Karpathy LLM Wiki Chinese Compilation
Provenance
- origin: chat
- note: user-provided Chinese compilation of the Karpathy
LLM Wiki idea plus selected high-quality community comments
Core Claims
- An LLM wiki compounds knowledge during ingest instead of rediscovering it during every query.
- The durable middle layer is markdown, not raw chunks or hidden retrieval state.
- The human chooses sources and asks questions; the agent performs compilation, cross-linking, updating, and bookkeeping.
- personalization should be first-class
- voice capture pipelines matter because collection often fails before synthesis
- source-type-aware ingest beats one-size-fits-all summarization
- layered context budgets matter once the wiki reaches hundreds of pages
- every substantive task should produce a user-facing answer plus a durable wiki update
- a divergence check helps prevent the wiki from becoming a reinforcement machine for existing bias
- graph structure is useful, but query and structure discipline matter more than visual polish
Adopted Into Vipin Wiki
Counterpoints and Gaps
- Several comments argue that pure markdown indexing eventually strains under scale and may need a more structured catalog or database layer.
- Voice-first workflows require transcription discipline, otherwise the wiki receives low-fidelity or hallucinated reconstructions.
- Stronger automation can increase drift unless source attribution and human validation stay explicit.