The video showcases Andre Karpathy’s method of building a persistent, interlinked knowledge base using large language models that incrementally ingest, summarize, and maintain information in a markdown wiki, enabling compounded understanding and efficient upkeep. This approach combines human curation with LLM-driven synthesis and maintenance, resulting in a dynamic, customizable knowledge system that grows smarter over time and can be applied across various domains.
The video explores Andre Karpathy’s innovative approach to building personal knowledge bases using large language models (LLMs). Unlike traditional retrieval-augmented generation (RAG) methods where LLMs fetch and synthesize information on demand without memory, Karpathy’s method involves the LLM incrementally constructing and maintaining a persistent, interlinked wiki. This wiki, composed of markdown files, accumulates knowledge over time, cross-references concepts, flags contradictions, and synthesizes information from multiple sources, effectively compounding understanding rather than rediscovering it with each query. The human’s role is to curate sources and ask insightful questions, while the LLM handles the summarizing, linking, and upkeep.
The architecture of this system consists of three layers: raw sources (immutable original documents like articles or datasets), the wiki itself (a directory of markdown files fully managed by the LLM), and a schema or configuration file that guides the wiki’s structure and workflows. This setup allows the wiki to evolve organically as new sources are ingested, summaries are created, and cross-references updated. The LLM performs three core operations: ingesting new sources to update the wiki, answering queries by synthesizing information from the wiki, and performing maintenance tasks such as identifying contradictions or gaps, ensuring the knowledge base remains accurate and comprehensive.
When a new source is ingested, the LLM reads it, extracts key concepts and data points, writes a summary page, updates related entity and concept pages, flags contradictions, updates the master index, and logs the changes. This process ensures that each new piece of information enriches the entire wiki, making it smarter and more interconnected over time. The division of labor is clear: humans provide direction and critical thinking, while the LLM manages the heavy lifting of knowledge organization and maintenance, overcoming the traditional problem of wikis becoming too burdensome to maintain manually.
The video then demonstrates building such a knowledge base focused on advanced trading strategies using Claude Code and Opus 4.6. The process involves setting up the directory structure, copying raw transcripts, defining the schema, seeding concept pages, and ingesting the raw sources. The resulting wiki is navigable and richly interlinked, viewable in tools like Obsidian. The LLM can answer detailed questions by referencing the wiki and, if needed, perform web searches to backfill missing information, continuously expanding and refining the knowledge base. This dynamic growth allows the wiki to become an increasingly powerful research tool.
In conclusion, Karpathy’s LLM-powered wiki approach offers a transparent, customizable, and interoperable way to build personal knowledge bases that grow smarter over time. By leveraging LLMs for maintenance and synthesis, it solves the traditional challenges of wiki upkeep and knowledge fragmentation. The video encourages viewers to experiment with this method, highlighting its applicability across domains such as research, personal development, business, and more. The approach’s flexibility and efficiency make it a compelling new paradigm for organizing and leveraging knowledge with AI assistance.