Andrej Karpathy recently shared an idea that caught the attention of the tech world. He outlined how to build a personal language model knowledge base. People often call this a second brain. Nick Spizak demonstrated how to bring this concept to life using a few simple tools. If you manage a team, this setup helps you organize scattered information and turn it into a compounding asset.
A second brain organizes your data so you work more effectively. You build it to hold information that matters to you or your business.
You handle large amounts of data every week. You read articles, study competitors, and attend meetings. But remembering all of this is impossible. A second brain stores this information in a way that an AI can read and understand. Instead of hunting for a specific file or trying to remember a conversation, you ask your AI a question. The AI searches your private knowledge base and gives you the exact answer.
The Three-Tier Architecture
Karpathy based his idea on a specific structure. The system relies on three tiers to manage information.
- Raw: This is your brain dump. You save everything here. It holds web clippings, PDF files, and meeting transcripts.
- Wiki: This is the organized section. An agentic AI reads your raw data, extracts the facts, and connects related concepts.
- Outputs: These are your final results. When you ask the system a question, it reads the wiki and generates a report or a decision file.
The Tools You Need
You need a few core components to make this work.
- Obsidian: This note-taking application stores your files locally on your computer. It features a graph view that visualizes the connections between your files.
- Obsidian Web Clipper: This browser extension saves web pages directly into your raw folder. It grabs both text and images.
- AI Framework: You need an AI to process the data. Nick uses a custom skill built for Claude Code. This acts as the engine to read and organize your local files.
How to Build the System
Setting up the knowledge base is straightforward. You start by running a setup wizard in your command line. The wizard asks what to name your vault and where to save it on your computer.
You also assign a domain to your vault. When you use it for team notes, you name the domain “Internal Meetings.” Giving the vault a focus helps the AI understand the context of the data.
You should create multiple second brains. It helps to keep personal data separate from business data. Mixing your fitness routines with your company financials makes the system inefficient.
The Daily Workflow
Once you set up the folders, you start gathering information. You use the web clipper to save Wikipedia pages, competitor websites, or industry news. You drop all of these files into the raw folder.
At this stage, the files just sit there. Obsidian calls these “orphan files” because they have no relationship to each other.
Next, you run the ingest command. This is where the agentic AI goes to work. The AI scans the raw folder and reads the new files. It summarizes the key points and moves the structured data into the wiki.
And as the AI does this, Obsidian builds a visual map.
You watch the graph populate in real time. If you save an article about a specific founder, the graph connects their name to their previous companies, their education, and their related projects. The orphan files become an interconnected web of knowledge.
You do not have to do this manually every time. You set the AI to loop. It automatically checks your raw folder and ingests new files every few hours or days.
Maintaining Your Knowledge Base
Information changes over time. Your knowledge base loses value when the data gets stale or contradicts itself.
To fix this, you run a lint command. The AI reviews the entire wiki. It spots conflicts, highlights outdated files, and identifies missing information. It tells you when you lack background details on a specific company. You then find that information, drop it into the raw folder, and let the AI update the system. This regular pruning keeps your data accurate.
The Business Value
This system solves a major problem for managers. Your team generates massive amounts of unstructured data. By saving meeting transcripts and project notes to a second brain, you create a private, searchable record.
You do not need to interrupt a coworker to ask what happened in a meeting last month. You just query your second brain.
On day one, the system is empty. But after thirty or ninety days, it becomes a powerful resource. This proprietary dataset acts as a competitive moat for your business. It is a custom intelligence layer that no other company has. As the technology grows, managing these private knowledge bases will become a critical business function.