Fable 5 + Karpathy’s LLM Wiki is Basically Cheating
The video demonstrates how to build a personal LLM-powered knowledge base using Obsidian and Claude, inspired by Andrej Karpathy's approach. By ingesting various data sources (YouTube transcripts, PDFs, URLs) into an interconnected wiki structure, users can create a 'second brain' that helps AI agents understand context and relationships across knowledge domains.
Summary
The speaker showcases their AIOS (AI Operating System) which uses LLM wikis to organize and interconnect knowledge from multiple sources. They begin by demonstrating a YouTube transcript wiki where videos are connected through relationships, allowing navigation between concepts like GitHub, Vercel, and Cloud Code. The speaker explains that what makes this powerful isn't just ingestion, but what the AI can do with the structured data—they use Fable to convert messy transcript connections into user-friendly HTML visualizations that beginners can understand.
The video reveals the speaker maintains multiple LLM wikis for different purposes: one for YouTube transcripts and another called 'Herk Brain' for meeting recordings. These wikis enable their AIOS to understand business context comprehensively, such as generating visual journey reports pulling stats, photos, and business metrics. The speaker credits Andrej Karpathy's concept of using LMs to build personal knowledge bases as the foundation.
The tutorial walks through implementation: install Obsidian, create a vault, open it in VS Code with Claude Code, and use Karpathy's LLM wiki gist as the system prompt. The system creates a folder structure with raw (input), wiki (processed), index (table of contents), and log (ingestion history). The speaker demonstrates ingesting two sources (Claude Fable 5 system card PDF and OpenAI GPT 5.6 article URL), which generated 20 cross-linked wiki pages and surfaced a non-obvious connection between the two documents regarding how OpenAI benchmarked different models differently than Anthropic.
The speaker emphasizes that wiki structure should match the data type and personal needs, not be one-size-fits-all. They stress that because the system uses plain markdown files with routing rules, it's not locked to any single AI tool—it can work with Hermes agents, CodeBox, or other systems. The underlying principle is that routing rules (in the claw.mmd schema) enable AI agents to efficiently navigate and find relevant information across the knowledge base.
Key Insights
- The speaker demonstrates that data sources automatically create unexpected connections in structured wikis—ingesting an OpenAI article and Claude system card revealed they used different benchmarking harnesses, meaning their model comparisons weren't directly comparable, a detail easy to miss when reading sources separately
- Fable demonstrates superior emotional reasoning and beginner-friendly interface generation compared to Claude Opus 4.8 when given the same underlying database and conceptual requirements
- The wiki system automatically structures itself based on the type of data ingested—YouTube transcripts generated organized folders (comparisons, concepts, techniques, tools) while meeting transcripts remained flat, suggesting the AI determines structure dynamicity based on source content
- The speaker's AIOS uses markdown files with routing rules in the claw.mmd schema to enable agents to efficiently navigate multiple wikis and find relevant information without wasting tokens crawling through all data
- Because LLM wikis use plain markdown files as the foundation, they remain tool-agnostic and can integrate with any AI agent (Hermes, CodeBox, etc.) rather than being locked to a single platform
Topics
Transcript
[0:00] What you're looking at right over here are a bunch of my YouTube videos being ingested into an LLM wiki. This LLM wiki, as you can see, if I zoom in, are different YouTube videos, and what's connecting them are different relations. So, we're starting to see this actual kind of like second brain of all of my YouTube videos and how they relate to each other. And all of this knowledge makes my AIOS so much smarter. And the coolest part about this is I didn't have to connect these concepts at all. I was able to just say, "Hey, Cloud Code, go grab my YouTube videos." And then ingest them into this wiki. And this thing continuously…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Nate Herk | AI Automation
Every Codex Concept Explained for Non-Coders
A comprehensive tutorial explaining 18 core Codex concepts for non-technical users, organized into four parts covering basics (projects, agents, agent loops, goals), environments (on-premises vs. cloud, working trees), customization (AI models, effort levels, permissions, skills), and scaling tools (browser, sites, subagents, scheduled tasks, voice mode).
Claude Code Mods Are Game Changers. This One Saves Me Money.
Claude Code now supports mods that allow customization of the Cloud desktop app. A user demonstrates a cache management mod that tracks usage data, displays costs, monitors session limits, and provides notifications before cache expiration to help optimize spending.
5 Claude Code Mods That Everyone Needs
The video demonstrates five useful Claude Code modifications (mods) that are plugins enabling customization of the Claude desktop application interface. The creator showcases practical mods including Cache Keeper for token management, Recording Mode for privacy, Goal tracking interface, and Collision Guard for file conflict detection, explaining how anyone can create mods using plain language.
How to Actually Build & Sell Software with AI as a Non-Techie
Dave Fabrikant, a 10+ year Python programmer and AI engineer, discusses how AI coding tools have transformed software development from a specialized skill into an accessible field for non-technical founders. He explains the progression from personal tools to scalable products, emphasizing architecture, security best practices, and the shift from specification-driven to intent-based development.
I Tested Codex's NEW $500/mo Ultrafast mode
A reviewer tests Codex's new $500/month Ultrafast mode, which claims to be 8x faster than standard mode. While results vary by task complexity, Ultrafast delivers significant speed improvements but at substantially higher resource consumption, making it suitable primarily for time-sensitive projects.