Wow, this tweet went very viral!
Andrej Karpathy ยท tweet ยท 2026-04-02 source โ Reader ๐ฅ๏ธ
LLMs should be used not just to write code but to compile, maintain, and query personal knowledge bases built from raw source documents -- and sharing the *idea* of such a system (rather than specific code) is the right unit of distribution in an era of LLM agents.
practical ยท expository / how-to sketch โ A practitioner's report on a personal workflow, framed as a shareable 'idea file' rather than a finished product. It sits between a tutorial and a product-vision sketch.
Structure
| Part | Role |
|---|---|
| Framing: the 'idea file' concept | Motivates why Karpathy shares an abstract idea rather than a repo |
| Data ingest | Describes the input pipeline: raw documents indexed, then LLM-compiled into a markdown wiki |
| IDE / Obsidian frontend | Specifies the viewing layer; emphasizes the human rarely edits the wiki directly |
| Q&A over the wiki | Shows the payoff: complex queries answered without fancy RAG at small scale |
| Output, Linting, Extra tools, Further explorations | Extend the core loop with feedback cycles, health checks, search tools, and a finetuning horizon |
| TLDR / product call | Summarises the pipeline and calls for a proper product |
Key terms
- idea file Knowledge Base
A shareable description of a concept/system that another person's LLM agent can customise and build, replacing the traditional open-source repo - compile (a wiki) Knowledge Base
Using an LLM to transform raw ingested documents into structured, interlinked markdown articles with summaries, backlinks, and categories - knowledge base Knowledge Base
A local directory of markdown files and images maintained by an LLM, serving as a personal research wiki - linting Knowledge Base
LLM-driven health checks that find inconsistencies, impute missing data, and suggest new connections within the wiki - token throughput Token Throughput
The volume of LLM input/output tokens consumed by a user's workflow, here shifting from code to knowledge manipulation
Propositions
- explicitIn the era of LLM agents, the right unit of sharing is the idea, not the code; another person's agent can customise and build from the idea.conf: high
- explicitcorroboratedLLMs are better used to manipulate knowledge (stored as markdown) than to manipulate code.conf: highsupports โTony Seale's Post ยท p4 โ Karpathy: LLMs are better used to manipulate knowledge (stored as markdown) than to manipulate code. Seale's claim that markdown is 'a beautiful substrate' for episodic memory and 'the prompt the model reads' supports the same premise โ markdown as the right substrate for LLM-manipulated knowledge.
- explicitAt the scale of ~100 articles / ~400K words, LLM-maintained indexes and summaries make dedicated RAG unnecessary.conf: high
- explicitdevelopedThe human should rarely write or edit the wiki directly; the wiki is the domain of the LLM.conf: highrefines โAn LLM wiki changed how I work ยท p2 โ Karpathy's idea file claims the human 'rarely writes or edits the wiki directly.' Newton's experience holds for wiki *content* โ he doesn't hand-write entries โ but not for the wiki *system*: he actively debugs a breaking pipeline and had to bring in a second LLM to rewrite the output style, so 'rarely edits' understates the ongoing human maintenance burden.
- explicitThere is room for a major product in this space, beyond the current 'hacky collection of scripts'.conf: high
- explicitAn idea file should be kept intentionally abstract/vague to allow many directions of implementation.conf: medium
- implicitPersonal knowledge queries and explorations should accumulate back into the knowledge base, making it richer over time.conf: high
- implicitThe natural evolution is to move knowledge from context windows into model weights via finetuning.conf: medium
Arguments
Share ideas, not code, in the age of LLM agents.
Explicit premises:
- LLM agents can take an abstract idea and customise/build it for a specific user's needs.
- Sharing specific code/apps is less necessary when agents can generate implementations.
Implicit premises:
- LLM agents are widely available and capable enough to build working software from natural-language descriptions.
- The value of open-source repos was primarily in saving implementation effort, not in the ideas themselves.
Inferences:
- The traditional open-source contribution model (share code, fork, PR) is partially displaced by idea-level sharing.
LLM-compiled markdown wikis are a practical personal knowledge management system. strength: high
Explicit premises:
- Raw documents can be indexed and compiled into interlinked markdown by an LLM.
- Obsidian provides a good viewing/rendering frontend.
- LLM-maintained indexes suffice for Q&A at ~400K-word scale without RAG.
- LLM linting can maintain data integrity.
Implicit premises:
- Markdown is a sufficiently expressive format for knowledge representation.
- The user trusts the LLM to produce accurate summaries and categorisations.
- The cost of LLM token usage for this workflow is acceptable.
Inferences:
- This approach works today with current models and tools, not just in theory.
Narratives
- Karpathy's personal journey from using LLMs primarily for code to using them for knowledge management, discovering that the shift in 'token throughput' from code to knowledge was more valuable.
โ Lends authority through personal experience and signals a broader trend. - The surprise that RAG was unnecessary: Karpathy expected to need it but found LLM-maintained indexes sufficient.
โ Lowers the perceived complexity barrier; encourages readers to try the approach without sophisticated infrastructure.
Questions addressed
- โ
How can LLMs be used for personal knowledge management beyond coding assistance?
Ingest raw documents, LLM-compile them into a markdown wiki, query and enhance iteratively via LLM agents, view in Obsidian. - โ
Do you need RAG for Q&A over a personal knowledge base?
Not at ~400K-word scale; LLM-maintained indexes and summaries suffice. - โ
What is the right way to share useful LLM workflows with others?
Share an 'idea file' -- an abstract description -- rather than specific code, so the recipient's agent can build a custom version. - โญ How does this approach scale beyond ~400K words?
Only hinted at via finetuning as a future direction; no concrete answer.
Gaps
- unstated assumptionThe argument that ideas replace code assumes LLM agents are reliable enough to produce correct, secure, and maintainable implementations from vague descriptions. No evidence or qualification is offered for this assumption.
Read next: Literature on LLM code generation reliability and failure modes; studies on specification ambiguity. - missing evidenceThe claim that RAG is unnecessary rests on a single data point (~100 articles, ~400K words). No discussion of what happens at 10x or 100x that scale, or with more heterogeneous source material.
Read next: Benchmarks on LLM retrieval accuracy vs. dedicated RAG at varying corpus sizes.closed byAn LLM wiki changed how I work โ Same evidence as above (operation at ~1,440 pages / a 12,000-word single page) is a practitioner data point at roughly an order of magnitude beyond the ~100-article baseline the original claim ('RAG becomes unnecessary at this scale') rested on, and reports it continuing to work without a dedicated vector-search layer. - unaddressed questionQuality and trustworthiness of LLM-generated wiki content is not discussed. How does the user verify that the LLM's summaries, categorisations, and imputed data are correct?
Read next: Work on LLM hallucination detection, human-in-the-loop verification, and provenance tracking in knowledge bases.closed byAn LLM wiki changed how I work โ Newton supplies a concrete answer to 'how does the user verify LLM-generated wiki content is correct': he routinely reopens original source documents before writing or speaking, specifically to catch hallucinations, rather than trusting wiki summaries outright. - tensionThe post says 'you rarely ever write or edit the wiki manually' but also describes a deeply personal, expert-driven workflow. The tension between full LLM autonomy over the wiki and the need for expert curation is unresolved.
Read next: Research on human-AI collaboration in knowledge curation; when does full delegation to AI degrade knowledge quality? - scope limitationThe 'idea file' concept is presented as a general replacement for code sharing, but the example is a single, relatively simple system. Whether idea files work for complex, multi-component systems is untested.
Read next: Case studies of complex system specifications; comparison of idea-level vs. code-level sharing for different project complexities.
Related documents
- LLM Knowledge Bases
- AI 101: Agentic Vector Databases โ What Is That?
- An LLM wiki changed how I work
- Graphify: Unifying Codebase Context to Streamline Agentic Software Engineering
- Someone just built the exact tool Andrej Karpathy said someone should build
- Enterprise Analytics Beyond Dashboards: Intelligent Data Orchestration with LLMs
- Presentation: From Retrieval to Reasoning: Building Production-Ready Agentic AI Systems with Knowledge Graphs
- Meta's Recipe for Building Agents as "Organizational Second Brains"
- Paweล Huryn's Post