Latest / AI Fire Daily / #63 Robin: Stop Babysitting Your Notes - Build Karpathy’s LLM Wiki with Obsidian & Fable 5
Transcript
- 0:00Imagine a digital brain that reads your chaotic
- 0:02notes, beat, and actually connects the dots for
- 0:06you. It sounds like science fiction at first
- 0:08glance, but it is entirely possible right now.
- 0:13Welcome to the Deep Dive. Today, we are exploring
- 0:16something that fundamentally changes how you
- 0:18interact with your own information. Yeah, I have
- 0:20been looking forward to this one. We are digging
- 0:22into a really fascinating stack of sources today.
- 0:25We have Andrzej Karpathy's original LLM Wiki
- 0:28Manifesto, and we are pairing that with some
- 0:31brand new case studies. Right. Those studies
- 0:34test this specific setup using Claude Code and
- 0:37the new Fable 5 model. The mission today is pretty
- 0:40clear. We are going to build a connected AI second
- 0:43frame. Exactly. And by the end of this deep dive,
- 0:46you will understand exactly how to set this up.
- 0:48It only takes about five minutes. And we are
- 0:49moving past the era of digital hoarding. This
- 0:52is not just about dumping files into a hard drive.
- 0:54Oh, absolutely not. We're talking about turning
- 0:56messy, disorganized knowledge into usable insights
- 0:59and doing it without that dreaded feeling of
- 1:02information overload. That is the dream, right?
- 1:05Yeah, it really is. We should probably start
- 1:07by defining the end. end goal, though. We need
- 1:09to know what we are actually building before
- 1:11we get into the mechanics. Okay, so an LLM wiki
- 1:14is essentially a personal Wikipedia, but it is
- 1:18built entirely from your own raw data. Taking
- 1:21your unique context into account. Exactly. Your
- 1:24files, your transcripts, your active projects.
- 1:27The crucial difference from a regular note -taking
- 1:29app is the manual labor. Usually you are the
- 1:32one highlighting things. Yeah, you are creating
- 1:34the folders and building the tags. Right. But
- 1:36here, you do not link things by hand. The AI
- 1:39handles the heavy lifting in the background.
- 1:41It reads your raw data autonomously. That means
- 1:44long YouTube transcripts. It means dense PDFs.
- 1:48It means those scattered meeting notes you completely
- 1:50forgot about. Exactly. It turns them into connected
- 1:52markdown pages. Instead of you doing the tedious
- 1:54work, the AI generates new pages for core concepts.
- 1:58It creates dedicated workflow documents. And
- 2:00it automatically wires all the links together.
- 2:03I still wrestle with hoarding PDS. I never actually
- 2:05read. I mean, it happens to everybody. You see
- 2:07a brilliant research paper. You download it to
- 2:09a desktop folder. And you tell yourself you will
- 2:11analyze it on Sunday. Right. But Sunday comes
- 2:14and goes. The folder just grows into a digital
- 2:16graveyard. Yeah. But this setup actively solves
- 2:19that friction. It creates natural pathways through
- 2:23that forgotten data. Let me pull an example from
- 2:25Karpathy's manifesto to illustrate that. Go for
- 2:28it. You might have a page about GitHub. That
- 2:31page naturally leads to Vercel. Then Vercel links
- 2:35over to Claude Code. And suddenly, that leads
- 2:37to a forgotten deployment workflow you saved
- 2:40six months ago. Exactly. Think of it like a tireless
- 2:43librarian. This librarian reads all your neglected
- 2:46books and writes a custom index just for you.
- 2:48Every new source you add gives the system a broader
- 2:51map. Does this mean the AI is just dumping summaries,
- 2:54or is it actually mapping relationships? It maps
- 2:57relationships. It finds the useful ideas, creates
- 3:00standalone pages, and weaves them together with
- 3:03backlinks. Got it. It builds an interconnected
- 3:05web, not just a pile of summaries. Exactly. That
- 3:09web becomes its foundational memory. When you
- 3:12ask it a question later, it follows those links
- 3:14to find precise answers. Okay, let's unpack this
- 3:17setup process. We need to move from the theoretical
- 3:19magic to the practical reality. Yeah, we should.
- 3:23You mentioned earlier that building this takes
- 3:25about five minutes. Let us walk through the mechanics
- 3:28of that. It sounds intimidating, but it's just
- 3:30four concrete steps. Step one focuses on the
- 3:33visual layer. You just need to install Obsidian
- 3:36on your computer. Obsidian is essentially a highly
- 3:38customizable text editor, and it is completely
- 3:42free to use. Right. You open Obsidian and you
- 3:44create a new vault. Just to demystify the jargon
- 3:47for a second, a vault is a simple folder on your
- 3:50computer that holds your text files. That is
- 3:53all it is, just a local directory on your hard
- 3:55drive. Now, the best practice from the source
- 3:56material is highly specific here. Okay. You should
- 3:59place this vault folder inside your main AI project
- 4:02directory. So it shares your active business
- 4:05context. You want your digital brain living right
- 4:08next to the code or projects you are currently
- 4:10building? Mm -hmm. They need to sit in the same
- 4:12neighborhood. Step two is where we bring in the
- 4:14tools. Right. You open that new vault folder
- 4:16in VS Code. You open up the integrated terminal.
- 4:19Then you run Claude code directly inside that
- 4:22empty folder. At this exact moment, your vault
- 4:24is still completely blank. There are no notes.
- 4:26There are no folders. Which brings us to step
- 4:29three. You have to turn Claude code into your
- 4:33wiki agent. Let us clarify that term quickly.
- 4:35An agent is an AI program that can independently
- 4:38execute tasks for you. Exactly. You initialize
- 4:41this agent using Andrej Karpathy's specific gist
- 4:45prompt. You paste in a command that says, you
- 4:47are now my LLM wiki agent. Implement this exact
- 4:51idea. The prompt gives the AI its marching orders.
- 4:53It tells the agent to construct the complete
- 4:55architecture of your second brain. Right. It
- 4:57generates a master file called clioed .mdgay.
- 5:01That file serves as the system schema. It lays
- 5:04out the rules. It sets up the index. It creates
- 5:06the ingestion log. Yeah, it defines all the naming
- 5:09conventions for future files. It is kind of like
- 5:11stacking Lego blocks of data. It builds the entire
- 5:14foundation autonomously. And the case studies
- 5:17point out something interesting here. What is
- 5:18that? You do not even need the newest, most expensive
- 5:22models for this initial setup. Opus is perfectly
- 5:25fine for this foundational phase. Oh, absolutely.
- 5:28You save the heavy reasoning models for later.
- 5:31Now you hit step four. You feed the machine.
- 5:34Right. You take those neglected PDFs and drop
- 5:37them into a newly created raw folder. Or you
- 5:40just paste a YouTube URL into the terminal. You
- 5:43simply tell Claude Code to ingest the new material.
- 5:46The AI reads the raw files. It synthesizes the
- 5:50information. It drafts the new markdown pages.
- 5:52And it updates every single relevant backlink
- 5:54across your entire wiki. Exactly. Why is it so
- 5:57crucial to run cloud code specifically inside
- 6:00the VS Code terminal for that folder? Because
- 6:02the AI needs direct local access to read the
- 6:06files, create new markdown pages, and physically
- 6:09update the wiki's folder structure on your machine.
- 6:12Makes sense. It needs direct access to actually
- 6:14build and organize the files. It is not a web
- 6:16app storing things on a distant server. It is
- 6:19physically writing data to your SSD. We are going
- 6:22to pause for a quick second. We will be right
- 6:24back. And we are back. Here is where it gets
- 6:28really interesting. We have all these files successfully
- 6:31ingested. The AI has built a web of notes. Yeah.
- 6:35But how does it actually navigate this web when
- 6:37you ask it a complex question? Well, it needs
- 6:40a very strict routing system. Without one, it
- 6:43gets confused. It also ends up wasting massive
- 6:46amounts of your computing power. The case studies
- 6:49break this architecture down into four core parts.
- 6:51First, we have the raw folder. That's exclusively
- 6:54for your unedited inputs. The transcripts, the
- 6:56lengthy articles, the raw meeting logs. The second
- 6:59is the wiki folder. This is where the polished
- 7:01AI -generated concept pages live. Right. Third
- 7:04is the index. That functions as your dynamic
- 7:07table of contents. Finally, you have the log.
- 7:10The log is arguably the most vital piece of the
- 7:13architecture. It acts as a permanent chronological
- 7:15record of every single ingestion. So the system
- 7:19always knows exactly what was added, when it
- 7:21was added, and what pages were generated from
- 7:23it? Exactly. The actual transformation of the
- 7:25data is fascinating. Yeah, the source material
- 7:28shows one single messy transcript being processed.
- 7:32The AI does not just summarize it into one document.
- 7:36No, it breaks it apart into five to ten. distinct,
- 7:38highly focused pages. It creates a dedicated
- 7:41page for the specific tools mentioned. It spins
- 7:43up another page for the workflows discussed.
- 7:46It isolates the people involved and the key decisions
- 7:49made. It turns a flat wall of text into a multidimensional
- 7:53relational database. But that brings up a major
- 7:55debate in the note -taking community. Oh, the
- 7:58folder debate. Yeah. People tend to obsess over
- 8:00perfect folder hierarchies. They want deep nested
- 8:03categories for everything. You certainly can
- 8:06use deep subfolders if you want to. You could
- 8:08separate your theoretical concepts from your
- 8:10raw sources. You could isolate people from companies.
- 8:12The system is flexible enough to handle that.
- 8:15But the guide strongly notes that a flat folder
- 8:16structure is actually better here. dumping all
- 8:19the polished pages into one single folder. To
- 8:22a human, that sounds like pure chaos. But for
- 8:25an AI, it is highly efficient. This is where
- 8:28we get to the secret sauce of the LLM wiki concept,
- 8:31routing. Routing dictates how the AI travels
- 8:34through your data without getting lost in the
- 8:36weeds. Normally, if you ask an AI a question
- 8:39about your files, it tries to read every single
- 8:42document from scratch. It brute forces the search.
- 8:45Yeah, which is terribly slow. In this system,
- 8:47it works systematically. It starts at the index
- 8:49file, then it checks the ingestion log for recent
- 8:51updates. Finally, it follows the specific backlinks
- 8:54to retrieve the exact answer. This targeted approach
- 8:58saves massive amounts of processing time. More
- 9:01importantly, it saves tokens. Right. Just to
- 9:04clarify the terminology, tokens are the basic
- 9:07chunks of text an AI reads or generates. If you
- 9:10force a language model to read your entire database
- 9:12for every single query, it burns through its
- 9:15token limit instantly. It gets sluggish. Yeah.
- 9:18Its context window gets cluttered and it starts
- 9:20hallucinating answers. But with this proper routing
- 9:23system, it stops guessing where the context is.
- 9:26Right. And it actually gives the knowledge a
- 9:28map. Exactly. It only reads the exact pathway
- 9:30required to answer your prompt. If I just dump
- 9:33everything in a flat folder, won't the AI get
- 9:35lost compared to neat subfolders? Actually, a
- 9:38flat structure can be faster. As long as the
- 9:41index and backlinks are clean, the AI can search
- 9:43across everything instantly without navigating
- 9:45a maze of subfolders. So for the AI, a flat,
- 9:49well -linked map beats deep, nested folders.
- 9:52Human brains prefer neat visual compartments.
- 9:55AI models prefer direct semantic links. The true
- 9:58map is the network of backlinks, not the visual
- 10:01folder hierarchy on your screen. So what does
- 10:03this all mean for the listener's daily workflow?
- 10:07We have built the foundational engine. Right.
- 10:09We have laid down the systematic tracks. Now
- 10:11we need to explore what happens when you introduce
- 10:13a high -speed train onto those tracks. We are
- 10:16finally talking about the Fable 5 model. Yes.
- 10:18The case studies provide three incredible demos
- 10:21showing what happens when advanced reasoning
- 10:23meets perfectly structured data. The contrast
- 10:26is staggering. The first demo focused on extracting
- 10:28clean resources from incredibly messy data. The
- 10:31researchers took a massive, disjointed database
- 10:34of technical transcripts. The stated goal was
- 10:36to transform this dense information into a beginner
- 10:39-friendly HTML resource. They initially attempted
- 10:42this with the Opus 4 .8 model. And Opus really
- 10:45struggled with the task. It took a full day of
- 10:48back -and -forth prompting to get anything usable.
- 10:51The output was rigid and overly academic. It
- 10:53was a mess. Why did Fable 5 succeed in creating
- 10:56the HTML resource in one shot when Opus 4 .8
- 10:59struggled for a whole day? Fable 5 is just vastly
- 11:02better at understanding your underlying intent
- 11:04and reasoning across multiple messy sources to
- 11:08generate a clean, usable output. It grasps the
- 11:11actual goal instead of just blindly processing
- 11:13the data. It intuitively understood what a human
- 11:16beginner actually needed to see on the screen.
- 11:18It created clickable content. It generated accurate
- 11:21source references without being asked. It built
- 11:24a clean, navigable interface on the first try.
- 11:26Demo 2 is where things get genuinely wild. They
- 11:30tasked the system with a comprehensive business
- 11:32recap. This perfectly illustrates the power of
- 11:34shared local context. They gave Fable 5 one single
- 11:37prompt. They asked it to create a visual business
- 11:39journey covering the first six months of 2026.
- 11:42It did not just write a generic corporate summary.
- 11:45It pulled real specific photos from the vault.
- 11:48It used the actual company logos. It generated
- 11:51the entire report matching their exact dark mode
- 11:54brand style. It could do all of that because
- 11:56the AI agent had persistent access to the entire
- 12:00business context inside that local folder. It
- 12:02pulled accurate subscriber growth charts. It
- 12:05identified revenue peaks. It calculated churn
- 12:07rates and conversion metrics. It even mapped
- 12:09out their specific sales funnel structure. It
- 12:12felt entirely bespoke because it genuinely knew.
- 12:15the operational history behind the business.
- 12:17It is fundamentally different from a standard
- 12:19chat bot hallucinating a business plan. It is
- 12:22accurately reporting on your documented reality.
- 12:25Then we arrive at demo three. Cross -source reasoning.
- 12:29The researchers added just two new technical
- 12:31sources into the raw folder. Two standard PDF
- 12:34files. In about 10 to 12 minutes, Fable 5 digested
- 12:37them and created 20 cross -linked concept pages.
- 12:40Yeah. But here's the most crucial detail of the
- 12:42entire study. The AI caught an incredibly obscure
- 12:46discrepancy buried in the data. The testing benchmark
- 12:48issue. Yes. It noticed that a major company benchmarked
- 12:52their GPT 5 .6 model against something called
- 12:55Mythos Preview. They did not benchmark it against
- 12:58the full released Mythos 5 model. Furthermore,
- 13:01Fable 5 noticed they used an entirely different
- 13:04testing harness to run the evaluation. Which
- 13:06means those performance numbers should not be
- 13:08directly compared at all. It is an apples to
- 13:11oranges comparison. Two secs silence. Whoa! Imagine
- 13:15it catching those invisible threads and benchmark
- 13:18discrepancies we completely miss. A human analyst
- 13:21would have deep fatigue by the time they reached
- 13:23page 40 of that report. It performs the deep
- 13:26analytical work of a senior researcher, and it
- 13:29does it automatically in the background while
- 13:31you grab a coffee. Which leads to the ultimate
- 13:33question. How does this practically scale for
- 13:36your daily workflow? It scales into a complete,
- 13:38personalized AI operating system. You are not
- 13:42limited to just one monolithic wiki. You can
- 13:45run multiple instances. You could have one dedicated
- 13:47solely to your YouTube scripts. You could have
- 13:49another strictly for client meeting follow -ups,
- 13:52a third for your business metrics, and a fourth
- 13:55for your academic research. They could all live
- 13:57in the same broader ecosystem. And the most important
- 14:00factor is the underlying format. Everything in
- 14:04this system is built entirely on Markdown. Markdown
- 14:06is a simple way to format text without needing
- 14:09complex software. Because your entire digital
- 14:11brain is just a collection of Markdown files,
- 14:14the system is completely portable. You are never
- 14:17locked into a proprietary platform. You are not
- 14:19trapped in a subscription ecosystem. Cloud Code
- 14:22can traverse it today. A new Codex model could
- 14:25use it tomorrow. Any future AI agent that can
- 14:28read basic text files can instantly connect to
- 14:30your second brain. It guarantees your accumulated
- 14:33knowledge is entirely future -proof. You own
- 14:35the files. You own the map. Exactly. Let us bring
- 14:38this all together into focus. The true power
- 14:41of an AI second brain isn't just about throwing
- 14:43thousands of files at a massive language model.
- 14:46We have been trying that brute force method for
- 14:49years. It is always clumsy, slow, and prone to
- 14:53errors. The real power lies entirely in the routing.
- 14:56By turning your raw, chaotic files into these
- 14:59cleanly linked markdown pages, you give your
- 15:02AI a clear, mapped path through your specific
- 15:04knowledge. You eliminate the guesswork. It stays
- 15:07incredibly simple at its core. And because it
- 15:10lives on your machine, it stays portable. And
- 15:12most importantly, it compounds in value over
- 15:14time. Every new PDF you drop in, every transcript
- 15:18you save, it physically makes the whole web smarter.
- 15:21The map gets more detailed. I highly encourage
- 15:24you to just try setting this up today. Download
- 15:26Obsidian. Grab Claude Code. Drop just two related
- 15:30PDFs into a raw folder. Run the Carpathy prompt.
- 15:33Sit back and watch the magic happen right before
- 15:35your eyes. It takes five minutes to change your
- 15:38workflow. It fundamentally changes how you see
- 15:40your own data. But looking at where this technology
- 15:43is heading, it leaves me with a rather profound
- 15:45question for you to ponder. What is that? If
- 15:48your AI second brain eventually holds your most
- 15:50private meeting notes, your detailed business
- 15:53metrics, your future strategies, and perfectly
- 15:55mimics your exact brand voice, at what point
- 15:59does it stop being just a second brain and start
- 16:01becoming a digital twin capable of making your
- 16:03decisions for you?