Andrei Karpathy appears to have devised a method for making Large Language Models (LLMs) work for users rather than the other way around. Instead of developing another data overlay, he has leveraged models to autonomously create and maintain personal knowledge bases. Raw data, such as articles, code repositories, and images, is fed into an LLM, which then compiles it into a structured wiki format using Markdown. This process extends beyond mere compilation; the model also generates summaries, interlinks related content, categorizes information, and even writes new articles. All of this knowledge is stored and managed within Obsidian, which Karpathy refers to as his "IDE" for knowledge management. The model is responsible for writing and maintaining the content, with user intervention required only occasionally.

As the knowledge base expands – Karpathy's currently comprises approximately 100 articles and around 400,000 words – the LLM agent begins to answer complex queries. Notably, this functionality is achieved without the reliance on cumbersome Retrieval-Augmented Generation (RAG) systems. The model independently manages indexes and summaries, retrieving the necessary information for its responses. Answers can be delivered not only as text but also as Markdown files, presentation slides (using Marp), or images, which are then integrated back into the wiki, further enhancing its intelligence. Each interaction thus contributes to the growth of the overall knowledge base.

This approach signifies a shift in how individuals and organizations can manage and utilize vast amounts of information. By enabling LLMs to build and curate knowledge repositories, Karpathy's method offers a pathway to more dynamic, self-sustaining knowledge systems that reduce the burden on human oversight and the complexity of traditional information retrieval methods. The ability to generate rich, interconnected content and respond to sophisticated inquiries without external retrieval mechanisms suggests a more integrated and efficient future for personal and professional knowledge management.

Large Language ModelsAI AgentsAutomationProductivityAI Tools