๐๐๐ The Tale of the Sheep Who Followed the Shepherd ๐๐๐
A Story of Markdown, Authority, and 23 Years of Proof
Sure, drop markdown notes into your “knowledge base” and call it a day… ๐๐๐
There is so much more to it. Images. Videos. PDFs. Spreadsheets. Emails. Database queries. Executable code. Voice recordings. Geospatial data. The real world is not made of markdown. It never was. It never will be.
But the Shepherd said: “Use markdown. Let the LLM write everything. You never have to write again.”
And the sheep looked at the Shepherd. The Shepherd had trained the machines that talk. The Shepherd had worked at the great temples of OpenAI and Tesla. Surely, the Shepherd knew the way.
So the sheep followed. ๐๐๐
The Shepherd’s Promise
“The LLM will maintain your wiki. It will write summaries. It will create cross-references. It will flag contradictions. You just curate sources and ask questions. The bookkeeping is near zero.”
The sheep were delighted. They threw their PDFs into the raw folder. The LLM read them. It wrote markdown files. Beautiful markdown files. Links everywhere. The Obsidian graph view looked like a constellation. โจ
Week 1: Heaven.
Month 1: 200 sources. 800 pages. The index.md is getting long, but the LLM still finds things.
Month 3: 500 sources. 2,000 pages. The LLM starts creating duplicates. “Machine Learning” and “ML” are separate pages. The index is now thousands of lines. The LLM’s context window cannot hold it all. The Shepherd said: “Use qmd. A search engine.”
Now the system is not one thing. It is two things. The LLM writes. The search engine retrieves. The wiki is no longer a seamless artifact. ๐๐
The Cracks Appear
Month 6: 1,500 sources. 8,000 pages. The LLM contradicts itself. It doesn’t remember what it wrote three months ago because each session starts fresh. The schema file (CLAUDE.md) has grown to 500 lines of instructions trying to enforce consistency. The LLM follows them imperfectly.
A sheep asks: “Who is the sister of my friend John?”
The LLM searches. It reads pages. It synthesizes. It answers โ maybe correctly, maybe not. Every time the sheep asks, the LLM does the work again. Nothing is cached. Nothing is indexed for this specific question.
Another sheep asks: “What documents are related to John?”
The LLM searches again. Reads again. Synthesizes again. Probabilistic. Expensive. Slow.
In a Dynamic Knowledge Repository, that question is a SQL query. Sub-second. Deterministic. Free.
But the sheep do not know this. The Shepherd did not tell them. ๐๐
Year One: The Mess
The wiki is now 20,000 pages. Contradictions everywhere. The lint operation finds 47 conflicts. The LLM tries to fix them, but it doesn’t remember why they existed in the first place. It overwrites. It guesses. It hallucinates.
Private notes about salaries, health records, and client NDAs are in the wiki. The LLM needs to read the wiki to answer questions. Now the LLM sees everything. There is no permission system in markdown. The Shepherd did not mention this problem. ๐ฆน
The sheep are spending more time linting and fixing than they ever spent writing. The promise of “near zero maintenance” has become a nightmare of constant supervision.
But they continue to follow. Because the Shepherd said it works. Because the Shepherd is an authority. ๐๐๐
The Tale of the Shepherd
Who is this Shepherd?
He trained neural networks. He wrote about software 2.0. He worked at OpenAI and Tesla. He is brilliant โ in his domain.
But his domain is not knowledge management.
He did not spend 23 years building a Dynamic Knowledge Repository. He did not read Engelbart. He does not know CODIAK. He never built an Open Hyperdocument System. He never designed a schema with 113 object types, 245,377 people, 95,211 hyperdocuments, and complete referential integrity.
He came up with a clever weekend hack โ markdown + LLM + Obsidian โ and wrote a gist about it.
And the world lost its mind. ๐๐๐๐๐
What the Shepherd Did Not Tell You
Sheep follow the Shepherd. ๐ But the Shepherd did not tell you:
LLMs have no persistent memory. Each session starts fresh. The wiki is just text files written by previous sessions. There is no guarantee of consistency across weeks or months.
Text files are not a database. No foreign keys. No referential integrity. No type safety. No permission system. No audit trail beyond git (which tracks files, not fields).
LLMs hallucinate confidently. When contradictions exist, the LLM picks a side and answers with confidence. It does not say “I’m not sure.” It does not flag the uncertainty.
Control drops over time. When your wiki has 10,000 pages, you cannot review every change. The LLM becomes the de facto authority because you have no efficient way to verify its work.
Private data cannot be protected. The LLM needs to read the wiki to answer questions. If the wiki contains private information, the LLM sees it. Markdown has no permissions.
Relationships are implicit, not explicit. “John’s sister” requires the LLM to infer from text, not query a structured edge. Fragile. Slow. Probabilistic. Expensive every time.
The Shepherd did not tell you these things. Maybe he did not know. Maybe he did not think they mattered. ๐
The Old Wizard in the Tower
While the sheep followed the Shepherd, an old wizard sat before his PostgreSQL database. 23 years he had been building. 245,377 people. 95,211 hyperdocuments. 113 object types. 25 person-object relationship types. 30+ person-person relationship types. Complete version control. Granular permissions. Deterministic metadata extraction.
The wizard used LLMs too. They generated descriptions. They summarized content. They accelerated his workflow. He got more money because he worked faster.
But the wizard never handed the keys to the LLM.
The wizard said: “The LLM is a refreshener, not the curator. A tool, not the master. Keep your hands on the wheel.” ๐ง
The sheep looked at the wizard. They looked at the Shepherd. They looked at their crumbling markdown wiki.
“But… but the Shepherd is an authority,” they bleated.
The wizard laughed. ๐
“Authority is not infallibility. The Shepherd trains neural networks. He did not spend 23 years building a Dynamic Knowledge Repository. He did not read Engelbart. He does not know CODIAK. He invented a weekend hack and you followed like sheep.”
๐โ๐
The Dynamic Knowledge Repository (DKR)
Doug Engelbart โ the real shepherd of knowledge work โ envisioned the Dynamic Knowledge Repository decades ago. Not as markdown files. Not as LLM-generated text. As a living, breathing, evolving collection of all knowledge assets: intelligence, dialog records, knowledge products. With global addressing. With backlinks. With structured documents. With human purpose at the center.
Engelbart’s CODIAK framework โ Concurrent Development, Integration, and Application of Knowledge โ is about humans analyzing, digesting, integrating, collaborating, developing, applying, and re-using knowledge.
These are human actions. A computer can assist. A computer cannot replace.
The LLM-Wiki pattern is not a DKR. It is not what Engelbart envisioned. It is a self-perpetuating LLM context generator. The wiki exists only to feed the LLM on the next query.
An LLM-Wiki without the LLM is just a bunch of files, without any organization.
A DKR without the LLM is still a fully functional, queryable, relational knowledge base with 23 years of data and complete referential integrity.
The LLM is optional. A nice interface. Not the engine.
The Verdict ๐ง
So go ahead. Run after the Shepherd. Throw your markdown notes into the machine. Let the LLM write your wiki. Let it hallucinate. Let it contradict itself. Let it leak your private data. Let it forget what it wrote last week. Let it answer every question with a probabilistic guess.
๐๐๐๐๐๐๐๐๐๐๐๐๐๐๐
Or…
Keep your hands on the wheel.
Use the LLM as a refreshener, not the curator.
Build a real Dynamic Knowledge Repository with deterministic programs, foreign keys, version control, permissions, and explicit relationships.
Read Engelbart. Learn CODIAK. Understand what a DKR actually is.
Not today. ๐ Not ever.
Sheep follow the shepherd. ๐๐๐
Wizards build their own towers. ๐ง
The Full Article
Hyperscope: Human-Curated Dynamic Knowledge Repositories vs. LLM-Wiki
“ The CODIAK capability is not only the basic machinery that propels our organizations, it also provides the key capabilities for their steering, navigating and self repair.”
โ Douglas C. Engelbart
“Every participant will work through the windows of his or her workstation into his or her group’s ‘knowledge workshop.’”
โ Douglas C. Engelbart
“What is new is a focus toward harnessing technology to achieve truly high-performance CODIAK capability.”
โ Douglas C. Engelbart
๐โ๐ Don’t be a sheep. ๐โ๐
Related pages
- Hyperscope: Human-Curated Dynamic Knowledge Repositories vs. LLM-Wiki
The text contrasts Andrej Karpathy's "LLM-Wiki" pattern, where an LLM autonomously maintains a markdown-based knowledge base, with the author's "Hyperscope" system, a PostgreSQL-based Dynamic Knowledge Repository. While the LLM-Wiki offers initial convenience, the author argues it inevitably degrades due to the LLM's lack of persistent memory, inability to enforce data integrity, and failure to protect privacy or scale effectively. In contrast, Hyperscope prioritizes human control by using deterministic programs for factual metadata and database constraints for relationships, reserving the LLM only for descriptive synthesis. Drawing on Douglas Engelbart's vision, the author concludes that a true "dynamic" knowledge repository relies on human judgment, purpose, and accountability rather than automated machine processes, asserting that the human remains the essential curator and the LLM is merely a powerful tool to accelerate workflow. - Karpathy's LLM-Wiki Is a Flawed Architectural Trap
The author sharply criticizes Andrej Karpathy's viral "LLM-Wiki" concept as a flawed architectural trap that mistakenly treats unstructured Markdown files as a robust database, arguing that relying on LLMs to autonomously generate and maintain knowledge leads to hallucinations, broken links, privacy leaks, and a loss of human cognitive engagement. While acknowledging the appeal of compounding knowledge, the text asserts that Markdown lacks essential database features like referential integrity, permissions, and deterministic querying, causing the system to collapse at scale and contradicting its own "zero-maintenance" promise. Ultimately, the author advocates for proven, structured solutions using real databases and human curation, positioning LLMs as helpful assistants rather than autonomous masters, and warns against blindly following a trend promoted by someone who has publicly admitted to being in a state of psychosis.