Andrej Karpathy Announces Shift from Coding to Building Personal Knowledge Base with LLMs
Former OpenAI co-founder and ex-Tesla AI director Andrej Karpathy stated that he has shifted his focus from "coding" to "building knowledge bases," managing personal knowledge systems with LLMs.
His complete workflow includes: collecting papers/articles/code/images in a raw directory, using Obsidian to clip web pages; LLM automatically compiles them into a Markdown Wiki with summaries, backlinks, and interconnections; Obsidian serves as the front-end for browsing and visualization; LLM Agent completes complex Q&A based on the Wiki and feeds back into the archive; automatic health checks to complete data.
Developers and researchers adopting the Karpathy-style LLM knowledge base workflow are shifting funding from traditional coding tools to AI-driven knowledge compilation and management platforms, benefiting from Karpathy-style prototype tools and the Obsidian ecosystem, while pure code generation agents face functional replacement pressure.
Source: Public Information
ABAB AI Insight
Andrej Karpathy previously promoted autonomous driving vision systems and the GPT series during his time at OpenAI and Tesla, and has shared his personal productivity system multiple times. He had earlier used LLMs for code generation, and this shift to knowledge base compilation continues his idea of "LLMs as general cognitive engines," upgrading from tool user to system builder.
On the capital path, Karpathy publicly shares the complete harness (raw collection → LLM compiles Wiki → Agent Q&A → feedback + health checks) through tweets and prototype scripts, guiding developer resources towards Obsidian plugins, custom LLM Agents, and synthetic data fine-tuning. Early adopters can quickly accumulate personal/team knowledge assets, forming a low-labor cost cycle where "humans only ask questions, LLMs read and write autonomously."
Similar to the evolution paths of knowledge management tools like Notion AI, Mem, and Reflect, Karpathy's workflow is in the early expansion phase of transforming personal/team knowledge management from "manual organization" to "LLM autonomous compilation and maintenance," having sparked widespread replication under 9.6 million views.
Essentially, this represents a technological replacement: traditional knowledge management work that relied on manual reading, organizing, and indexing is completely replaced by the LLM's automatic compilation of Wiki + Agent cycles. Karpathy upgrades LLMs from "code generation tools" to "knowledge engines," reconstructing personal productivity from a "human-led" to an "AI fully autonomous maintenance" mechanism, paving the way for the next generation of productized knowledge operating systems.