Build a Second Brain That Learns With You Using AI
For self-directed learners · Based on Karpathy Self-Improving AI Knowledge Base
// TL;DR
Self-directed learners can build a self-improving AI knowledge base to stop hoarding articles and highlights they never revisit. You dump book notes, saved essays, and course material into a Raw folder, and an AI librarian compiles a cross-linked Wiki on your subject. You then query it to test your understanding, save answers back, and ask the AI for your three biggest knowledge gaps — which it drafts new articles to fill during the monthly health check. It turns passive saving into an active, compounding learning loop that gets genuinely valuable around day 100.
Why does saving articles never make you smarter?
Most self-directed learners are compulsive savers — bookmarked essays, Kindle highlights, YouTube transcripts, course notes — that pile up and never get revisited. Saving feels like learning but isn't. The Karpathy self-improving AI knowledge base breaks this by making an AI your librarian: you dump everything into a Raw folder, and the AI organises it into a cross-linked Wiki, surfaces connections you missed, and actively identifies what you don't yet understand.
How do I set up a learning knowledge base?
Create a top-level second brain folder and a domain folder for your subject — say `investing-kb` or `cognitive-science-kb`. Add Raw, Wiki, and Outputs subfolders plus a Claude MD schema file. In the Claude MD, define three to five themed focus areas within your subject so the AI knows where to deepen coverage. Also add an anti-AI writing style guide — generate one by pasting Wikipedia's AI writing style article into the AI and asking for rules to never do any of that — so your Wiki reads clearly rather than like generic AI slop.
How do I turn saved material into a learning tool?
Dump everything into Raw: book highlights, saved articles via the Obsidian web clipper, course notes, video transcripts pasted straight into the chat and saved as MD. Don't organise it. Then prompt the AI to read Raw and build the Wiki per the Claude MD — index first, one file per major topic, cross-linked. About 30 minutes later you have a structured map of everything you've been consuming, with connections between ideas made explicit.
Now use it actively. Query it like a study partner: 'What does my knowledge base say about compound interest and risk?' The AI answers from your own material with sources. Save that report to Outputs. Then run the killer question: 'Based on everything in the Wiki, what are the three biggest gaps in my understanding of this topic?' Save that gap report. This turns passive saving into a directed learning loop.
How does the system make me learn faster over time?
Each month, run the seven-stage health check. It flags contradictions between your sources, stale notes over 90 days, and — most valuably for a learner — coverage gaps and new article candidates. In interactive mode, the AI drafts articles to fill the exact holes in your understanding, which you then review and ingest. This is the compounding loop: every query and gap report feeds the next, so your learning accelerates rather than plateaus.
Be patient. The system is deliberately weak on day one — it's just an organised version of what you already had. The real value emerges around day 100 of consistent dumping, querying, and re-ingesting. That's when it becomes a genuine thinking partner that knows what you know and, crucially, what you don't.
Keep the rules: never edit the Wiki by hand, always save outputs back, apply the anti-AI style guide, and maintain the Change Log so ingestion stays clean.
Next step: Pick one subject you're actively learning, create its knowledge base folder, and spend 15 minutes dumping your best book highlights and saved articles into Raw before running your first Wiki build.
// FREQUENTLY ASKED QUESTIONS
How is this better than just re-reading my highlights?
Re-reading highlights is passive and rarely reveals what you're missing. This system makes the AI cross-link your notes, surface connections you didn't see, and — most importantly — identify your three biggest knowledge gaps on demand. It then drafts articles to fill them during the monthly health check, turning passive review into an active, directed learning loop.
Which subjects work best for a learning knowledge base?
Any subject where you accumulate lots of source material works well — investing, a technical field, a language, or a discipline like cognitive science. Give each subject its own knowledge base folder under one second brain container with three to five themed focus areas. You can query them independently or together for cross-disciplinary connections.
How long until this actually makes me smarter?
Day one gives you an organised, cross-linked version of what you already saved — useful but basic. The genuine compounding value emerges around day 100 with consistent use: dumping new material, querying, saving outputs back, and running monthly health checks that fill your gaps. Treat the early output as a starting point, not the final product.