# Please explain the idea laid out below. It's from a recent tweet by Andrei Karpathy. LLM Knowledg... ## Retrieval Notes - Session id: `019d5de9-1abb-7043-b422-243dee5b06be` - Source: `Codex raw session` - Last activity: `2026-04-05T13:51:51.592Z` - Model: `gpt-5.4` - CWD: `/Users/nickyoung` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User Please explain the idea laid out below. It's from a recent tweet by Andrei Karpathy. LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts. ### Assistant I’m going to load the vault context the way your workspace asks me to: first the Obsidian CLI workflow, then `Dashboard.md` and today’s daily note, and after that I’ll explain Karpathy’s idea in plain terms. ### Assistant I’ve loaded the Obsidian CLI instructions and I’m waiting on the actual vault context now. Once I have `Dashboard.md` and today’s note, I’ll keep the explanation focused on what the workflow is really doing and why it matters. ### Assistant `Dashboard.md` is mostly a live dashboard shell, so the daily note is the more useful bit of actual context here. I’m opening today’s note next, then I’ll give you the explanation. ### Assistant Karpathy is describing a shift from using LLMs mainly as chatbots or coding assistants to using them as ongoing research maintainers. The basic idea is this: 1. You collect raw source material into one place. That includes papers, articles, repos, datasets, images, and notes. 2. You ask an LLM to turn that messy archive into a structured markdown wiki. The model summarizes documents, creates topic pages, adds links and backlinks, groups related ideas, and keeps the whole thing organized. 3. You keep querying that wiki with an LLM agent. Instead of asking questions against the open web every time, the agent works over your accumulated research environment. 4. The outputs go back into the same system. New answers, slide decks, diagrams, and notes become part of the knowledge base, so your past work compounds. What makes this interesting is the “compile” metaphor. He is treating the LLM a bit like a compiler: - `raw/` is the source material - the wiki is the compiled artifact - “linting” means running checks for gaps, inconsistencies, or weak links - extra CLI tools act like utilities around that compiled knowledge system So the point is not just “store notes in Obsidian.” The point is to build a semi-autonomous research loop: - ingest sources - synthesize them into structured knowledge - query that structure - generate outputs - feed those outputs back in - periodically clean and enrich the system Why he says this can work without fancy RAG is that at a moderate scale, an LLM can often navigate a well-maintained markdown corpus directly if: - summaries exist - index files exist - links are good - the corpus is not enormous In other words, structure can substitute for infrastructure. A carefully maintained wiki can be enough before you need embeddings, vector databases, retrieval pipelines, and so on. Obsidian matters here because it gives a human-readable interface to the whole thing. You can inspect sources, browse the wiki, open generated slides, and view images, while the LLM does most of the writing and upkeep. The larger vision is a new kind of product: not just note-taking, not just chat, not just RAG, but a research environment where the LLM continuously builds and maintains a living knowledge system for you. The main upside is accumulation: your questions do not disappear into chat history; they become durable structure. The main risk is epistemic drift: if the LLM is doing most of the writing, the wiki can become confident, tidy, and wrong unless you have good source-tracing and periodic checks. *L’idea centrale è trasformare la ricerca in un ciclo cumulativo in cui ogni lettura, domanda e risposta diventa parte di una memoria organizzata e riusabile.*