Knowledge Graphs: Advancing Karpathy’s LLM Wiki for Deeper Insights
Clip title: Fix Karpathy’s LLM Wiki with a Knowledge Graph | Claude Code + Obsidian + InfraNodus Author / channel: Nodus Labs URL: https://www.youtube.com/watch?v=yYSTsKo8moU
Summary
The video introduces Andrej Karpathy’s “LLM Wiki,” a framework designed for building personal, persistent knowledge bases using Large Language Models (LLMs). Unlike traditional LLM interactions that often involve “retrieval-augmented generation” (RAG) — where an LLM retrieves information from uploaded files and generates an answer, effectively “rediscovering knowledge” each time — LLM Wiki creates a structured, interlinked collection of markdown files. This wiki incrementally builds and maintains an evolving knowledge base, allowing the LLM to read, extract, and integrate new information into existing concepts, questions, data, sources, and systems, fostering accumulation rather than episodic rediscovery.
However, the video identifies a critical missing piece in Karpathy’s original framework: while the LLM Wiki provides a structured knowledge base, simply asking an LLM to produce results from it can still lead to generic answers. LLMs are trained to generate the most probable outcome, not necessarily to uncover novel insights or identify gaps in the existing knowledge. The presenter argues that a complementary tool is needed to navigate and understand the underlying structure of the wiki, and, more importantly, to identify areas of underdeveloped knowledge or disconnected concepts that can spark new ideas.
To address this, the video proposes integrating knowledge graphs, specifically using the InfraNodus tool. InfraNodus visualizes the LLM Wiki’s content as a dynamic network of interconnected concepts, applying graph science metrics to identify main ideas, clusters, and, crucially, structural “gaps” in the knowledge. By highlighting these gaps, InfraNodus enables users to generate new, pertinent research questions and insights by explicitly bridging these disconnected clusters. The tool can be used as a plugin in integrated development environments like Cursor or Obsidian, or directly via an MCP (Micro-Service Communication Protocol) server that allows the LLM agent to interact with the knowledge graph analysis directly, thereby steering the LLM’s “thinking” process towards more original and targeted outcomes.
The presenter demonstrates this workflow by first initializing a new LLM Wiki on a chosen research topic (e.g., finance), where the LLM asks guiding questions about the domain, sources, and goals. It then automatically ingests raw research papers and notes, converts them into markdown, and structures them into concept pages, connection pages, data, questions, sources, and strategies. As this structured wiki is built, InfraNodus provides a live, visual representation, allowing the user to identify key topics, less-developed areas, and conceptual gaps. This augmented approach transforms the LLM Wiki into a powerful “second brain” that not only stores and organizes information but actively helps researchers discover new connections, formulate original hypotheses, and direct their future intellectual exploration in a far more coherent and interconnected way than traditional LLM methods alone.
Related Concepts
- Knowledge Graphs — Wikipedia
- LLM Wiki — Wikipedia
- Large Language Models — Wikipedia
- Retrieval-Augmented Generation (RAG) — Wikipedia
- InfraNodus — Wikipedia
- Graph Science — Wikipedia