, causing high token costs.
Karpathy's 'LLM Wiki' compiles data once and stores it in memory, eliminating repeated reloads.
Designing Claude Code Skills and CLAUDE.md carefully already implements this structure.

Where this lives now

This signal is part of the DigestOps evidence archive. The living answers built on evidence like this are at AI Coding & Agents; agents can query it as JSON at /api/v1/cards/38 or over MCP.