The full technical and philosophical case — from the researcher who published the pattern, to why RAG fails, to what AI IBIZA built on top of it.
In April 2026, Andrej Karpathy — co-founder of OpenAI, former Director of AI at Tesla, now a Member of Technical Staff on the Pretraining Team at Anthropic — published a pattern he called the LLM Wiki.
Karpathy is not a theorist. He built the computer vision system behind every Tesla vehicle's self-driving capability (2017–2022). He co-founded OpenAI in 2015 as one of the original research team. He coined "vibe coding" — the term that defined a generation of AI-native development. When he describes an architecture, it is a production-ready idea from someone who has shipped systems at scale.
The LLM Wiki is his answer to the fundamental problem with how most organisations use AI today.
Co-founded OpenAI — original research team, first generation transformer models
Director of AI & Autopilot Vision, Tesla — neural networks behind every Tesla self-driving system
Coined "vibe coding" — the defining term for AI-native development
Introduced the LLM Wiki pattern — compounding knowledge architecture vs stateless RAG
Member of Technical Staff, Pretraining Team — Anthropic (makers of Claude)
Traditional RAG is stateless by design. Every query starts from zero. The AI rediscovers the same relationships, makes the same connections, draws the same inferences — over and over — charging you tokens for work it already did.
For a law firm or tax advisory practice processing thousands of documents, this is not an inconvenience — it is a structural failure of the technology.
"Knowledge is compiled once and kept current, not re-derived on every query. The wiki is a persistent, compounding artifact. Every new source added makes the entire wiki smarter — cross-references, contradictions, and synthesis already exist."
Karpathy's system is structurally simple. The power is in how the layers relate to each other — and how a single new document can touch 10–15 knowledge pages simultaneously.
Immutable ground truth. Every document ever ingested, preserved exactly as received. PDFs, emails, regulatory publications, case law, client files. Nothing is deleted. Everything is auditable.
LLM-generated structured knowledge. Concept pages. Entity pages. Cross-references. Contradiction flags. The AI builds and maintains this layer. A new document arrives — the AI reads it, extracts, and updates all relevant pages. No human intervention.
Rules that tell the AI how to think about and structure knowledge for this specific domain. For a law firm: how to classify case law, how to cross-reference regulatory sources, how to flag contradictions between GDPR and AEPD guidance.
Karpathy on his own system: "I rarely touch it directly — the AI maintains it autonomously." That is the definition of agentic. Not a search box. Not a chatbot. A system that keeps improving without instruction.
The system monitors official sources — BOE, BOIB, ATIB, BORME, INE — and processes every new publication without prompting. When a new fiscal circular is published at 7am, Hermes reads it, structures it, updates every related wiki page, and flags any contradiction with prior law. Before the first client call. No human scheduled this.
Every document ingested does not just add a new note — it rewrites the relationships across the entire knowledge graph. A new ATIB circular touches existing pages on ITP rates, IEET compliance, client precedents, and regulatory cross-references. The vault does not get bigger. It gets smarter. Each document permanently improves every answer the system can give.
Weekly, the system runs a full knowledge audit without human instruction: scanning for contradictions between sources, flagging stale data superseded by newer law, surfacing knowledge gaps where coverage is thin. The system identifies what it does not know and flags it. No checklist. No human review cycle. The system audits itself.
This is what separates an agentic system from a search interface. The system is not waiting for a query to do work. It is continuously ingesting, structuring, cross-referencing, and auditing — autonomously — so that when a query does arrive, the answer is already compiled, not re-derived.
Karpathy's own LLM Wiki, reported April 2026. Built from raw source documents into structured, cross-referenced knowledge.
Not raw documents. Synthesised, organised, cross-referenced wiki content — maintained autonomously by the AI librarian.
"I rarely touch it directly — the AI maintains it autonomously." No human editor. No manual tagging. No re-indexing.
"Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase."
AI IBIZA's architecture for law firms, despachos, and gestorías is a production implementation of the LLM Wiki pattern — hardened for professional legal and tax environments in Spain and beyond.
The entire stack — inference, embeddings, index, orchestrator, and keys — is designed to run on hardware physically inside the client's building. This directly addresses the AEPD's February 2026 Operational Sovereignty doctrine. Not EU residency (where the bytes sit). Sovereignty (who can cause the processing to stop, be inspected, or be subpoenaed). There is no US-headquartered entity in the chain.
Third parties — including AI IBIZA itself — are designed not to materially access the firm's knowledge base in normal operation. This is built to satisfy Art. 542.3 LOPJ and the secreto profesional del abogado. The architecture is designed to be the compliance mechanism. Not a DPA. Not a data residency clause. The physical and logical design.
BOE daily feed. AEPD doctrine. EU regulations. AEAT communications. Case law. Client files. Firm precedents. All feed into the same compounding vault. When the BOE publishes a new circular at 7am, the system has processed, structured, and cross-referenced it before your first client call.
Every claim is anchored to its source with numbered references. The system cannot assert something it cannot attribute. This makes every AI-assisted output auditable — essential for professional liability in legal and tax work, where a wrong answer has consequences.
Spanish and English knowledge coexist in the same vault with proper cross-referencing. BOE (ES), AEPD doctrine (ES), EU regulations (EN/ES), international case law (EN), client communications (both). The system does not translate — it understands in both languages natively.
A firm that starts building its LLM Wiki today is building an asset. Every document it processes, every BOE circular it ingests, every case it closes — all of it permanently enriches the knowledge base.
A competitor firm that starts twelve months later does not start at the same level. They start twelve months of compounding knowledge behind.
The firms that adopt this architecture now are not just buying an AI tool. They are building an institutional intelligence layer that strengthens with every working day, permanently — while their competitors are still uploading files one query at a time into stateless systems that forget everything.
Karpathy is now at Anthropic — Claude's maker — validating the direction. The LLM Wiki pattern is not a research project. It is where the most influential AI researchers are betting.
"ANDREJ KARPATHY JUST DESCRIBED THE EXACT SYSTEM THIS SILICON VALLEY PROFESSOR HAS ALREADY BEEN BUILDING FOR CLIENTS. Karpathy described the theory — this is the production version."
"You drop any file in — and a dedicated librarian agent structures it and adds it to the knowledge vault automatically — zero manual work. The entire base then shows up as a 3D mind graph where every node is a concept and every connection shows how things relate."
"The structured wiki that grows over time idea is more interesting than basic RAG. Memory needs shape, not just more chunks."
"Karpathy drew the map, you're building the road — when can we test drive it"
I'll show you exactly what the system looks like with your type of documents.
No technical knowledge required. No commitment.
Private clients only. Available remotely worldwide.
We take a small number of new clients per quarter. If you are considering working with AI Ibiza, the conversation starts here.