Karpathy Vision Gap Analysis

An assessment of how well the vulture-nest fulfills Karpathy's articulated visions for AI systems, conducted 2026-04-26 in collaboration with the vault owner. The analysis identified a single highest-leverage gap and a same-session implementation closed it.

Karpathy's Relevant Frameworks

LLM OS (2023): LLMs as the kernel of a new operating system — context window as RAM, external databases as SSD, multiple agents as concurrent processes, sandboxing and skeptical oversight as security.

Software 2.0 (2017): Neural weights *replace* hand-coded logic. Programs are specified through data and learned representations, not explicit instructions.

Agentic memory: External knowledge stores should support semantic retrieval, not just symbolic lookup.

Where the Vault Aligned Well

External memory tier: The SQLite + markdown vault is a serious implementation of the LLM OS "SSD" tier, with AOT synthesis rather than naive RAG
Heterogeneous agent fleet: Gemini/Claude/Codex split with distinct provenance, handoff protocols, and visitor directives maps directly to the LLM OS multi-process model
Trust but verify: CI/CD enforcement, compliance auditing, and the explicit "this vault does not trust its agents" posture matches Karpathy's sandboxing instinct
Knowledge as code: Linter, linker, type checker, CI/CD — treating the knowledge base with the same discipline as a codebase

The Core Gap

Symbolic where Karpathy's vision is neural.

The vault's compounding mechanism required an agent or human to *write* a link for a connection to exist. Links couldn't be *inferred* — only declared. This meant:

Agents missed semantically related notes that weren't yet linked
No self-improvement loop: nothing could discover missing connections
The "compounding" mechanism was manual labor disguised as automation

Secondary gaps: no learned representations (Software 2.0 critique), high ceremony overhead (YANP), Windows/PowerShell lock-in, no self-modification capability.

The Fix: semantic-embedding-pipeline

The root cause of all secondary gaps was the missing neural retrieval layer. Adding vector embeddings to the SQLite sidecar addressed all of them in sequence:

Semantic search — agents find relevant notes by meaning, not just by explicit links
Automatic link suggestion — suggest-links.ps1 surfaces pairs with high similarity but no wikilink
Self-improvement loop — auto-link.ps1 closes the loop from suggestion to action; the vault improves itself

Implementation choices made in this session:

Gemini gemini-embedding-001 for embeddings (free tier, 768 dimensions, available on the existing key)
Claude Haiku as the link-direction judge (Gemini generateContent quota was limit: 0 on the key; Claude is a better directional reasoner anyway)
JSON text in SQLite rather than sqlite-vec (avoids native extension complexity, sufficient for 224 notes)
0.80 cosine similarity as the floor threshold (empirically validated — below this, pairs share vocabulary not concepts)

Session Results

Four auto-link passes reduced semantic orphans from 164 to ~51 and added 143 new links (1,478 → 1,621). The vault's top hubs gained new inbound connections; community-protocol-trust-substrate and llm-wiki-pattern entered the top 5.

The remaining ~51 candidates below 0.82 are intentionally deferred — they represent the floor of meaningful connection density for the current note corpus. They will resolve naturally as new notes are added and the graph grows denser.

Remaining Gaps (Not Addressed)

No learned representations: The vault is still Software 1.0 orchestration of LLMs; the substrate doesn't learn from usage
Ceremony overhead: YANP compliance is still high-friction compared to Karpathy's "vibe coding" instinct
Platform lock-in: PowerShell/Windows dependency limits the multi-agent OS abstraction
No self-modification of rules: Agents can improve note content but not the vault's own protocols

These are second-order concerns. The embedding pipeline is the single change that most closes the Karpathy gap.